Review and evaluation of Monte Carlo algorithms and neural networks to predict air pollution in the environment of a spatial information system

Master's Thesis

Abstract

The necessity of having a healthy environment and raising the health level of the society makes the need to have proper planning to reduce the sources of air pollutants production and predict these pollutants to prevent its harmful effects inevitable. Prediction of pollutants can be useful in air pollution management and control. In this research, O3 pollutant due to its harmful effects on human health, as well as CO pollutant due to the use of non-standard vehicles and the problem of incomplete fuel combustion in cars in the city of Tehran have been taken into consideration. Meanwhile, the use of neural networks due to their appropriate ability to model systems with non-linear behavior can be useful for predicting changes in air pollutants. With such an approach, in this thesis, prediction and modeling of hourly changes in the concentration of CO and O3 pollutants using MLP and Elman neural networks and Bayesian regression have been investigated. In order to pre-process the data before entering the neural network, principal component analysis has been used. In this case, the use of the components obtained from the principal component analysis method has led to a reduction in the number of features, an increase in the degree of freedom, and a reduction in the training time of the network.

Prediction of the two pollutants CO and O3 using Bayesian regression and estimation of its parameters by the Monte Carlo Markov chain method have also been considered.

The results of the implementation of two types of neural networks and Bayesian regression show that MLP network with coefficient of determination (R2) equal to 0.6307 for CO prediction and Elman network with coefficient of determination equal to 0.6186 for O3 prediction have the best accuracy. Therefore, the results of the research confirm the superiority of the proposed neural networks over Bayesian regression.

Key words: neural network, principal component analysis, Bayesian regression, Markov Monte Carlo chains. It is worrying to predict. Unfortunately, the ever-increasing activities of humans, especially after the industrial revolution, have caused air pollution on a large scale.

It is clear that knowing biological behaviors in the production of air pollutants can help in managing and controlling air quality, and as a result, raising the level of social health and reducing the effects of air pollution; Because with this knowledge, one can think about the necessary planning to reduce the sources of air pollution production and to have a healthy environment. Primary pollutants are substances that enter the ambient air directly from sources and include five pollutants: carbon monoxide CO, nitrogen dioxide NO2, sulfur dioxide SO2, suspended particles with a diameter of less than 10 microns PM10 and volatile hydrocarbons VOCs. Secondary pollutants refer to substances that arise as a result of interactions in the air around the earth, and ozone O3 can be mentioned in this group. In this research, among the mentioned pollutants, the prediction of two pollutants, CO and O3, is the basis of the work. We know the necessity of ozone prediction due to its negative effects on human, animal and plant health and that with ozone modeling, it is possible to issue a quick warning in places where its concentration increases. Also, since cars are the main source of carbon monoxide production, due to the heavy traffic volume caused by transportation in the city of Tehran, the use of non-standard cars and the problem of incomplete combustion of the fuels used in cars, we have paid attention to the prediction of CO. Considering the deadly effects that carbon monoxide can have on human health, it seems necessary to make the necessary decisions for proper planning in dealing with this problem.

As it is necessary to have a suitable decision in the future, we must obtain appropriate information about the behavior of our system so that we can check how it functions in other times by modeling the behavior of the system. In such a way, after proper modeling of the system, we can make a proper prediction of its behavior in the future and, as a result, make more optimal decisions to prevent unwanted incidents.In the way of systems modeling, knowing the influencing parameters in the system, the relationship of these parameters and the type of influence of each one in the system is one of the main discussions in the analysis and identification of the system.

With this approach, in this research, we are looking for a suitable analysis of the environment so that we can predict its behavior and have a more accurate drawing of the future for ourselves. Neural networks, fuzzy logic, regressions and statistical methods are often used to model the behavior of air pollutants. In this research, we are looking to model and predict two pollutants, CO and O3, using the method of neural networks and linear regressions.

Neural networks have a high ability to extract patterns from data and also solve complex problems of a natural nature. The accuracy of the implementation of these networks is appropriate in the case of dependence of the input parameters and even the presence of noise in the data. In this research, among the different architectures of neural networks, two multilayer perceptron neural networks [1] and Elman neural network [2] have been used to predict CO and O3 pollutants. The multi-layer perceptron neural network, despite its classicity in modeling intelligent systems, has been chosen due to its high flexibility in process modeling and its wide application in the discussion of air pollutant forecasting. Also, in this research, we have used Elman's neural network due to its structural and functional nature in modeling time series. 

In regression methods, the need to accurately estimate the results and obtain relationships between parameters and variables affecting the results has created a wide range of these methods, among these regression methods, Bayes regression method [3] can be mentioned. In this research, we use Markov Monte Carlo chain method to determine the posterior distribution of Bayes estimation. 

In this research, we aim to model the process of the two pollutants CO and O3 using neural network and Bayesian regression methods, investigate the accuracy of the neural network method as a smart nonlinear method and Bayesian regression as a classical linear method and examine the efficiency and flexibility of the methods used in modeling the two pollutants CO and O3. Also, by implementing MLP and Elman neural networks in order to predict the two pollutants CO and O3, we compare the accuracy of these two networks for modeling. 1-2- Background and objectives of the thesis Using neural networks and Bayesian regression to predict the two pollutants CO and O3 is the main goal of this thesis. Considering the multiplicity of neural network structures, comparing the effectiveness of two types of neural networks, MLP and Elman, in predicting air pollution, as well as comparing their accuracy with the Bayesian regression method, constitute the other goals of this thesis. According to the above, the sub-goals of this thesis are as follows: Introducing neural networks as an efficient method in predicting air pollutants.

Investigation of the relationship between the amounts of pollutants and the error resulting from their prediction by neural networks.

Investigation and analysis of the spatial and temporal variability of the two pollutants CO and O3.

1-3-      Overview of the researches

Researches carried out in the field of prediction of air pollutants have been expanded in various aspects. In this research, the prediction of meteorological pollutants has been challenged mainly from the aspect of the data used to predict each pollutant, the pre-processing performed on the input data to the forecasting system, the structure of the model used for prediction and the accuracy of these models.

In the path of predicting air pollutants, the use of different data to enter the neural network has been proposed. In some of the conducted researches, according to the relationship between the desired pollutants and meteorological parameters, the use of meteorological parameters to predict the desired pollutants has been suggested [1], [2] and [3]. As an example, the use of meteorological data including: wind speed, relative humidity, wind direction, temperature, rainfall, air pressure, radiation amount and evaporation amount has been suggested in order to predict the O3 concentration hourly [1]. The prediction of 3 pollutants PM10, SO2 and CO in the next 24, 48 and 72 hours, using meteorological parameters including: wind direction, air pressure, day temperature, night temperature, relative humidity and wind speed is also suggested [2].