Received 5 June 2023, accepted 21 June 2023, date of publication 30 June 2023, date of current version 7 July 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3291146

Air Quality Index Forecasting via Genetic

Algorithm-Based Improved Extreme
Learning Machine
CHUNHAO LIU , (Graduate Student Member, IEEE), GUANGYUAN PAN , (Member, IEEE),
DONGMING SONG , AND HAO WEI
School of Automation and Electrical Engineering, Linyi University, Linyi 276000, China
Corresponding author: Dongming Song ([email protected])
The research conducted in this study was supported by National Natural Science Foundation of China under Grant no. 62103177. This
funding played a crucial role in facilitating data collection, conducting experiments, and analyzing the results.

ABSTRACT Air quality has always been one of the most important environmental concerns for the general
public and society. Using machine learning algorithms for Air Quality Index (AQI) prediction is helpful
for the analysis of future air quality trends from a macro perspective. When conventionally using a single
machine learning model to predict air quality, it is challenging to achieve a good prediction outcome under
various AQI fluctuation trends. In order to effectively address this problem, a genetic algorithm-based
improved extreme learning machine (GA-KELM) prediction method is enhanced. First, a kernel method
is introduced to produce the kernel matrix which replaces the output matrix of the hidden layer. To address
the issue of the conventional limit learning machine where the number of hidden nodes and the random
generation of thresholds and weights lead to the degradation of the network learning ability, a genetic
algorithm is then used to optimize the number of hidden nodes and layers of the kernel limit learning machine.
The thresholds, the weights, and the root mean square error are used to define the fitness function. Finally, the
least squares method is applied to compute the output weights of the model. Genetic algorithms are able to
find the optimal solution in the search space and gradually improve the performance of the model through an
iterative optimization process. In order to verify the predictive ability of GA-KELM, based on the collected
basic data of long-term air quality forecast at a monitoring point in a city in China, the optimized kernel
extreme learning machine is applied to predict air quality (SO2 , NO2 , PM10 , CO, O3 , PM2.5 concentration
and AQI), with comparative experiments based CMAQ (Community Multiscale Air Quality), SVM (Support
Vector Machines) and DBN-BP (Deep Belief Networks with Back-Propagation). The results show that the
proposed model trains faster and makes more accurate predictions.

INDEX TERMS Time series, air quality forecasting, machine learning, extreme learning machine, genetic
algorithm.

I. INTRODUCTION (GB3095-2012), there are six conventional air pollutants

Air pollution is a prevalent environmental problem in the used to measure air quality: sulfur dioxide (SO2 ), nitrogen
twenty-first century. In light of the rapid industrialization and dioxide (NO2 ), particulate matter with a particle size less than
urbanization, air pollution is getting worse, which greatly 10 microns (PM10 ), particulate matter with a particle size less
affects our living environment and health [1]. Li et al. came to than 2.5 microns (PM2.5 ), ozone (O3 ), and carbon monoxide
the conclusion that outdoor physical activity poses numerous (CO) [4], [5], [6]. These pollutants have adverse effects on
health risks due to ambient air pollution in China. [2], [3]. human health. The International Energy Agency estimates
According to the Chinese Ambient Air Quality Standards that air pollution causes 6.5 million premature deaths per
year, while long-term exposure to pollutants, such as fine
The associate editor coordinating the review of this manuscript and particles (e.g.,PM2.5 ) or traffic-related pollutants, is linked
approving it for publication was Frederico Guimarães . to higher rates of lung cancer, coronary heart disease, and
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
67086 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
C. Liu et al.: AQI Forecasting via GA-KELM

other illnesses [7], [8]. Therefore, studies on air quality performance and outperforms some advanced methods cur-
prediction are particularly important and are considered a rently in use. The main contributions of this paper are:
key factor for environmental protection. In order to more (1) modifying the ELM activation function or using the kernel
comprehensively assess the health effects of air pollution, function to improve the prediction accuracy, (2) optimizing
numerous air quality monitoring stations have been set up in the ELM using GA to improve the stability of the results and
major cities. Air quality predictions can be made based on further enhance the prediction accuracy, and (3) obtaining
the data collected from these stations. Air quality monitoring, the correlation analysis results of atmospheric environmental
modeling, and accurate predictions are important for having quality prediction parameters by comprehensively consider-
a clear understanding of future pollution levels and their ing each relevant factor in line with the actual situation. The
associated health risks. remainder of this paper is organized as follows. Section II
Recently, the inherent property of machine learning presents related work. Section III describes ELM and the
algorithms to automatically learn features at multiple levels of proposed GA-KELM, and illustrates the improvements using
abstraction has become increasingly important in providing the model. Section IV discusses experimental results where
solutions to this challenging task [9], [10]. However, the GA-KELM is compared with several other methods in terms
model only forecasts PM10 and SO2 levels, and it is of prediction results. The last section concludes the entire
also challenging to obtain measurement values needed to work and presents directions for future research.
construct the dataset [11]. Wu Q. et al. proposed an
optimal-hybrid model for daily AQI prediction considering
air pollutant factors, with the model’s inputs being the six II. RELATED WORK
atmospheric pollutants. However, neural networks typically Air quality prediction has been extensively researched in
struggle with slow learning, a tendency to fall into local the literature [17]. In recent years, numerous researchers
minima, and a complex network training process. Based have made significant contributions to the field by leveraging
on the generalized inverse matrix theory, Huang et al. quantitative studies and the latest techniques to identify
proposed an extreme learning machine (ELM) algorithm various air quality patterns and their underlying trends [18].
with a feedforward neural network that includes a single Existing work in this area relies on statistical methods and
hidden layer, such that the problems of conventional neural shallow machine learning models to address the problem of
network algorithms are circumvented. The ELM algorithm air quality prediction [19].
used to predict the AQI outperformed neural networks in Agarwal and Sahu [20] conducted air quality prediction
terms of parameter selection, training speed, and prediction studies by employing statistical models. Lary et al. [21] com-
accuracy [12]. However, the parameters of the hidden layer bined remote sensing and meteorological data with ground-
nodes and the number of nodes in the test hidden layer are based PM2.5 observations. Zheng et al. [22] proposed a hybrid
selected at random, which puts the prediction accuracy to a prediction method that combines a linear regression- based
great test. temporal prediction method with an ANN-based spatial pre-
In order to solve the aforementioned problems, we propose diction method for pollutant concentrations. Zheng et al. [23]
to optimize the number of ELM hidden layer nodes, used a data-based approach for the next 48 hours of
thresholds, and weights, along with an improved genetic PM2.5 prediction, implementing a prediction model based
algorithm (GA) that uses root mean square error (RMSE) as on linear regression and neural network. They combined
the fitness function, to obtain the optimal network structure meteorological data, weather forecast data, and air quality
for air quality prediction [14]. The number of hidden layer data from monitoring stations. Rajput and Sharma [24] used
nodes is updated by continuous coding discretization, the a multiple regression model to represent the changes in
input weights and hidden layer thresholds are updated by air quality index (AQI), considering ambient temperature,
continuous coding, and the update thresholds and weights relative humidity, and barometric pressure as the main
are selected with the number of updated layers to form a parameters in the regression model for AQI calculation [25].
hierarchical control structure [15]. The proposed GA-based These classical methods and models all have the advantages
improved extreme learning machine (GA-KELM) algorithm of simple algorithms, easy processing, and acceptable
is applied to air quality prediction, and its performance is prediction results. However, obtaining precise and specific air
compared with that of community multiscale air quality quality prediction values remains challenging [26].
modeling system (CMAQ), support vector regression (SVR), Elbaz et. al. [27] proposed a novel deep learning approach
and deep belief network-back propagation (DBN-BP). The that extracts high-level abstractions to capture the spatiotem-
results show that the accuracy of the proposed GA-KELM poral characteristics of NEOM city in Saudi Arabia at
algorithm is reliable for air quality prediction [16]. hourly and daily intervals. Campbell et al. [28] described
In this study, an improved extreme learning machine the development of FV3GFSv16 coupled with the ‘‘state-
model based on a genetic algorithm is designed and of-the-art’’ CMAQ model version 5.3.1. Jin et al. [29]
applied to AQI prediction. To verify the effectiveness of proposed an interpretable variational Bayesian deep learning
the model, we conducted tests on three real-world datasets. model with self-filtering capability for PM2.5 prediction
The results confirmed that the proposed method has superior information, which effectively improves prediction accuracy.

VOLUME 11, 2023 67087

 C. Liu et al.: AQI Forecasting via GA-KELM

Zhou et al. [30], [31], [32] proposed a method based on an The AQI of each pollutant, or individual AQI (IAQI), is the
improved Grasshopper optimization algorithm to classify the largest value of the day. AQI is calculated as in (1):
color difference of dyed fabrics using kernel extreme learning
machine. In this study, the classification of color differences AQI = max {LAQI1 , L AQI2 , L AQI3 , · · · , L AQIn } (1)
in dyed fabric images is performed using the kernel
limit learning machine, and the kernel function parameters where LAQI1 , L AQI2 , L AQI3 , · · · , L AQIn represent the
are optimized by the improved Grasshopper optimization sub-index corresponding to each pollutant item. In this
algorithm to achieve color difference classification of dyed study, the calculation of AQI involves only six pollu-
fabric images. Xue et al. [33] proposed a GA-based air tants.Therefore, (1) can be expressed as:
quality prediction model to optimize the parameters of the
weighted extreme learning machine (WELM). Despite the AQI = max{IAQISO2 , IAQIN2 , IAQIO3 ,
progress made by the aforementioned methods, they also
IAQIPM10 , IAQIPM2.5 , IAQICO } (2)
exhibit limitations; their training efficiency is relatively low,
and deep learning algorithms are not yet fully mature. These
To meet our research goals, this paper proposes
challenges present greater obstacles for the application of
GA-KELM, a method for air quality prediction based on
deep learning, necessitating improvements to existing mod-
an improved extreme learning machine, which in turn is
els, the development of new models, and the enhancement of
based on an improved GA [40]. Aiming at the problem of
their predictive capabilities [34], [35].
network instability caused by the randomly generated input
The use of statistical or numerical forecasting techniques
layer weights and hidden layer thresholds of KELM, a GA is
is subject to several limitations. Neural networks are widely
used to optimize the KELM weights and thresholds, thereby
used because of their unique associative abilities, memory,
improving the model’s performance in terms of prediction
and distinctive learning [36], [37]. Given the highly nonlinear
accuracy, which is the main objective of this algorithm.
nature of AQI changes and the strong generalization and
In each iteration of the GA, a new offspring population
nonlinear characterization abilities of neural networks, the
is generated by selection, crossover, and mutation, and the
nuclear limit learning machine neural network model,
individual with good fitness value is selected. The GA stops
also known as kernel extreme learning machine (KELM),
iterating when the stopping criteria are satisfied. The GA is
is employed to investigate air quality prediction using a real
used to determine the optimal weights and threshold values,
dataset. The weights and threshold values of KELM are
which overcomes the instability of KELM and reduces
optimized using a genetic optimization algorithm [38].
prediction errors, thus resulting in a more reliable prediction
model and improved air quality prediction accuracy. Details
III. METHODOLOGY
of the model will be discussed in the following sections.
In this section, AQI is first introduced, the ELM and KELM
algorithms are presented next, and a new GA-KELM learning
method for AQI prediction is then proposed. A. EXTREME LEARNING MACHINE
Air quality forecasting has been a key issue in early ELM was first proposed by Huang. It is characterized by
warning and control of urban air pollution. Its goal is to its fast training and high training accuracy. Feedforward
anticipate changes in the AQI value at observation points over neural networks are mainly based on the gradient descent
time. The observation period, which is decided by the ground- method [41]. Their main drawbacks are the slow training, the
based air-quality monitoring station, is usually set for one tendency to fall into a local minimum point that cannot reach
hour. the global optimum, and the high sensitivity to the learning
Furthermore, a location’s air quality value is largely influ- rate η (if the selected rate is improper, it might cause slow
enced by the weather conditions prevailing at that location. convergence and the training process will thus take a long
Air quality monitoring stations measure air temperature, time, or it becomes unstable). FIGURE 2 shows the network
wind direction, atmospheric pressure, relative humidity, wind structure of an ELM.
speed and other meteorological parameters, as well as air Consider N different samples (xi , ti ) , i ∈ 1, 2, · · · N ,
pollutant concentrations [39]. Air quality prediction is also where xi denotes the input and ti represents the target, L
challenging due to the rapid changes in pollutant emission hidden layer neurons, and an activation function g(x).The
and weather conditions. Numerous variables, such as wind mathematical expression for ELM output is:
speed, temperature, humidity, and pollutants themselves, are X
β j g wj xi + bj


highly nonlinear, dynamic, and have inherent interdependen- yi = (3)
cies, making it more challenging to accurately predict air
quality at a specific time and place. Therefore, it is essential where j ∈ 1, 2, · · · , L, wj are the weights of the input and
to figure out how to deal with these factors and exploit hidden layer neurons,bj are the thresholds of the hidden layer
them from multivariate time-series data related to air quality. neurons, βj is the weight matrix
of neurons in the hidden and
A typical meteorological factor sequence diagram is shown output layers, and g wj xi + bj is the output of the hidden
in FIGURE 1. layer neurons.

67088 VOLUME 11, 2023

C. Liu et al.: AQI Forecasting via GA-KELM

FIGURE 1. Sequence diagram of meteorological factors.

where H + is the Moore-Penrose generalized inverse of H .

When the number of training observations is greater than the
number of hidden nodes, the expression β of the calculated
output weight is given by:
 −1
I
β̂ = + HT H HT T (8)
C

where I is the K-dimensional identity matrix and CI is

the regularization parameter. When the number of training
samples is less than the number of hidden nodes, the output
weight β̂ expression is written as:
 −1
I
β̂ = H T + HH T T (9)
FIGURE 2. Schematic diagram of ELM structure. C

B. IMPROVED ELM BASED ON KERNEL METHOD

In order to approximate the ideal output result, there must
Inspired by SVM, the kernel method is introduced into
be a set of wj , bJ and βj such that:
ELM, namely KELM. KELM maps linearly inseparable
N
X patterns to a high-dimensional feature space to achieve
β j g wj xi + bj = ti , i = 1, 2, · · · , N


(4) linear separability and improve the accuracy and robustness
i=1 of the model [42]. Therefore, it ensures that the network
Like the single hidden layer neural network, (4) can be produces good generalization and learns fast, and it avoids
expressed as: the shortcomings of gradient descent training algorithms that
are prone to falling into local optima and necessitate a large
Hβ = T (5) number of iterations.
where H is the output matrix of the hidden layer neurons and ELM is a single hidden layer feedforward neural network,
T is the target output of the ELM network, i.e.,T = {ti }N and its learning objective (10) can be expressed as a matrix:
i=1 ,
so (5) can be transformed into the least squares problem of
F(x) = h(x)β = H β = L (10)
solving for the weight matrix, namely:
∥Hβ̂ − T∥ = min ∥Hβ− | T∥ (6) Turning network training into a problem of solving a linear
β system, β is determined by β = H ∗ L,where H ∗ is a
generalized matrix of H .In order to enhance the stability of
The output weight matrix β̂ can be obtained as:
the model, the regularization coefficient C and the identity
β̂ = H + T (7) matrix I are introduced, and the output weights can be

VOLUME 11, 2023 67089

 C. Liu et al.: AQI Forecasting via GA-KELM

calculated by the least square’s solution in (11).

I −1
 
β = H T HH T + L (11)
c
In addition to the many advantages of the ELM algorithm,
KELM combines the kernel function to achieve linear
separability by non-linearly mapping the linearly inseparable
pattern to the high-dimensional feature space, which further
improves the judgmental accuracy. The kernel function is
introduced into ELM, and the kernel matrix is:
ELM = HH T = h (xi ) h xj = K xi , xj



(12)
where xi and xj represent test input vector. Therefore,(10) can
be expressed as:
 −1
I
F(x) = [K (x, x1 ) ; . . . ; K (x, xn )] + ELM L
C FIGURE 3. GA-KELM algorithm flowchart.
(13)
where x1 , x2 , · · · , xn is a given training sample, n is the
number of samples, and K (x1 , xi ) is the kernel function,
i = 1, 2, · · · , n. One of the widely used kernel functions is
the radial basis function (RBF). The RBF kernel can be
defined as:
 
K (x, y) = exp −γ ∥x − y∥2 (14) FIGURE 4. Model architecture diagram.

where y is the kernel parameter. Since the results of the

KELM model strictly depend on the regularization parameter Step 2: The input weights w and thresholds b of the hidden
C and the choice of the kernel parameter y, these two layer of the ELM network are randomly generated as the
parameters need to be optimized effectively. initial population of GA.
In conventional ELM, input weights and hidden layer Step 3: The KELM model is trained using the training
neurons are assigned at random, and output weights are then samples, and the trained network is used to predict test
obtained on the basis of input weights and hidden layer samples. The testing result is sent to the GA as objective
neurons. Thus, human experiments are essentially important function.
in customary training, which often results in inaccurate Step 4: The improved GA is used to optimize weights and
predictions [43]. To overcome this shortcoming, this paper model structure.
considers using a GA to optimize input weights and model Step 5: Steps 3-4 are repeated until convergence is reached.
structure. In this section, as the input weights are continuous Step 6: The KELM is assigned the optimal hidden layer
variables and the model structure is discrete, the GA is input weights w and thresholds b for the final training and
improved to handle the difference between continuous and testing.
discrete variables of KELM in parallel optimization. The
improved GA-based KELM process is shown in FIGURE 3. C. OVERVIEW OF THE DEEP AIR QUALITY FORECASTING
The steps for using a GA to optimize the parameters of FRAMEWORK
KELM are as follows: The flowchart of applying GA-KELM for AQI prediction is
Step 1: Data normalization. Since the observed data has shown in FIGURE 4.
different properties, dimensions, and obvious differences in a) Data collection and preparation. In GA-KELM, before
numerical values, normalization is performed so that the performing GA, the fitness value obtained during the
distribution range of the processed data is kept between prediction process of the initial KELM is considered.
[0, 1], in order to avoid learning errors caused by the bias. b) Prediction. In the experiment, two-thirds of the collected
The normalization formula is given by: data are used to train the KELM network, and the remaining
one-third is used as a test set.
x − xmin
X= (15) c) Optimization. The GA is applied through the processes
xmax − xmin of selection, crossover and mutation, until the optimal result
where x is the observed value, X is the input data of is obtained. The weights and structure parameters of KELM
the prediction model,xmin and xmax are the minimum and are then updated using the values produced by the GA. This
maximum of the observed values, respectively [44]. process is repeated until the optimal result is achieved.

67090 VOLUME 11, 2023

C. Liu et al.: AQI Forecasting via GA-KELM

All the spatiotemporal relationships that dominate the

input data are not entirely helpful in predicting the output.
Each feature has a level of association with the predicted
output. The goal of GA-KELM is to identify spatiotemporal
features and the degree of their association with the predicted
output learned through its weight matrix [45].

IV. EXPERIMENTAL RESULTS

In this section, experiments using real-world air quality
datasets are conducted to analyze and evaluate the perfor-
mance of the proposed model. The predictive performance of
the model is verified by comparing with other baseline deep
learning models.

A. DATASETS DESCRIPTION AND DATA PREPROCESSING

1) DATA COLLECTION AND DATA BASICS
For the study in this paper, we will use 3 real air
quality datasets. Our dataset is available on the following FIGURE 5. Before data processing.
link: https://github.com/chunhaoliu/AQI_Data.git. It includes
air quality data from July 2019 to July 2021, cover- outliers in the observed data, statistical analysis methods are
ing all seasons and different weather conditions. The usually used. To extract multiple observations and remove
dataset contains multidimensional air pollutant monitoring outliers, statisitcal criteria such as the Lajda criterion and the
µg/m 3 , NO µg/m3 , PM µg/m 3 ,




data, such as: SO 2 2 10 Grubbs criterion are commonly employed.
PM2.5 mg/m3 , O3 µg/m3 and CO mg/m3 , hourly aver-




The Lajda criterion may not be highly accurate in removing
age concentration data and day-by-day detection data. and identifying outliers when the number of measurements
In the preprocessing stage, the data are initially examined is small (n<10). This method follows a three-step process
for missing value outliers. To ensure data completeness for determining outliers. First, it assumes that the dataset
and accuracy, linear interpolation and 3σ are applied for contains only chance errors. Second, it computes the standard
processing. The original data are then sorted by time series deviation of this set of data, and determines the interval
and divide it into a training set and a test set. This dataset range by certain probability. As data within the range
will be used to train and test our Improved Extreme Learning of (µ-3σ , µ+3σ ) account for 99.74% of the total data,
Machine (ELM) model. With this approach, we can offer the probability of exceeding this range is less than 0.3%.
crucial information regarding future air quality conditions to Therefore, the distribution characteristics of the vast majority
the government and the public. This data-driven insight can of the data align with the 3σ criterion, conforming to a normal
assist in the development and implementation of appropriate distribution.The standard deviation was calculated as follows:
environmental policies and measures. s
Pn
2
l Pi − P̄
Sg = (16)
2) DATA PREPROCESSING n−1
When dealing with missing data, various methods such as where Pi is the i th observation, Xg is the standard deviation
nearest neighbor filling, plural filling, median filling, and of the observation, P̄ is the average of n observations, and
mean filling have been widely discussed in the literature. n is the number of observations. For a suspicious data Xi ,
After conducting a thorough literature review and considering if Vi = Xi − X̄ > 3Sg , then X contains coarse deviation
the characteristics of the data at hand, the interpolation and is rejected. The data are then processed using MATLAB
method is selected as the preferred approach for filling the after using outliers, as shown in FIGURE 6, and they were
missing data. This method identifies the closest sample point obviously smoothed.
as the fill value by searching sample points around the
missing data. The interpolation method offers advantages B. EVALUATION METRICS
such as reduced calculation, high processing efficiency, The GA-KELM framework is compared with three base-
high flexibility and ease of implementation. Moreover, this line deep learning models: CMAQ, SVR, and DBN-BP.
approach has minimal impact on the correlation between the CMAQ (Community Multiscale Air Quality) is a computer
data and ensures the integrity of the experimental data. The model used to simulate and predict atmospheric pollutant
interpolation effect is shown in FIGURE 5. concentrations and air quality. SVR (Support Vector Regres-
However, as observed in the aforementioned figure, the sion) is a machine learning algorithm used to model and
presence of many outliers in the data is still evident. There- predict regression problems. It is an extension based on the
fore, outlier processing becomes necessary. To eliminate the support vector machine (SVM) algorithm designed to solve

VOLUME 11, 2023 67091

 C. Liu et al.: AQI Forecasting via GA-KELM

AQI, as shown in Table 2.The prediction performance of

GA-KELM is assessed by calculating the accuracy of the
predicted values in the actual categories.

C. CORRELATION WITHIN THE METEOROLOGICAL

PARAMETERS
This section examines the correlation between variables to
make more accurate predictions. When pollutant sources are
identified, pollutant concentrations are basically determined
by meteorological conditions without analyzing the changes
brought by the pollutant emissions to the AQI. Therefore,
a change in pollutant concentration inevitably causes a
change in AQI, and meteorological factors in turn determine
the pollutant concentration. Meteorological factors have a
significant role in AQI change. The main meteorological
factors that affect the AQI are climatic conditions, weather
conditions, temperature, wind, humidity, and precipitation.
Temperature is one of the most basic meteorological factors
FIGURE 6. After data processing.
among them. Higher temperatures increase the water vapor
content in the air. Increased wind speed accelerates air flow,
regression problems. DBN-BP (Deep Belief Network with making it easier for pollutants to be transported to higher
Backpropagation) is a neural network model that combines altitudes, which in turn causes increased pollution at ground
deep confidence network (DBN) and backpropagation (BP) level. Relative humidity is related to the moisture in the
algorithms. In the experiments, 70% of the data were used air. Turbulence in the atmosphere has a diffusion effect on
for training and 30% for testing. The predictive performance pollutants, causing a decrease in the AQI in the region.
of the proposed model was assessed using several evaluation Furthermore, temperature and humidity significantly affect
metrics, such as RMSE, mean square error (MSE), and the pollutant chemical reaction process. The wind direction
correlation coefficient (R2 ). affects the diffusion of pollutants, which in turn causes
RMSE quantifies the difference between predicted and changes in pollutant concentrations and the AQI.
actual data. A low value of RMSE underlines a good forecast. Therefore, for the impact of meteorological factors on AQI,
v the correlation between the five variables of temperature,
u N X − X̂ 2
uP  
t i=1 i i humidity, air pressure, wind speed, wind direction and the
RMSE = (17) monitored concentrations of SO2 , NO2 , CO, PM10 , PM2.5 ,
N
and O3 can be considered. Typically, to study the correla-

groups of variables x1 , x2 , x3 , · · · , xp ,


MSE is calculated as the squared difference between the tion between two
true value and the predicted value, which is then averaged. y1 , y2 , y3 , · · · yq , the p ∗ q correlation coefficients between
It is defined over the range [0,+∞), where it is 0 when the the two sets of variables can be calculated, using an idea
predicted value is exactly the same as the true value, and the similar to that of principal components. A certain linear
larger the error, the larger the value. combination of each of the two sets of variables is identified
n and the correlation between the linear combinations is
1X
2
MSE = ŷi − yi (18) discussed. In order to have maximum correlation, it is
n necessary to find the first pair of linear combinations in each
i=1
set of variables:
R2 is a correlation coefficient that defines how well a
u1 = a11 x1 + a21 x2 + . . . ap1 xp

predicted value matches the original one. The higher the value (20)
of R2 , the better the predictive model. R2 is given by: v1 = b11 y1 + b21 y2 + . . . bq1 yq

P 2 Similarly, the second pair of linear combinations with the

Xi − X̂i next largest correlation is found, so that they are respectively
R2 = 1 − P
2 (19) uncorrelated with the first linear combination pair within this
Xi − X̄
group:
Air quality is measured using the AQI obtained by calcu-
u2 = a12 x1 + a22 x2 + . . . ap2 xp

lating the concentration of each pollutant being monitored. (21)
v2 = b12 y1 + b22 y2 + . . . bq2 yq
The classification of pollutants in the AQI can help in
understanding the status of air pollutants. From the measured where u2 and v2 are correlated, and u1 is uncorrelated
concentration values, the AQI classification range for each with u2 . In this way, r steps are carried out until the
pollutant is strictly determined according to the Ambient correlation between the two groups of variables is extracted,

67092 VOLUME 11, 2023

C. Liu et al.: AQI Forecasting via GA-KELM

TABLE 1. Correlation of various pollutants with five meteorological conditions.

and the r group of variables can be obtained, with

r ≤ min(p, q). To study the correlation between two sets of
random variables, the full correlation coefficient is used. The
correlation model selects temperature, humidity, barometric
pressure, wind speed, and wind direction as meteorological
factor indicators, assuming that the correspondence between
NO2 and SO2 (fuel combustion) is not taken into account.
SO2 , NO2 , CO, PM10 , PM2.5 , and O3 are selected as the FIGURE 7. End-to-end architecture of GA-KELM.
source indicators in the model, and all factors are selected
based on daily average values. The influence of each weather
factor on each pollutant is divided into positive and negative
correlations through separate statistical correlations.
Table 1 shows that the atmospheric water-holding capacity
is enhanced with temperature rise in typical times, the atmo-
spheric movement slows, the accumulation rate of pollutants
exceeds the diffusion rate, and pollutants are not easily
diffused, leading to an increase in pollutant concentration and
in AQI. The increase in humidity reflects the enhanced water
holding capacity of the atmosphere, which also makes the
accumulation rate of pollution higher than the diffusion rate,
thus causing the pollutant concentration to increase and the
AQI to rise. However, increased humidity can also be caused
by rainfall that can decrease pollutant concentrations, and
consequently, AQI is decreased. Pressure affects air quality
mainly as a result of seasonal changes rather than day-to- FIGURE 8. Results of the change in fitness for GA optimization.
day variations. For instance, low-pressure weather dominates
in winter and the atmosphere moves slowly, which is output is the pollutant concentration to be predicted at this
unfavorable to pollutant dispersion. Wind speed is consistent moment (i.e., at time t), as shown in FIGURE 7.
with wind direction. The analysis of the relationship between In the GA, the number of iterations is set to 100,
meteorological factors and pollutants reveals that the total the population size is 10, the number of hidden layer
amount of pollutant emissions varies greatly under different neurons is 30, the number of output layer neurons is
meteorological conditions, and pollutant emissions decrease 1 representing the predicted AQI, the number of training
with an increasing wind speed under the same meteorological iterations is 1000, the learning rate is 0.05, and the training
conditions. Conceptually, an increase in wind speed causes a error target is set to 0.001. Additionally, the crossover
decrease in pollution concentration and a decrease in AQI at probability is varied in the range [0.6 1.0] in increments
the designated monitoring stations. However, the wind speed of 0.05, and the variance probability is varied in the range
increases in some direction whereas the AQI decreases at the [0 1] with an increment of 0.01. RMSE is chosen as the
monitoring stations. fitness parameter for the optimization problem. For each
parameter set, 20 independent replications are performed.
D. POLLUTANT CONCENTRATION FORECASTING RESULTS For each case, the RMSE values converge rapidly to a stable
ANALYSIS point.
In the experiment, a GA is used to optimize the parameters The parameters to be optimized, in this case the penalty
of KELM. The fitness calculated by the GA is a weighted factor and the kernel parameters, are first set at random.
combination of AQI and primary pollutant error. Data The data are divided into two parts: training and prediction
from five previous hours (actual collected values including of the remaining days. The training samples and the set
pollutants and weather) are used as input. For instance, parameters are used to train the model, and the test samples
assuming that the moment to be predicted is t, then data at are used to produce prediction results for AQI and major
t − 1, t − 2, t − 3, t − 4, and t − 5 are taken as input. The pollutants. A weighted combination of the maximum relative

VOLUME 11, 2023 67093

 C. Liu et al.: AQI Forecasting via GA-KELM

FIGURE 9. Annual air quality forecast comparison (Dataset A).

FIGURE 10. Seasonal air quality forecast trends (Dataset B).

TABLE 2. Category range of air pollutants (technical regulation on ambient air quality index, HJ633-2012).

error of AQI and the average relative error of major pollutants generation is set to 40, and the kernel function selected
is used as our fitness function. Then the GA is used to in KELM is the Gaussian RBF. The fitness curve of GA
optimize the fitness function, and the penalty factor and optimization is shown in Fig.8. It can be observed that the
kernel parameter are used as the optimization parameters. fitness keeps decreasing and finally reaches convergence.
According to previous theory, the standard GA will complete At this point, the optimal penalty factor and kernel function
convergence at an early stage. The maximum evolutionary parameters can be obtained.

67094 VOLUME 11, 2023

C. Liu et al.: AQI Forecasting via GA-KELM

FIGURE 11. Monthly air quality forecast accuracy (Dataset C).

TABLE 3. Comparison of RMSE of GA-KELM with other baseline to total predictions is estimated to measure the accuracy of
approaches.
the model in predicting the correct AQI category.
The performance of the model in predicting air pollutant
concentrations is measured based on the RMSE, MSE
and R2 values of all predicted pollutants. The results are
compared with baseline methods for predicting air pollutant
concentrations, such as CMAQ, SVR, and DBN-BP, with
meteorological parameters as auxiliary inputs. However, the
TABLE 4. Comparison of MSE of GA-KELM with other baseline
approaches. factors considered in model evaluation are different.
The proposed and baseline models were initially trained
using the training dataset, and from Table3-5, we infer that
GA-KELM air has smaller RMSE and MSE values and
higher R2 values compared to baseline methods for predicting
concentrations of various air pollutants.The reduced RMSE
TABLE 5. Comparison of R2 of GA-KELM with other baseline approaches. and MSE values elucidate the reliability of the GA-KELM
air predictions, while the reduced RMSE and improved
R2 values infer the specificity of its mean predictions. The
superior performance of GA-KELM illustrates its efficiency
in accurately capturing spatiotemporal relationships and their
impact on predicted values.

The obtained optimal parameters are then used to train V. CONCLUSION

the limit learning machine. The actual values are compared The economic development achieved by the country through
with the predictions of GA-KELM from July 2020 to July rapid urbanization is polluting the environment in an alarming
2021; the red line represents the predicted values, and the way and putting people’s lives in danger. Therefore, a correct
blue color represents the actual values. The results of three analysis and accurate prediction of air quality remains a
different datasets A, B, and C are shown in FIGURE 9, primary condition to achieve the objective of sustainable
FIGURE 10, and FIGURE 11, respectively. As can be seen development. This paper focuses on the problem of prediction
from the figures, the training effect is relatively good. model design, and investigates the problems related to the
In China, the National Air Quality Index sets breakpoints optimization of the model parameters. A GA-KELM model
for all air pollutants under 6 categories, namely good, is designed, implemented, and tested. It is experimentally
satisfactory, moderately polluted, poor, extremely poor, and proven to be more efficient than the classical shallow learning
severe, as shown in Table 2. During the validation process, and can effectively explore and learn the interdependence
the predicted values are grouped under each AQI category of multivariate air quality correlation time series such
based on the predicted value. The ratio of correct predictions as temperature, humidity, wind speed, SO2 , and PM10 .

VOLUME 11, 2023 67095

 C. Liu et al.: AQI Forecasting via GA-KELM

Therefore, the GA-KELM model developed in this study can [9] A. C. Cosma and R. Simha, ‘‘Machine learning method for real-time
be used to provide valuable support to vulnerable groups and non-invasive prediction of individual thermal preference in transient
conditions,’’ Building Environ., vol. 148, pp. 372–383, Jan. 2019.
trigger early warning of adverse air quality events. [10] M. Bhowmik, K. Deb, A. Debnath, and B. Saha, ‘‘Mixed phase
However, there are still areas for further investigation and Fe2 O3 /Mn3 O4 magnetic nanocomposite for enhanced adsorption of
improvement. In recent years, numerous advanced algorithms methyl orange dye: Neural network modeling and response surface
methodology optimization,’’ Appl. Organometallic Chem., vol. 32, no. 3,
and optimization methods based on genetic algorithms and p. e4186, Mar. 2018.
population intelligence have emerged. Therefore, future [11] V. Chaudhary, A. Deshbhratar, V. Kumar, and D. Paul, ‘‘Time series based
research should explore the underlying significance and value LSTM model to predict air pollutant’s concentration for prominent cities in
of combinatorial intelligence optimization algorithms such as India,’’ in Proc. Int. Workshop Utility-Driven Mining (UDM), Aug. 2018,
pp. 1–9.
the Limit Learning Machine. Additionally, we acknowledge [12] M. Chen, J. Yang, L. Hu, M. S. Hossain, and G. Muhammad, ‘‘Urban
the need to address the issue of manually setting the number healthcare big data system based on crowdsourced and cloud-based air
of hidden layer nodes in the optimal Limit Learning Machine. quality indicators,’’ IEEE Commun. Mag., vol. 56, no. 11, pp. 14–20,
Nov. 2018.
Although the Dynamic Extreme Learning Machine (DELM) [13] R. Chen, X. Wang, W. Zhang, X. Zhu, A. Li, and C. Yang, ‘‘A hybrid
algorithm offers adaptive determination of hidden layer CNN-LSTM model for typhoon formation forecasting,’’ GeoInformatica,
nodes without human intervention, further work should be vol. 23, no. 3, pp. 375–396, Jul. 2019.
dedicated to this aspect. Moreover, to enhance the accuracy [14] S. Du, T. Li, Y. Yang, and S. Horng, ‘‘Deep air quality forecasting using
hybrid deep learning framework,’’ IEEE Trans. Knowl. Data Eng., vol. 33,
and validity of air quality measurement and assessment, no. 6, pp. 2412–2424, Jun. 2021.
it is crucial to integrate pollutant emission factors and [15] R. Feng, H.-J. Zheng, H. Gao, A.-R. Zhang, C. Huang, J.-X. Zhang, K. Luo,
meteorological factors into the evaluation system. This and J.-R. Fan, ‘‘Recurrent neural network and random forest for analysis
and accurate forecast of atmospheric pollutants: A case study in Hangzhou,
integration will enable a more precise and comprehensive China,’’ J. Cleaner Prod., vol. 231, pp. 1005–1015, Sep. 2019.
evaluation of air quality. [16] B. S. Freeman, G. Taylor, B. Gharabaghi, and J. Thé, ‘‘Forecasting air
In conclusion, our study highlights the significance of quality time series using deep learning,’’ J. Air Waste Manage. Assoc.,
vol. 68, no. 8, pp. 866–886, Aug. 2018.
the GA-KELM model in predicting air quality. We have
[17] S. Mahajan, H.-M. Liu, T.-C. Tsai, and L.-J. Chen, ‘‘Improving the
addressed the optimization challenges and demonstrated its accuracy and efficiency of PM2.5 forecast service using cluster-based
superiority over traditional methods. However, there is still hybrid neural network model,’’ IEEE Access, vol. 6, pp. 19193–19204,
room for improvement and further research. Future studies 2018.
[18] J. Jin, J. Gubbi, S. Marusic, and M. Palaniswami, ‘‘An information
should delve into advanced optimization algorithms based on framework for creating a smart city through Internet of Things,’’ IEEE
genetic algorithms and population intelligence, explore the Internet Things J., vol. 1, no. 2, pp. 112–121, Apr. 2014.
potential of the Limit Learning Machine, and strive for adap- [19] A. Grover, A. Kapoor, and E. Horvitz, ‘‘A deep hybrid model for
weather forecasting,’’ in Proc. 21st ACM SIGKDD Int. Conf. Knowl.
tive determination of hidden layer nodes. Furthermore, the DiscoveryData Mining, Aug. 2015, p. 379–386.
integration of pollutant emission factors and meteorological [20] A. Agarwal and M. Sahu, ‘‘Forecasting PM2.5 concentrations using
factors into the evaluation system will advance the accuracy statistical modeling for Bengaluru and Delhi regions,’’ Environ. Monit.
and reliability of air quality measurement and assessment. Assessment, vol. 195, p. 502, Mar. 2023.
[21] D. J. Lary, T. Lary, and B. Sattler, ‘‘Using machine learning to estimate
global PM2.5 for environmental health studies,’’ Environ. Health Insights,
REFERENCES vol. 9, no. 1, pp. 41–52, 2015.
[22] Y. Zheng, X. Yi, M. Li, R. Li, Z. Shan, E. Chang, and T. Li,
[1] X. Li, L. Jin, and H. Kan, ‘‘Air pollution: A global problem needs local ‘‘Forecasting fine-grained air quality based on big data,’’ in Proc. 21th
fixes,’’ Nature, vol. 570, no. 7762, pp. 437–439, Jun. 2019. ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, New York, NY,
[2] Y. Han, J. C. K. Lam, and V. O. K. Li, ‘‘A Bayesian LSTM model to USA, Aug. 2015, pp. 2267–2276.
evaluate the effects of air pollution control regulations in China,’’ in Proc. [23] Y. Zheng, L. Capra, O. Wolfson, and H. Yang, ‘‘Urban computing:
IEEE Big Data Workshop (Big Data), Dec. 2018, pp. 4465–4468. Concepts, methodologies, and applications,’’ ACM Trans. Intell. Syst.
[3] L. Bai, J. Wang, X. Ma, and H. Lu, ‘‘Air pollution forecasts: An overview,’’ Technol., vol. 5, no. 3, p. 38, Sep. 2014.
Int. J. Environ. Res. Public Health, vol. 15, no. 4, p. 780, 2018. [24] T. S. Rajput and N. Sharma, ‘‘Multivariate regression analysis of air quality
[4] Y. Ding and Y. Xue, ‘‘A deep learning approach to writer identification index for Hyderabad city: Forecasting model with hourly frequency,’’ Int.
using inertial sensor data of air-handwriting,’’ IEICE Trans. Inf. Syst., J. Appl. Res., vol. 3, no. 8, pp. 443–447, 2017.
vol. E102-D, no. 10, pp. 2059–2063, 2019. [25] Z. Kang and Z. Qu, ‘‘Application of BP neural network optimized by
[5] S.-Q. Dotse, M. I. Petra, L. Dagar, and L. C. De Silva, ‘‘Application of genetic simulated annealing algorithm to prediction of air quality index
computational intelligence techniques to forecast daily PM10 exceedances in Lanzhou,’’ in Proc. IEEE Comput. Intell. Appl. (ICCIA), Sep. 2017,
in Brunei Darussalam,’’ Atmos. Pollut. Res., vol. 9, no. 2, pp. 358–368, pp. 155–160, doi: 10.1109/CIAPP.2017.8167199.
Mar. 2018. [26] B. Liu, S. Yan, J. Li, G. Qu, Y. Li, J. Lang, and R. Gu, ‘‘A sequence-to-
[6] M. Jia, A. Komeily, Y. Wang, and R. S. Srinivasan, ‘‘Adopting sequence air quality predictor based on the n-step recurrent prediction,’’
Internet of Things for the development of smart buildings: A review of IEEE Access, vol. 7, pp. 43331–43345, 2019.
enabling technologies and applications,’’ Automat. Construct., vol. 101, [27] K. Elbaz, I. Hoteit, W. M. Shaban, and S.-L. Shen, ‘‘Spatiotemporal air
pp. 111–126, May 2019. quality forecasting and health risk assessment over smart city of NEOM,’’
[7] S. Abirami, P. Chitra, R. Madhumitha, and S. R. Kesavan, ‘‘Hybrid Chemosphere, vol. 313, Feb. 2022, Art. no. 137636.
spatio-temporal deep learning framework for particulate matter (PM2.5 ) [28] P. C. Campbell, Y. Tang, P. Lee, B. Baker, D. Tong, R. Saylor,
concentration forecasting,’’ in Proc. Int. Conf. Innov. Trends Inf. Technol. A. Stein, J. Huang, H.-C. Huang, E. Strobach, J. McQueen, L. Pan,
(ICITIIT), Feb. 2020, pp. 1–6. I. Stajner, J. Sims, J. Tirado-Delgado, Y. Jung, F. Yang, T. L. Spero, and
[8] Y. Cheng, S. Zhang, C. Huan, M. O. Oladokun, and Z. Lin, ‘‘Optimization R. C. Gilliam, ‘‘Development and evaluation of an advanced national air
on fresh outdoor air ratio of air conditioning system with stratum quality forecasting capability using the NOAA Global Forecast System
ventilation for both targeted indoor air quality and maximal energy version 16,’’ Geosci. Model Develop., vol. 15, no. 8, pp. 3281–3313,
saving,’’ Building Environ., vol. 147, pp. 11–22, Jan. 2019. Apr. 2022.

67096 VOLUME 11, 2023

C. Liu et al.: AQI Forecasting via GA-KELM

[29] X.-B. Jin, Z.-Y. Wang, W.-T. Gong, J.-L. Kong, Y.-T. Bai, T.-L. Su, CHUNHAO LIU (Graduate Student Member,
H.-J. Ma, and P. Chakrabarti, ‘‘Variational Bayesian network with infor- IEEE) was born in Linyi, Shandong, China,
mation interpretability filtering for air quality forecasting,’’ Mathematics, in 1997. He received the B.S. degree in automation
vol. 11, no. 4, p. 837, Feb. 2023. from Linyi University, China, in 2019, where
[30] J. Li, W. Shi, and D. Yang, ‘‘Color difference classification of dyed fabrics he is currently pursuing the Graduate degree in
via a kernel extreme learning machine based on an improved grasshopper electronics and information technology with the
optimization algorithm,’’ Color Res. Appl., vol. 46, no. 2, pp. 388–401, School of Automation and Electrical Engineering.
Apr. 2021.
His current research interests include machine
[31] Z. Zhou, Y. Hu, J. Ji, Y. Wang, Z. Zhu, D. Yang, and J. Chen, ‘‘Robot
learning and deep learning, data mining, and
manipulator visual servoing via Kalman filter-optimized extreme learning
control systems.
machine and fuzzy logic,’’ KSII Trans. Internet Inf. Syst., vol. 16, no. 8,
pp. 2529–2551, 2022.
GUANGYUAN PAN (Member, IEEE) was born
[32] Z. Zhou, W. Deng, Z. Zhu, Y. Wang, J. Du, and X. Liu, ‘‘Fabric defect
detection based on feature fusion of a convolutional neural network and in Linyi, Shandong, China, in 1987. He received
optimized extreme learning machine,’’ Textile Res. J., vol. 92, nos. 7–8, the B.S. degree in automation from the Qingdao
pp. 1161–1182, Apr. 2022. University of Science and Technology, China,
[33] Y. Xue, Q. Wu, and H. Sun, ‘‘Air quality prediction model based on in 2009, and the M.S. and Ph.D. degrees in
genetic algorithm and weighted extreme learning machine,’’ Proc. SPIE, control science and engineering from the Beijing
vol. 12294, Oct. 2022, Art. no. 1229454. University of Technology (Dr. Junfei Qiao’s team),
[34] X. Liang, G. Wang, M. R. Min, Y. Qi, and Z. Han, ‘‘A deep spatio-temporal Beijing, China, in 2012 and 2016, respectively.
fuzzy neural network for passenger demand prediction,’’ in Proc. SIAM Int. From 2013 to 2014, he was a joint-training Ph.D.
Conf. Data Mining (SDM), 2019, pp. 100–108. student with the University of Victoria (Dr. Nikitas
[35] Y. Liu, C. Lyu, A. Khadka, W. Zhang, and Z. Liu, ‘‘Spatio-temporal Dimopoulos’ team), BC, Canada. He was a Postdoctoral Researcher with
ensemble method for car-hailing demand prediction,’’ IEEE Trans. Intell. the Department of Civil and Environmental Engineering, University of
Transp. Syst., vol. 21, no. 12, pp. 5328–5333, Dec. 2020. Waterloo, ON, Canada, from 2016 to 2019. He is currently an Associate
[36] J. Ma, Y. Ding, J. C. P. Cheng, F. Jiang, and Z. Wan, ‘‘A temporal-spatial Professor with the Department of Automation and Electrical Engineering,
interpolation and extrapolation method based on geographic long short- Linyi University, China. His current research interests include machine
term memory neural network for PM2.5 ,’’ J. Cleaner Prod., vol. 237,
learning and deep learning, data mining and signal processing, transportation
Nov. 2019, Art. no. 117729.
safety, and intelligent systems.
[37] J. Ma, Y. Ding, V. J. L. Gan, C. Lin, and Z. Wan, ‘‘Spatiotemporal
prediction of PM2.5 concentrations at different time granularities using
DONGMING SONG was born in Yantai,
IDW-BLSTM,’’ IEEE Access, vol. 7, pp. 107897–107907, 2019.
Shandong, China, in 1976. He received the
[38] S. Maji, S. Ghosh, and S. Ahmed, ‘‘Association of air quality with
respiratory and cardiovascular morbidity rate in Delhi, India,’’ Int. bachelor’s degree in mechanical design and
J. Environ. Health Res., vol. 28, no. 5, pp. 471–490, Sep. 2018. manufacturing from the China University of
[39] M. Wang, H. Chen, B. Yang, X. Zhao, L. Hu, Z. Cai, H. Huang, and Petroleum (East China), in 1999, and the master’s
C. Tong, ‘‘Toward an optimal kernel extreme learning machine using a degree in mechanical and electronic engineering
chaotic moth-flame optimization strategy with applications in medical and the Ph.D. degree in computer application
diagnoses,’’ Neurocomputing, vol. 267, pp. 69–84, Dec. 2017. technology from the Nanjing University of Science
[40] M. A. E. Aziz, A. A. Ewees, and A. E. Hassanien, ‘‘Whale optimization and Technology, in 2004 and 2009, respectively.
algorithm and moth-flame optimization for multilevel thresholding image He is currently an Associate Professor with the
segmentation,’’ Expert Syst. Appl., vol. 83, pp. 242–256, Oct. 2017. Department of Automation and Electrical Engineering, Linyi University,
[41] Y. Yang, Y. Wang, and X. Yuan, ‘‘Parallel chaos search based incremental China. His current research interests include energy internet, energy big data,
extreme learning machine,’’ Neural Process. Lett., vol. 37, no. 3, electrical, and control systems.
pp. 277–301, Jun. 2013.
[42] F. Jiang, J. He, and Z. Zeng, ‘‘Pigeon-inspired optimization and HAO WEI was born in Linyi, Shandong, China,
extreme learning machine via wavelet packet analysis for predicting bulk in 1989. He received the bachelor’s degree in
commodity futures prices,’’ Sci. China Inf. Sci., vol. 62, no. 7, pp. 1–19, computer science from Imperial College London,
Jul. 2019. U.K., in 2011, and the master’s degree in software
[43] Q. Li, H. Chen, H. Huang, X. Zhao, Z. Cai, C. Tong, W. Liu, and X. Tian, engineering and the Ph.D. degree in computer
‘‘An enhanced grey wolf optimization based feature selection wrapped
science from the University of York, U.K., in
kernel extreme learning machine for medical diagnosis,’’ Comput. Math.
2013 and 2018, respectively. He is currently a
Methods Med., vol. 2017, Jan. 2017, Art. no. 9512741.
[44] G.-B. Huang, D. H. Wang, and Y. Lan, ‘‘Extreme learning machines: Lecturer with the Department of Automation and
A survey,’’ Int. J. Mach. Learn. Cybern., vol. 2, no. 2, pp. 107–122, Electrical Engineering, Linyi University, China.
Jun. 2011. His current research interests include swarm
[45] G. Huang, H. Zhou, X. Ding, and R. Zhang, ‘‘Extreme learning machine robotics, genetic algorithm, and system development related to artificial
for regression and multiclass classification,’’ IEEE Trans. Syst., Man, intelligence and big data.
Cybern. B, Cybern., vol. 42, no. 2, pp. 513–529, Apr. 2012.

VOLUME 11, 2023 67097