Renewable Energy 186 (2022) 742e754

Contents lists available at ScienceDirect

Renewable Energy
journal homepage: www.elsevier.com/locate/renene

A hybrid approach to multi-step, short-term wind speed forecasting

using correlated features
Fei Sun a, Tongdan Jin b, *
a
Material Sciences, Engineering and Commercialization Program, Texas State University, San Marcos, TX, 78666, USA
b
Ingram School of Engineering, Texas State University, San Marcos, TX, 78666, USA

a r t i c l e i n f o a b s t r a c t

Article history: Wind power is becoming a main alternative energy source to meet the growing electricity needs.
Received 22 September 2020 Forecasting wind speed is important to mitigate generation uncertainty and optimize asset utilization.
Received in revised form This paper proposes a hybrid wind speed prediction model with multivariate input and multi-step
29 December 2021
output capability. The model synthesizes linear time series regression with nonlinear machine
Accepted 11 January 2022
Available online 13 January 2022
learning algorithm. The input neurons of the hybrid model are determined by the number of lag ob-
servations in autoregressive integrated moving average (ARIMA), and also by correlated meteorological
features, such as wind direction, air pressure, humidity, dew point, and temperature. The output neurons
Keywords:
Neural network
are further derived based on the forecasting horizon. The hybrid model is trained, validated, and tested
Supervised learning by using 1.73 million hourly meteorological records from three cities with diverse wind profiles. The
Time series performance of the model is compared with several existing methods based on root mean square error
Meteorological features and mean absolute error. Though the hybrid model does not show obvious advantage in 1-h ahead
Hybrid model prediction, it outperforms persistence model, ARIMA, and univariate neural network models in 3-to-24 h
Wind rose ahead prediction. The hybrid model is able to reduce the prediction error by 20% in comparison with
univariate neural networks.
© 2022 Published by Elsevier Ltd.

1. Introduction into four categories: 1) physical method, 2) statistical model, 3)

artificial intelligence technique, and 4) hybrid forecasting method.
The installed capacity of wind generation has increased rapidly A physical method is often used to establish a rigorous mathe-
in the past two decades. According to the report of the World Wind matical formula based on the principles of geophysical fluid dy-
Energy Association [1], the global cumulative wind power capacity namics and thermodynamics. The numerical weather prediction
reached 597 Gigawatt (GW) at the end of 2018, whereas it was only (NWP) model is one type of physical model. Carvalho et al. [4] point
6.1 GW in 1996. Wind generation is intermittent, and the power out NWP requires a considerable amount of computational time
output is time-varying due to the uncertainty of wind speed. Thus, due to the complex mathematical formula that typically needs
wind speed becomes a crucial parameter for predicting and man- topography information. Hence NWP is often used by meteorolo-
aging this type of variable generation in the smart grid era. In fact, gists for long-term weather prediction in a large-scale area. How-
wind speed forecasting plays an important role in allocating gen- ever, the model does not guarantee the accuracy in short-term
eration capacity, scheduling power dispatch, and managing asset prediction. Statistical models, such as the autoregressive integrated
utilization and maintenance. moving average (ARIMA), can explicitly reveal the linear relation-
A variety of wind speed forecasting models have been proposed ship between the input and output wind speeds [5,6]. Quite often
for both short-term and long-term predictions. Short-term wind the performance of statistical models often exceeds NWP in short-
speed forecasting is carried out in a range of minutes, hours or term prediction. Since the model structure is limited to the linear
several days. Based on review works [2,3], the many methods being form, statistical models become less effective in handling wind data
developed to forecast the short-term wind speed can be classified with large variations. Most of the statistical models also assume the
wind speed follows the normal or the Weibull distribution.
Artificial Intelligence (AI) techniques such as feedforward neural
* Corresponding author. network (FNN) and support vector machine are also adopted to
E-mail address: [email protected] (T. Jin).

https://doi.org/10.1016/j.renene.2022.01.041
0960-1481/© 2022 Published by Elsevier Ltd.
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

Notation Description wkj weight between node k in the output layer and node j
p the number of lag observations in the model, called in the hidden layer
the lag order xi input value
d the number of times that the raw observations are bj, bk bias value
different, called the degree of differencing
q the size of the moving average window, called the ACRONYM
order of moving average ANN Artificial Neuron Network
4i the ith autoregressive coefficient ARIMA Autoregressive Integrated Moving Average
4j the jth moving average coefficient ARMA Autoregressive Moving Average
εt error term at t FNN Feedforward Neural Network
m constant term MAE Mean Absolute Error
^k
Y the predicted value for node k in the output layer, for NWP Numerical Weather Prediction
k ¼ 1, 2, …, K RMSE Root Mean Square Error
gj() activation function of hidden layer RNN Recurrent Neural Network
gk() activation function of output layer WRF Weather Research and Forecast
wji weight between node j in the hidden layer and node i
in the input layer

forecast wind speed [7,8]. FNN is a multi-layer perceptron neural characterizes the wind speed time series data. This section also
network. It differs from physical and statistical methods in that it discusses the important features of meteorological variables. Sec-
can learn the input-output relation from historical data, recognize tion 4 illustrates the principle of the forecasting models and the
hidden patterns based on past observations, and use them to approach to model performance evaluations. Section 5 presents the
forecast future values. FNN can also tolerate data errors and ease in comparisons of the forecasting models. Finally, Section 6 concludes
adaptability to online measurements. Therefore, AI-based predic- the paper.
tion methods turn out to be more capable of handling non-linear
relationship with no assumptions on the statistical distribution of 2. Literature review
underlying data. However, FNN could become computationally
intensive with the growth of the number of neurons or hidden The physical model for wind prediction is usually built upon
layers. To achieve the minimum forecasting error, FNN model needs meteorological features (e.g., temperature, pressure, and humidity)
to be trained by a large amount of data. This could be difficult if no and geographical information (e.g., surface roughness and latitude).
sufficient training data are available. To combine the advantages of Ren et al. [10] and Zhang et al. [11] utilize sophisticated meteoro-
AI-based algorithm with statistical method, Cadenas and Rivera [9] logical data for wind speed and wind generation prediction. A
propose a hybrid forecasting approach that consists of time series considerable amount of computational time is required due to the
model and ANN model. The former characterizes the linear com- complexity of the underlying mathematical formulas. Numerical
ponents and the latter captures the nonlinear relation between the weather prediction (NWP) method is a common physical model
input and the output data. However, most of hybrid models ignore and can generate satisfactory results for a relatively long prediction
the impact from the other meteorological features, such as air horizon up to several days. Weather research and forecast (WRF)
pressure, humidity, and temperature. They are often suitable to 1-h model is one of the most popular NWP approaches [12]. It has been
ahead forecasting. used to forecast air chemistry, hydrology, wildland fires, hurricanes,
This paper proposes a hybrid forecasting method to provide and regional climate [13]. For instance, Carvalho et al. [4] apply
multi-step, short-term wind speed prediction based on multiple WRF model in an area of Portugal under different numerical and
meteorological features. The hybrid model is called ARIMA-FNN in physical options. The area contains complex terrain and is charac-
which FNN stands for feedforward neural network, and ARIMA terized by significant wind energy resource. However, WRF usually
represents autoregressive integrated moving average. The contri- requires comprehensive terrain information in order to achieve
bution of the study is twofold. First, the proposed hybrid model better forecasting results. As pointed by Wu and Hong [14], NWP
synthesizes different meteorological features with statistical models including WRF have two drawbacks: 1) they do not update
inference and machine learning techniques. These features include the predictions very frequently; and 2) they require a large amount
wind direction, temperature, humidity, among others. Data ana- of computing resources.
lytics is applied to extract the key meteorological features to Time series models including the ARIMA family are widely used
enhance the forecasting performance. Second, the hybrid model in wind speed forecasting due to the robustness of the model [15].
simultaneously generates multiple speed prediction data corre- For instance, Kavasseri and Seetharaman [5] propose a fractional-
sponding to different time intervals, varying from 1 to 24 h ahead. ARIMA model to predict the wind speed in one- or two-day
The proposed model is extensively examined using 11-year climate ahead, respectively. The expected wind energy throughput is
data from three cities: Wellington, New Zealand; San Francisco, further derived from the predicted wind speed along with the
USA and Phoenix, USA. These cities represent a broad range of wind power curve of the wind turbine. Fang and Chiang [6] derive a
profiles in terms of geographical locations and hourly wind speed. multivariate wind power prediction model that accommodates
Over 1.73 million meteorological data are used to test the proposed multiple weather features, including wind speed, temperature and
model. The numerical experiments show that the proposed model weather condition. The forecasting accuracy is further improved by
outperforms classical time series method and univariate neural considering the interdependency of different features. In general,
networks in terms of forecasting accuracy and robustness. time series models, such as ARIMA, perform well for the short-term
The rest of the paper is organized as follows. Section 2 reviews prediction due to the use of Box-Jenkins methodology for model
the state-of-art of the wind speed prediction methods. Section 3 construction. The limitation is that the model structure is linear,
743
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

hence the accuracy is not guaranteed in long-term prediction. In concludes that the direct method yields better results than the
addition, the trend and seasonal factors in the time series data must iterative method. This conclusion is also supported by the study of
be decomposed prior to the model application. Hamzaçebi et al. [28]. In our paper, the proposed hybrid model is
Considering the existence of nonlinearity in wind speed data, capable of making multi-output predictions by incorporating other
research interests are shifting towards machine learning algo- meteorological features in addition to wind speed. The prediction
rithms. Artificial neural network (ANN) has been recognized as a quality is improved because the correlations of these features with
powerful tool to solve non-linear problems like weather fore- wind speed are also considered. This is crucial for improving the
casting. For instance, FNN is a class of the multilayer neural network security and economic benefits of wind power generation.
consisting of an input layer, an output layer, and one or more hid-
den layers. Each layer can host one or multiple neurons that are 3. Wind data and meteorological features
connected to the neurons of the previous layer using appropriate
weights. The neurons in the same layer are mutually independent, 3.1. Seasonal and diurnal variations
and each neuron has a transfer function. The studies by Khashei
et al. [16] and Zhang et al. [17] show that ANN is more effective than To compare the performance of forecasting models, multiple
the statistical prediction techniques, especially for the input data forecasting horizons are adopted in this study. Without loss of
with non-linearity relation. Amellas et al. [18] predict the short- generality, five forecasting horizons are considered and shown in
term wind speed based on multi-layer perceptron and nonlinear Fig. 1. These are 1-, 3-, 8-, 12-, and 24-h, respectively. The 24-h is
autoregressive exogenous model with multivariable features. The also known as day-ahead forecasting. An out-of-sample forecast
model aims to assist wind farms and utility companies in maxi- method is adopted, meaning the prediction data outside the data-
mizing the economic benefits. Liu et al. [19] design a hybrid ARIMA- set are used to validate and test the model. Croonenbroeck and
ANN and an ARIMA-Kalman method for hourly wind speed fore- Stadtmann [29] emphasize that the out-of-sample forecasting
casting. Both methods result in good prediction because the dy- strategy is essential in constructing a wind forecasting model. The
namics and non-linearity of the wind profile is effectively captured. sliding window technique is used for wind speed forecasting. The
Damousis et al. [20] use a genetic algorithm-based learning scheme input data are normalized because different meteorological fea-
to predict the wind speed and power generation for a wind park. tures have a different range. Min-max scaling is applied to trans-
Their model focuses on short-term forecasting in a range from 0.5-h form the original data into a common range between 0 and 1. The
to 2-h ahead. Comparative studies are also carried out between proposed hybrid model allows us to generate multiple outputs
ANN and time series models by Mohandes et al. [21], showing that corresponding to different horizons at the same time. This can be
ANN is superior to the autoregressive model in multi-step easily realized through the implementation of multiple neurons in
prediction. the output layer. For instance, in the 1-h ahead prediction model,
ANN can conveniently accommodate different meteorological the output layer only has one neuron. However, in the 12-h ahead
features such as air pressure, temperature, and wind direction to forecasting model, the output layer has 12 neurons, each repre-
improve the wind forecasting accuracy. For instance, Li et al. [7] senting the wind speed of upcoming 12 h, respectively.
utilize a three-layer model with 4-, 8-, and 1-neuron, respectively, For comparison purposes, three cities representing diverse wind
for input, hidden, and output layers to estimate wind power. The profiles and weather conditions are chosen to test the proposed
four input features are the wind speed and wind directions of two methods. They are Wellington in New Zealand, San Francisco in the
meteorological towers, respectively. Moreover, compressing func- USA, and Phoenix in the USA. Wellington has a very strong wind
tions for the four input valuables have been used to accelerate the profile in four seasons, while Phoenix has strong sunshine with
ANN training process. Mabel and Fernandez [22] use wind speed, weak wind across the year. San Francisco has abundant wind re-
relative humidity, and hourly generation as input variables to sources in summer and fall seasons, but the wind speed drops in
construct an ANN wind speed model. They conclude that a better spring and winter seasons. The meteorological data, including
performance can be obtained by incorporating additional features. hourly wind speed, are retrieved from the web portal of Weather
Peng et al. [23] develop prediction methods for short-term wind Underground (WU, 2017) between 2004 and 2014. These data are
power generation forecasting by considering temperature, wind recorded and collected by Automated Weather Observing System at
speed, and wind direction features. The study finds that the hybrid 10-m height above the ground. This study considers hourly obser-
model incorporating multiple features yields more accurate results vations of wind speed, wind direction, dew point, and ambient
than the univariate ANN model. Zhou et al. [24] propose a hybrid temperature, resulting in about 90,000 records over 11-year for
wind forecasting method comprised of four modules: data analysis, each city.
model selection, multi-criteria forecasting, and performance eval- Fig. 2 plots the annual wind profiles of testing cities in 2014. The
uation. Given a set of wind speed data, their approach allows for average wind speed appears uncertain and varies from month to
selecting an optimal forecasting model without prior knowledge. In month. The maximum wind speed in Wellington reaches 28.3 m/s,
other literature, Cao et al. [25] and Mohandes [8] leverage recurrent and the average wind speed is around 6.8 m/s. The average wind
neural network (RNN) and support vector machines (SVM) to speed in San Francisco and Phoenix is 5.4 m/s and 2.3 m/s,
forecast the wind speed. Compared with the multilayer perceptron respectively. Similar observations are found in other years.
neural networks, more computational time is required to train RNN The random behavior of wind speed can be characterized by
and SVM models in exchange for slightly improved prediction normal distribution [30] or Weibull distribution [31]. Fig. 3(a) plots
quality. the hourly wind speed histogram in 2014 with 8760 data points per
The multi-output forecasting strategy involves the development city. The corresponding Weibull wind speed distributions for the
of short-term, multi-step forecasting model that can predict a three cities are depicted in Fig. 3(b). The distribution has a bell
sequence of wind data in a one-shot manner. Kline et al. [26] shape with extended right tail, and distribution parameters are in
construct a multi-output neural network where each output node Table 1. Note that c and k are the scale and shape parameter of
corresponds to one prediction horizon. This approach prevents the Weibull distribution, respectively. For a given c, a larger k makes the
accumulation of prediction errors and avoids the performance distribution curve more concentrated around the mean value. For a
degradation as the forecasting horizon increases. Zhang [27] de- given k, a larger c pushes the distribution curve to the right side.
velops an ANN model for multi-period time series forecasting and The seasonality of wind speed varies in different cities. The
744
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

Fig. 1. Multiple steps (or horizons) forecasting.

Fig. 2. Annual wind speed of testing cities in 2014.

Fig. 3. (a) Hourly wind speed frequency histogram, and (b) Weibull Distributions.

Table 1
Weibull distribution parameters in three cities.

City and Country c (m/s) k

Wellington, New Zealand 7.685 2.022

San Francisco, CA, USA 5.067 1.434
Phoenix, AZ, USA 3.152 1.836

average monthly wind speed of San Francisco is plotted in Fig. 4

based on 11-year data. The wind speed exhibits a strong season-
ality with the largest speed occurred in the second quarter and the
lowest speed in the fourth quarter. In other words, the wind profile
in each year swings from spring to winter. The windy season spans
from April to August, and May is the peak wind period. A similar
Fig. 4. Seasonal variation of wind speed of 12 months in San Francisco.

745
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

pattern also appears in Phoenix where the windy season is from

April to August, yet the average wind speed in Phoenix is much
smaller than San Francisco. Wellington does not show a strong
seasonality, rather the wind speed patterns remain homogeneous
across 12 months.
The wind profiles of Wellington, San Francisco, and Phoenix also
show significant diurnal variations. Fig. 5 plots the average hourly
wind speed data in a day of Wellington and San Francisco. In
Wellington, the peak speed appears around 2 a.m. The wind speed
decreases and reaches the valley at 1 p.m. The curve and trend of
wind speed in San Francisco and Phoenix are opposite to
Wellington. In San Francisco, the windiest time is around 4 p.m. and
the speed declines to the lowest level in a day around 5e7 a.m. The
wind speed tends to increase after sunrise and decrease after
sunset in San Francisco. Fig. 6. Correlation between meteorological features.

3.2. Correlation with meteorological features

magnitude of the correlation coefficient are represented by two
The seasonal and diurnal variations are indeed related to the colors and the intensity of the colors. It shows the wind speed has a
local meteorological features, such as temperature, air pressure, positive correlation with wind direction and a negative correlation
and humidity. Understanding the correlation between wind speed with humidity.
and other meteorological features is important and useful to pre- A wind rose is a graphic tool that gives a succinct view of how
dict wind speed. One of the tasks of this paper is to determine the wind speed, direction, and frequency are typically distributed in
number of input neurons for the proposed model because these one location. The length of the bars indicate frequency in per-
neurons represent different meteorological features. Fig. 6 graphi- centage units. The wind roses of the three cities are presented in
cally shows correlation matrix between any two of the features by Fig. 7. Note that 0 , 90 , 180 , and 270 denote north, east, south,
the heat map. These features include air pressure, temperature, and west directions, respectively. In Wellington and Phoenix, the
humidity, dew point, and wind speed direction. The sign and dominant wind direction is from the north (N). For example, in

Fig. 5. Diurnal wind speed in Wellington and San Francisco.

746
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

Fig. 7. The wind rose with frequency distribution of hourly wind speed (unit: m/s).

Wellington 19.1% of the wind blowing between 6.1 and 11.6 m/s is respectively, the observed and forecasted values at time t, or
from the north. In Phoenix, the average wind speed blowing from indexed by n for n ¼ 1, 2, …, N.
the northeast is over 3.9 m/s. In San Francisco, the dominant wind MAE is a measure of the average of the absolute error. The
direction is from the southwestenorthwest (SWeNW) in the range advantage is that the metric is relatively easy for non-specialists to
between 260 and 300 . The wind rose of these cities shows that understand. MAE is defined as follows
the wind speed is highly correlated with the wind direction.
Table 2 summarizes the basic statistics of meteorological fea- 1 XN 
b n
tures for three cities. In Wellington, about 50% of wind directions
MAE ¼ jYn  Y (2)
N n¼1
occur between the north and the northwest. In San Francisco, about
40% of wind directions occur between the west and the northwest.
The wind direction is a circular function resulting in a discontinuity
between 0 and 360 . Thus it is decomposed into two continuous
4.2. Persistence model
input features, sine and cosine functions, using trigonometric
equations.
Before we present our hybrid model called ARIMA-FNN model,
existing forecasting approaches including persistence model,
4. Forecasting methodology ARIMA, FNN, and RNN are briefly reviewed as these models are
used for comparing with the hybrid model.
4.1. Performance measures The persistence method is adopted as the benchmark to
compare the forecasting performances with existing models such
To compare the performance of the proposed forecasting model as ARIMA, FNN and the proposed ARIMA-FNN model. The persis-
with existing methods, two error measures are employed for model tence method is a simple way to forecast the wind speed and often
evaluation: root mean square error (RMSE) and the mean absolute used as a reference to evaluate the performance of other methods.
error (MAE). RMSE is a quadratic scoring rule that also measures This method assumes that the wind speed at time tþDt is the same
the average magnitude of the error, which is more often used in as it was at time t where Dt is the time increment or step size. This
error evaluation. Both RMSE and MAE have the same unit as the ideal is derived from the fact that a high correlation exists between
observation data. RMSE is defined as present and future wind speed. Let Yt and Y b be the observed
tþDt
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi wind speed at t and predicted speed at tþDt, respectively. The
u
u1 X N
persistence model is given as follows,
RMSE ¼ t ðYn  Y b n Þ2 (1)
N n¼1
b
Y tþDt ¼ Yt ; for t ¼ 1; 2; /; T: (4)
b n are,
where N is the number of wind speed observations, Yn and Y

Table 2
Statistics for wind speed (m/s), temperature ( C), and dew point ( C).

Category Wellington San Francisco Phoenix

Dir Speed Temp Dew Dir Speed Temp Dew Dir Speed Temp Dew
Point Point Point

Median 180 7.9 14 10 280 4.1 13.3 8 200 2.5 25 2.2

Mean 191 6.8 13.6 9.4 253 4.6 24.5 8.6 188 2.8 24.8 3.2
Max 360 28.3 29 19 360 20.6 50 19.4 360 17.4 47.2 23
25% 40 4.1 11 7 260 2.1 10.6 5.6 80 2.1 17.8 3.2
50% 180 7.1 14 10 280 4.1 13.2 8 200 2.6 25 2.2
75% 340 9.7 16 12 300 6.7 16.7 10.6 300 4.1 32.2 10
Std Dev 128 3.7 3.4 3.6 84 3.3 4.2 4 117 1.6 9.2 8.9

Note: Dir ¼ wind direction, Temp ¼ temperature, Std Dev ¼ standard deviation.

747
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

4.3. ARIMA model tools. Fig. 9 shows the ACF and PACF graphs presented for wind
speed time series data of Wellington. If both ACF and PACF show
ARIMA(p, d, q) model where the integers represent the param- exponential decay and damped sinusoid, other techniques such as
eters of a particular model is built upon the Box-Jenkins method- Akaike's information criterion, Bayesian information criterion, and
ology. It best fits the time series data with non-seasonal behavior. A grid search could be used to optimize the model structure [33,34].
time series model reproduces future value based on the prior Step 2. Parameter Estimation. After the model structure is
pattern of variables. The univariate method employs ARIMA model determined, the parameters resulting in the lowest residual need to
and takes single feature data set as the input variable. Let yt be the be estimated. YuleeWalker estimation or maximum likelihood
value under prediction at time t. The ARIMA model with d ¼ 0 is estimation can be applied to estimate the parameter values, and
expressed as, uncorrelated residuals can be identified using non-significant P-
values.
X
p X
q Step 3. Diagnostic Checking. After the model structure and pa-
yt ¼ 4i yti þ m þ 4j εtj þ εt rameters are determined, diagnostic checking is used to examine
i¼1 j¼1 the model adequacy and make further improvements. If the model
is a good fit to the data, the residuals should be white noise with no
¼ m þ 41 yt1 þ 42 yt2 þ ; :::; þ 4p ytp þ εt þ 41 εt1 autocorrelation.
þ 42 εt2 þ ; :::; þ 4q εtq (5)

4.4. Feedforward neural network model

where.

FNN is widely applied for time series data forecasting. A typical

p ¼ the number of lag observations in the model, also called the
FNN architecture is shown in Fig. 10. It consists of an input layer, one
lag order.
output layer, and one or more hidden layers. The backpropagation
d ¼ the number of times that the raw observations are different,
(BP) algorithm is the most common learning algorithm to train the
also called the degree of differencing.
FNN model. The learning involves two phases: Phase 1 is a feed-
q ¼ the size of the moving average window, also called the order
forward process in which the information at the input nodes is
of moving average.
propagated to compute the output signal at the output neurons;
4i ¼ the ith autoregressive coefficient.
Phase 2 is called backward process in which the weighted
4j ¼ the jth moving average coefficient.
εt ¼ error term at time t.
εt-j ¼ the random error of a prior points at time t-j.
m ¼ constant term.

Since wind speed often possesses the so-called seasonality, the

measured data exhibit non-stationary behavior. They should be
differenced and converted into a stationary dataset prior to fitting
the ARIMA model. This result in an “integrated” data, and d is the
order of differencing. Fig. 8 shows one-month wind speed data of
Wellington upon differencing. The differenced data possess a good
stationary behavior compared with the original data.
According to the work in [32], three steps are involved in con-
structing an ARIMA model: 1) model identification, 2) parameter
estimation, and 3) diagnostic checking. These steps are elaborated
next.
Step 1. Model Identification. Autocorrelation function (ACF) and
partial autocorrelation function (PACF) can be applied to obtain the
preliminary orders of the ARIMA model. This can be done after the
original data is transformed with stabilized mean and variance.
Graphing the original and the differenced data series, along with
autocorrelation and partial autocorrelation functions, are the useful
Fig. 9. ACF and PACF of stationary wind speed data of Wellington.

Fig. 8. Actual and differenced wind speed data of Wellington.

748
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

Fig. 10. Structure of a three-layer FNN model.

Fig. 11. Flowchart of FNN model development process.

connection and biases are adjusted through gradient descent

optimization. Note input vector is X ¼ [x1, x2, xi], and output vector considered. The proposed model belongs to a class of discrete time,
is Y ¼ [Y ^ 1, Y
^ 2, Y
^ k].
non-linear techniques that can represent a variety of non-linear
Note that X is the input vector of the FNN model, and Y is the dynamic systems, as in the case of wind speed time series.
output vector, or the forecast values. The output of each layer be- It is well documented in the literature that ARIMA and neural
comes the input of the next layer, except for the last layer. An networks have their own advantages, and none of them can always
appropriate transfer function is stored in each hidden layer and outperform the other in terms of the forecasting accuracy. ARIMA
used for processing the data from the input nodes. The function of (p, d, q) model can capture linearity information and decide the
hidden neurons is to connect the input and output layers. Analyt- number of input neurons. The latter is closely related to wind speed
ically, the output is a combination of the weighted input signals and feature selection in this study. Compared with FNN, ARIMA has less
the bias under activation function gk () defined as follows, capability in handling the data with nonlinear characteristics, such
0 !1 as wind speed. Thus, the performance of both forecasting methods
X
J X
I
largely depends on the given data. A hybrid methodology is ex-
b ¼ g @b þ w
Y gj wij xi þ bj A (6)
k k k jk
j¼1 i¼1
pected to capture both linear and nonlinear characteristics of the
input data.
where. As shown in Fig. 12, the development of the hybrid forecasting
model consists of four steps: (1) capturing wind speed autocorre-
I, J, K ¼ the number of the nodes in the input, hidden, and output lation by ARIMA model; (2) estimating the correlations between
layers. different feature pairs; (3) training various input features models
^ k ¼ the predicted value for the node k in the output layer, for
Y and identifying the one with the minimum error; and (4) gener-
k ¼ 1, 2, …, K. ating the forecasted data based on the selected model.
g() ¼ activation function.
wij, wjk ¼ weights between layers.
xi ¼ input for the node i in input layer.
bj, bk ¼ bias values for hidden and output layer, respectively.

The task of the backpropagation algorithm is to minimize the

cost function Jw often expressed as mean square error as follows,

1 XN
b n Þ2
Jw ¼ ðYn  Y (7)
N n¼1

However, the determination of the number of input and hidden

nodes of a neural network structure is vitally important, yet
cumbersome. Architectures with different numbers of input and
output neurons need to be compared to identify the one with the
best performance. To avoid overfitting issue, the available data set
should be divided into three parts: training data, validation data,
testing data. The steps for developing FNN model are illustrated in
Fig. 11.

4.5. Hybrid Forecasting Model

This paper proposes a hybrid model integrating multiple

meteorological features, time series techniques, and FNN algorithm
for short-term wind speed prediction. In the proposed model, not
only the wind speed feature but also other correlated features are Fig. 12. Flowchart for the hybrid model development.

749
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

Table 3
ARIMA models for three cities.

City ARIMA (p, d, q) Model

Wellington ARIMA(4, 0, 2) yt ¼ 2:12yt1  1:59yt2 þ 0:36yt3 þ 0:02yt4 þ þ εt  1:3εt1 þ 0:56εt2

San Francisco ARIMA(2, 0, 1) yt ¼ yt1  7:4yt2 þ εt  2:3εt1
Phoenix ARIMA(4, 0, 2) yt ¼  0:5yt1  0:12yt2 þ 0:76yt3 þ 0:08yt4 þ þ εt þ 1:1εt1 þ 0:99εt2

Table 4 within p2{1, 2, …, 8}, q2{1, 2, …, 8}, and d2{0, 1}. The best values
Error comparison of ARIMA for different forecasting steps. of p, d, and q are determined based on the minimum RMSE or MAE.
City Error (m/s) ARIMA Table 3 lists the ARIMA model with the best forecasting preference
1-h 3-h 8-h 12-h 24-h
for each city. The sign and coefficient indicate the direction and the
strength with pairs of wind speed variable between different
Wellington RMSE 0.93 1.81 1.92 2.12 3.01
hourly measurement. For instance, in Wellington the current wind
MAE 0.82 1.22 1.68 1.75 2.32
San Francisco RMSE 0.91 1.39 1.61 1.91 2.56 speed yt has the positive relation with yt-1, yt-3, yt-4 and negative
MAE 0.96 1.22 1.42 1.68 2.28 relation with yt-2. Hence the ARIMA model with (4, 0, 2) appears to
Phoenix RMSE 0.40 0.84 0.85 0.87 0.96 be the best fit for the wind speed forecasting in Wellington. ARIMA
MAE 0.41 0.73 0.74 0.75 0.84
(2, 0, 1) has been found to be the best for San Francisco, and ARIMA
(4, 0, 2) is the best for Phoenix.
According to Table 4, the prediction accuracy decreases with the
5. Wind speed forecasting and comparison increase of the forecasting horizon, but the degradation of accuracy
is non-linear in terms of RMSE and MAE. For example, in
5.1. Forecasting using ARIMA Wellington the MAE value for 1-h ahead forecasting based on
ARIMA is 0.82 m/s, while the values for 3-h, 8-h, 12-h, and 24-h
Wellington, San Francisco and Phoenix are selected to demon- ahead forecasting are 1.2 m/s, 1.68 m/s, 1.75 m/s, and 1.9 m/s,
strate the effectiveness of the proposed forecasting methods. The respectively. Similar trends can be observed in San Francisco and
11-year meteorological dataset of each city is divided into three Phoenix.
parts. The data from 2004 to 2012 are used for model training. The To compare the gap between the ARIMA outcome and the actual
data of 2013 serve as the model validation. The last year data (i.e. wind speed observation, Fig. 13 shows 1-h and 24-h ahead pre-
2014) are for model testing and forecasting generation. In other diction results along with the actual wind speed data over 24 h in
words, for each city 78,910 observations are used for training. Wellington. The solid line is the wind speed observation, the dotted
Validation and testing use 8760 records, respectively. line is the 1-h ahead forecasting and the dashed line is the day-
ARIMA model is used to detect the existence of the relation ahead forecasting. The 1-h ahead forecasting is quite close to the
between the current and the previous observations. It further de- actual observation, whereas the day ahead result becomes more
termines the optimal neuron number for the input layer that is flatten and smoother. For instance, the lowest wind speed in a day
associated with the autocorrelation of wind speed data. To ensure happens at 11am and the occurrence time based on 1-h ahead
the best performance of the ARIMA model, parameters (p, d, q) forecasting is 12 noon.
should be optimized. In this study, the order of the model varies
5.2. Comparison with ARIMA, FNN and RNN

In this section, the meteorological information of each city is

considered as multivariable in the proposed forecasting model.
Fig. 6 shows there exist significant correlations between wind
speed and other feature variables. For instance, wind speed and
wind direction are correlated because they are influenced by the
Earth's rotation. In addition to wind direction, our approach also
considers air pressure, temperature, and dew point as predictors or
input variables. Table 5 presents five candidate models for
Wellington in order to select the optimal input features. Similar
candidate models can also be constructed for San Francisco and
Phoenix.
Among these models, Model 4 has the best performance
because of the smallest RMSE and MAE as shown in Fig. 14. If more
Fig. 13. 1- and 24-h ahead wind speed forecasting by ARIMA in Wellington. input features are incorporated, it may not improve the prediction

Table 5
Five candidate forecasting models for Wellington with multiple input features.

Model Predictors

Model 1 ‘WindSpd(t-4), ‘WindSpd(t-3)', ‘WindSpd(t-2)', ‘WindSpd(t-1)', ‘WindDir_sin(t-1)', ‘WindDir_cos(t-1)'

Model 2 ‘WindSpd(t-4), ‘WindSpd(t-3)', ‘WindSpd(t-2)', ‘WindSpd(t-1)', ‘WindDir_sin(t-1)', ‘WindDir_cos(t-1)', “AirPressure(t-1)"
Model 3 ‘WindSpd(t-4), ‘WindSpd(t-3)', ‘WindSpd(t-2)', ‘WindSpd(t-1)', ‘WindDir_sin(t-1)', ‘WindDir_cos(t-1)', “AirPressure(t-1)", “Humidity(t-1)"
Model 4 ‘WindSpd(t-4), ‘WindSpd(t-3)', ‘WindSpd(t-2)', ‘WindSpd(t-1)', ‘WindDir_sin(t-1)', ‘WindDir_cos(t-1)', “AirPressure(t-1)", “Humidity(t-1)", ‘Temperature(t-1)'
Model 5 ‘WindSpd(t-4), ‘WindSpd(t-3)', ‘WindSpd(t-2)', ‘WindSpd(t-1)', ‘WindDir_sin(t-1)', ‘WindDir_cos(t-1)', “AirPressure(t-1)", “Humidity(t-1)", ‘Temperature(t-1)', ‘Dewpoint(t-1)'

750
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

accuracy such as Model 5. Therefore, we conclude that wind di-

rection, humidity, air pressure, and temperature are among the
most important features for wind speed forecasting for Wellington
because they lead to the minimum RMSE and MAE errors.
Finally, the neurons at the input and output layers are shown in
Fig. 15(a) for Wellington and Phoenix with nine input neurons.
Fig. 15(b) is for San Francisco with seven input neurons. Since wind
direction is a circular variable, it is decomposed into sine and cosine
parts to ensure the continuity between 0 and 360 .
The comparison in Table 6 shows that the performance of
ARIMA is still better than the proposed hybrid model in 1-h ahead
forecasting. Compared with the persistence model, ARIMA is able to
reduces RMSE by 6.1% as opposed to 5.6% reduction by the hybrid
model.
Fig. 16 compares the 1-h ahead forecasting of ARIMA, FNN, and
Fig. 14. Performance comparison with various input features for Wellington.
the proposed hybrid model with the actual observations in
Wellington. The hybrid model and the ARIMA follow the actual
wind speed quite well, meanwhile a delay exists in the predictions
from the FNN model. Fig. 17 displays the values of RMSE and MAE
for five models: persistence, ARIMA, FNN, RNN and the hybrid
models over five forecasting horizons. It is observed that the per-
formance of the hybrid model remains the best if the forecasting
horizon falls between 3 and 24 h ahead.
The inputs for FNN and RNN are also multiple features, including
the wind speed, wind direction, air pressure, humidity, and tem-
perature. These input features are selected based on their correla-
tion with wind speed. Since the feature selection is similar to our
proposed hybrid model, the detailed process and their structure is
not shown in this paper.
Fig. 18 depicts the prediction intervals (PI) in a day-ahead
forecasting for the wind speed of Wellington. The solid and
dashed lines represent the actual observation and the forecasting,
respectively. The shaded areas display prediction intervals at 95%
confidence for each model. It shows the hybrid model yields the
smallest interval.
The comparisons of four different models indicate that: (1) The
prediction accuracy degrades with the forecasting horizon. (2) The
hybrid model yields a better performance, compared with others in
multi-step prediction. The major performance drop of the hybrid
model occurs between 1 and 3 h as shown in Fig. 17. The perfor-
mance degradation tends to be flat as the horizon increases from 3
to 24 h (3) On the other hand, the proposed hybrid method does not
outperform the ARIMA model in 1-h ahead forecasting according to
the MAE and RMSE values. The reason is that wind speed is time
series data with strong correlation. This means the current wind
speed is likely the same as the previous hour value. Hence the
proposed hybrid model does not outperform ARIMA or persistence
model in very short term, such as the 1-h ahead forecasting in
Wellington and San Francisco.

5.3. Comparison with other hybrid models

Existing hybrid forecasting models often adopt various tech-

niques to improve the prediction quality, such as feature selection
Fig. 15. Hybrid wind speed forecasting models for three cities. [35e37], feather extraction [38,39], model optimization [40], and
residual error analysis [9,41,42]. For instance, Liu et al. [37] utilize
the feature correlation between consecutive wind speed data to
Table 6 determine the number of the model inputs. In their study, besides
Comparison with persistence model in 1-h ahead forecasting for Wellington.
wind speed, both the autocorrelation lags and the wind directions
Improvement ARIMA model Proposed hybrid model are also taken as the input features. Zhou et al. [40] propose a model
RMSE 6.1% 5.6% selection strategy based on multi-criteria optimization with the
MAE 3.4% 1.3% purpose of retaining the non-linearity feature of neural network
models while eliminating unsuitable forecasting models. The
models in [37,40] are chosen to compare the performance with our
proposed hybrid model. This is because our method also uses
751
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

Fig. 16. Comparison of 1-h ahead forecasted results for Wellington.

Fig. 17. RMSE and MAE comparisons with five models for three cities.

feather correlation and model selection to determine the optimal meteorological data of Wellington, San Francisco and Phoenix. It
inputs of the neural networks. shows that our model and the model in [40] exceed the model in
Table 7 summarizes the average value of MAE and RMSE of three [37] as both result in lower MAE and RMSE. Our model also exceeds
models, and the comparisons are made based on the 11-year the model in [40] for cities of San Francisco and Phoenix. But the

752
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

compares the performance of the proposed hybrid model with

persistence model, neural networks, ARIMA and two other hybrid
methods in three different cities. It would be also interesting to
examine the model performance using meteorological data from
other locations of the world with more diverse wind speed profiles.

CRediT authorship contribution statement

Fei Sun: Data curation, Formal analysis, Investigation, Method-

ology, Software, Validation, Visualization, Original manuscript
draft. Tongdan Jin: Conceptualization, Funding acquisition, Project
administration, Supervision, Manuscript review and editing, as well
as Communication with the journal editors.

Declaration of competing interest

The authors declare that they have no known competing

Fig. 18. Prediction intervals in a day-ahead forecasting for Wellington. financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

Table 7
Comparison with other hybrid models.
Acknowledgement

MAE RMSE
This project is supported by the United States National Science
City Xu et al. Zhou et al. Our Xu et al. Zhou et al. Our Foundation (NSF) under the Chemical, Bioengineering, Environ-
[37] [40] Model [37] [40] Model mental and Transport Systems (CBET) program with grant No.
Wellington 1.01 0.75 0.80 1.31 1.04 1.07
1704933.
San 1.22 0.89 0.87 1.50 1.02 0.97
Francisco
Phoenix 0.9 0.73 0.68 1.03 0.89 0.69 References

[1] WWEA, Wind Power Capacity Worldwide Reaches 597 GW, 50.1 GW Added
in, 2018. Available at, https://wwindea.org/information-2/information/.
(Accessed 10 November 2019).
latter yields lower MAE and RASE than ours in Wellington, indi- [2] A. Costa, A. Crespo, J. Navarro, G. Lizcano, H. Madsen, E. Feitosa, A review on
cating that the model in [40] is more capable of handling larger the young history of the wind power short-term prediction, Renew. Sustain.
Energy Rev. 12 (6) (2008) 1725e1744.
wind speed variations like Wellington. [3] M. Lei, L. Shiyan, J. Chuanwen, L. Hongling, Z. Yan, A review on the forecasting
of wind speed and generated power, Renew. Sustain. Energy Rev. 13 (4)
(2009) 915e920.
6. Conclusion [4] D. Carvalho, A. Rocha, M. Go mez-Gesteira, C. Santos, A sensitivity study of the
WRF model in wind simulation for an area of high wind energy, Environ.
Model. Software 33 (2012) 23e34.
This paper proposes a hybrid wind speed forecasting model [5] R.G. Kavasseri, K. Seetharaman, Day-ahead wind speed forecasting using f-
with multivariate inputs and multi-step outputs. The model ARIMA models, Renew. Energy 34 (5) (2009) 1388e1393.
[6] S. Fang, H.D. Chiang, Improving supervised wind power forecasting models
structure is determined through optimal feature selection and
using extended numerical weather variables and unlabeled data, IET Renew.
autoregression analysis. Supervised learning algorithm is used to Power Gener. 10 (10) (2016) 1616e1624.
train the model and determine the model parameters. A set of 11- [7] S. Li, D.C. Wunsch, E.A. O'Hair, M.G. Giesselmann, Using neural networks to
estimate wind turbine power generation, IEEE Trans. Energy Convers. 16 (3)
year wind data from Wellington, San Francisco, and Phoenix are
(2001) 276e282.
used to study the correlations between six meteorological features, [8] M.A. Mohandes, T.O. Halawani, S. Rehman, A.A. Hussain, Support vector ma-
including wind speed. wind direction, air pressure, humidity, dew chines for wind speed prediction, Renew. Energy 29 (6) (2004) 939e947.
point, and temperature. To assess the model performance, multi- [9] E. Cadenas, W. Rivera, Wind speed forecasting in three different regions of
Mexico using a hybrid ARIMAeANN model, Renew. Energy 35 (12) (2010)
step forecasted results are generated and compared with four 2732e2738.
existing models: persistence model, ARIMA, feedforward neural [10] Y. Ren, P.N. Suganthan, N. Srikanth, A comparative study of empirical mode
network, and recurrent neural network. Several findings are ob- decomposition-based short-term wind speed forecasting methods, IEEE
Trans. Sustain. Energy 6 (1) (2014) 236e244.
tained from the numerical comparisons. First, wind direction is [11] C. Zhang, H. Wei, J. Zhao, T. Liu, T. Zhu, K. Zhang, Short-term wind speed
found to be the most significant feature in the multivariate model. forecasting using empirical mode decomposition and feature selection,
It improves the prediction performance and leads to a much better Renew. Energy 96 (2016) 727e737.
[12] W. C. Skamarock, J. B. Klemp, J. Dudhia, D. O. Gill, D. Barker, M. G. Duda, J. G.
forecasted result. Second, the proposed hybrid model outperforms Powers, A Description of the Advanced Research WRF Version 3 (2008) (No.
the benchmark results of four existing models from 3- to 24-h NCAR/TN-475þSTR). University Corporation for Atmospheric Research. doi:
ahead prediction. Third, the persistence and ARIMA models 10.5065/D68S4MVH.
[13] J.G. Powers, J.B. Klemp, W.C. Skamarock, C.A. Davis, J. Dudhia, D.O. Gill,
generally outperform univariate neural network models in 1-h J.L. Coen, D.J. Gochis, R. Ahmadov, S.E. Peckham, G.A. Grell, The weather
ahead forecasting, yet all these models result in good predictions if research and forecasting model: overview, system efforts, and future di-
wind speed involves less variations between consecutive hours, rections, Bull. Am. Meteorol. Soc. 98 (8) (2017) 1717e1737.
[14] Y.K. Wu, J.S. Hong, A literature review of wind forecasting technology in the
such as in Phoenix. There are two interesting directions worthy of
world, in: Proceedings of 2007, IEEE Lausanne Power Tech, 2007,
further investigation. First, wind and solar are complementary in pp. 504e509.
terms of renewable power generation. Utility companies often [15] F. Ziel, C. Croonenbroeck, D. Ambach, Forecasting wind poweremodeling
purchase both wind and solar energy from the day-ahead market, periodic and non-linear effects under conditional heteroscedasticity, Appl.
Energy 177 (2016) 285e297.
hence it is desirable to develop an integrated model that jointly [16] M. Khashei, M. Bijari, An artificial neural network (p, d, q) model for time
forecasts wind speed and sunshine conditions. Second, this study series forecasting, Expert Syst. Appl. 37 (1) (2010) 479e489.

753
F. Sun and T. Jin Renewable Energy 186 (2022) 742e754

[17] G.P. Zhang, M. Qi, Neural network forecasting for seasonal and trend time Weibull wind speed distribution for wind energy analysis, J. Wind Eng. Ind.
series, Eur. J. Oper. Res. 160 (2) (2005) 501e514. Aerod. 85 (1) (2000) 75e84.
[18] Y. Amellas, O. El Bakkali, A. Djebli, A. Echchelh, Short-term wind speed pre- [32] G. Box, G. Jenkins, G. Reinsel, Time Series Analysis: Forecasting and Control,
diction based on MLP and NARX networks models, Indones. J. Electr. Eng. fourth ed., John Wiley & Sons, Hoboken, NJ, 2008.
Comput. Sci. 18 (1) (2020) 150e157. [33] J. Kuha, AIC and BIC: comparisons of assumptions and performance, Socio.
[19] H. Liu, H.Q. Tian, Y.F. Li, Comparison of two new ARIMA-ANN and ARIMA- Methods Res. 33 (2) (2004) 188e229.
Kalman hybrid methods for wind speed prediction, Appl. Energy 98 (2012) [34] K.H. Pho, S. Ly, S. Ly, T.M. Lukusa, Comparison among Akaike information
415e424. criterion, Bayesian information criterion and Vuong's test in model selection:
[20] I.G. Damousis, M.C. Alexiadis, J.B. Theocharis, P.S. Dokopoulos, A fuzzy model a case study of violated speed regulation in Taiwan, J. Adv. Eng. Comput. 3 (1)
for wind speed prediction and power generation in wind parks using spatial (2019) 293e303.
correlation, IEEE Trans. Energy Convers. 19 (2) (2004) 352e361. [35] J. Yan, G. Huang, X. Peng, Y. Li, Q. Yang, A novel wind speed prediction
[21] M.A. Mohandes, S. Rehman, T.O. Halawani, A neural networks approach for method: hybrid of correlation-aided DWT, LSSVM and GARCH, J. Wind Eng.
wind speed prediction, Renew. Energy 13 (3) (1998) 345e354. Ind. Aerod. 174 (2018) 28e38.
[22] M.C. Mabel, E. Fernandez, Analysis of wind power generation and prediction [36] N. Bokde, A. Feijo  o, K. Kulat, Analysis of differencing and decomposition
using ANN: a case study, Renew. Energy 33 (5) (2008) 986e992. preprocessing methods for wind speed prediction, Appl. Soft Comput. 71
[23] H. Peng, F. Liu, X. Yang, A hybrid strategy of short-term wind power predic- (2018) 926e938.
tion, Renew. Energy 50 (2013) 590e595. [37] W. Xu, P. Liu, L. Cheng, Y. Zhou, Q. Xia, Y. Gong, Y. Liu, Multi-step wind speed
[24] Q. Zhou, C. Wang, G. Zhang, Hybrid forecasting system based on an optimal prediction by combining a WRF simulation and an error correction strategy,
model selection strategy for different wind speed forecasting problems, Appl. Renew. Energy 163 (2021) 772e782.
Energy 250 (2019) 1559e1580. [38] J. Wang, Y. Wang, Y. Li, A novel hybrid strategy using three-phase feature
[25] Q. Cao, B.T. Ewing, M.A. Thompson, Forecasting wind speed with recurrent extraction and a weighted regularized extreme learning machine for multi-
neural networks, Eur. J. Oper. Res. 221 (1) (2012) 148e154. step ahead wind speed prediction, Energies 11 (2) (2018) 321.
[26] D.M. Kline, G.P. Zhang, Methods for multi-step time series forecasting with [39] F. Davo , S. Alessandrini, S. Sperati, L.D. Monache, D. Airoldi, M.T. Vespucci,
neural networks, in: Book: Neural Networks in Business Forecasting, 2004, Post-processing techniques and principal component analysis for regional
pp. 226e250, https://doi.org/10.4018/978-1-59140-176-6.ch012. wind power and solar irradiance forecasting, Sol. Energy 134 (2016) 327e338.
[27] G. Zhang, B.E. Patuwo, M.Y. Hu, Forecasting with artificial neural networks: [40] Q. Zhou, C. Wang, G. Zhang, Hybrid forecasting system based on an optimal
the state of the art, Int. J. Forecast. 14 (1) (1998) 35e62. model selection strategy for different wind speed forecasting problems, Appl.
[28] C. Hamzaçebi, Improving artificial neural networks' performance in seasonal Energy 250 (2019) 1559e1580.
time series forecasting, Inf. Sci. 178 (23) (2008) 4550e4559. [41] Z. Liu, R. Hara, H. Kita, Hybrid forecasting system based on data area division
[29] C. Croonenbroeck, G. Stadtmann, Renewable generation forecast and deep learning neural network for short-term wind speed forecasting,
studieseReview and good practice guidance, Renew. Sustain. Energy Rev. 108 Energy Convers. Manag. 238 (2021), 114136, https://doi.org/10.1016/
(2019) 312e322. j.enconman.2021.114136.
[30] C.G. Justus, W.R. Hargraves, A. Mikhail, D. Graber, Methods for estimating [42] Q. Han, F. Meng, T. Hu, F. Chu, Non-parametric hybrid models for wind speed
wind speed frequency distributions, J. Appl. Meteorol. 17 (3) (1978) 350e353. forecasting, Energy Convers. Manag. 148 (2017) 554e568.
[31] J.V. Seguro, T.W. Lambert, Modern estimation of the parameters of the

754