IEEE Computational Intelligence Magazine - August 2021 - 51

efficiently capture the nonlinear correlations among distinct
features and dendritic branches. Extensive experimental
results and statistical tests demonstrate that compared with
other state-of-the-art prediction techniques, DNR can
achieve highly competitive results in wind speed forecasting.
I. Introduction
ith the rapid development of society, it has
W
become challenging for the remaining fossil
fuel supply to meet societal requirements worldwide;
thus, the demand for sustainable renewable
energy is growing. In recent years, wind energy, as a sustainable
renewable energy source, has attracted particular attention.
Wind is an abundant, pollution-free energy source that
is widely distributed, has few geographical restrictions, and
can be easily used anywhere. Building wind farms is one of
the main challenges in the use of wind energy. Nevertheless,
the cost of building wind farms is low, no additional energy
source is required, and such farms do not impact the surrounding
environment.
Since the energy supplied by wind farms mainly depends
on the wind power, precise wind power forecasting is imperative
to enable reliable power system planning and wind
farm operation [1, 2]. Wind power is closely related to wind
speed, and these two parameters share several characteristics,
such as randomness, uncontrollability and intermittency [3].
Because of these characteristics, wind farm management is
extremely challenging. Calculating the energy production
of a wind farm is essential for assessing the economic feasibility
of such a project prior to construction planning [4].
Therefore, accurate wind speed prediction is being strongly
prioritized in this context [5]. More accurate forecasting
capabilities correspond to larger reductions in the construction
costs of wind farms [6].
To date, various models have been used for wind speed
prediction. These methods can be classified into three categories:
physical models, statistical models, and artificial
intelligence models
[7]. Examples of physical models
include the Mesoscale Model Version 5 [8] and the Weather
Research and Forecasting Model [9]. These numerical prediction
models can achieve satisfactory performance in
long-term wind speed prediction; however, such models
require complex atmospheric information pertaining to
pressure, temperature and other environmental factors and
exhibit high computational complexity [10], [11].
Compared to physical models, statistical models are more
widely used to forecast wind speed. Examples of such models
include direct random time series models [12], autoregressive
models [13] and autoregressive integrated moving average
(ARIMA) models [14]. ARIMA models are regarded as a
typical class of statistical models, and their prediction performance
in short-term wind speed forecasting has been verified
[15]. However, in general, the prediction performance of
statistical models is flawed [16] because most statistical models
are based on the assumption that the wind speed series
follows a normal distribution, although this assumption is not
valid in all cases. Moreover, statistical models have linear
correlation structures and always yield large errors when
applied to intermittent and stochastic wind speed series [17].
The third category pertains to artificial intelligence models,
which can effectively overcome the shortcomings of the
abovementioned methods. Considering the nonlinear nature
of wind speed series, artificial intelligence algorithms that are
designed for effectively solving nonlinear problems, such as
artificial neural networks (ANNs) [16] and support vector
machines (SVMs) [18], are suitable for wind speed forecasting.
A previous comparison of prediction performance has
demonstrated that artificial intelligence algorithms are faster
and more accurate than statistical models [19]. SVMs are
commonly used in prediction frameworks, and they can outperform
ANNs in certain cases. However, the performance
of an SVM is limited by its penalty settings and kernel
parameters; consequently, algorithms for tuning these hyperparameters
are necessary [20]. For example, a genetic algorithm
was employed to enhance the prediction results of an
SVM in [21]; a reduced SVM with feature selection, trained
using the particle swarm optimization algorithm, was used to
optimize the parameters of an SVM in [22]; and the performance
of an SVM was enhanced using the cuckoo search
algorithm in [23]. In addition to SVMs, an increasing number
of ANNs and their variants have been proposed for wind
speed prediction. For instance, a backpropagation (BP) neural
network was employed to forecast a wind speed series in
[24], a combination of an ANN and Markov chains was proposed
for forecasting in [25], and a functional network was
utilized for multistep wind speed prediction in [26]. Furthermore,
a fine-tuned long short-term memory (LSTM) neural
network hybridized with the crow search algorithm, the
wavelet transform and feature selection was applied for shortterm
wind speed forecasting in [27]. In [28], a hybrid model
involving a causal convolutional network and a gated recurrent
unit architecture was used in wind speed prediction.
In general, physical models and traditional statistical models
both have several limitations pertaining to the precision
and robustness of wind speed time series prediction, whereas
artificial intelligence models can effectively overcome these
problems to offer more powerful prediction performance.
Based on these considerations, the dendritic neuron model
(DNM), which was recently developed based on inspiration
from biological neurons in vivo [29], is adopted in this study
for the prediction of wind speed time series. In the DNM,
synaptic nonlinearity is implemented in a dendritic structure
to effectively solve linearly inseparable problems, and this
model has been applied to a variety of complex continuous
functions [30]-[32]. The original DNM was specifically
designed for classification problems. By discarding the unnecessary
synapses and dendritic branches in the DNM, diverse
dendritic structures can be produced to pursue an extremely
high classification speed for each task. Notably, however, the
structure of the original DNM is extremely simple, and it
AUGUST 2021 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 51
IEEE Computational Intelligence Magazine - August 2021

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - August 2021
