IEEE Computational Intelligence Magazine - August 2021 - 52

The demand for sustainable renewable energy is
growing. Wind energy, as a sustainable renewable
energy source, has attracted particular attention.
ignores the mechanism by which the signal transformation
strengths of the dendritic branches vary with their thickness
[33]. Therefore, this study proposes the use of a variant of the
DNM called dendritic neural regression (DNR), in which the
intensity of each dendritic signal is considered to enhance the
regression ability.
The neural architecture considerably influences the model
performance of ANNs [34]. Accordingly, evolutionary algorithms
have been widely used to realize neural architecture
search tasks for performance optimization. For instance, Sun
et al. proposed an automatic method of designing convolutional
neural network (CNN) architectures for solving
image classification problems based on genetic algorithms
[35]. Lu et al. used a multiobjective evolutionary algorithm
for CNN architecture design [36]. In contrast to that of
conventional ANNs, the neural architecture of a DNR
model is determined by the values of the weights and
thresholds rather than by hyperparameters. Subsequently, an
inherent pruning mechanism can be implemented to simplify
the DNR model to produce unique neural architectures
for particular real-world tasks, as described in our previous
research [29], [37]. To further enhance the performance of
wind speed forecasting, the recently proposed states of matter
search (SMS) algorithm [38] can be used to optimize the
neural architecture of the DNR model. The SMS algorithm
is a global search algorithm that can escape from local minima
more effectively than gradient-based optimization algorithms.
In fact, many researchers have attempted to employ
evolutionary algorithms to enhance the performance of
ANNs for chaotic time series prediction. For example, particle
swarm optimization has been introduced into echo
state networks as a pretraining tool
to optimize the
untrained weights to address time series forecasting problems
[39]. Furthermore, a genetic algorithm has been introduced
into the Elman neural network to optimize the
connection weights and thresholds to prevent the optimization
from becoming trapped in local minima and enhance
the training speed and success rate [40]. Moreover, a modified
cuckoo search algorithm has been used to optimize
wavelet neural networks to achieve higher generalization
capabilities in chaotic time series forecasting [41]. In addition,
an evolving fuzzy neural network predictor has been
proposed to effectively capture the dynamic properties of
multidimensional datasets and accurately track the system
characteristics [42], while in [43], a novel strategy was proposed
for evolving the structure of deep recurrent neural
networks by means of ant colony optimization.
Compared with other ANNs, a DNR model can help
enhance the performance of wind speed prediction because
52 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | AUGUST 2021
its plastic neural architecture can efficiently
capture the nonlinear correlations among
distinct features and different dendritic
branches. In addition, a one-dimensional
wind speed time series typically appears to
exhibit intermittent and random features
since it contains complex information projected from a higher-dimensional
space. Thus, the phase space should be reconstructed
based on Takens' theorem [44]. The time delays and
numbers of embedding dimensions for a wind speed time
series can be calculated using a mutual information (MI)
approach and the false nearest neighbors (FNN) algorithm,
respectively. To evaluate the forecasting results of the DNRSMS
method, a comprehensive experiment has been conducted
in which the performance of the DNR approach has
been compared with that of other commonly used time series
prediction algorithms.
The main contributions of this study are as follows:
First, a novel DNR model
that considers
the dendrite
strength of each individual branch is proposed. This
approach can significantly enhance the regression performance
for wind speed forecasting. Second, due to its powerful
search ability, the SMS algorithm is used to optimize
the neural architecture of the DNR model. The SMS algorithm
can escape from local minima more effectively than
gradient-based optimization algorithms can. Third, the
results of extensive experiments demonstrate that a DNR
model trained using the SMS algorithm can achieve highly
competitive performance in wind speed forecasting compared
with other state-of-the-art prediction techniques.
The remainder of the paper is organized as follows: Section
II describes the DNR approach. Section III introduces the
process of wind speed prediction. Section IV presents and
discusses the experimental results. Finally, concluding
remarks are presented in Section V.
II. Dendritic Neural Regression
In this study, we demonstrate the utilization of DNR for
wind speed prediction and evaluate its performance at various
time scales. As illustrated on the left side of Fig. 1, the
neural architecture for DNR includes four layers. The first
layer is the synaptic layer, which represents the specific tissue
that receives electrical or chemical signals from other neurons.
The second layer is the dendrite layer, which consists of
many branches that integrate the output signals from the
synaptic layer. The third layer is the membrane layer, which
sums the outputs of the dendritic layer and transfers the result
to the next layer. The final layer is the cell body (soma),
which compares the signal against a given threshold. If the
signal is larger than the threshold, the soma fires; otherwise,
no action is performed.
A. Synaptic Layer
Synapses are a kind of neural tissue that conveys information
between dendrites and axons or among dendrites of different
IEEE Computational Intelligence Magazine - August 2021

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