Atmospheric Pollutant Concentration Prediction Based on KPCA BP

International Journal of Trend in Scientific Research and Development (IJTSRD)
Volume 6 Issue 5, July-August 2022 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470
@ IJTSRD | Unique Paper ID – IJTSRD51746 | Volume – 6 | Issue – 5 | July-August 2022 Page 2008
Atmospheric Pollutant Concentration
Prediction Based on KPCA-BP
Xin Lin, Bo Wang, Wenjing Ai
School of Information, Beijing Wuzi University, Beijing, China
ABSTRACT
PM2.5 prediction research has important significance for improving
human health and atmospheric environmental quality, etc. This paper
uses a model combining nuclear principal component analysis
method and neural network to study the prediction problem of
meteorological pollutant concentration, and compares the
experimental results with the prediction results of the original neural
network and the principal component analysis neural network. Based
on the O3, CO, PM10, SO2, NO2 concentrations and parallel
meteorological conditions data of Beijing from 2016 to 2020, the
PM2.5 concentration was predicted. First, reduce the latitude of the
data, and then use the KPCA-BP neural network algorithm for
training. The results show that the average absolute error, root mean
square error and expected variance score of the combined model are
relatively good, the generalization ability is strong, and the extreme
value prediction is the best, which is better than that of the single
model.
KEYWORDS: KPCA; prediction of atmospheric pollutants; BP neural
network; PCA
How to cite this paper: Xin Lin | Bo
Wang | Wenjing Ai "Atmospheric
Pollutant Concentration Prediction
Based on KPCA-BP" Published in
International Journal
of Trend in
Scientific Research
and Development
(ijtsrd), ISSN: 2456-
6470, Volume-6 |
Issue-5, August
2022, pp.2008-
2016, URL:
www.ijtsrd.com/papers/ijtsrd51746.pdf
Copyright © 2022 by author (s) and
International Journal of Trend in
Scientific Research and Development
Journal. This is an
Open Access article
distributed under the
terms of the Creative Commons
Attribution License (CC BY 4.0)
(http://creativecommons.org/licenses/by/4.0)
1. INTRODUCTION
Beijing air pollution index has remained high in
recent days, with PM2.5 As the primary pollutant, due
to its small particle size, it comes with a large number
of toxic and harmful substances, and it is suspended
in the air for a long time, which is a greater harm to
human health. Looking at the governance of PM2.5 in
the past five years, the concentration of PM2.5 is due
to the proposal of relevant policies and the
enhancement of people's awareness of environmental
protection It has been significantly reduced, but the
control of PM2.5 and other related pollutants should
continue to be strengthened[1]
. According to the results of
the national air quality forecast consultation published
by the Ministry of Ecology and Environment of the
People's Republic of China from November 2020 to
October 2021[2]
. In Beijing-Tianjin-Hebei and some
surrounding areas, the air quality between April and
October is mainly good to mild pollution, and its
primary pollutant is PM2.5. Air quality is relatively
poor in autumn and winter, so being able to predict
PM2.5 concentration more accurately is an important
issue.
Common prediction methods for PM2.5 concentration
include: artificial neural networks [3]
, wavelet-neural
networks[4]
, multiple linear regression models[5]
,
LSTM algorithm [6]
et al. In these prediction methods,
artificial neural network algorithms are most
commonly used for complex nonlinear relationships,
such as water resource prediction, traffic route
prediction, etc. Due to the characteristics of the
artificial neural network algorithm itself, it is a model
of distributed parallel processing algorithm, and it is
difficult to consider the influence of each factor in the
multi-factor problem on the predicted value, so this
paper adopts a dimensionality reduction method to
analyze the multi-factor problem.
In the method of reducing the dimension, the concept
of kernel function (kernel) is introduced in this paper
because the dimensionality reduction effect of
processing linear data in principal component analysis
needs to be improved. To date, neural network
algorithms based on nuclear principal component
analysis have been used less in related studies on
atmospheric pollutant concentration prediction. There
are many influencing factors for the concentration of
IJTSRD51746

International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
n
X
X
MAE
n
i
i
el
i
obs
∑
=
−
= 1
,
mod
,
( )
n
X
X
RMSE
n
i
i
el
i
obs
∑
=
−
= 1
2
,
mod
,
( ) { }
{ }
i
obs
i
el
i
obs
i
el
i
obs
X
Var
X
X
Var
X
X
iance
lained
,
,
mod
,
,
mod
, 1
,
var
exp
−
−
=
−
pollutants in the atmosphere, and a hot issue in the
current research is how to effectively extract the
relevant information between the main factors, which
is of great significance for the subsequent
improvement of the accuracy of the PM2.5 prediction
model.
In this paper, when analyzing the influence of
multiple factors on the concentration of PM2.5
pollutants in the atmosphere, based on the prediction
method of traditional neural network, a variety of
dimensionality reduction methods are used to predict
it. The experimental results of various prediction
methods were analyzed and compared, and a
relatively good method was proposed to improve the
accuracy of predicting PM2.5 concentration values. In
this paper, based on the prediction method based on
the traditional neural network, the TensorFlow-BP
algorithm is used to PM the influencing factors of 5
factors (pollutant concentration) and 12 factors
(pollutant concentration and meteorological
conditions), respectively Concentration prediction;
Principal component analysis and nuclear principal
component analysis are mainly used to extract
components and input them into tensorFlow-BP
neural network model to predict PM2.5 concentration.
Finally, the prediction results of each model are
analyzed and compared by using relevant indicators
such as MAE and RMSE.
2. Data collection and evaluation indicators
2.1. data source
According to the big data information released by the
National Meteorological Science Data Center,
download the monitoring data of the Ecological
Environment Monitoring Center Station in Beijing
from 2016 to 2020, and the main pollution factors in
the atmosphere are O3 and CO, PM10, PM2.5, SO2,
NO2, etc., and parallel to the daily wind speed, air
temperature, surface temperature, sunshine hours,
humidity, barometric pressure and cumulative
precipitation. This article is selected from January 1,
2016 to December 2019 The PM2.5 concentration
value for the period on the 31st is used as a training
sample set, from January 1, 2020 PM2.5 concentration
values for this period on December 31, 2012 were
used as a test sample set.
2.2. Data processing
In the 2016-2019 data collected, some small amounts
of missing data were populated with the mean of the
data from the adjacent observatory. The data outliers
were then processed using the boxplot and
3σ principles, and the z-score data for each feature in
the training set and the test set were normalized,
taking into account that the physical meaning and
dimensions of the air pollutant concentration data and
the meteorological condition data were not the same.
2.3. Evaluation indicators
This paper uses the mean absolute error (MAE), root
mean square error (RMSE), and explained variance
score as evaluation indicators to compare the degree
of difference between the predicted values and the
measured values of each model Do not show this in
Equations (1)-(3).
（1）
where: Xobs, i represents the forecast, Xmodel, i
represents the measured data, and n represents the
number of predictions. The smaller the value of the
MAE, the better the fit between the predicted data and
the real data, so the smaller the indicator, the better.
（2）
where: Xobs, i represents the ith forecast, Xmodel, i
represents the measured data, and n represents the
number of predictions. The smaller the value of
RMSE, the smaller the error between the model's
prediction data and the real data, so the smaller the
indicator, the better.
（3）
where: Xobs, i represents the ith predicted value, Xmodel, i represents the true value, n represents the number of
predictions, and it's The range of values is [0,1]. When the Application variance score is closer to 1, it shows that
the independent variable can explain the variance change of the dependent variable. The better the support vector
regression model is built, so the closer the value of the Deployed variance score is to 1, the better.
3. Experimental model
3.1. BP neural network model
BP Neural Network (BP Neural Network) belongs to the nonlinear dynamic information processing system of
backpropagation algorithm, which is one of the most widely used models in meteorological forecasting
applications [7]
, the algorithm does not need to clarify the functional relationship between input and output, and
can make predictions about new data by adjusting parameters inside the network[8]
。

)
( T
b
X
W
g
Y +
=
This paper uses the BP neural network as a predictive model, and its structure mainly includes three layers,
which are the input, output and implicit layers of the structure. First, the relevant data is input through the input
layer, and then the data is passed to the implicit layer, and after the data is activated and enlarged, it is passed to
the output layer and output by the output layer, the model for:
（4）
During operation, when the actual error is larger than the expected error, the backpropagation of the error will
begin, and the corresponding values will be adjusted, that is, the weighting value and the threshold value The
network is continued to be trained repeatedly so that the model parameters move in the direction of reduced loss
values until the desired error is met, the mapping between the input and output is determined. ABP neural
network implemented by the Keras library under the tensorflow framework in deep learning. Figure 3 depicts the
computational flow of the relevant network data, and uses this as the basic computing node of the framework,
responsible for maintaining and updating the node state.
Figure 1 TensorFlow calculation graph
3.2. PCA-BP neural network model
Principal Component Analysis is a multivariate statistical method that transforms multiple variables that
originally had a certain correlation into a few unrelated principal components through dimensionality reduction
techniques[9].
The relevant process can be roughly divided into the following steps:
1. Standardization processing, the purpose of which is to eliminate variable dimensional relationships;
2. Establish a Pearson coefficient matrix;
3. Calculate the eigenvalues and eigenvectors corresponding to the Pearson coefficient matrix and sort them by
size;
4. Calculate the matrix of cumulative contribution rate and principal component score coefficient.
First of all, in order to reduce the multiple correlations between various factors, the method of principal
component analysis was used to select new indicators to predict the concentration of PM2.5 more accurately;
Then, the reduced complexity dataset is combined with the BP neural network to improve the running speed of
the neural network algorithm, solve the nonlinear problem between multiple data, reduce the redundancy of the
input data, and improve the accuracy of the prediction result.
3.3. KPCA-BP neural network model
The Nuclear Principal Component Analysis (KPCA) method is a nonlinear extension of the Principal
Component Analysis (PCA) method, which first introduces a nonlinear mapping function Φ whose purpose is to
better process the relevant data, mapping N
R sample vectors in the original space k
x to high-dimensional space
F, that is, ( )
i
k
N
x
x
F
R Φ
→
→ , the relevant data can be converted from linearly indivisible to linearly separable,
and the corresponding principal component analysis can be performed in high-dimensional space F.
This article makes the following assumptions: the centralized set of samples is denoted X, { }
N
x
x
x ,
,
, 2
1 L which
is N
R the set of samples in space, where the total number of samples is N, and the dimension of each sample is d,
and then passes through in high-dimensional space The mapping in F can be obtained ( )
X
Φ , where

)
2
exp(
)
,
( 2
2
σ
i
i
x
x
x
x
K
−
−
=
)
x
x
(
))
x
(
)
x
(
(
))
x
(
( i
i
new
k
，
K
k
i
k
i ∑
∑ =
Φ
⋅
Φ
=
Φ α
α
α
, ( ) 0
1
=
Φ
∑
=
N
i
i
x . The corresponding eigenvectors in the high-dimensional space F are denoted d
i
vi ,
,
2
,
1
, L
= , and
the corresponding eigenvalues are denoted d
i
i ,
,
2
,
1 L
=
，
λ , where the eigenvectors can be represented linearly
by the collection of samples in the space, that is, there is a set of parameters ( )
N
α
α
α
α ，
，
， L
2
1
= satisfied:
( ) ( )
X
x
v i
N
i
i
i Φ
=
Φ
= ∑
=
α
α
1
（5）
The use of PCA in high-dimensional space F, availabl:
( ) ( ) i
i
i v
v
X
X λ
=
Φ
Φ
T
（6）
After the operation, the equation 9 is obtained.
( ) ( ) ( ) ( ) ( ) ( )α
λ
α X
X
X
X
X
X i Φ
Φ
=
Φ
Φ
Φ
Φ
T
T
T
（7）
In Equation (9), both sides of the equation contain it, ( ) ( )
X
X
T
Φ
Φ and it is replaced with a kernel matrix K. In
general, kernel functions satisfy Mercer's theorem, and this article uses Gaussian kernel functions.
（8）
where, σ is the argument to the kernel function.
At this point, any test sample vector mapped to a high-dimensional space new
x has:
（9）
After that, the PCA method of extracting principal components is used to calculate the projection of each data
point on the corresponding characteristic vector to obtain the nuclear principal components. The KPCA method
is then used, which is divided into four steps [10].
1. Extract the factors that affect the load, and use the generated variable matrix as the initial input matrix;
2. Select the corresponding kernel function to generate a kernel matrix by transforming and mapping;
3. The eigenvalues and eigenvectors corresponding to the computed kernel matrix;
4. Calculates the numerical value of the cumulative contribution rate to determine the number of input variables
of the neural network.
4. Experimental results and analysis
4.1. PM2.5 based on BP neural network model predictions
Establish a BP neural network model under the TensorFlow framework according to Section 2.1, and enter the
normalized data into the BP neural network for 5 contaminant factors and one for each 12 meteorological and
pollutant factors[11]
are trained. In the test set, January 1, 2020 to December 31, 2020 PM2.5 concentration is
predicted. After debugging, use ReLU to activate the function.
By O3, CO, PM10, SO2, NO2 predict the result of PM2.5 concentration in the data of the influencing factors of
various pollutants is recorded as (PM2.5-5). By O3, CO, PM10, SO2, NO2 These five pollutants, as well as the
average wind speed, average temperature, average surface temperature, sunshine hours, average relative
humidity, average station pressure, 20-20 hours cumulative precipitation and other meteorological influencing
factors data on PM 2.5 concentration prediction result model is denoted as (PM2.5-12)[12]
.
The mean squared error (MSE) of the two models is smaller than that of the PM2.5-12 model, indicating PM2.5-12
The model is trained with higher accuracy.

Figure 2 PM2.5-5 Loss Function Diagram Figure 3 PM2.5-12 Loss Function Diagram
Statistics were performed on three evaluation indicators of PM2.5-5 model and PM2.5-12 model. As can be seen
from Table 1, the prediction accuracy of the BP neural network has been greatly improved after increasing the
relevant meteorological factors.
Table 1 Comparative analysis of neural network prediction accuracy
MAE RMSE Explained variance score
PM2.5-5 models 0. 3963 0. 3595 0. 5636
PM2. 5-12 models 0. 3188 0. 2025 0. 7414
The prediction of the fitting curve results of the PM2.5-5 model and the PM2.5-12 model are shown in Figure 4
and Figure 5, respectively, and the fitting effect is good, which is confirmed the generalization and effectiveness
of the model proposed in this article are discussed. Through comparative analysis, the PM2.5-12 model predicts a
relatively better prediction effect.
Figure 4 PM2.5-5 Model Prediction Results
Figure 5 PM2.5-12 Model Prediction Results
4.2. PM2 based on PCA-BP neural network model 5 predictions
4.2.1. The eigenvalue is determined with the principal component
In general, principal component analysis methods require that the various factors have a certain correlation with
each other. The Pearson correlation coefficient |r| > 0.35 is usually required. Programmed using SPSS software,
the correlation analysis is performed after standardizing the relevant 12 factors considered, the results of which
are shown in Table 2, and two factors are excluded according to the above rules: sunshine hours, Cumulative
rainfall and SO2 concentration.

Table 2 Pearson correlation coefficients
Average
wind
speed
average
tempera
ture
Surface
tempera
ture
Hours
of
sunshi
ne
Averag
e
humidit
y
Average
air
pressure
Cumula
tive
precipit
ation
PM25 PM10 SO2 CO NO2
Average
wind speed
1 -0.064 -0.039 0.211 -0.444 0.029 -0.021 -0.26 -0.09 -0.005 -0.404 -0.37
Average
temperatures
-0.064 1 0.984 0.026 0.414 -0.88 0.121 0.153 0.141 -0.374 -0.169 0.067
Surface
temperatures
-0.039 0.984 1 0.098 0.352 -0.864 0.106 0.144 0.159 -0.359 -0.179 0.03
Hours of
sunshine
0.211 0.026 0.098 1 -0.573 0.03 -0.071 -0.292 0.009 0.019 -0.038 -0.357
Average
humidity
-0.444 0.414 0.352 -0.573 1 -0.4 0.087 0.477 0.05 -0.23 0.072 0.541
Average air
pressure
0.029 -0.88 -0.864 0.03 -0.4 1 -0.097 -0.178 -0.17 0.316 0.189 -0.087
Cumulative
precipitation
-0.021 0.121 0.106 -0.071 0.087 -0.097 1 0.115 0.077 -0.026 0.074 0.103
PM25 -0.26 0.153 0.144 -0.292 0.477 -0.178 0.115 1 0.718 0.303 0.494 0.812
PM10 -0.094 0.141 0.159 0.009 0.05 -0.171 0.077 0.718 1 0.432 0.583 0.551
SO2 -0.005 -0.374 -0.359 0.019 -0.23 0.316 -0.026 0.303 0.432 1 0.517 0.394
CO -0.404 -0.169 -0.179 -0.038 0.072 0.189 0.074 0.494 0.583 0.517 1 0.574
NO2 -0.37 0.067 0.03 -0.357 0.541 -0.087 0.103 0.812 0.551 0.394 0.574 1
Next, principal component analysis is performed on the remaining 10 factors. The tangency quantity of KMO
sampling amount is 0.721, and the Sig value is less than 0.05, so the obtained results are reference and scientific,
see Table 3 below[13]。 The rotation process is performed using the Kaiser normalized maximum variance
method, so that the factor load values deviate from 0 and 1, and some of the indicator content that has no
obvious correlation is removed. The amount of data in this article is large, so it is set up multiple iterations. After
300 iterations, the results will converge. The characteristic values and contribution rates of each principal
component are shown in
components with a cumulative contribution rate of > 85% are extracted and determined as the final indicators,
and the cumulative contribution rates of the four items are determined 88.109%, basically can reflect the cause
sub-information, recorded as F1 ~ F4, for the original indicators of the load status see
ingredients
Initial eigenvalue Extract the sum of squares of the loads
total
Percentage
of variance
Cumulative
/%
total
Percentage
of variance
Cumulative /%
1 3.559 35.592 35.592 3.559 35.592 35.592
2 3.227 32.274 67.867 3.227 32.274 67.867
3 1.299 12.986 80.852 1.299 12.986 80.852
4 0.726 7.257 88.109 0.726 7.257 88.109
5 0.451 4.515 92.624
6 0.289 2.887 95.511
7 0.168 1.682 97.192
8 0.155 1.547 98.739
9 0.113 1.134 99.873
10 0.013 0.127 100.000
ingredients
total
Percentage
of variance
Cumulative
/%
total
Percentage
of variance
Cumulative /%
1 3.559 35.592 35.592 3.559 35.592 35.592
2 3.227 32.274 67.867 3.227 32.274 67.867
3 1.299 12.986 80.852 1.299 12.986 80.852

Table 。
Table1 KMO and Bartlett tests
Number of KMO sample tangents 0.721
Bartlett spherical degree test
Approximate chi-square 12732.239
degree of freedom 45
Significance 0.000
Table 4 Principal component characteristic values and contribution rates
Table 5 Principal component load matrix
Index 1 2 3 4
PM2. 5 0.809 0.374 0.060 0.309
NO2 0.768 0.465 -0.134 0.259
Average relative humidity (1%) 0.669 -0.200 -0.567 0.318
PM10 0.655 0.406 0.493 -0.062
The average temperature (0.1°C). 0.593 -0.752 0.172 -0.138
Average surface temperature (0.1°C). 0.568 -0.748 0.225 -0.162
SO2 0.122 0.738 0.367 -0.081
The average air pressure of this station (0.1hPa). -0.581 0.705 -0.205 0.060
CO 0.468 0.684 0.027 -0.386
Average wind speed (0.1m/s). -0.438 -0.180 0.675 0.504
4.2.2. Neural network construction and prediction results
After factor loading calculation, 4 principal components F1 to F4 are used as inputs. The model operation effect:
the average absolute error is 0.2476, the root mean square error is 0.1150, and the interpretation expected score
is 88.15%. Using the PCA-BP neural network model, the predicted values of PM2.5 concentration in 2020 are
obtained and plotted the actual concentration of PM2.5 in 2020 and the model calculate a fitted curve for the
predicted concentration, as shown in Figure 10. It can be seen from the fitting curve of Figure 10 that the
experimental method of PM2.5-PCA is used to predict, which is basically consistent with the actual PM2.5
concentration value.
4 0.726 7.257 88.109 0.726 7.257 88.109
5 0.451 4.515 92.624
6 0.289 2.887 95.511
7 0.168 1.682 97.192
8 0.155 1.547 98.739
9 0.113 1.134 99.873
10 0.013 0.127 100.000
ingredients
total
Percentage
of variance
Cumulative
/%
total
Percentage
of variance
Cumulative /%
1 3.559 35.592 35.592 3.559 35.592 35.592
2 3.227 32.274 67.867 3.227 32.274 67.867
3 1.299 12.986 80.852 1.299 12.986 80.852
4 0.726 7.257 88.109 0.726 7.257 88.109
5 0.451 4.515 92.624
6 0.289 2.887 95.511
7 0.168 1.682 97.192
8 0.155 1.547 98.739
9 0.113 1.134 99.873
10 0.013 0.127 100.000

Figure 6 Fitted curve of PM2.5 concentration in 2020
Comparing the prediction effects of PM2.5-5, PM2.5-12, and PM 2.5-PCA, it can be seen that PM The 2.5-PCA
model has the lowest RMSE, the highest interpretation expectation score, and the best fit curve effect. It can be
obtained that after principal component analysis, multiple variables are reduced to 4 unrelated components, and
neural network prediction is performed There will be a significant improvement in the model's predictive
performance.
4.3. PM2 based on KPCA-BP neural network model 5 predictions
The experimental environment is Python 3.8. The original data is first preprocessed, then the kernel function is
introduced, the dimensionality reduction processing is carried out using the PCA method, and after debugging,
the obtained data is used as training samples and test samples. In the KPCA processed data, the data from 2016
to 2019 are used as training data[14]
, 2020 as test data. After many trainings, the results are output. Table 6 shows
the parameters corresponding to the model.
Table 6 KPCA model parameter values
Training parameters Ways and values
Forecast duration 2021. 1. 1-2020. 12. 31
Output layer activation function ReLU
The number of neurons in the output layer 128
The number of neurons in the hidden layer 64
Loss function Mean variance (MSE).
Optimize iterative algorithms Adam
Epoch 1000
batch_size 128
To compare the prediction accuracy of PCA-BP neural network and KPCA-BP neural network, see Table.
Table 7 Model Fitting Effects
MAE RMSE Explained variance score
PM2.5-PCA Neural Network 0. 2476 0. 1150 0. 8815
PM2.5-KPCA Neural network 0.2358 0. 0978 0. 8921
To make the results clearer, a small amount of data is randomly selected. The experiment takes the forecast data
for the 22nd of each month in 2020, and the prediction results of the above two models are shown in Figure 12.
The dot data marker points represent the true values, the triangular data marker points represent the predicted
values of the KPCA-BP neural network, and the cross data marker points represent the predictions of the PCA-
BP neural network Value. As can be seen from the figure, the distance from the triangular data marker point to
the dot data marker point is closer, which illustrates the prediction of the KPCA-BP neural network model The
effect is better than other predictive models.

Figure 7 2020 PM2.5 concentration fitting curve
Figure 8 Comparison of KPCA and PCA model predictions
5. Conclusion
In this paper, three neural network models were
studied by predicting PM2.5 concentrations in 2020
Accuracy in predicting the annual concentration of
meteorological pollutants. A comprehensive
evaluation of the comparative analysis of the above
models:
1. The K PCA-BP neural network model predicts
better results than other models. Compared with
the BP neural network and the PCA-BP neural
network prediction method, the three evaluation
indicators are lower and the stability is better.
2. Based on the prediction results, both combination
models have better prediction results, indicating
that extracting components can effectively learn
the impact factor.
3. Based on the predictions, nuclear principal
component analysis is superior to traditional
principal component analysis in predicting
pollutant concentrations. When the factor
increases, the latitude of the data can be lowered
more reasonably to improve the accuracy of the
prediction results.
With the development of science and technology, the
instantaneous acquisition processing power of big
data will become more and more mature. The
prediction method presented in this paper will have a
promising use and is suitable for prediction of the
concentration of the other 5 pollutants. Further
analysis based on the forecast results will provide
support for the decision-making of the prevention and
control department.
Bibliography
[1] [1] China Environment News. Transcript of the
regular press conference of the Ministry of
Ecology and Environment in February
[EB/OL](2020-02-26) [2021-11-01].
https://www.cenes.com.cn/news/202102/t20210
226_970791.html?from=singlemessage&isappi
nstalled=0
[2] Ministry of Ecology and Environment of the
People's Republic of China. Air Quality
Forecast. 京政办發 [A/OL](2021-08-31)[2021-
11-01].
https://www.mee.gov.cn/hjzl/dqhj/qgkqzlzk/ind
ex_1. shtml
[3] Nitin Tiwari, Neelima Satyam. Coupling effect
of pond ash and polypropylene fiber on strength
and durability of expansive soil subgrades: an
integrated experimental and machine learning
approach[J]. Journal of Rock Mechanics and
Geotechnical Engineering, 2021, 13(05): 1101-
1112.
[4] Xu Yixin, Ren J, Feng Lei, Liang Yinglu, Liu
Yiming. Optimization of PM_(2. 5)
concentration prediction model based on
wavelet analysis[J]. Environmental Monitoring

Management and Technology, 2021, 33(02):
24-28+34.
[5] Ilaboya I R. Performance of Multiple Linear
Regression (MLR) and Artificial Neural
Network (ANN) as Predictive Tool for Rainfall
Modelling. 2019.
[6] Zhao YM. LSTM algorithm based on spatio-
temporal correlation and application to PM_(2.
5) concentration prediction [J]. Computer
Applications and Software, 2021, 38(06): 249-
255+323.
[7] ZJ, Hu HP, Bai YP, Li Q. Research on urban
precipitation prediction based on BP, PCA-BP
and PLS algorithms[J]. Journal of North
Central University (Natural Science Edition),
2016, 37(02): 181-186.
[8] Jiang Shi. Application of PSO-BP neural
network in the safety inventory prediction of a
coal machine enterprise [J]. Coal Technology,
2017, 36(10): 305-307.
[9] Xie, Zhonghua. Matlab statistical analysis and
applications: 40 case studies [M]. Beijing:
Beijing University of Aeronautics and
Astronautics Press, 2010.
[10] Ren Jaguar, Gong K, Ma FJ, Gu QB, Wu
Qianqian. Prediction of heavy metals and PAHs
in contaminated site soil based on BP neural
network [J]. Environmental Science Research,
2021, 34(09): 2237-2247.
[11] Liu, Lanfang, Tan, Binglin, Zhang, Ke, Wu,
Jinyu. Risk assessment of urban haze disasters
in Hunan Province based on principal
component analysis [J]. Disaster Science, 2021,
36(01): 76-81.
[12] Guo S. K., Jane T., Dong Y. L. Radar one-
dimensional distance image recognition based
on PSO-KPCA-LVQ neural network [J].
Electro-Optics and Control, 2019, 26(06): 22-
26.
[13] Zhou Q. Y., Zhang S. R., Lai X. T., Wang J. B.
Research on fault diagnosis algorithm based on
PCA/KPCA initial dictionary optimization[J].
Journal of Sensing Technology, 2020, 33(11):
1599-1607.
[14] Li Xinning, Wu Hu, Yang Xianhai. Multi-view
recognition of fruit packing boxes based on
features clustering angle[J]. High Technology
Letters, 2021, 27( 02): 200-209.

Atmospheric Pollutant Concentration Prediction Based on KPCA BP

More Related Content

Similar to Atmospheric Pollutant Concentration Prediction Based on KPCA BP (20)

More from ijtsrd (20)

Recently uploaded (20)

Atmospheric Pollutant Concentration Prediction Based on KPCA BP