www.nature.

com/scientificreports

OPEN Optimizing solar power efficiency

in smart grids using hybrid machine
learning models for accurate
energy generation prediction
Muhammad Shoaib Bhutta 1*, Yang Li 1, Muhammad Abubakar 2, Fahad M. Almasoudi 3
,
Khaled Saleem S. Alatawi 3, Mohammad R. Altimania 3 & Maged Al‑Barashi 4
The fourth energy revolution is characterized by the incorporation of renewable energy supplies
into intelligent networks. As the world is shifting towards cleaner energy sources, there is a need
for efficient and reliable methods to predict the output of renewable energy plants. Hybrid machine
learning modified models are emerging as a promising solution for energy generation prediction.
Renewable energy generation plants, such as solar, biogas, hydropower plants, wind farms, etc. are
becoming increasingly popular due to their environmental benefits. However, their output can be
highly variable and dependent on weather conditions, making integrating them into the existing
energy grid challenging. Smart grids with artificial intelligent systems have the potential to solve
this challenge by using real-time data to optimize energy production and distribution. Although by
incorporating sensors, analytics, and automation, these grids can manage energy demand and supply
more efficiently, reducing carbon emissions, increase energy security, and improve access to electricity
in remote areas. However, this research aims to enhance the efficiency of solar power generation
systems in a smart grid context using machine learning hybrid models such as Hybrid Convolutional-
Recurrence Net (HCRN), Hybrid Convolutional-LSTM Net (HCLN), and Hybrid Convolutional-GRU Net
(HCGRN). For this purpose, this study considers various parameters of a solar plant such as power
production (MWh), irradiance or plane of array (POA), and performance ratio (PR). The HCLN model
demonstrates superior accuracy with the RMSE values of 0.012027 for MWh, 0.013734 for POA
and 0.003055 for PR, along with the lowest MAE values of 0.069523 for MWh, 0.082813 for POA,
and 0.042815 for PR. The obtained results suggest that the proposed machine learning models can
effectively enhance the efficiency of solar power generation systems by accurately predicting the
required measurements.

Keywords Renewable energy resources, Hybrid machine learning models, Smart grids, Energy generation
prediction, Solar power generation

Abbreviations
AI Artificial intelligence
ARIMA Autoregressive integrated moving average
AutoGCN Automatic graph convolutional networks
CNN Convolutional neural networks
CNN-RNN Convolutional neural network-recurrent neural network
CNN-GRU​ Convolutional Neural Network-Gated Recurrent Unit
CNN-LSTM Convolutional neural network-long short-term memory
DL Deep learning
DNN Deep neural network

1
School of Automobile Engineering, Guilin University of Aerospace Technology, Guilin 541004, China. 2Key
Laboratory of Smart Grid of Ministry of Education, Tianjin University, Tianjin 300072, China. 3Department of
Electrical Engineering, Faculty of Engineering, University of Tabuk, 47913 Tabuk, Saudi Arabia. 4School of
Aeronautics and Astronautics, Guilin University of Aerospace Technology, Guilin 541004, China. *email:
[email protected]

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 1

Vol.:(0123456789)
 www.nature.com/scientificreports/

DDQN Double deep Q-network

F-RAN Fog radio access network
GRU​ Gated Recurrent Unit
GCN Graph convolutional network
HCRN Hybrid convolutional-recurrence net
HCLN Hybrid convolutional-LSTM net
HCGRN Hybrid convolutional-GRU net
IoT Internet of Things
KHSS Keyed-hashing-based self-synchronization
LSTM Long short-term memory
MPPT Maximum power point tracking
MWh Mega-watt hour
POA Plane of array
PR Performance Ratio
QoS Quality of Service
RMSE Root mean square error
RNN Recurrent neural network
RQGPR Rational quadratic Gaussian process regression
RAS Recirculating aquaculture systems

Recent advancements in artificial intelligence (AI) and the Internet of Things (IoT) have spurred innovative
approaches in various domains. In the field of service recommendation and Quality of Service (QoS) prediction,
where traditional methods encounter accuracy limitations, the introduction of a two-stream deep learning model
incorporating user and service graphs has demonstrated improved QoS prediction a­ ccuracy1. Exploring the inter-
section of deep learning and defense mechanisms in Digital Twins and Deep-Learning-as-a-Service Computing
Systems, the study emphasizes the use of deep learning algorithms for data analysis to address security chal-
lenges in service computing systems, highlighting the crucial role of Digital Twins in enhancing network defense
­capabilities2. In the domain of pattern recognition, a proposed Attention-Driven Framework for Unsupervised
Pedestrian Re-identification aims to address challenges posed by variations in pose, occlusion, and lighting
­conditions3. The study underscores the importance of efficient sampling strategies and introduces a clustering
optimization approach to enhance the performance of Convolutional Neural Networks (CNN) in pedestrian
re-identification3. Exploring the integration of intrusion detection, Deep Learning (DL), and Digital Twins for
city network security, a model utilizing Deep Neural Network (DNN) is presented to enhance network security
defense systems, incorporating a trust model based on Keyed-Hashing-based Self-Synchronization (KHSS)2.
Addressing the challenges of inconsistent file sizes in fog radio access networks (F-RANs), the investigation
into joint edge caching and content recommendation proposes a Double Deep Q-Network (DDQN) based dis-
tributed edge caching algorithm, demonstrating increased net profit and caching e­ fficiency4. In the context of
IoT and edge computing, an intelligent model is introduced for supporting edge migration of virtual function
chains, focusing on exploiting computational power at the edge and supporting demanding services through
features like auto-healing and Quality of Service m ­ onitoring5.
Recent advances in artificial intelligence (AI) and the Internet of Things (IoT) have led to innovative solutions
across various domains. Specifically in aquaculture, the control of water quality in Recirculating Aquaculture
Systems (RAS) is critical for the survival and growth of aquatic organisms. To overcome the limitations of
conventional methods in water quality prediction, a hybrid deep learning framework is introduced, integrating
Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), and Attention mechanisms to enhance
the efficiency and accuracy of water quality p ­ rediction6. In exploring power systems, particularly in power line
outage identification, Graph Convolutional Networks (GCNs) are applied to frame the problem as a graph sig-
nal classification challenge. Utilizing spatial and spectral GCN architectures, the study incorporates graph shift
operators and frequency filters for effective convolution, successfully classifying abnormal signal patterns and
demonstrating efficacy in power line outage i­ dentification7. Addressing limitations in graph convolutional net-
works, Automatic Graph Convolutional Networks (AutoGCN) are introduced to capture the entire spectrum of
graph signals, autonomously updating filter bandwidth and avoiding the low-pass filter approach. This innovative
approach, grounded in graph spectral theory and spatially localized, outperforms baseline methods, providing
a more comprehensive understanding of graph-structured d ­ ata8.
So, all the above discussion about the technological landscape necessitates the development of smart grids,
which are intricate systems designed to facilitate this integration. Smart grids are complex infrastructures that
provide the seamless integration of renewable energy resources into the existing power grid. Hybrid models of
machine learning, such as Convolutional Neural Network-Recurrent Neural Network (CNN-RNN), Convolu-
tional Neural Network-Gated Recurrent Unit (CNN-GRU), and Convolutional Neural Network-Long Short-
Term Memory (CNN-LSTM), have shown a great deal of promise in the area of anticipating the demand for
energy and optimizing the amount of energy that is used by smart g­ rids9. In addition, these models could
potentially reduce our reliance on fuels and enhance the long-term sustainability of our energy system by maxi-
mizing the generation and storage of renewable e­ nergy10. Machine learning based models have been utilized to
forecast the production of energy and address short term electric power ­demands11–13. Additionally advanced
machine learning techniques, like the CNN RNN CNN GRU and CNN LSTM have been utilized to forecast traffic
­patterns14,15. In the context of smart grids, deep learning can be used to enable demand response by predicting
energy demand, and machine learning models may be used to improve the accuracy of these p ­ redictions12.
Furthermore, deep learning has the potential to support demand response by incorporating reinforcement

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 2

Vol:.(1234567890)
www.nature.com/scientificreports/

learning. These techniques can also aid in the real time management of the power grid, aiming to boost the share
of renewable energy sources in the grid’s energy mix while reducing its carbon footprint. The integration and
management of renewable energy sources may be improved via the use of machine learning and deep learning
models in smart grids. This will ultimately result in energy systems that are more effective and sustainable. For
instance, a hybrid CNN-LSTM model has been proposed for the prediction of the price of ­electricity16, a novel
hybrid model combining Random Forest and LSTM has been developed for the forecasting of traffic ­flow17, and a
hybrid model that combines CNN and Autoregressive Integrated Moving Average (ARIMA) has been developed
for the forecasting of the electricity ­demand18.
In addition, a technique that combines CNN and SVR has been suggested to predict wind s­ peed19. The hybrid
machine learning models that are being used possess the capability to enable the seamless integration of renew-
able energy sources into smart grids, thereby supporting the global shift towards sustainable energy systems.
Researchers have employed a mix of advanced machine learning techniques like CNN RNN, CNN GRU and
CNN LSTM to forecast various factors such as electricity prices, traffic patterns, wind power output in the short
term, electricity demand fluctuations, wind speeds, the health of wind turbines, long term electricity demand
trends, short term load predictions and residential energy ­usage16–24. These models can take into account the data’s
temporal dependencies as well as their non-linear correlations. For short-term load forecasting, a hybrid CNN
model that incorporates residual learning and an attention mechanism has been ­suggested23. Forecasting energy
usage has also been accomplished with the assistance of multi-hybrid ensemble l­ earning25. These models have the
potential to increase the accuracy of forecasts and to assist in the incorporation of renewable energy resources
into smart grids. Wind turbine monitoring and malfunction diagnosis is another area in which they may be of
­use21. Current research projects have focused on constructing these hybrid machine-learning models for different
applications relating to energy, such as predicting renewable energy, monitoring the state of wind turbines, and
forecasting the power demand. For instance, ­research21,24, ­and26 have all applied CNN-based models to estimate
the condition of wind turbines, household energy consumption, and wind speed forecasts, respectively.
In the meanwhile, ­research27,28, ­and29 have employed hybrid models that combine LSTM with other method-
ologies to obtain accurate forecasts of power demand and load. These hybrid models have the potential to find
applications in smart grids, which are an important aspect of the fourth industrial revolution. This is due to its
ability to facilitate the efficient integration of renewable energy sources into the power grid, hence optimizing
energy use. A technique based on frequency modulation is used i­ n30 to evaluate and determine the most effec-
tive way to coordinate energy storage devices. An updated gravitational search algorithm is used to construct a
dual-stage optimization strategy to optimize the economic operation of energy storage i­ntegration31. This was
done to maximize the effectiveness of the optimization. The authors of ­reference32 put out a proposition for a
voltage stability index that incorporates a unique single-port equivalent and is based on the distinctive attributes
and enduring sensitivity of components. Using methods from machine learning, the authors o ­ f33 examined the
operational efficiency of large-scale solar power facilities.
Also, ­in34, Machine learning algorithms perform better than statistical techniques in accurately estimating
solar power output. The Coarse Tree model has the lowest Root Mean Square Error (RMSE) when utilized com-
bined with a Maximum Power Point Tracking (MPPT) controller. However, Rational Quadratic Gaussian Process
Regression (RQGPR) provides the lowest Root Mean Square Error (RMSE) without involving the application
of a Maximum Power Point Tracking (MPPT) controller. Studies ­in35–38 utilizes sophisticated deep learning and
machine learning models, including GRU, LSTM, SVM, and SNN, to predict the operating conditions of power
grids. This research demonstrates notable improvements by training these models with synthetic data and their
real-time practicality. Furthermore, it emphasizes the improved accuracy of a Deep Neural Network (DNN)
model compared to Bagged Tree and ARIMA models in predicting solar power generation. The DNN model
exhibits the lowest RMSE and MSE values, both with and without the MPPT controller. All of these studies
highlight the significance of optimizing energy storage and renewable energy systems in smart grids through
the application of sophisticated machine learning models to improve the effectiveness and dependability of
these systems, as well as to facilitate the incorporation of renewable energy sources during the fourth industrial
revolution. These applications may be made smarter by using more advanced hybrid machine learning models,
which are successful in predicting power demand and short-term load forecasting. These models can also be
used to make these applications.
In this study, the novelties and the main contribution are discussed as follows:

1. The hybrid models are carefully designed by altering their structures to incorporate the useful features of
both convolutional and recurrent neural networks.
2. The significant implications of these modifications are demonstrated by the real-time data gathered from a
solar plant, which surpasses the prediction capability of individual models.
3. The study demonstrates a significant improvement in accuracy rates and a notable decrease in error values.
The primary achievement of this study is the progress made in predictive modeling for renewable energy
systems, representing a significant enhancement in this area. Additionally, the study has refined hybrid
machine learning models, specifically HCRN, HCGRN, and HCLN, which were designed to overcome the
limitations of traditional models when dealing with complex datasets and intricate patterns. Through the
integration of supplementary techniques and the optimization of algorithms, these sophisticated hybrid
models exhibit exceptional performance by outperforming simple base models. The study emphasizes the
enhanced effectiveness and accuracy attained by these hybrid models, specifically in tasks necessitating
complex pattern identification and sequence manipulation. Moreover, the study highlights the significance
of utilizing these advanced models as a primary focus in deep learning research and development, specifically
for tasks involving sequential data and image processing.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 3

Vol.:(0123456789)
 www.nature.com/scientificreports/

Section "Proposed frame work/methodology" is the description of the proposed framework/methodology that
they have developed for analyzing the real-time data parameters acquired from the solar plant. The framework
involves creating mathematical models and utilizing customized machine learning algorithms to study the data
and recognize patterns and trends. Section "A comparative analysis between developed HCRN, HCGRN, HCLN
models, and the basic state of the art machine learning models" is a case study that discusses the data parameters
acquired from a solar plant. Various visualization methods are employed to analyze the data, such as box plots,
pair plots, heat maps and histograms. The methods mentioned are utilized to spot patterns and tendencies, in
the data with the goal of enhancing the efficiency of the facility. In section "Case study", the results of the analy-
sis of the real-time data parameters acquired from the solar plant using the proposed framework/methodology
are discussed in terms of patterns and trends identified in the data and how these can be used to improve the
solar plant’s performance. Lastly, section "Results and discussion" provides concluding remarks about the study
summarizes the key findings discusses the limitations of the study, and provides suggestions for future research.

Proposed frame work/methodology

Figure 1 shows the overall methodology of this research paper, which depicts the comprehensive framework of
the study. A real-time data collected from a solar plant, comprising three essential parameters: power production
(MWh), the plane of array (POA), and Performance ratio (PR). To develop the predictive models, 80% of the
collected data is utilized for training, while the remaining 20% is employed for testing and validation process. The
study employs three hybrid models: Hybrid Convolutional-Recurrence Net (HCRN), Hybrid Convolutional-GRU
Net (HCGRN), and Hybrid Convolutional-LSTM Net (HCLN). These models are the modified improved ver-
sions of CNN-RNN, CNN-GRU, and CNN-LSTM respectively which is the novelty of this research work. These
modifications are made by changing the hyperparameters and layer structures of the basic models to create new

Figure 1.  A framework for identifying the optimal predictive model for solar plants.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 4

Vol:.(1234567890)
www.nature.com/scientificreports/

models. The data is split into two parts: the predicted output of the parameters and the projected errors RMSE
and MAE generated by these modified hybrid models. The results indicate that Hybrid Convolutional-LSTM Net
(HCLN) outperformed the other two hybrid models, with lower RMSE and MAE values compared to Hybrid
Convolutional-Recurrence Net (HCRN) and Hybrid Convolutional-GRU Net (HCGRN).
To begin the process, the first dataset is imported, and the data’s quality is assessed before proceeding to the
validation stage to find the independence of the data. Figure 2 illustrates the validation process, which involves
checking whether the imported data is in numerical form and meets the requirements of time series analysis.
Reliable and accurate data is crucial, so any data in the form of strings, containing negative numbers or outliers,
is filtered out. If the data is confirmed to be accurate, the process moves on to the next phase. However, if the data
does not meet the validation criteria, it undergoes another round of filtering, as depicted in Fig. 2. The flowchart
describes the methodology of this research study starting with the importation of the dataset, which is then
immediately followed by a stringent validation stage. If the data fails the validation, it goes through a process of
recycling; otherwise, it is transferred to the modified codes and continues to its final destination. After the dataset
has been verified effectively, it is then split up into several subsets, such as testing and training. After that, the
data from these subsets is used as input for the updated models, which include the Hybrid Convolutional-GRU
Net (HCGRN), the Hybrid Convolutional-LSTM Net (HCLN), and the Hybrid Convolutional-Recurrence Net
(HCRN). These models were developed to improve the system’s capacity for accurate prediction. The dataset is
run through these models, which then produces findings that may be used for prediction. A thorough comparison
of prediction outcomes, taking into account important metrics such as root mean squared error (RMSE) and
mean absolute error (MAE) helps to identify the model that displays greater performance. This solid technique
guarantees that the study will use the most accurate prediction model possible.
Convolutional neural networks (CNNs) have proven to be highly effective, in tasks involving the analysis
of time series data, such, as signal processing, speech recognition and image interpretation. Convolutional lay-
ers, known for their weight matrices serve as the principle of Convolutional Neural Networks (CNNs) and are
utilized to execute convolutions on the input data. The data provided includes multiple data points linked with
specific timestamps. The models’ layers are crafted to uncover insights and details, from this time sensitive data.
The basic setup of a Convolutional Neural Network (CNN) is depicted in Fig. 3.
In the Convolutional Layer, the input sequence is subjected to convolutional operations involving a series of
filters, resulting in the generation of feature outputs. By applying a sliding mechanism to the filter throughout
the input sequence and calculating the dot product between the filter and the relevant subsequence of the input,
the corresponding output can be obtained. The non-linearity of the model is achieved by the application of non-
linear activation functions to the outputs of the components within each layer. As a result of this phenomenon,
the Convolutional Neural Network (CNN) can effectively acquire and comprehend complex patterns within
the dataset. To effectively decrease the dimensionality of spatial data while preserving crucial information, a
pooling layer may be used to down sample the outputs of the layer. This can be achieved by the utilization of
functions such as max pooling or average pooling. One possible approach to mitigate overfitting is the use of
a regularization method known as dropout, which may be implemented on the output of the pooling layer.

Figure 2.  Flowchart illustrating the step, by step process of generating predicted values and visualizing errors.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 5

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 3.  Convolution neural network structure.

To promote the acquisition of more intricate features, the technique of dropout is used, whereby a fraction of the
activation values are randomly set to zero over the whole of the training phase. Convolutional layers produce a
vector, which is subsequently fed into one or more fully connected layers tasked with performing classification
or regression processes.
Recurrent Neural Networks (RNNs) are renowned for their ability to proficiently handle time-series and
sequential data. The architecture of the RNN cell comprises a succession of interconnected ‘memory cells’ as
illustrated in Fig. 4a29. These memory cells exhibit a sequential connection, resembling a chain-like structure.
At discrete time intervals, the input undergoes processing by the initial memory cell, leading to alterations in its
internal state and the subsequent transmission of information to the subsequent memory cell. The output of each
memory cell depends on both its internal state and the input it receives at a given time step. A distinctive feature
of recurrent neural networks (RNNs) is their unique property which is the output of a memory cell at each time
step is recursively fed into the subsequent memory cell in the sequence. Using this feedback mechanism, the
network adeptly stores and retrieves past inputs, facilitating the prediction of future inputs. Equations (1) and
(2) provide a mathematical representation of RNN internal structure.
Ht−1 = σ (Ph ∗ Ht−1 + Px ∗ Xt + Ba ) (1)

Yt = tanh(Po ∗ Ht + Bo ) (2)
The hidden state of RNN at time t−1, computed from the input at time t−1 and the previous hidden state at
time t−2, is denoted by ­Ht-1 in the given equation. Input at time t is denoted by the vector ­Xt, which may be a
visual feature. Both the input at time t and the previous hidden state can have an effect on the current hidden
state, which is controlled by the weight matrices ­Ph and P ­ x. Before applying the sigmoid function (σ), which
modifies the output based on the input values and the last hidden state. The bias vector ­Ba is multiplied by the
products of P ­ h and H
­ t-1 and P
­ x and X
­ t. The sigmoid function is used to determine how much the current input
and the previous hidden state matter when calculating the current hidden state. Po is also a weight matrix, but it
calculates how much the current hidden state contributes to the output at time t. Before applying the tanh activa-
tion function, which scales the output to a value between -1 and 1, the bias vector Bo is added to the product of Po
and Ht−1. By using the present concealed state and the matrix of weights Po, the output at time t is denoted by Yt.
Gated Recurrent Units, commonly referred to as GRUs, stand as a distinct variant of recurrent neural net-
works, initially developed as a substitute for the long-term memory (LSTM) a­ pproach29. The internal configu-
ration, as depicted in Fig. 4b29, mirrors that of an LSTM while employing fewer parameters. GRUs, similar to
LSTM, include a set of memory cells that are specifically engineered to store information for a duration of time.
In contrast to Long Short-Term Memory (LSTM) models, Gated Recurrent Units (GRUs) include a gating
mechanism that effectively controls the flow of information inside the memory cells.
Similarly, a GRU encompasses two crucial gates: one is reset gate and the other is update gate. The initial
gate, known as reset gate, dictates the degree of retention of prior memory states. Subsequently, the update gate,
the second gate, assumes the role of determining the preservation of the previous state and the assimilation of
fresh information from the current input. Before calculating these gates, the input at each time step undergoes a
series of linear transformations. The collective transformations are responsible for determining the states of the
gates. Once they are formed, the update and reset gates work together to alter the state of the current memory
cell. The resultant state is formed by a combination of its previous condition and the present input. The result-
ant output of the Gated Recurrent Unit (GRU) is fundamentally governed by the condition of the memory cell
at each sequential time step. This result demonstrates usefulness in both predicting and classificatory tasks.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 6

Vol:.(1234567890)
www.nature.com/scientificreports/

RNN Cell
ban

ph*ht-1
ht-1 tanh ht
px*xt
xtn ytn

(a)

GRU Cell

ht-1 ht

rt 1-
Reset Gate
zt ht
σ
σ
h't
GateUpdate tanh
xt

(b)

LSTM Cell
yt
ct-1 ct

Forget Gate it tanh

ft
c’t
Output Gate
Input Gate σ σ tanh

ht-1 ot ht
σ
xt

(c)

Figure 4.  (a) Basic RNN ­structure29. (b) Basic GRU ­structure29. (c) Basic LSTM s­ tructure29.

To have a thorough comprehension of the internal mechanisms of the Gated Recurrent Unit, one may consult
Eqs. 3–6, shown below:
Rt = σ (Wrh ∗ Ht−1 + Wrx ∗ Xt ) (3)

Zt = σ (Wzh ∗ Ht−1 + Wzx ∗ Xt ) (4)

H ′ t = tanh{Wh′ h ∗ (Rt ∗ Ht−1 ) + Wh′ x ∗ Xt } (5)

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 7

Vol.:(0123456789)
 www.nature.com/scientificreports/

Ht = {(1 − Zt )Ht−1 + Zt ∗ H ′ t } (6)

At time step t, the update gate Rt determines the proportion (ranging between 0 and 1) of the prior hidden
state to retain and update with fresh data. The sigmoid function (σ), denoted as Sigma, compresses inputs into
a range of 0 to 1. The weight matrix Wrh pertains to the previous hidden state Ht−1 influencing the amount by
which it should persist. Ht−1 is a vector capturing the preceding GRU output. Another weight matrix Wrh relates
to the current input Xt , modulating its contribution to updating the hidden state. Xt is a vector embodying novel
information destined for integration into the hidden state.
The internal structure of the Long Short-Term Memory (LSTM), a subtype of recurrent neural network
(RNN), is visualized in Fig. 4c29. This design is adept at capturing extensive temporal relationships within sequen-
tial data and mitigating the issue of vanishing gradients. The internal configuration of the Long Short-Term
Memory (LSTM) network is distinguished by an array of memory cells, each equipped with a specific ensemble
of gate mechanisms that oversee the flow of information within the cell. Comprising three distinct types of input,
output, and forget gates, the internal structure of the Long Short-Term Memory (LSTM) network leverages linear
transformations to process input data at each successive time step. These transformations calculate the activa-
tion levels of the gates. The forget gate determines the exclusion of certain information from the prior memory
state, while the input gate dictates the assimilation of new data into the current memory state. Concurrently,
the output gate regulates the selection of information to be extracted from the current memory state. The Long
Short-Term Memory (LSTM) cell’s inner state undergoes updates by combining the previous state with fresh
information, overseen by the forget and input gates. Subsequently, the output gate decides the pertinent informa-
tion for transmission to the subsequent cell in the sequence. LSTMs have showcased their efficacy in handling
prolonged dependencies across a spectrum of sequential tasks, encompassing natural language processing, speech
recognition, and time-series forecasting. Equation 7–12 delineates the mathematical representation portraying
the internal architecture of the Long Short-Term Memory Unit (LSTM).
Ft = σ {WF (Ht−1 , Xt) } (7)

It = σ {WI (Ht−1 , Xt) } (8)

C′t = tanh {WC ∗ (Ht - 1 , Xt )} (9)

Ct = (Ft ∗ Ct−1 ) + Ct ∗ C ′ t (10)

Ot = σ {WO .(Ht−1 , Xt )} (11)

Ht = Ot ∗ tanh C t (12)
 

At the current time step, denoted as t the components of the Long Short-Term Memory (LSTM) architecture
interact to manage information flow. The forget gate, labeled as Ft , decides the relevance of the previous cell
state, considering both the input Xt and the preceding hidden state Ht−1, determining what to retain or discard.
Simultaneously, the "input gate," denoted as It , assesses the extent of alteration that the incoming cell state should
undergo, factoring in the variables Xt and the preceding hidden state Ht−1. The current cell state, Ct , is then
updated using the outcome of the forget gate Ft , the input gate It , and a new potential value, C′t , computed from′
Xt and Ht−1. This process unfolds against the backdrop of the prior cell state, Ct−1. The candidate value C t is
derived by applying the hyperbolic tangent to the linear combination of Xt and Ht - 1. The "output gate," labeled
as Ot , intervenes to establish the portion of present state of the cell that should be conveyed as updated hidden
state, Ht , factoring in Xt and the preceding hidden state Ht−1. Ultimately, the current hidden state, Ht , emerges
through the element-wise multiplication of the output gate Ot and the hyperbolic tangent of the current cell state
Ct . This intricate interplay enables LSTMs to effectively manage and process sequential data.

Hybrid convolutional‑GRU net (HCGRN)

A neural network model for sequence data processing is shown in Fig. 5a. The model has layers for extracting
features, modifying sequences, and predicting outputs. Convolutional operations on input data are performed
by two one-dimensional convolutional layers, conv1d and conv1d_1. These provide output tensors with forms
(None, 3, 256) and (2, 128). After the convolutional layers, max_pooling1d reduces the spatial dimensions of
the output. Its tensor shape is (None, 1, 128). A single-dimensional tensor with 128 elements is produced by
the flattening layer, which may be fed to subsequent recurrent layers. A repeat vector layer copies this tensor 30
times, creating an output shape of (None, 30, 128) for the recurrent neural network (RNN) layers. The model has
three GRU layers: gru, gru_1, and gru_2. Each layer analyses 30-length sequences and contains 100 hidden states.
Dropout layers after GRU layers prevent overfitting during training. The model’s temporal linkages are improved
with a bidirectional layer that handles sequence data forward and backward. Another dropout layer precedes the
heavy layers. The "dense" layer has 100 units and is completely connected. Next is "dense_1", a single-unit dense
layer that outputs regression tasks. The model has 459,073 trainable parameters and no non-trainable parameters.

Hybrid convolutional‑LSTM net (HCLN)

Figure 5b shows the architecture of a neural network model for sequential data analysis. The system has layers
that extract distinctive features, transform sequences, and predict potential outcomes. Two one-dimensional

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 8

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 5.  (a) Layer structure of hybrid convolutional-GRU Net (HCGRN), (b) Hybrid convolutional-LSTM
Net (HCLN), (c) Hybrid convolutional-recurrence net (HCRN).

convolutional layers, conv1d and conv1d_1, convolve the input data. It produces output tensors of dimensions
(None, 3, 256) and (None, 2, 128). After that, max_pooling1d decreases the spatial dimensions of the output
from the previous layer to (None, 1, 128). To prepare the output for the next recurrent layer, the flattening layer
converts it into a one-dimensional tensor of 128 elements. An output shape of (None, 30, 128) is suitable for input
into Long Short-Term Memory (LSTM) layers after a repeat vector layer duplicates this tensor 30 times with a
single dimension. The model has three LSTM layers, lstm, lstm_1, and lstm_2—that handle 30-length sequences
with 100 hidden states. Dropout layers after each LSTM layer prevent model overfitting during training. This
model better captures temporal links by handling sequence data in both directions using a bidirectional layer.
Another dropout layer precedes the heavy layers. The "dense" layer has 100 units and is completely connected.
Next, "dense_1" has one unit and is the regression-specific output layer. The model has 579,129 trainable param-
eters and 0 non-trainable parameters.

Hybrid convolutional‑recurrence net (HCRN)

The layers, output forms, and number of parameters for each layer in a neural network model are listed in Fig. 5c.
The model starts with two one-dimensional convolutional layers, conv1d, and conv1d_1, that extract input
data properties. Output tensor sizes are (None, 3, 256) and (None, 2, 128) for these layers. A max-pooling layer,
max_pooling1d, reduces the spatial dimensions of the previous layer’s output to (None, 1, 128). The flattened

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 9

Vol.:(0123456789)
 www.nature.com/scientificreports/

layer creates a one-dimensional tensor of 128 elements from the output. A repeat vector layer copies this tensor
30 times with a single dimension, creating an output shape of (None, 30, 128) for the recurrent neural network
(RNN) layers. The model uses three Simple RNN layers: simple rnn, simple_rnn_1, and simple_rnn_2. Each
layer analyses 30-length sequences and contains 100 hidden states. Dropout layers after each Simple RNN layer
stochastically change input units to zero during training to prevent overfitting. A bidirectional layer evaluates
sequence data forward and backward to enhance the model’s temporal linkages. Before the thick layers, another
dropout layer is added. The "dense" layer has 100 units and is completely connected. Next is "dense_1", a single-
unit dense layer that outputs regression jobs. The model has 213,957 trainable parameters and no untrainable
parameters.

A comparative analysis between developed HCRN, HCGRN, HCLN models,

and the basic state of the art machine learning models
A comparative analysis between the developed HCRN, HCGRN, and HCLN models and other state-of-the-art
machine learning models reveals notable distinctions in their capabilities and applications. While simple basic
models like CNN, RNN, GRU, and LSTM utilize a combination of Convolutional Neural Networks (CNNs) and
Recurrent Neural Networks (RNNs) to extract features and handle sequential data, advanced hybrid models
such as HCGRN, HCLN, and HCRN further enhance efficiency by incorporating additional methods and refin-
ing algorithms. While basic models may excel at simpler tasks, they often face challenges when dealing with
complex datasets, whereas sophisticated models demonstrate superior performance, particularly in managing
intricate patterns and sequences. However, the increased sophistication of models like HCGRN, HCLN, and
HCRN necessitates more significant computational resources and longer training periods. Moreover, while basic
models offer simplicity and ease of use, advanced hybrid models provide greater customization possibilities and
are well-suited for tasks requiring high accuracy and intricate pattern recognition, representing the forefront of
research and development in deep learning for sequential data.
The comparison Table 1 shows some differences between basic hybrid machine learning models like CNN-
RNN, CNN-GRU, and CNN-LSTM and more complex ones like HCGRN, HCLN, and HCRN. For feature
extraction, simple models use Convolutional Neural Networks (CNN), and for handling sequential data, they
use basic Recurrent Neural Networks (RNN), GRU, or LSTM. These simple models are easy to use and offer a
clear method, but they may not be able to capture long-range relationships in data and can’t handle complex pat-
terns well. Instead, more advanced models use more complex types of RNN, GRU or LSTM, along with maybe
some other methods for improving performance. However, advanced models typically require more computa-
tional resources, extensive training, and a deeper understanding of advanced deep learning concepts. The choice
between these models hinges on the specific requirements of the problem at hand, with simple models sufficing
for basic tasks and advanced models excelling in complex and challenging applications.

Advanced hybrid machine learning modified models (HCGRN,

Aspect Simple basic models (CNN-RNN, CNN-GRU, CNN-LSTM) HCLN, HCRN)
Utilizing a Convolutional Neural Network (CNN) to extract features in Advanced hybrid models that merge CNN with RNN variations, like
Model Architecture combination, with a Recurrent Neural Network (RNN) GRU or LSTM GRU or LSTM along, with incorporating other methods to boost overall
for handling data efficiency
Adapting to types and levels of data presents some challenges due, to the Enhanced versatility resulting from the use of versions and further
Flexibility
constraints, in flexibility refinement methods
It can work well with data that comes in a sequence. Might find it chal- More adept, at recognizing connections and managing data in a
Handling Temporal Sequences
lenging to capture relationships that span over a distance sequence
Sophisticated models may necessitate resources and extended training
Training Complexity Training may seem simpler when compared to models
periods
Basic models might excel at tasks. They could face challenges when deal- Sophisticated models often excel over models, in tasks particularly when
Performance
ing with more intricate problems handling complex datasets
The level of prediction accuracy varies moderately influenced by the Overall, there is usually accuracy, in predictions because of the enhanced
Predictive Accuracy
complexity of the issue and the dataset used ability to capture patterns
Well-suited for tasks demanding high accuracy and the capture of intri-
Use Cases Ideal, for image or sequence processing assignments
cate patterns in images or sequences
A common practice that is well known and comprehensible, within the Represent the forefront of research and development in deep learning
Research and Development
field of machine learning for sequential data and image processing
More complex to implement, requiring a deeper understanding of
Complexity of Implementation Easier to set up and use. Not much room, for personalization
advanced deep learning concepts
Greater customization possibilities, allowing fine-tuning for specific
Model Customization Not much room, for personalization
applications
It can understand features. It might struggle with grasping intricate
Learning Representations Better at learning intricate representations and relationships in data
connections
May require more significant computational resources, especially for
Computational Resources Computational resources needed are minimal
large datasets and deep architectures

Table 1.  A comparison table between simple CNN-RNN, CNN-GRU, CNN-LSTM and advanced hybrid
machine learning modified models (HCGRN, HCLN, HCRN).

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 10

Vol:.(1234567890)
www.nature.com/scientificreports/

Case study
Data collected for this study is from a large-scale solar plant. Details discussion of the plant is given in subsec-
tion "Capacity/Structure of solar plant" and data collection, analysis and visualization are provided in subsection
"Data collection process" and "Data analysis and visualization" respectively.

Capacity/Structure of Solar Plant

The current solar plant was developed in different phases and each phase utilizes different equipment. The fur-
ther details of this solar plant have been presented in Table 2. The first phase of this project had an installation
capacity of 100 MW and was completed in 2015. It consisted of more than 400,000 solar panels from various
manufacturers, including JA Solar, Jinko, and Canadian Solar companies. These panels were mounted on fixed-
tilt structures. The inverters from Huawei and SMA companies were used to convert the DC power generated
by the solar panels into AC power which later can be used by the main grid. These inverters were connected to
the solar panels in a centralized architecture which allows better control and monitoring of the overall system.
To manage the flow of power from the solar plant to the grid, the project used a substation that contained trans-
formers from ABB company.
The substation also contained other electrical equipment such as switchgear, protection relays, and control
­systems26. The second phase of the project had an installation capacity of 300 MW and was completed in 2017. It
used more than 1.1 million solar panels from Trina Solar and Canadian Solar companies, which were mounted
on single-axis structures. The third phase of the project was completed in 2019 and had an installed capacity of
600 MW. It used more than 2.4 million solar panels from different manufacturers, including Canadian Solar,
Trina Solar, and Jinko, which were also mounted on fixed-tilt structures. The solar power plant was a significant
achievement for Pakistan and helped the country to further reduce its dependence on fossil fuels. The scale of
the project and size demonstrate the potential for large-scale solar projects in the country and the use of multiple
manufacturers and suppliers shows that Pakistan is diversifying its sources of renewable energy. Figure 6 shows
the fundamental structure of the solar plant.

Data collection process

To gather data from a solar power plant, a diverse range of sensors and monitoring devices must be used as
depicted in Fig. 7. These sensors are used for the measurement of several factors, including daily power produc-
tion, array orientation, and performance ratio. Energy meters are used to gather daily data on the quantity of
electricity produced at the solar power plant. These meters quantify the aggregate amount of energy generated by
the plant. Analyzing the data allows for monitoring the plant’s performance and identifying any potential issues
or malfunctions. Weather sensors mounted on solar panels can identify the plane of the array, which refers to the
angle of the solar panels’ positioning. These figures are crucial for optimizing energy production and achieving
peak performance at the facility. The performance ratio of the plant may be calculated by dividing the actual
energy production by the projected energy output.
This quantifies the plant’s operational efficiency. The inverter monitoring devices are used to measure the
DC voltage and current output of the solar panels and to collect the data. The plant operators have the capability
to optimize the functioning of the solar power plant and ensure that it is generating the maximum quantity of
energy.

Installed Number of solar Mounting MW generation Monitoring

Phase capacity (MW) Solar panels panels inverters systems Transformer Storage capacity (MW) system
JA Solar, Jinko, More than Arctech Solar,
1 100 Huawei,SMM ABB N/A Up to 100 Centralized
Canadian Solar 400,000 Clenergy
Trina Solar, More than 1.1 Huawei, Sun- Nextracker, Array
2 300 Siemens N/A Up to 300 Decentralized
Canadian Solar million grow Technologies
Canadian Solar, More than 2.4 Huawei, Sun- Arctech Solar,
3 600 Siemens N/A Up to 600 Decentralized
Trina Solar, Jinko million grow Nextracker

Table 2.  Present configuration and prospective growth strategies of solar power facilities.

Figure 6.  Fundamental framework and key components of solar power plants.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 11

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 7.  Data collection process and monitoring system of solar plant.

Data analysis and visualization

Data visualization is a multifaceted procedure that entails comprehending the data, the objective of the repre-
sentation, the intended viewers, and the accessible tools and design ideas. By taking into account these variables,
you may generate impactful and significant visual representations that effectively and properly convey your facts.
The correlation heat maps in Fig. 8a and the density heat map in Fig. 8b illustrate the correlation coefficients
between three characteristics of a solar plant’s performance during a year. These parameters are power output
(MWh), plane of array (POA), and performance ratio (PR). The correlation coefficient is a statistical measure of
the strength of the relationship between two variables with a range of negative one to positive one. Here negative
one indicates a perfect negative correlation, zero indicates no correlation and a positive one indicates a perfect
positive correlation. The numbers shown in the heat map indicate the correlation coefficients among these three
parameters. A correlation coefficient of positive one in the center of the map represents a perfect positive rela-
tion between POA and itself. The color scale on the right side of the heat map displays the range of correlation
coefficients. The lowest value of -0.4 indicates a mild negative relation while the maximum value of 1.0 indicates
a perfect positive correlation. The heat map observations indicate a moderate positive relation between MWh
(megawatt hours) and POA (plane of the array), a weak negative correlation between MWh and PR (performance
ratio), and a moderate negative correlation between POA and PR. These correlations may provide valuable
insights into the overall efficiency of the solar plant and facilitate the identification of areas that can be improved.

Figure 8.  (a) Correlation Heat Map, (b) Density Heat Map.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 12

Vol:.(1234567890)
www.nature.com/scientificreports/

A histogram is a useful tool for analyzing the features of a dataset by giving information about its shape, center,
and dispersion. It can detect data patterns like skewness or normal distribution, show outliers or uncommon
characteristics, and compute statistical measures like mean, median, and standard deviation. In general, a histo-
gram represents the data distribution in a clear and simple manner, making it easier to compare various datasets
and acquire insights that may not be immediately evident from raw data. Figure 8 shows the data parameters
solar power generation in (MWh), plane of array (POA) and performance ratio (PR) on the x-axis represents
range values, divided into a set of bins, and the Y-axis represents the frequency of occurrence for each bin. In
Fig. 9 it can be seen clearly that the data used in this study is symmetrical, almost equally distributed, and with
a minimum outlier which indicates that the data set is predictable and consistent to draw valid conclusions.

Innovation and improvements

In this research, notable advancements and innovations were achieved by elevating the conventional CNN-
LSTM, CNN-GRU, and CNN-RNN models to a new echelon through the introduction of hybrid architectures.
These enhanced models, specifically the hybrid CNN-LSTM, hybrid CNN-GRU, and hybrid CNN-RNN, were
meticulously crafted by modifying their structures to amalgamate the strengths of convolutional and recurrent
neural networks. The transformative impact of these modifications was assessed through the extraction of results
derived from real-time data obtained from a solar plant. Significantly, the results exceeded the predictive capa-
bilities of individual models, demonstrating a significant increase in accuracy rates and a noticeable decrease in
error values. This is a significant achievement. Through the presentation of newly developed and enhanced hybrid
models that demonstrate higher performance in forecasting energy output in solar plants, this study represents
an important improvement in this field. As a result, it contributes to the development of predictive modeling
in renewable energy systems. In the next section "Case study", graphical results from the modified models are
discussed with charts and tables.

Results and discussion

The comparison between actual loss and validation loss for the prediction of the MWh parameter using the
Hybrid Convolutional-GRU Net (HCGRN) has been shown in Fig. 10a. The actual loss which represents the loss
on the training dataset and exhibits a consistent downward trend as the number of epochs increases. Starting
at 1.9785 it gradually decreases to 0.0841 by the 99th epoch. This declining pattern signifies that the model is
proficiently assimilating knowledge from the training dataset, resulting in a reduction of error when forecasting
the MWh parameter. On the other hand, the validation loss starts at 2.0746 and varies until it reaches 0.0468
in the last epoch.
Figure 10b is the comparison between actual MAE loss and validation MAE loss for MWh. The actual MAE
loss on the training dataset shows a consistent trend of reduction as the number of epochs increases. Starting

Figure 9.  A quantitative analysis and distribution patterns, advanced histograms depicting key data parameters
in solar power plant.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 13

Vol.:(0123456789)
 www.nature.com/scientificreports/

2.5 1.5
Actual MSE Loss of MWh
Actual Loss of MWh Actual MAE Loss of MWh 2.0
Validation MAE Loss of MWh
Validation MSE Loss of MWh
Validation Loss of MWh
2.0 1.2
1.6

1.5

MSE Loss
0.9

MAE Loss
Loss 1.2

1.0 0.6 0.8

0.5 0.4
0.3

0.0 0.0
0.0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Epochs Epochs Epochs
(a) (b) (c)

Figure 10.  (a) Actual loss vs. Validation loss, (b) MAE loss vs. validation MAE loss and (c) MSE loss vs.
validation MSE loss of MWh through HCGRN.

at 1.3593, it steadily decreases to 0.2139 by the 99th epoch. This shows that the model is efficiently acquiring
knowledge from the training dataset, as the MAE loss decreases significantly. In contrast, the validation MAE
loss, which assesses the model’s performance starts at 1.4264, fluctuates, and ends at 0.1204 in the final epoch.
These fluctuations suggest that the model’s performance on unseen data varies during training.
Figure 10c is the actual MSE loss and validation MSE loss of the MWh parameter. The actual MSE loss, rep-
resenting the mean squared error on the training dataset, steadily decreases as the number of epochs increases.
Starting at 1.9785, it consistently reduces to 0.0841 by the 99th epoch. This implies that the model is proficiently
acquiring knowledge and refining its predictive capabilities through training data, as evidenced by a substantial
reduction in the Mean Squared Error (MSE) loss. Similarly, the validation MSE loss begins at 2.0746, fluctuates
and ends at 0.0468 in the final epoch. These fluctuations suggest that the model’s performance on unseen data
varies during training.
In the context of optimizing the POA parameter prediction through Hybrid Convolutional-GRU Net
(HCGRN), it is crucial to examine the dynamics of actual loss and validation loss over epochs. The training
process in Fig. 11a unfolds with notable trends in both losses. Initially, at epoch 0, the actual loss stands at
2.14545989, while the validation loss is slightly higher at 2.456800461, indicating a disparity between training
and generalization. However, as training progresses, these losses steadily converge. By epoch 3, there is a sub-
stantial drop in both actual loss (0.357102871) and validation loss (0.050259765), signifying improved model
performance and its ability to generalize well. Throughout subsequent epochs, the actual loss and validation
loss maintain a relatively close relationship, underscoring the robustness of the HCGRN model in maintaining
low prediction errors. Notably, both losses exhibit fluctuations but generally remain aligned, indicating that
the model doesn’t overfit the training data. Towards the end of the training process, at epoch 99, the actual loss
is 0.083061673, and the validation loss is 0.056649759, demonstrating a successful optimization of the POA
parameter prediction model through HCGRN.
In Fig. 11b, a comparison is shown between the actual MAE loss and the validation MAE loss across numerous
epochs. The actual MAE loss, starting at 1.41453886 at epoch 0, it consistently decreased, eventually reaching
0.214511037 at epoch 99. This signifies that the model is becoming more adept at minimizing prediction errors
during training. In contrast, a similar pattern was shown by the validation MAE loss, which measures the model’s
performance on untested data. It started at epoch 0 at 1.550364137 and decreased steadily, ending at epoch 99
at 0.165685818. The fact that the validation and actual MAE losses synchronize shows that the HCGRN model
avoided overfitting by successfully extending its predictions to previously unobserved data. Figure 11c of the

3.0 2.5
1.5 Actual MAE Loss of POA
Actual Loss of POA Actual MSE Loss of POA
Validation MAE Loss of POA
Validation Loss of POA Validation MSE Loss of POA
2.4 2.0
1.2

1.8 1.5
MAE Loss

0.9
MSE Loss
Loss

1.2 1.0
0.6

0.6 0.5
0.3

0.0 0.0
0.0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Epochs Epochs Epochs
(a) (b) (c)

Figure 11.  (a) Actual loss vs. Validation loss, (b) MAE loss vs. validation MAE loss and (c) MSE Loss vs.
validation MSE loss of POA through HCGRN.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 14

Vol:.(1234567890)
www.nature.com/scientificreports/

Hybrid Convolutional-GRU Net (HCGRN) displays a comparison between the actual Mean Squared Error (MSE)
loss and the validation MSE loss for the POA parameter.
At epoch 0, the actual mean squared error (MSE) loss was 2.14545989, whereas the validation MSE loss was
2.456800461. During the training process, it was seen that both losses consistently decreased, indicating that
the HCGRN model successfully learned from the data. During epoch 3, there was a notable decrease in both
the actual and validation mean squared error (MSE) losses. The actual MSE loss was 0.357102871, while the
validation MSE loss was 0.050259765. These data suggest that the model is improving in accuracy. The tendency
persisted consistently throughout the training procedure, with the real and validation losses maintaining an
accurate alignment. The convergence of these loss curves indicates that the HCGRN effectively optimized the
prediction of the POA parameter, showcasing its potential for accurate forecasting in real-world situations.
In Fig. 12a, analyzing the performance of the Hybrid Convolutional-GRU Net (HCGRN) for predicting the PR
parameter, a comparison between the actual loss and the validation loss across various epochs is conducted. The
actual loss, which represents the model’s error during training, shows a decreasing trend throughout the training
process. Starting at 1.129761338 during the initial epoch, it consistently diminished, reaching 0.025812617 at
epoch 99. The observed decline in performance suggests that the HCGRN model is effectively acquiring infor-
mation from the given training data, as shown by its gradual reduction of errors. Likewise, the validation loss,
which evaluates the model’s performance on data it hasn’t seen before, displayed a synchronized pattern. Starting
at 0.586367011 in the first epoch, it progressively decreased to 0.01203534 by the 99th epoch. The correlation
between the actual and validation loss indicates that the HCGRN model has the capability to make accurate
predictions on new data, demonstrating its ability to prevent overfitting on the training dataset.
Figure 12b depicts a comparison between the empirical Mean Absolute Error (MAE) loss and the validation
MAE loss over many epochs. The actual MAE loss represents the error incurred by the model during its training
phase, while the validation MAE loss assesses the performance of the model on unseen data. The actual MAE loss
started at a relatively high value of 1.034575105 during the initial epoch and exhibited a consistent downward
trend. It gradually decreased with each subsequent epoch, ultimately reaching a low value of 0.104480341 at
epoch 99. This consistent reduction demonstrates that the HCGRN model is effectively learning and minimizing
errors during the training process, indicating its ability to capture patterns in the training data. Simultaneously,
the validation MAE loss, which assesses the model’s ability to generalize to unknown data, also exhibited a syn-
chronized pattern. The value started at 0.753925204 on the first epoch and gradually declined to 0.085371122
by the 99th epoch. The observed consistency between the mean absolute error (MAE) losses of the real and
validation data indicates that the HCGRN model is not affected by overfitting and has the capability to provide
accurate predictions for unseen data.
The Mean Squared Error (MSE) loss and the validation MSE loss for the PR parameter in the Hybrid Con-
volutional-GRU Net (HCGRN) model can be shown in Fig. 12c. At the beginning of the training process, the
mean squared error (MSE) loss is much larger, namely at 1.129761338. This suggests a significant difference
between the model’s predictions and the actual target values in the training data. Nevertheless, as the training
advanced, the mean squared error (MSE) loss consistently dropped, ultimately achieving a much lower value of
0.025812617 by the 99th epoch. The constant decline seen indicates that the HCGRN model successfully reduced
the squared errors between its predictions and the actual training data, demonstrating its strong learning ability.
Simultaneously, the validation mean squared error (MSE) loss, which evaluates the model’s accuracy on new
data, showed a same trend. The value started at 0.586367011 in the first epoch and consistently decreased, finally
reaching 0.01203534 by epoch 99. The convergence found in the real and validation Mean Square Error (MSE)
losses suggests that the HCGRN model does not suffer from overfitting and has the capacity to make correct
predictions for data that was not included in the training set.
Similarly in Fig. 13a the training of Hybrid Convolutional-LSTM Net (HCLN) model for prediction the
MWh parameter, the comparison between actual loss and validation loss across epochs offers significant infor-
mation into the efficiency and modification capabilities of the model. At the start, at epoch 0, the true loss was
1.95187974, although the validation loss was marginally higher at 2.034267902. This suggests that the model’s
performance on the training data was not ideal, since there was a notable difference between its predictions and
the actual MWh values. As the training progressed, both the actual and validation loss continually reduced. As

1.5
1.0 Actual MAE Loss of PR Actual MSE Loss of PR
Actual Loss of PR Validation MAE Loss of PR Validation MSE Loss of PR
1.0
Validation Loss of PR
1.2
0.8
0.8
0.9
MAE Loss

MSE Loss

0.6 0.6
Loss

0.6
0.4 0.4

0.3
0.2 0.2

0.0 0.0
0.0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Epochs Epochs Epochs
(a) (b) (c)

Figure 12.  (a) Actual loss vs. Validation loss, (b) MAE loss vs. Validation MAE loss and (c) MSE loss vs.
Validation MSE loss of PR through HCGRN.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 15

Vol.:(0123456789)
 www.nature.com/scientificreports/

2.5 1.5
Actual MSE Loss of MWh
Actual Loss of MWh Actual MAE Loss of MWh 2.0
Validation MAE Loss of MWh
Validation MSE Loss of MWh
Validation Loss of MWh
2.0 1.2
1.6

1.5

MSE Loss
0.9

MAE Loss
Loss 1.2

1.0 0.6 0.8

0.5 0.4
0.3

0.0 0.0
0.0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Epochs Epochs Epochs
(a) (b) (c)

Figure 13.  (a) Actual loss vs. Validation loss, (b) MAE loss vs. Validation MAE loss and (c) MSE loss vs.
Validation MSE loss of MWh through HCLN.

the model iteratively processed the data, the actual loss consistently decreased to 0.076261148 at the 99th epoch.
The significant reduction indicates that the HCLN model successfully enhanced its capacity to reduce discrepan-
cies between its forecasts and the real MWh values during the training phase. Simultaneously, the validation loss
exhibited a similar downward trajectory, ultimately reaching a value of 0.044133689 at epoch 99. The convergence
of both the actual and validation loss indicates that the HCLN model effectively learnt from the training data and
shown excellent generalization skills by reaching low loss values on unseen validation data.
Figure 13b displays a comparison examination of the Actual Mean Absolute Error (MAE) loss and validation
MAE loss throughout epochs, providing vital insights into the model’s performance and its ability to generalize to
unseen data. Initially, at epoch 0, the observed real loss is 1.350197196, but the validation loss is slightly greater
which is equal to 1.412231326. At epoch 99, the actual loss dropped significantly to 0.202950507, demonstrat-
ing that the model has greatly improved its capacity to reduce errors when predicting the MWh parameter on
the training data. It is important to mention that the validation loss closely followed the real loss throughout
the training process. At epoch 99, the validation loss decreased to 0.12600483, which nearly matched the actual
loss. The convergence of the real and validation losses indicates that the HCLN model effectively learnt from
the training data and exhibited robust generalization capabilities by reaching minimal loss values on unseen
validation data.
The comparison of the Mean Squared Error (MSE) loss and the validation MSE loss of the MWh parameter
in Fig. 13c during the training of the Hybrid Convolutional-LSTM Net (HCLN) model provides valuable insights
into the model’s performance and its ability to generalize. During the initial training phase (epoch 0), the mean
squared error (MSE) loss is 1.95187974, whereas the validation MSE loss is 2.034267902. The first loss values
indicate that the model is encountering difficulty in accurately capturing the fundamental patterns in the train-
ing data, resulting in some significant errors in both the training and validation sets. During the course of the
training, as each epoch passed, both the mean squared error (MSE) losses for the real data and the validation
data consistently decreased. This trend indicates that the model is learning from the training data and gradually
improving its performance in terms of minimizing the mean squared error. By the end of the training process,
at epoch 99, the actual MSE loss had decreased significantly to 0.076261148, and the validation MSE loss also
reached a low value of 0.044133689. This convergence of the two loss values suggests that the HCLN model
learned from the training data and demonstrated strong generalization capabilities by achieving similarly low
MSE values on the unseen validation data.
Figure 14a is the actual loss to the validation loss of the POA parameter during the training of the Hybrid
Convolutional-LSTM Net (HCLN). At the beginning of training (epoch 0), the actual loss is 2.147356033, while

3.0
1.5 Actual MAE Loss of POA 2.5 Actual MSE Loss of POA
Actual Loss of POA Validation MAE Loss of POA Validation MSE Loss of POA
Validation Loss of POA
2.4 2.0
1.2

1.8
MAE Loss

MSE Loss

0.9 1.5
Loss

1.2 0.6 1.0

0.6 0.5
0.3

0.0 0.0
0.0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Epochs Epochs Epochs
(a) (b) (c)

Figure 14.  (a) Actual loss vs. Validation loss, (b) MAE loss vs. Validation MAE loss and (c) MSE loss vs.
Validation MSE Loss of POA through HCLN.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 16

Vol:.(1234567890)
www.nature.com/scientificreports/

the validation loss is 2.503918409. These initial loss values indicate that the model has a relatively high error rate
on both, training as well as validation datasets, suggesting that the model’s performance needs improvement. As
training proceeds through subsequent epochs, both the actual and validation losses exhibit a consistent decreas-
ing trend. This graph indicates that the model is learning from the training data, since the error is decreasing.
By the end of the training process, at epoch 99, the actual loss had decreased significantly to 0.077837422, and
the validation loss had also decreased to 0.065957859. The convergence of the actual and validation loss func-
tions signifies that the model has effectively captured the inherent data patterns, allowing it to produce accurate
predictions, not only on the training dataset but also on previously unseen validation data.
In Fig. 14b, at the beginning of training (epoch 0), the actual MAE loss is 1.414865255, and the validation
MAE loss is 1.565488696. These initial losses indicate that the model has a relatively high error rate on both the
training and validation datasets. As training starts through subsequent epochs, both the actual and validation
MAE losses consistently decrease. By the end of the training process, at epoch 99, the actual MAE loss has signifi-
cantly decreased to 0.204322264, and the validation MAE loss has also reduced to 0.198790282. The alignment
between the actual and validation MAE losses signifies that the model has effectively acquired the inherent data
patterns, enabling it to generate accurate predictions, not only within the confines of the training dataset but
also when applied to previously unseen validation data.
Figure 14c is the comparison of the actual Mean Squared Error (MSE) loss versus the validation MSE loss
for the POA parameter during the training of the Hybrid Convolutional-LSTM Net (HCLN) gives significant
information about the model’s accuracy and capacity to generalize. At the beginning of training (epoch 0), the
actual MSE loss is 2.147356033, while the validation MSE loss is 2.503918409. These initial losses indicate a
relatively high error rate on both the training and validation datasets, suggesting that the model has substantial
room for improvement. As the training starts through subsequent epochs, both the actual and validation MSE
losses consistently decreased. This pattern shows that the model is improving its performance as a result of being
exposed to training data. By the end of the training process, at epoch 99, the actual MSE loss has significantly
decreased to 0.077837422, and the validation MSE loss has also reduced to 0.065957859. This convergence of
the actual and validation MSE losses highlights the model’s successful learning of the underlying patterns in
the data and its capacity to make accurate predictions, not only on the training dataset but also on previously
unseen validation data.
Figure 15a is the actual loss versus the validation loss for the PR through training the Hybrid Convolutional-
LSTM Net (HCLN). At the outset of training (epoch 0), the actual loss was 1.183701396, and the validation loss
is 0.638390422.
These initial losses indicate that the model’s performance on the training dataset is relatively higher compared
to the validation dataset. Such a discrepancy is common in the early stages of training, as the model has yet to
fully adapt to the data’s underlying patterns. As training progresses, the actual loss consistently decreases with
each epoch, suggesting that the model is learning from the training data and optimizing its parameters effectively.
Meanwhile, the validation loss also shows a similar decreasing trend. This alignment between actual and valida-
tion losses indicates that the model is generalizing well to unseen data, which is a positive sign of its robustness.
By the end of the training process, at epoch 99, both the actual loss and validation loss had converged to similar
values, with the actual loss at 0.026409347 and the validation loss at 0.01381046. This convergence signifies that
the model had effectively learned the underlying patterns in the data and can make accurate predictions, not
only on the training data but also on unseen validation data.
Similarly, Fig. 15b is the actual Mean Absolute Error (MAE) loss versus the validation MAE loss for the PR
through training the Hybrid Convolutional-LSTM Net (HCLN). At the beginning of training (epoch 0), the
actual MAE loss is 1.059277177, while the validation MAE loss is 0.787348688. These initial losses showed that
the model performed better on the training data than on the validation data, which is a common occurrence as
the model starts learning from the training set. As training proceeds, both the actual MAE loss and validation
MAE loss consistently decrease with each epoch. This decrease indicates that the model is efficiently acquir-
ing the fundamental patterns in the data and is improving its performance on both the training and validation
datasets. The congruence between the actual and validation losses is an encouraging indication, indicating
that the model is effectively extrapolating to unfamiliar data. At epoch 99, the training process reached a point

1.5
1.0 Actual MAE Loss of PR Actual MSE Loss of PR
Actual Loss of PR Validation MAE Loss of PR Validation MSE Loss of PR
1.0
Validation Loss of PR
1.2
0.8
0.8
0.9
MAE Loss

MSE Loss

0.6 0.6
Loss

0.6
0.4 0.4

0.3
0.2 0.2

0.0 0.0
0.0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Epochs Epochs Epochs
(a) (b) (c)

Figure 15.  (a) Actual loss vs Validation Loss, (b) MAE Loss vs Validation MAE Loss and (c) MSE loss vs.
Validation MSE Loss of PR through HCLN.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 17

Vol.:(0123456789)
 www.nature.com/scientificreports/

where both the actual MAE loss and validation MAE loss had converged to identical values. The real MAE loss
is 0.10242752 and the validation MAE loss is 0.094468474. The convergence demonstrates that the model has
effectively understood the essential data patterns, allowing it to make accurate predictions, not just on the train-
ing dataset but also on unseen validation data.
Figure 15c displays the plot of the Mean Squared Error (MSE) loss compared to the validation MSE loss for
the parameter PR during the training of the Hybrid Convolutional-LSTM Net (HCLN). During the first stage
of training (epoch 0), the mean squared error (MSE) loss is 1.183701396 for the real data, but the validation
data has an MSE loss of 0.638390422. The first discrepancy suggests that the model’s performance is superior on
the training data compared to the validation data, which is a typical occurrence when the model starts to learn
from the training set. During the training process, the mean squared error (MSE) loss for both the real data and
the validation data continuously decreases with each epoch. The declining pattern indicates that the model is
successfully acquiring the fundamental patterns in the data and improving its performance on both the training
and validation datasets. The convergence of the real and validation losses indicates that the model is effectively
generalizing to unseen data. At epoch 99, the real mean squared error (MSE) loss and the validation MSE loss
both reach identical levels, indicating convergence towards the conclusion of the training procedure. The current
mean squared error (MSE) loss is 0.026409347, whereas the validation MSE loss is 0.01381046. The convergence
showcases the model’s proficiency in understanding the underlying data patterns, allowing it to provide accurate
predictions not only on the training data but also on the previously unknown validation data.
In Fig. 16a the evaluation of the MWh parameter through the Hybrid Convolutional-Recurrence Net
(HCRN), it is evident that there are significant differences between the Actual Loss and Validation Loss over the
course of training. The Actual Loss, which represents the model’s performance on the training data, decreases
steadily as the number of epochs increases. Commencing at 0.3257 and progressively diminishing to 0.0385,
this trend signifies the model’s effective learning from the training dataset and performance enhancement. In
contrast, the validation Loss, evaluating the model’s generalization to unseen data, exhibits a more oscillatory
pattern. Initiating at 0.1145, it initially decreases but subsequently displays fluctuations during training. This
fluctuation suggests the possibility of overfitting, since the model’s consistent performance on the validation
data is not sustained. Significantly, there are occurrences when the validation loss experiences sudden increases,
namely around epoch 58, reaching a value of 0.1421, which suggests possible problems with the model’s ability
to generalize. In general, the comparison between Actual Loss and Validation Loss indicates that the model is
effectively acquiring knowledge from the training data. However, there is still potential for enhancing its ability
to apply this knowledge to unfamiliar data.
In the same way, Fig. 16b presents the mean absolute error (MAE) loss and validation MAE Loss for the
MWh parameter estimated by the Hybrid Convolutional-Recurrence Net (HCRN) model. The MAE loss, which
measures the average absolute difference between the predicted and actual values on the training data, steadily
decreases from 0.3858 to 0.1357 as the number of epochs increases. This indicates that the model is effectively
minimizing errors and improving its accuracy on the training data. While the validation MAE loss which evalu-
ates the model’s performance on unseen data, also demonstrates a decreasing trend from 0.2800 to 0.2109 but
exhibits more fluctuations compared to the MAE Loss. Initially, the validation MAE loss is higher than the MAE
loss, indicating that the model struggles with generalization. However, as training progresses, the gap between
the two losses narrows, suggesting that the model is learning to predict better over time.
Figure 16c depicts the comparative analysis between the Mean Squared Error (MSE) Loss and the Validation
MSE Loss concerning the MWh parameter determined by the Hybrid Convolutional-Recurrence Net (HCRN)
model. The MSE Loss, representing the average squared difference between the predicted and actual values on
the training data, decreases progressively from 0.3257 to 0.0385 as the number of epochs increases. This steady
decline indicates that the model is effectively minimizing the errors and improving its predictive accuracy on
the training dataset. In contrast, the Validation MSE Loss, which assesses the model’s performance on unseen
data, follows a similar decreasing trend but shows more fluctuation compared to the MSE Loss. Initially, the
Validation MSE Loss is higher than the MSE Loss, suggesting that the model struggles with generalization,
which is not uncommon during early training epochs. However, as training progresses, the Validation MSE
Loss gradually converges toward the MSE Loss, indicating an improvement in the model’s ability to generalize.

1.0 1.2
1.0 Actual Loss of MWh Actual MAE Loss of MWh Actual MSE Loss of MWh
Validation Loss of MWh Validation MAE Loss of MWh Validation MSE Loss of MWh
1.0
0.8 0.8

0.8
0.6 0.6
MAE Loss

MSE Loss

0.6
Loss

0.4
0.4
0.4
0.2
0.2 0.2
0.0
0.0
0.0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Epochs Epochs Epochs
(a) (b) (c)

Figure 16.  (a) Actual loss vs Validation Loss, (b) MAE Loss vs Validation MAE Loss and (c) MSE loss vs.
Validation MSE Loss of MWh through HCRN.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 18

Vol:.(1234567890)
www.nature.com/scientificreports/

Hence, the decreasing trends in both MSE Loss and Validation MSE Loss signify that the model is learning and
enhancing its performance over time. The fluctuations in Validation MSE Loss suggest that there may be some
initial overfitting, but the model eventually learns to generalize better.
The comparative analysis of the Mean Squared Error (MSE) Loss and Validation MSE Loss for the MWh
parameter during the training of the Hybrid Convolutional-Recurrence Net (HCRN) provides insightful obser-
vations into the model’s performance and its generalization capability, as shown in Fig. 17a. At the outset of
training, specifically in epoch 0, the actual loss records 1.03837204, while the validation loss is notably lower at
0.049138479. This initial contrast indicates the model’s superior performance on the training dataset, a common
occurrence as it begins to grasp the training-specific data patterns. With the progression of training, both actual
loss and validation loss exhibit consistent reduction in each subsequent epoch. This diminishing trend signifies
the model’s effective acquisition of underlying data patterns, resulting in enhanced performance on both the
training and validation datasets. The convergence of these loss metrics strongly suggests the model’s proficiency
in generalizing to previously unseen data. By the conclusion of the training process, at epoch 99, the actual loss
and validation loss have both reached comparable values of 0.080057055 and 0.076434441, respectively. This
convergence underscores the model’s successful mastery of underlying data patterns, enabling accurate predic-
tions not only on the training dataset but also on novel validation data.
Figure 17b offers a comparative analysis of the Mean Absolute Error (MAE) losses, both actual and validation,
pertaining to the MWh parameter during the training of the Hybrid Convolutional-Recurrence Net (HCRN).
This illustration provides valuable insights into the model’s performance and its ability to generalize to previously
unencountered data. At the commencement of training, precisely in epoch 0, the actual MAE loss is recorded
at 0.862085044, while the validation MAE loss is notably lower at 0.123599067. The substantial gap observed
at this stage strongly implies that the model is exhibiting signs of overfitting to the training data, resulting in a
considerably lower MAE for the training dataset compared to the validation dataset. Overfitting transpires when
a model excessively molds itself to the training data, impairing its capacity to effectively generalize to novel, unob-
served data. Nevertheless, as the training regimen advances, both actual and validation MAE losses consistently
ameliorate with the progression of epochs, signifying that the model is progressively assimilating the latent data
patterns and improving its predictive prowess. The consistent diminishing of these loss metrics indicates the
model’s diminishing overfitting tendencies and its increasing capacity for generalization. By the culmination of
the training process, occurring at epoch 99, both the actual and validation MAE losses have reached congru-
ent values of 0.211874843 and 0.201278687, respectively. This convergence underlines the model’s triumphant
acquisition of the underlying data patterns, endowing it with the capability to render precise predictions not
only on the training dataset but also on hitherto unexplored validation data.
Figure 17c elucidates a comparative analysis involving the Mean Squared Error (MSE) losses, encompassing
both actual and validation aspects, pertaining to the MWh parameter. This analysis is conducted within the con-
text of training the Hybrid Convolutional-Recurrence Net (HCRN), offering invaluable insights into the model’s
performance and its adeptness at generalization. Commencing with the initial stages of training, specifically
during epoch 0, the actual MSE loss registers a notably higher value, specifically 1.03837204, in contrast to the
validation MSE loss, which stands at 0.049138479. This pronounced disparity serves as a significant indicator
that the model, during its inception, is prone to overfitting to the training data. This results in a substantially
lower MSE when applied to the training dataset as opposed to the validation dataset.
Overfitting is a recurrent issue where the model excessively focuses on the noise and intricacies inherent
within the training data, thus impeding its ability to effectively generalize its predictions to unexplored data.
Nevertheless, as the training regime progresses, both the actual and validation MSE losses consistently record
declines with each passing epoch. This trajectory highlights the model’s ongoing process of actively acquiring
the underlying data patterns and simultaneously improving its forecasting abilities. The simultaneous reduc-
tion in both loss measures indicates that the model is effectively reducing overfitting and improving its ability
to adapt. At epoch 99, after completing the training procedure, the actual mean squared error (MSE) loss and
validation MSE loss show a convergence towards remarkably similar values. The current mean squared error
(MSE) loss is 0.080057055, whereas the validation MSE loss is 0.076434441. The convergence indicates that the
model has successfully identified the underlying data patterns and is effective at providing accurate predictions.

1.0 1.5
1.2 Actual Loss of POA Actual MAE Loss of POA Actual MSE Loss of POA
Validation Loss of POA Validation MAE Loss of POA Validation MSE Loss of POA
0.8 1.2
0.9

0.6 0.9
MAE Loss

MSE Loss
Loss

0.6
0.6
0.4
0.3
0.3
0.2

0.0 0.0
0.0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Epochs Epochs Epochs
(a) (b) (c)

Figure 17.  (a) Actual loss vs Validation Loss, (b) MAE Loss vs Validation MAE Loss and (c) MSE loss vs.
Validation MSE Loss of POA through HCRN.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 19

Vol.:(0123456789)
 www.nature.com/scientificreports/

This ability goes beyond the scope of the training dataset, demonstrating its usefulness in accurately predicting
unknown validation data.
Figure 18a illustrates a comparison of loss metrics, including both actual and validation, especially within
the PR parameter domain. This study occurs during the training phase of the Hybrid Convolutional-Recurrence
Net (HCRN), offering deep insights into the model’s performance and its ability to successfully generalize the
new data. Starting from the beginning, both the real loss and validation loss begin at rather high levels. The first
loss for the real data is 0.325732052, whereas the validation data begins with a loss of 0.114503279. These early
points indicate a clear restriction in the model’s ability to provide accurate predictions for the PR parameter. As
the training process progresses, both loss indicators consistently show a lower trend. This gradual decline sug-
gests that the model is actively identifying and absorbing the underlying data patterns. Moreover, throughout
the whole training process, the real loss and validation loss closely resemble each other. This alignment serves
as a strong signal that the model avoids overfitting to the training data, as seen by the agreement between the
validation loss and the actual loss. The progressive decrease seen in both loss metrics highlights the model’s
effectiveness in both acquiring information from the training data and later using this knowledge to make cor-
rect predictions when tested on unknown validation data. As the training process nears its end, namely around
epoch 99, both the real loss and validation loss converge towards lower values that are equivalent to each other.
The loss value reaches 0.038496502, and the validation loss stabilizes at 0.089028001. This convergence demon-
strates the model’s successful incorporation of the underlying data patterns and its intrinsic capacity to provide
accurate predictions for the PR parameter.
Figure 18b displays the Mean Absolute Error (MAE) loss for the PR parameter in the training process of the
Hybrid Convolutional-Recurrence Net (HCRN) model. This loss includes both the real and validation parts.
This comparison provides a deep perspective for understanding the effectiveness of the model and its ability to
extrapolate to new data that has not been seen before. At the nascent stage of the training process (epoch 0), the
actual MAE loss manifests a relatively higher value, registering at 0.385757178, in contrast to the slightly lower
figure of 0.279971421 observed in the validation MAE loss. This initial disparity proffers an indication that the
model might be moderately susceptible to overfitting in the initial phases of training, as it exhibits superior
performance concerning the training data when juxtaposed with its performance on the concealed validation
data. Nonetheless, with the maturation of the training process, both the actual MAE loss and validation MAE
loss consistently witness a decline, denoting that the model is in the process of honing its capability to formulate
more precise predictions concerning the PR parameter.
Notably, these loss metrics closely shadow each other, delineating the model’s congruence in terms of perfor-
mance on the validation data as compared to its performance on the training data. This alignment emphasizes
the idea that the model, instead of being affected by overfitting, is actively focused on understanding and incor-
porating important patterns in the data. As the training trip nears its completion (at epoch 99), both the real
MAE loss and validation MAE loss converge towards similar reduced levels. More precisely, the mean absolute
error (MAE) loss reaches a value of 0.13570945, whilst the validation MAE loss stabilizes at 0.210953534. This
convergence demonstrates the model’s successful incorporation of the underlying data patterns, confirming its
ability to provide accurate predictions about the PR parameter.
In Fig. 18c, the contrast between the genuine Mean Squared Error (MSE) loss and the validation MSE loss,
both pertinent to the PR parameter, unfolds an array of salient insights into the performance of the Hybrid
Convolutional-Recurrence Net (HCRN) during its training phase, particularly regarding its capacity to general-
ize effectively. At the onset of the training process, precisely at epoch 0, the actual MSE loss registers a relatively
higher value of 0.325732052, while the validation MSE loss concurrently manifests a lower figure of 0.114503279.
This initial divergence hints at the possibility of a moderate degree of overfitting, where the model’s performance
appears notably superior when gauged against the training data in comparison to its performance with respect to
the concealed validation data. However, as the training regimen advances, both the actual MSE loss and validation
MSE loss consistently exhibit a diminishing trend. This observation denotes the model’s ongoing refinement in
formulating increasingly accurate predictions concerning the PR parameter. Significantly, this downward trajec-
tory of the losses underscores the model’s active learning from the training data and its remarkable proficiency in
generalizing its acquired knowledge to the validation data. The harmonization observed between the training and
validation losses stands as an encouraging indicator, suggesting that the model is not afflicted by a conspicuous

0.4 0.5 0.5

Actual Loss of PR Actual MAE Loss of PR Actual MSE Loss of PR
Validation Loss of PR Validation MAE Loss of PR Validation MSE Loss of PR
0.4 0.4
0.3

0.3 0.3
MAE Loss

MSE Loss
Loss

0.2
0.2 0.2

0.1
0.1 0.1

0.0 0.0 0.0

0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Epochs Epochs Epochs
(a) (b) (c)

Figure 18.  (a) Actual loss vs Validation Loss, (b) MAE Loss vs Validation MAE Loss and (c) MSE loss vs.
Validation MSE Loss of PR through HCRN.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 20

Vol:.(1234567890)
www.nature.com/scientificreports/

overfitting dilemma. As the training approaches its culmination, precisely at epoch 99, both the actual MSE loss
and validation MSE loss converge to lower values. More precisely, the mean squared error (MSE) loss reaches a
convergence point of 0.038496502, whereas the validation MSE loss stabilizes at 0.089028001. This convergence
clearly demonstrates the model’s ability to accurately capture the underlying data patterns, hence confirming its
capability to provide accurate predictions about the PR parameter.
Figure 19 presents the findings in a graphical format. The y-axis represents energy values, ranging from 0 to
700 MWh. The x-axis represents the number of days, ranging from 0 to 400. The first 40 days are allocated for
test results, shown by the red line on the graph. The blue line represents the MWh estimates made by HCGRN,
offering a visual representation of the model’s forecasting skills. Similarly, the black line depicts the MWh fore-
casts obtained via HCLN, demonstrating an alternate method of prediction.
Finally, the pink line represents the MWh estimations produced by HCRN, providing an additional viewpoint
on energy forecasts. This graph is a useful tool for evaluating the precision and efficiency of several prediction
models over a period of time. The HCGRN, HCLN, and HCRN columns of the dataset are individually analyzed
and their statistical measurements are reported in Table 3. The HCGRN column exhibits a range of data points,
with the lowest value estimated to be around 85.26 and the highest value estimated to be about 574.38. The aver-
age value for this model is almost equal to 463.71. The column data for HCLN model ranges from a minimum
value of roughly 113.08 to a highest value of around 627.37 and has a mean value of approximately 481.71.
The column for HCRN model includes values ranging from around 79.06 to almost 592.90, with a median
value of approximately 467.89. These statistical metrics provide an understanding of the characteristics of the
dataset giving insights on the data’s range and central tendency.
Figure 20 is the graphical representation of POA testing and predictive performance through HCGRN,
HCLN, and HCRN Models, the y-axis measures the Plane of Array (POA) ranging from 0 to 8, while the x-axis
represents the number of days spanning from 0 to 400. The initial 40 days are dedicated to test results, indicated
by the red line. Subsequently, the blue line charts the POA predictions generated by the HCGRN model, the
black line illustrates predictions from the HCLN model, and the magenta line displays predictions produced by
the HCRN model. This graph provides a clear visual overview of how these various models forecast the Plane of
Array values over time, aiding in the assessment and comparison of their predictive capabilities.
For each of the three neural network models, the Table 4 provides daily predictions of some parameter
(presumably the Point of Arrival, or POA) for a PV system. HCGRN (Hybrid Convolutional-GRU Net): The
maximum POA prediction in this model is approximately 7.15, the minimum is around 2.47, and the mean (aver-
age) POA prediction is roughly 6.13. HCLN (Hybrid Convolutional-LSTM Net): For this model, the maximum
POA prediction is about 7.55, the minimum is around 1.04, and the mean POA prediction is approximately 6.12.
HCRN (Hybrid Convolutional-Recurrence Net): In this model, the maximum POA prediction is about 7.41, the
minimum is approximately 0.70, and the mean POA prediction is around 6.34. These statistics provide insights
into the performance and variability of each neural network architecture in predicting the POA parameter for
the PV system over the 404-day period. It seems that HCRN had the lowest minimum forecasts while HCLN
had the greatest maximum predictions. For each model, the mean values provide an overall impression of the

700

600

500
MWh

400

300
Test
200
MWh through HCGRN
MWh through HCLN
100 MWh through HCRN

0
0 100 200 300 400
Days

Figure 19.  Graphical representation of MWh testing and predictive performance using enhanced HCGRN,
HCLN, and HCRN models.

Metric HCGRN HCLN HCRN

Maximum 574.37842 627.36517 592.89856
Minimum 85.2552 113.08041 79.05732
Mean 463.711082 481.713238 467.891581

Table 3.  Maximum, minimum, and average values of MWh parameter after testing and predictive
performance using enhanced HCGRN, HCLN, and HCRN models.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 21

Vol.:(0123456789)
 www.nature.com/scientificreports/

8
7
6
5

POA
4
3 Test
POA through HCGRN
2 POA through HCLN
1 POA through HCRN

0
0 100 200 300 400
Days

Figure 20.  Graphical representation of POA testing and predictive performance using enhanced HCGRN,
HCLN, and HCRN models.

Metric Maximum Minimum Mean

HCGRN 7.15084 2.47586 6.13084
HCLN 7.55692 1.04586 6.12293
HCRN 7.40918 0.69586 6.34151

Table 4.  Maximum, minimum and average values of POA parameter after testing and predictive performance
using enhanced HCGRN, HCLN, and HCRN models.

central tendency of the forecasts. To find out which model performed the best or if their performance differs
noticeably depending on the scenario and for this further investigation would be required. Table 4 displays the
mean, minimum, and maximum values for each of the three prediction models (HCRN, HCLN, and HCGRN)
according to the provided dataset.
Figure 21 depicts a visual representation illustrating the testing and prediction performance of the parameter
PR by the models HCGRN, HCLN and HCRN. The Performance Ratio (PR) values range from 60 to 90 on the
y axis, while the x axis represents the number of days from 0 to 400. The initial forty days are allocated for test
results denoted by the red line. Additionally, PR forecasts derived from the HCGRN model are represented
by the blue line, those from the HCLN model by the black line and predictions from the HCRN model by the
magenta line.
This graph facilitates a direct comparison of performance and prediction accuracy across these various mod-
els. Furthermore, it offers a comprehensive visual depiction of how these models forecast Performance Ratio
(PR) values over time. Table 5 presents the statistics for these models namely, HCGRN, HCLN, and HCRN. In
the Maximum Value column, HCGRN achieved a peak of 85.64155, HCLN reached 87.13593 and HCRN hit
87.01641 as their highest value. Conversely, in the Minimum Value column, HCGRN scored a minimum of

90

85

80
PR

75

70 Test
PR through HCGRN
PR through HCLN
65 PR through HCRN

60
0 100 200 300 400
Days

Figure 21.  Graphical representation of PR testing and predictive performance using enhanced HCGRN,
HCLN, and HCRN models.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 22

Vol:.(1234567890)
www.nature.com/scientificreports/

Statistic HCGRN HCLN HCRN

Maximum value 85.64155 87.13593 87.01641
Minimum value 70.31578 61.31177 61.3588
Mean value 75.83582791 75.87467544 75.97012276

Table 5.  Maximum, Minimum, and Average values of POA parameter after Testing and Predictive
Performance Using Enhanced HCGRN, HCLN, and HCRN Models.

70.31578, HCLN at 61.31177, and HCRN at 61.3588. The Mean Value column indicates the average values over
a specific period, approximately 75.836 for HCGRN, around 75.875 for HCLN, and about 75.970 for HCRN.
These data points provide insights into the performance and variability of each model. Table 5 represents the
highest, lowest, and average values of the datasets derived from the prediction models.
These results section also provides a tabular form that outlines the computational load of each hybrid
machine-learning model. Table 6 presents data on the duration and memory consumption of three different
models: HCRN, HCGRN, and HCLN. The HCRN model stands out for its very quick time elapsed of 120.1 s,
while also using a comparatively minimal amount of memory resources, namely 0.8 Mbytes. By contrast, the
HCGRN model has a little longer processing time of 128.5 s, while also significantly increasing memory use to
0.9 Mbytes. Conversely, the HCLN model has the highest computing demands, taking 189.4 s to complete and
using 0.9 Mbytes of memory resources. This table provides a concise summary of the computational expense
associated with each hybrid machine-learning model, making it easier to compare and analyze different models
based on their computational efficiency.
The prediction performance comparison of the models HCGRN, HCLN, and HCRN across three parameters
(MWh, POA, PR) using the error metrics MSE and MAE reveals distinct performance differences. HCLN con-
sistently exceeds the other models, obtaining the lowest MSE values of 0.012027 for MWh, 0.013734 for POA,
and 0.003055 for PR, as well as the lowest MAE values of 0.069523 for MWh, 0.082813 for POA, and 0.042815
for PR. HCGRN shows moderate performance, with MSE values of 0.012259 for MWh, 0.014046 for POA, and
0.003170 for PR, and MAE values of 0.082857 for MWh, 0.085446 for POA, and 0.043418 for PR. Furthermore,
HCRN generally performs the least efficient, having the highest MSE values of 0.012613 for MWh, 0.014500
for POA, and 0.003287 for PR, as well as the highest MAE values of 0.084669 for MWh, 0.085291 for POA, and
0.043226 for PR.
Upon careful examination and analysis of the results obtained from these models, it has been determined
that they possess certain limitations. There are numerous hurdles involved in using these systems for real-time
data prediction purposes. Firstly, these models require considerable computational resources, which may restrict
their practicality in real-time applications that require prompt decision-making. Moreover, the longer duration
of training required for these advanced models might affect their ability to provide real-time predictions, hence
impacting their practical applicability. Likewise, the complexity linked to the implementation and deployment
of these models presents additional challenges, particularly in situations with limited resources where simplicity
and efficiency are of the highest priority. In order to maximize the capabilities of advanced hybrid models for
real-time prediction tasks, it is essential to address these issues in the future.

Conclusion
The integration of multiple machine learning models, such as the Hybrid Convolutional-GRU Net (HCGRN),
Hybrid Convolutional-LSTM Net (HCLN), and Hybrid Convolutional-Recurrence Net (HCRN), provided a
promising prospect for accurate energy production forecasts. The models demonstrated outstanding predictive
efficiency via comprehensive assessments of important characteristics of a solar plant, such as electricity output
(MWh), plane of array (POA), and performance ratio (PR). The models provided an opportunity to greatly
improve the efficiency of solar power generating systems, with an average output of MWh ranging from around
463.71 to 592.90. The examination of POA values further shown their capacity to consistently forecast values
within a range of 1.04 to 7.55 (HCLN), 0.70 to 7.41 (HCRN), and 2.47 to 7.15 (HCGRN). The models HCGRN,
HCLN, and HCRN showed great potential for smart grids. They were able to achieve high intended PR outcomes,
with maximum values of 87.13593, 87.01641, and 85.64155, while still retaining relatively low minimum values of
61.31177, 61.3588, and 70.31578. Therefore, these results demonstrated that HCLN performed better and high-
lighted the ability to effectively simplify energy demand management and accelerate the shift towards a cleaner
and more sustainable energy environment. We will propose future research directions to further and broaden the
use of these promising models, using their revolutionary capacities in the field of integrating renewable energy.

Model Time elapsed (s) Memory resource utilized (Mbytes)

HCRN 120.1 0.8
HCGRN 128.5 0.9
HCLN 189.4 0.9

Table 6.  The computational cost of various models.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 23

Vol.:(0123456789)
 www.nature.com/scientificreports/

Data availability
The datasets used and/or analysed during the current study available from the corresponding author on reason-
able request.

Received: 18 March 2024; Accepted: 18 July 2024

References:
1. Zhang, P., Huang, W., Chen, Y. & Zhou, M. Predicting Quality of Services based on a two-stream deep learning model with user
and service graphs. IEEE Trans. Serv. Comput. 16, 4060 (2023).
2. Lv, Z., Chen, D., Cao, B., Song, H. & Lv, H. Secure deep learning in defense in deep-learning-as-a-service computing systems in
digital twins. IEEE Trans. Comput. (2021).
3. Wang, X., Sun, Z., Chehri, A., Jeon, G. & Song, Y. a novel attention-driven framework for unsupervised pedestrian re-identification
with clustering optimization. Pattern Recognit. 146, 110045 (2024).
4. Yan, J. et al. Double Deep Q-Network based joint edge caching and content recommendation with inconsistent file sizes in fog-
RANs. IEEE Trans. Vehicular Technol. (2023).
5. Tsakanikas, V., Dagiuklas, T., Iqbal, M., Wang, X. & Mumtaz, S. An intelligent model for supporting edge migration for virtual
function chains in next generation Internet of Things. Sci. Rep. (2023).
6. Yang, J. et al. Prediction and control of water quality in Recirculating Aquaculture System based on hybrid neural network. Eng.
Appl. Artif. Intell. 121, 106002 (2023).
7. He, J. & Cheng, M. Graph convolutional neural networks for power line outage identification. 25th International Conference on
Pattern Recognition (ICPR) (2021).
8. Wu, Z., Pan, S., Long, G., Jiang, J. & Zhang, C. Beyond low-pass filtering: Graph convolutional networks with automatic filtering.
IEEE Trans. Knowl. Data Eng. (2022).
9. Amin, M. R. & Wollenberg, B. F. Toward a smart grid: Power delivery for the 21st century. IEEE Power Energy Mag. 3(5), 34–41.
https://​doi.​org/​10.​1109/​MPAE.​2005.​15001​34 (2005).
10. Hamidian, M. & Khorram, E. Smart grid architecture and applications: A review. Renew. Sustain. Energy Rev. 99, 43–58. https://​
doi.​org/​10.​1016/j.​rser.​2018.​10.​027 (2019).
11. Patnaik, S. & Singh, V. K. Predicting renewable energy generation: A review of machine learning models. Renew. Sustain. Energy
Rev. 107, 407–438. https://​doi.​org/​10.​1016/j.​rser.​2019.​02.​011 (2019).
12. Abubakar, M., Che, Y., Faheem, M., Bhutta, M. S. & Mudasar, A. Q. Intelligent modeling and optimization of solar plant production
integration in the smart grid using machine learning models. Adv. Energy Sustain. Res. 5, 2300160 (2024).
13. Sun, L., Ren, H., Yang, S. & Zhao, X. Short-term electricity demand forecasting based on machine learning: A review and evalu-
ation. Renew. Sustain. Energy Rev. 81, 1548–1568. https://​doi.​org/​10.​1016/j.​rser.​2017.​06.​068 (2018).
14. Du, Y. & Wang, J. A novel hybrid model of convolutional neural network and recurrent neural network for traffic flow prediction.
Expert Syst. Appl. 96, 246–255. https://​doi.​org/​10.​1016/j.​eswa.​2018.​01.​013 (2018).
15. Qi, L. & Zhang, Y. A novel hybrid model combining convolutional neural network and long short-term memory for traffic flow
prediction. Neurocomputing 291, 87–93. https://​doi.​org/​10.​1016/j.​neucom.​2018.​02.​014 (2018).
16. Xu, X., Li, Y., Wang, D. & Wang, C. A hybrid CNN-LSTM model for electricity price prediction. Appl. Energy 242, 728–738. https://​
doi.​org/​10.​1016/j.​apene​rgy.​2019.​03.​160 (2019).
17. Shen, Y., Zhang, Q. & Zhang, X. A novel hybrid model combining random forest and long short-term memory for traffic flow
forecasting (2019).
18. Xu, Z., Sun, Y. & Liu, S. A hybrid model combining convolutional neural network and autoregressive integrated moving average
for electricity demand forecasting. Appl. Energy 242, 673–686. https://​doi.​org/​10.​1016/j.​apene​rgy.​2019.​03.​090 (2019).
19. Tavana, M., Abbasi, A. & Ebrahimzadeh, A. A hybrid CNN-SVR approach for wind speed forecasting. Renew. Energy 116, 105–116.
https://​doi.​org/​10.​1016/j.​renene.​2017.​09.​011 (2018).
20. Yin, X. & Wu, J. Short-term wind power forecasting using a novel hybrid model. Appl. Energy 240, 371–384. https://​doi.​org/​10.​
1016/j.​apene​rgy.​2019.​02.​089 (2019).
21. Khosravi, A., Nahavandi, S., Creighton, D. & Swiegers, G. Wind turbine condition monitoring and fault diagnosis using convo-
lutional neural networks and signal processing techniques. Energy 157, 796–807. https://​doi.​org/​10.​1016/j.​energy.​2018.​05.​192
(2018).
22. Zafar, A. et al. Machine learning autoencoder-based parameters prediction for solar power generation systems in smart grid. IET
Smart Grid. 7, 328 (2024).
23. Xue, Y., Zhang, X., Zeng, Z. & Sun, X. Short-term load forecasting using convolutional neural network with residual learning and
attention mechanism. Int. J. Electr. Power Energy Syst. 107, 172–180. https://​doi.​org/​10.​1016/j.​ijepes.​2018.​11.​047 (2019).
24. Zhang, G., Guo, Y. & Huang, B. A novel hybrid model for residential energy consumption prediction using optimized deep neural
network and LSSVM. Energy 174, 361–370. https://​doi.​org/​10.​1016/j.​energy.​2019.​02.​060 (2019).
25. Wu, J. & Zhang, W. Energy consumption forecasting with multi-hybrid ensemble learning. Energy 181, 969–985. https://​doi.​org/​
10.​1016/j.​energy.​2019.​06.​156 (2019).
26. Liao, H., Zhang, L., Chen, Q. & Sun, Y. A hybrid model combining convolutional neural network and extreme gradient boosting
for wind speed forecasting. Renew. Energy 146, 2476–2488. https://​doi.​org/​10.​1016/j.​renene.​2019.​07.​044 (2020).
27. Ren, Z., Wei, Y. & Wu, J. Hybrid model based on CNN and LSTM for predicting residential electricity demand. J. Clean. Prod. 220,
914–926. https://​doi.​org/​10.​1016/j.​jclep​ro.​2019.​02.​132 (2019).
28. Feng, Y., Huang, B. & Zhou, Z. Short-term load forecasting based on deep residual network and LSTM. Energies 12(5), 906. https://​
doi.​org/​10.​3390/​en120​50906 (2019).
29. Feng, Y., Huang, B., Zhou, Z. & Li, X. A novel hybrid model based on long short-term memory and extreme learning machine for
short-term load forecasting. Appl. Energy 236, 1333–1350. https://​doi.​org/​10.​1016/j.​apene​rgy.​2018.​12.​051 (2019).
30. Bhutta, M. S. et al. Neuro-fuzzy based high-voltage DC model to optimize frequency stability of an offshore wind farm. Processes
11(7), 2049 (2023).
31. Nazir, M. S. et al. Optimized economic operation of energy storage integration using improved gravitational search algorithm and
dual stage optimization. J. Energy Storage 50, 104591 (2022).
32. Bhutta, M. S. et al. Voltage Stability Index using new single-port equivalent based on component peculiarity and sensitivity per-
sistence. Processes 9(10), 1849 (2021).
33. Abubakar, M., Che, Y., Ivascu, L., Almasoudi, F. M. & Jamil, I. Performance analysis of energy production of large-scale solar plants
based on artificial intelligence (machine learning) technique. Processes 10(9), 1843 (2022).
34. Mukherjee, D., Chakraborty, S., Guchhait, P. K. & Bhunia, J. Machine learning based solar power generation forecasting with and
without MPPT controller. In 2020 IEEE 1st International Conference for Convergence in Engineering (ICCE), Kolkata, India 44–48
(2020). https://​doi.​org/​10.​1109/​ICCE5​0343.​2020.​92906​85.

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 24

Vol:.(1234567890)
www.nature.com/scientificreports/

35. Mukherjee, D., Chakraborty, S., Ghosh, S. & Mishra, R. K. Application of deep learning for power system state forecasting. Int.
Trans. Electr. Energy Syst. 31(9), e12901. https://​doi.​org/​10.​1002/​2050-​7038.​12901 (2021).
36. Mukherjee, D., Chakraborty, S. & Ghosh, S. Power system state forecasting using machine learning techniques. Electr. Eng. 104(1),
283–305. https://​doi.​org/​10.​1007/​s00202-​021-​01328-z (2022).
37. Mukherjee, D., Chakraborty, S. & Ghosh, S. Deep learning-based multilabel classification for locational detection of false data
injection attack in smart grids. Electr. Eng. 104(1), 259–282. https://​doi.​org/​10.​1007/​s00202-​021-​01278-6 (2022).
38. Mukherjee, D., Chakraborty, S. & Abdelaziz, A. Y. Deep learning assisted prediction for generation of power from solar PV. In
2022 IEEE 10th Region 10 Humanitarian Technology Conference (R10-HTC), Hyderabad, India 382–386 (2022). https://​doi.​org/​
10.​1109/​R10-​HTC54​060.​2022.​99295​59.
39. Akhtar, U. & Mahmood, T. A novel hybrid CNN-SVR model for electricity demand forecasting. Energy 170, 511–521. https://​doi.​
org/​10.​1016/j.​energy.​2019.​06.​107 (2019).

Acknowledgements
This work was supported by the Department of Education of Guangxi Autonomous Region under grant number
2023KY0826.

Author contributions
Conceptualization, M.S.B. and Y.L.; methodology, M.S.B.; software, M.A.; validation, M.S.B., M.A. and M.R.A.;
formal analysis, M.A.B.; investigation, Y.L.; resources, M.S.B.; data curation, M.A.; writing original draft prepara-
tion, M.S.B.; writing review and editing, K.S.S.A.; visualization, F.M.A.; supervision, K.S.S.A.; project administra-
tion, M.R.A.; funding acquisition, F.M.A. and K.S.S.A.

Competing interests
The authors declare no competing interests.

Additional information
Correspondence and requests for materials should be addressed to M.S.B.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-
NoDerivatives 4.0 International License, which permits any non-commercial use, sharing,
distribution and reproduction in any medium or format, as long as you give appropriate credit to the original
author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the
licensed material. You do not have permission under this licence to share adapted material derived from this
article or parts of it. The images or other third party material in this article are included in the article’s Creative
Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the
article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the
permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by-​nc-​nd/4.​0/.

© The Author(s) 2024

Scientific Reports | (2024) 14:17101 | https://doi.org/10.1038/s41598-024-68030-5 25

Vol.:(0123456789)