See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/320353848

Machine Learning in PV Fault Detection, Diagnostics and Prognostics: A

Review

Conference Paper · June 2017

CITATION READS

1 1,614

3 authors:

Sandy Rodrigues Helena G. Ramos

Madeira Interactive Technologies Institute Technical University of Lisbon
27 PUBLICATIONS   156 CITATIONS    169 PUBLICATIONS   1,698 CITATIONS   

SEE PROFILE SEE PROFILE

F. Morgado-Dias
Universidade da Madeira
163 PUBLICATIONS   946 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Several projects on Artificial Neural Networks, Sleep and Renewable Energy View project

Marítimo Training LAB View project

All content following this page was uploaded by Sandy Rodrigues on 15 February 2018.

The user has requested enhancement of the downloaded file.

 Machine Learning in PV Fault Detection, Diagnostics and Prognostics:
A Review
Sandy Rodrigues1,3, Helena Geirinhas Ramos3, and F. Morgado-Dias1,2
1
Madeira Interactive Technologies Institute, 2Universidade da Madeira, 3Telecommunications Institute from
the Superior Technical Institute of the University of Lisbon.

Abstract — Photovoltaic (PV) system malfunctions cause Alongside research on PV system FDD which predicts
output efficiency to lower which consequently lowers the return current faults, there is also ongoing research on Fault
of the investment (ROI) and delays investment payback times.
These malfunctions can be limited by implementing Photovoltaic Prognostics that predict the performance over the remaining
System Monitoring (PVSM) solutions. Recently, Machine useful life (RUL) of the PV system and consequently
Learning Techniques (MLT) have been implemented to improve predicting possible future faults by using regression MLT. The
PVSM results and aid in PV performance and PV fault detection, RUL of a PV system is determined by estimating the
identification, diagnostics and prognostics. This paper provides a degradation rate of the PV modules through degradation
review of the work done in the MLT PVSM research field,
provides an organized list of MLT solutions used in PVSM, and models that consider UV radiation, temperature, humidity,
provides a list of opportunities and challenges to further research state-of-health models and end of life models [2], [3], [4].
in the PVSM field. By combining PV system FDD and Fault Prognostics
Index Terms — Diagnostics, Machine Learning, Prognostics, research areas it is possible to develop PV system Prognostics
PV Fault, PV System Monitoring, Return of the Investment. and Health Management (PHM) [4] approaches which helps to
plan preventive, corrective and/or condition-based
I. INTRODUCTION maintenance trips contributing to lower the O&M costs by an
estimated 20% and consequently shorten investment payback
Over the past few years, photovoltaic (PV) systems have times on the ROI [5]. Condition-based maintenance is the new
become very popular amongst homeowners and solar plant term associated to the prognostics approach, in which the
investors since they are viewed as low risk investments that faults are predicted and the condition-based maintenance trips
can achieve high return rates and attractive investment are made when the faults occur and therefore lowers the
payback times [1]. frequency of preventative measures and reduces the impacts
The PV systems are composed by photovoltaic modules and and costs of the corrective measures by anticipating failures or
the balance of system (BOS) that includes wiring, switches, a catching them early [6]. Fig. 1 illustrates how the Machine
mounting system and solar inverters. In order to avoid PV Learning Techniques are used in the PVSM field of research.
system downtime and maintain the output efficiency at high
levels to shorten investment payback periods, it is essential for
the investors, owners and maintenance operators to have
access to the input and output data of the PV system. This data
can only be acquired through monitoring systems and should
be accessible in real-time for the PV system faults to be
addressed accordingly to avoid production loss. The PV
system real-time information is useful to continuously keep
investors and owners informed about the Return on the
Investment (ROI), and keep the maintenance operators aware
of the PV system status to easily schedule maintenance trips.
Research on PV system Fault Detection and Diagnostics
(FDD) has made it possible to detect PV system faults and
correctly diagnose them when using Machine Learning
Techniques (MLT). Research on MLT is popular and has
become very accurate due to available software and
computing capacity of the new computers. The two main
model categories used in MLT are classification and
regression. Classification is used to detect and identify PV
faults while regression is used in PV system diagnostics which
allows for PV system performance analysis or solar output
estimations. Fig. 1. MLT in the PVSM field of research
 According to the thesis written by Zhao, PV system faults After this introduction section, a methodology description is
may remain hidden due to the connection infrastructure of the provided in section II. The results are presented and discussed
PV system and MLT can successfully detect and identify these in section III, which is then followed by a conclusion of the
types of faults, also known as “blind-spots” [7]. Thus, work in section IV.
demonstrating the importance of the use of data-driven models
or MLT in PV fault detection, diagnostics and prognostics
over model-based techniques. Data-driven approaches are also II. METHODOLOGY
recommended to deal with non-linear, unpredictable and This section describes the methodology that was applied, in
complex settings such as the PV system solar production that order to understand which MLTs are being studied in the PV
deal with incoherent weather conditions that impact the solar system monitoring (PVSM) research field and which should
production of the PV system. be studied further. First, a total of 90 scientific papers were
The Machine Learning research area is vast, but recently identified as being related to the MLT PVSM research field.
Pedro Domingos shared in his book [8] (recommended by Bill Secondly, an experiment count was made according to
Gates) that MLTs can be organized into five “tribes” and that specific parameters. These parameters are namely the type of:
each one is related to a specific master algorithm as shown in  MLT tribe (Tribe 1, Tribe 2, Tribe 3, Tribe 4, Tribe 5);
Table I. This concept helps to organize and narrow down the  PVSM technique (Model-based, Data-Driven, Hybrid
MLTs that are used in PV System Monitoring (PVSM) and MLT);
indicate which MLTs that are not being used and may have  MLT models (Classification, Regression);
the potential to present good results in the PVSM research  PVSM method (Detection, Identification, Diagnostics,
field. Pedro Domingos defends the idea of combining various Prognostics);
master algorithms presented in Table I to make one master  PVSM papers (PV performance, PV fault);
algorithm. This idea of combining different master algorithms  PV system testbed (Grid PV system, lab PV system,
is already being adopted in the PVSM research field but in a simulation); In this paper, a grid PV system is assumed
to have more than 3kWp of installed power which
smaller scale by combining two or three Master Algorithms
represents a typical roof-top household PV system.
and are referred to as hybrid MLTs. Two examples of hybrid
In addition, the experiment count is also applied to the
MLT can be found in [9] and [10] where the first analysis a
parameters that do not involve MLT and that influence the
hybrid MLT that uses the Genetic Algorithm and an Artificial PVSM research analysis such as the type of
Neural Network which combines the master algorithms of  PV system faults (arc, short-circuit, faulty condition,
tribes 3 and 2. The latter work tests a hybrid MLT that uses a degradation, shading, Line to Line, soiling, open-circuit);
sparse Bayesian learning theory and Support Vector Machines  PV system scale (roof, commercial, plant);
which combines the master algorithms of tribes 4 and 5.  PVSM level (array, string, module);
 PV system measured inputs (irradiance, module
TABLE I temperature, ambient temperature, current, voltage, etc.)
MLT TRIBES AND RESPECTIVE MASTER ALGORITHMS [8] and outputs (power, energy, I-V curve, decision).
Tribe Master Algorithm The following four subsections present a brief description of
Tribe 1 Symbolists Inverse deduction / Induction all the PVSM techniques that were studied in this work which
Tribe 2 Connectionists Backpropagation include the model-based approach, data-driven regressive
Tribe 3 Evolutionaries Genetic Programming approach, machine learning techniques and the hybrid MLTs.
Tribe 4 Bayesians Probabilistic Inference
Tribe 5 Analogizers Kernel Machines
A. Brief Description of the Model-Based Approach
Based on this information, the motivation was to understand The Model-Based (MB) approach in PVSM is achieved by
how the different MLT have been used in the different using sensor data associated to mathematical and visual
research areas of the PVSM field such as PV Fault Detection, methods that have indicators and thresholds to aid in decision
Identification, Diagnostics and Prognostics. The main goal of making [11]. Some examples of this type of approach
this work is to analyze 90 research papers and make a list of identified during this work include the Persistence Model,
the MLT and organize them into their associated MLT tribe. and the Physical Model.
The other goal of this work is to present the number of The Persistence model may be used as a forecasting
experiments associated to a parameter (PV system faults, PV technique which assumes that the weather conditions remain
system scale, Hybrid MLTs, etc.) related to the MLT PVSM the same between one sample of the time-series and the next
field. This experiment count provides information about the one. The Physical model consists of a set of equations that
areas of the PVSM research field that are more popular as well describe the behavior based on the physical principles. This
as those that show good results but need more research. The model may consist of a sensor and stored data which is then
challenges and opportunities that are available in the PVSM analyzed by indicators and thresholds that were determined
field are also provided in this work. through trial and error experiments [11].
B. Brief Description of the Data-Driven Regressive Approach Tribe 5: Analogizers (Kernel Machines)
The Analogizers MLTs recognize the similarities between
The Data-Driven Regressive (DDR) approach in PVSM
the old and the new data/knowledge by using the nearest
includes all the data-driven algorithms that are not learning
neighbour kernel machine algorithms which analyse its
algorithms but are regressive algorithms and depend on
surroundings and try to generalize by similarity. These results
historical data for forecasting [11]. The most popular DDR
are similar to those of the neural network models [8]. The
algorithms that were identified in this work are the ARIMA
most popular analogizer MLTs are the Support Vector
(Auto-Regressive Integrated Moving Average), the linear
Machines and the K-Nearest Neighbour algorithms [11].
regression models and the Principal component analysis
(PCA).
The forecasting values of ARIMA are based on its inertia, D. Brief Description of the Hybrid MLTs
while the regression models analyses the relationship between
Hybrid MLTs is the name given to the combination between
the dependent and independent variables. The PCA is used to
two or more PVSM techniques (MB, DDR, and MLTs). This
select the most relevant variables [11].
combination contributes to strengthening the accuracy of the
forecasting results [11]. The most popular hybrid MLTs that
C. Brief Description of the Machine Learning Techniques were identified in this work are namely the Radial Basis
Function Neural Network (RBFNN), the ANN + Physical
This section briefly describes the MLT Tribes and indicates
Model, the Genetic Algorithm-Artificial Neural Network
which of the MLT are most popular in each tribe. Table I
(GA-ANN), the ANFIS (adaptive neuro-fuzzy inference
presents a description of the five tribes that follow.
system), and the Relevance Vector Machine (RVM).
Tribe 1: Symbolists (Inverse deduction / Induction)
The RBFNN is a MLT combination between tribe 5 (RBF)
The MLTs that belong to this tribe attempt to find the
and tribe 2 (NN). The ANN+Physical model is a PVSM
missing knowledge by applying inverse deduction which starts
technique combination between a tribe 2 (ANN) MLT and a
with some properties and conclusions and works its way
model-based technique. The GA-ANN is a MLT combination
backward to find the gaps in knowledge by using analysis and
between tribe 3 (GA) and tribe 2 (ANN). The ANFIS is
data sets [8]. The most popular MLT of this tribe are Decision
another MLT that combines tribe 2 (neuro) and tribe 1 (fuzzy).
Trees, Random Forests and the Fuzzy Logic [11].
Tribe 2: Connectionists (Backpropagation)
The MLT of this tribe are composed by algorithms that III. RESULTS AND DISCUSSION
claim to emulate the functions of the brain by creating
artificial neurons and connecting them in a neural network by Over 150 PV systems monitoring (PVSM) research papers
using an input layer, one or various hidden layers and an were identified and analysed for this work. Only 90 of them
output layer. The inputs are taken by the hidden neurons were considered since they include the topics that are
which then generate an output (new knowledge) that can be discussed in this work such as MLTs, Hybrid MLTs, power
read by other neurons that perform the same function [8]. The and energy estimation, as well as PV performance and PV
Artificial Neural Networks and Extreme Learning Machine fault analysis. The papers that only present the framework and
are some the most popular tribe 2 MLTs [11]. do not include an experimental research were not considered.
Tribe 3: Evolutionaries (Genetic Programming) The graphs that follow present the number of times a given
Here the algorithms of the MLTs try to mimic the parameter is mentioned in the 90 papers that were studied
evolutionary process of genomes and DNA where the (blue). The graphs also present the number of times that a
performance is measured by the fitness of the offspring. The given parameter is mentioned in the 32 papers that study the
algorithms are divided into a set of sub programs that mutate PV fault analysis (orange). Some of these papers focus only
and combine with other sub programs to determine if the on the PV performance analysis, others only on the PV fault
prediction results improve with multiple iterations of the analysis and others on both the PV performance and PV fault
program. New knowledge is determined by simulating analysis.
evolution [8]. The popular evolutionary MLTs include the Fig. 2 presents the various MLT tribes and how many times
Genetic Algorithm and the Genetic Programming [11]. they were used in all the papers and in the PV fault papers.
Tribe 4: Bayesians (Probabilistic Inference) This graph also presents the experiment count of the Model-
These MLTs systematically attempt to reduce the
Based, Data-Driven Regressive and hybrid MLT approaches.
uncertainty of the new knowledge by applying the
The MLTs of Tribe 2 and 3 are used the most in the PVSM
probabilistic inference algorithm to events. The events that are
research field which include the ANN and SVR respectively.
known to occur are assigned prior probabilities, and as more
evidence is observed, the priors either lose their importance or The Regression Trees and the Bayesian probability from the
become more important [8]. The Monte Carlo Method is a Tribes 1 and 4 respectively are used similarly. The MLTs
very popular Bayesian MLT [11]. from Tribe 3 (Genetic Programming and Genetic Algorithm)
are the least used and not used at all in the PV fault papers.
Fig. 2 also presents the experiment count that use the three
approaches namely the MB approach, the DDR approach and latter testbeds are not as reliable because the nonlinear nature
the Hybrid MLT approach. The hybrid MLT approach is used of the weather conditions may not be considered in the
more in the PV performance analysis compared to the PV fault analysis. Models always have limitations compared to the true
analysis. PV systems. The PV fault analysis research is mainly done on
lab and simulation testbeds which might not result in reliable
findings. The PV installed power is also reflected in the graph
that follows by differentiating the PV scale into plant,
commercial and roof-top sizes. There is an even amount of
work that is carried out in all three scales. The monitoring
level of the PV systems is normally made on the array level
due to cost control, however it is interesting to see that the PV
fault analysis research tends to use the module-level
monitoring which in turn provides very accurate results.

Fig. 2. MLT Tribes and PVSM Technique approaches

Fig. 3 presents the number count of the papers for the

classification and regression MLT models as well as the four
PV system monitoring methods such as the Detection,
Identification, Diagnosis and Prognosis. Fig. 3 also presents
the number count of the papers that analyse the PV
performance as well as those that analyse the PV faults. There
are double the papers that consider regression as opposed to
classification MLT models in the PV system monitoring field
Fig. 4. PV System Testbeds, Scale and Monitoring Level
of research. Classification is mostly used in the PV fault
analysis. PV fault Detection is the most popular PVSM
Fig. 5 presents the input features that are most often used in
method and prognosis the least popular, while PV fault
the research papers studied in this work. Irradiance, ambient
identification and diagnosis are similar in use. We can see
temperature and power are the input features that are used the
that PV performance analysis was used in 87 out of the 90
most followed by the wind speed, module temperature, current
papers studied in this work while 29 of the 32 PV fault papers
and voltage values. Not many papers consider the energy
studied in this work performed PV performance analysis. PV
values.
performance research is more popular than PV fault analysis.

Fig. 3. MLT Models, Methods and PV System Analysis Fig. 5. PVSM Input Features

Fig. 4 presents the parameters of the testbed used in the Fig. 6 presents the output features that are mostly used in
research papers. The end results of the experiments are always the research papers studied in this work. Power prediction is
more reliable when using a grid connected PV system. From the most popular while energy and I-V curve prediction is also
the graph below, it is possible to see that there are many used. Classification decisions are widely used in the PV fault
experiments carried out on PV grid connected systems. analysis as well as the power and I-V curve predictions. Not
However, there are still a fair amount of experiments that are many papers consider the standalone voltage and current
carried out in lab testbeds (a few PV modules) and by predictions.
simulating a PV system using software (Matlab). These two
 followed by Regression Tree, Decision Tree, and Functional
Tree algorithms. The ANN algorithm from Tribe 2 is highly
used in the PVSM research field since it was identified 41
times followed by the ELM algorithm. Tribe 3 includes the
Genetic Programming and Genetic algorithms which were not
used much in the PVSM research. The Gaussian Process
algorithm is the most popular MLT out of the Tribe 4
algorithms. The K-Nearest Neighbor algorithm is the most
popular out of the Tribe 5 MLTs followed by the SVM, the
Fig. 6. PVSM Output Features SVR, SVR with Radial kernel, Radial Basis Function,
Polynomial SVM or SVR, and the K-Means algorithm.
Fig. 7 presents the experiment count of the PV fault analysis Table IV presents 4 of the 69 Hybrid MLT algorithms that
papers that are associated to eight different PV fault
parameters. Shading, short-circuit and degradation faults are TABLE III
researched the most while faulty conditions, line to line, LIST OF MACHINE LEARNING ALGORITHMS ORGANIZED INTO
soiling and open-circuit faults are less popular in PV fault FIVE MACHINE LEARNING TRIBES
research. The arc fault analysis using MLT present the least Tribes MLT #
research papers. Random Forests (RF) 4
Regression Tree (RT) 2
Tribe Classification and Regression Decision Tree (CART) 2
1 Functional Tree 2
C.45 Decision Tree 1
# Best First Decision Tree 1
15 Gradient Boosting 1
Binary Tree Bagging 1
Fuzzy Logic (FL) 1
Artificial Neural Network (ANN) 41
Extreme Learning Machine (ELM) 4
Extension Neural Network 2
Recurrent Neural Network (RNN) 2
Neural Network Ensemble (NNE) Bagging Or
Tribe 2
Bootstrap Based Neural Network (BNN)
Fig. 7. PV System Faults 2 Group method of data handling neural network
1
(GMDH-type NN)
Table II presents some of the most popular model-based Dynamic Recurrent Neural Network (DRNN) 1
#
(MB) and data-driven regressive approaches (DDR) identified Deep Belief Network (DBN) 1
58
Long Short-Term Memory networks (LSTM) Based on
in the 90 papers studied in this work. The persistent and Recurrent Neural Network (RNN)
1
physical models were mostly used in the papers when using Auto-LSTM Long Short-Term Memory networks 1
the model-based approach. While the ARIMA, Linear Model Entropy Extreme Learning Machine (EELM) 1
and PCA were mostly applied when using the data-driven Online Sequential Extreme Learning Machine 1
regressive approach. The complete list of MB and DDR Tribe
3 Genetic Programming (GP) 1
approaches can be found in the following reference [12].
#2 Genetic Algorithm (GA) 1
TABLE II Tribe Gaussian Process 6
LIST OF MODEL-BASED AND DATA-DRIVEN ALGORITHMS AND 4 Monte Carlo Method 1
RESPECTIVE P APER N UMBER COUNT Graph-based Semi-Supervised Learning 1
# Naïve Bayes Classification 1
Model-Based Data-Driven Regressive 10 Bayesian Belief Network 1
Approach # Approach #
(14 experiments) (33 experiments) K – Nearest Neighbor (K-NN) 9
Support Vector Machine (SVM) 6
ARIMA (Auto-Regressive
Persistence Model 6 7 Support Vector Regression (SVR) 5
Integrated Moving Average)
SVR with Radial kernel 4
Linear Model or Linear Tribe
Physical Model 2 4 Radial Basis Function 3
Regression 5
Hampel Identifier and Polynomial SVM or SVR 2
PCA (Principal component
1 4 K-Means Algorithm 2
Boxplot rule analysis) # Linear SVM 1
15 Local Outlier Factor (LOF) 1
Locally Weighted Learning 1
Table III presents all the MLT algorithms that were Local Linear Fitting / Regression 1
identified in the 90 papers studied in this work. The Random Clustering by fast search and find of density peaks 1
Forests algorithm is very popular amongst Tribe 1 MLTs
were identified in the 90 papers studied in this work. The IV. CONCLUSION
RBFNN, the ANN+Physical model, the GA-ANN, and
This paper highlights the key parameters, challenges, and
ANFIS were the most popular Hybrid MLTs. The complete
research opportunities that can further research in the MLT
list of the hybrid MLTs can be found in the following
PVSM field. A list of MLTs organized by tribes is provided to
reference [12].
narrow down the types of MLT that should be considered in
TABLE IV future research. Hybrid MLT research and MLT comparison
LIST OF HYBRID MACHINE LEARNING ALGORITHMS studies are encouraged. PV system Prognostics research is
Hybrid MLT (69 algorithms) Tribes # encouraged to provide condition-based maintenance
Radial Basis Function Neural Network (RBFNN) T5 + T2 5 scheduling and improve ROI payback times. PV module-level
ANN + Physical Model T2 + MB 5 monitoring provides accurate results which is useful to explore
Genetic Algorithm – ANN (GA- ANN) T3 + T2 4 PV faults. Real-world testbed research is encouraged, as
ANFIS (adaptive neuro-fuzzy inference system) T1 + T2 3 accurate experimental results are presented.

A. Challenges and Opportunities for the PVSM research field ACKNOWLEDGEMENT

The MLTs suggested in Table V, present the best results Acknowledgments to the LUSO-AMERICAN
comparison studies, however they have only been considered DEVELOPMENT Foundation (Fundação Luso-Americana
in between 1 and 3 experiment papers and therefore need para o desenvolvimento – FLAD) for the travel grant. The
support from the Portuguese Foundation for Science and
further research. Based on the results, there are some
Technology for their support through Projeto Estratégico LA 9
opportunities and challenges that were identified to help
- UID/EEA/50009/2013. The support from ARDITI – Agência
further research in the PVSM field which are discussed as Regional para o Desenvolvimento e Tecnologia under the
follows. scope of the Project M1420-09-5369-000001 – PhD
TABLE V Studentship, is also gratefully appreciated.
OPPORTUNITIES AND CHALLENGES
Tribe 1 MLT present good results but few experiments,
therefore should be further researched, particularly with the REFERENCES
Random Forests, Decision Trees, Regression Trees and [1] S. Rodrigues, et al, “Economic feasibility analysis of small scale
Fuzzy Logic. PV systems in different countries,” Solar Energy, vol. 131, pp.
Tribe 2 presents a high experiment count due to the use of 81–95, 2016.
the ANN MLT, which indicates that the PVSM researchers are [2] A. Chokor, et al, “A Review of PV DC Systems Prognostics and
hesitant to try other MLT such as the ELM and RNN which Health Management: Challenges and Opportunities,” Prognostics
have shown to present very accurate results.
and Health Management (PHM) Society, 2016.
Only two research experiments have dealt with the Tribe 3
(genetic programming and genetic algorithms) MLT.
[3] N. Laayouj, H. Jamouli, and M. El Hail, “New Prognostic
Opportunities

Tribe 4 MLT present very good results when using the Framework for Degradation Assessment and Remaining Useful
Gaussian Process, the Monte Carlo and the Graph-based Life Estimation of Photovoltaic Module,” in Control and
Semi-Supervised Learning. More research is encouraged. Automation (MED), 24th Conference IEEE, 2016, pp. 378–383.
Amongst the Tribe 5 MLTs, the K-NN MLT is more popular [4] N. Clements, “Introduction to Prognostics,” PHM Society, 2011.
than the SVM, and therefore SVM work is encouraged. [5] S. Vohnout, et al “Uptime improvements for photovoltaic power
69 Hybrid MLT algorithms were identified but most explored inverters,” Technical Program for MFPT The Prognostics and
only in one paper. More research should be done on Health Management Solutions Conference, pp. 1–15, 2012.
combining MLT to discover the most accurate Hybrid MLTs.
[6] T. Keating, A. Walker, and K. Ardani, “Best Practices in PV
PV system module-level monitoring improves the accuracy
of the results even though this solution is costly, however this
System Operations and Maintenance” NREL, p. 57, 2015.
should be further explored. [7] Y. Zhao, “Fault Detection, Classification and Protection in Solar
PV fault diagnostics and prognostics research is initiating, PV Arrays,” Northeastern University Boston, PhD Thesis 2015.
but more emphasis should be put on PVSM prognostics to [8] P. Domingos, The Master Algorithm: How the Quest for the
improve M&O scheduling and lower ROI payback times. Ultimate Learning Machine Will Remake Our World. 2016.
There is a vast number of MLTs, therefore many researchers [9] M. Russo, et al, “Genetic programming for photovoltaic plant
tend to use the same ones as they rely on research results output forecasting,” Solar Energy, vol. 105, pp. 264–273, 2014.
Challenges

from previous works and do not try new MLT. [10] N. Laayouj, et al, “Photovoltaic Module Health monitoring and
Research on PVSM should be done on real-world PV system Degradation Assessment,” no. September, pp. 570–577, 2016.
testbeds to obtain credible results that consider real-world
[11] J. Antonanzas, et al, “Review of photovoltaic power
disturbances caused by environmental factors and signal
noise coming from the measurement devices. forecasting,” Solar Energy, vol. 136, pp. 78–111, 2016.
[12] S. Rodrigues, “TABLES,” Available: https://goo.gl/0mIhm8.

View publication stats