Nitrogen dioxide is one of the pollutants that has the most bad effects on human health. Machine learning technologies can perform predictive analyzes of the concentration of this pollutant with higher accuracy. One of the most advanced machine learning algorithms is the adaptive fuzzy-neural inference system, which is used in this article to predict nitrogen dioxide concentration. To increase the accuracy in this prediction, the crow learning algorithm has been used in the training of the adaptive neural fuzzy inference system and the results have been compared with the radial basic neural network. The results indicate the better performance of the proposed method.

Keywords: pollution prediction, nitrogen dioxide, fuzzy-adaptive neural inference system, crow learning algorithm.



1- Introduction

Today, air pollution due to continuous urbanization has become a global issue in both social and environmental fields, researches have been conducted in this field, Lim et al. in the capital region of Korea through regression modeling. The results indicate a relatively high concentration of NO2 in winter in the present and future forecasts, which is caused by the high use of fossil fuels in steam boilers and showed climate changes [1]. In 2021, Shams et al. evaluated the accuracy of multi-linear regression and multi-layer perceptron neural networks in predicting the concentration of NO2 in the air of metropolises. The results show that the multi-layer perceptron neural network had a more accurate prediction than the multi-linear regression [2].









2- An overview of algorithms

2-1- Crow's learning algorithm

In this algorithm, crows are trained based on two more optimal solutions which are parents. Another learning is the learning of each crow from its brothers and sisters, and the behavior of crows to hunt worms that are inside the tree trunk is used for modeling.

In Crow's algorithm, parents X1, X2 reward their behaviors according to the following matrix.

(1) F=[█((X_11,X_12,X_13,…,X_1d )=>X_1@(X_21,X_22,X_23,…,X_2d )=>X_2@(X_31,X_32,X_33,…,X_3d )=>X_3@⋮@(X_n1,X_n2,X_n3,…,X_nd )=>X_n )]

The competence of each crow is calculated according to the behavioral characteristics according to the following matrix:

ذخیره کردن ترجمه

(2) Cost=[█(Cost(X_11,X_12,X_13,…,X_1d )@Cost(X_21,X_22,X_23,…,X_2d )@Cost(X_31,X_32,X_33,…,X_3d )@⋮@Cost(X_n1,X_n2,X_n3,…,X_nd ) )]



2-2- adaptive neural fuzzy inference system

ANFIS structure has a good capability in training, construction and classification. Its learning rule is based on the error backpropagation algorithm by minimizing the mean squared error between the network output and the real output. [3].





Figure 1- simple diagram of ANFIS [3]





2-3- Basal-radial neural network

Radial-based neural network is used for non-parametric estimation of multidimensional functions from a limited set of training information. In this network, the hidden layer plays an important role in converting non-linear patterns into linear separable patterns. which is in the form of relation (3):

(3) "f" ("x" )"=" ∑_"i=1" ^"p" ▒〖"w" _"i" "φ(" 〖"Xc" 〗_"i" "-x)" 〗



3- Steps of the proposed method

All steps of the proposed method include pre-processing (cleaning, normalization and feature selection) and post-processing (proposed method).

In this article, the data of Tehran meteorological station is used, which includes 1000 data samples with 23 features.

Then, the fuzzy-neural adaptive inference system is used to predict the amount of nitrogen dioxide pollution. Crow learning algorithm is used to train this system.





Figure 3- The structure of ANFIS neural-fuzzy inference system

To select the parents based on the competence of the population members, the two crows that have the most competence are considered as parents.





Figure 4- The learning phase and the new position of the crow after the learning phase

Then, in the evaluation stage, the objective function is called and the mean square error is calculated. Finally, the termination conditions of the iteration are checked based on the lower mean square error.



4- Simulation results

4-1- Prediction results with the proposed method

The parameters of the population size of crows are 50 and the maximum number of repetitions is 500, the type of fuzzy inference system is Sogno type and Gaussian input membership functions are considered.



Table 1- Types of errors in the proposed method in predicting NO2





Figure 5- Error histogram for training and testing data in the proposed method



Figure 6- Target outputs and outputs of the proposed method for training data



Figure 7- Outputs of the target and the proposed method for the total experimental data





Figure 8- Outputs of the target and the proposed method for the total data

4-2- Forecasting results with the basic-radial neural network

The maximum number of neurons is 20, the type of radial functions is Gaussian, and the value of the dispersion parameter is 1.3.



Table 2- Types of errors in the basic-radial neural network in predicting NO2







Figure 9- Error histogram for training and testing data in radial-basal neural network



Figure 10 - Target outputs and radial-basis neural network for training data





Figure 11 - Target outputs and radial-basis neural network for the entire test data





Figure 12 - Target outputs and base-radial neural network outputs for the whole data

Conclusion:

This article is based on predicting the amount of nitrogen dioxide pollution using machine learning methods. According to tables (1) and (2), the fuzzy-adaptive neural inference system trained with the crow learning algorithm and the radial basis neural network performed the prediction with mean square error of 0.0081 and 0.0101, respectively. Therefore, the best performance belongs to the adaptive neuro-fuzzy inference system trained with the crow learning algorithm.



Reference:.

[1]. N. O. Lim, J. Hwang, S.-J. Lee, Y. Yoo, Y. Choi, and S. Jeon, "Spatialization and Prediction of Seasonal NO2 Pollution Due to Climate Change in the Korean Capital Area through Land Use Regression Modeling," International Journal of Environmental Research and Public Health, vol. 19, no. 9, p. 5111, 2022.

[2]. S. R. Shams, A. Jahani, S. Kalantary, M. Moeinaddini, and N. Khorasani, "Artificial intelligence accuracy assessment in NO2 concentration forecasting of metropolises air," Scientific Reports, vol. 11, no. 1, pp. 1-9, 2021.

[3]. A. Sabziparvar and M. B. Varkeshi, "Accuracy evaluation of ANN and Neuro-Fuzzy in global solar radiation," Iranian Journal of Physics Research, vol. 10, pp. 347-357, 2010.