Research On Prediction Of Air Pollutant Concentration In Jinchang City

Jinchang City of Gansu Province is well known as the"Nickel Capital of the Motherland"due to its abundance of nickel,and a large state-owned enterprise-Jinchuan Group Company is located in this city.Jinchuan Group is the country’s largest producer of nickel cobalt group metals and the third largest copper producer.Its production process uses pyrometallurgical technology,whose emissions contains sulfur dioxide,nitrogen oxides,soot and some other pollutants.In terms of geography and climate,Jinchang City is located in the eastern part of the Hexi Corridor.The year-round climate is mainly sunny,with a temperate continental climate,more frequent sand and dust activities in spring and autumn,a certain concentration of respirable particles in the atmosphere,low average annual precipitation,high solar radiation,and easy ozone generation in the air.In summary,various pollutants exist in the air of Jinchang city at different levels,including sulfur dioxide,nitrogen oxides,soot,respirable particulate matter and ozone,etc.The prediction research on their air pollution conditions can help people to prevent the air pollution in advance,and can also investigate more thoroughly the patterns behind the variation of pollutant concentrations.The prediction of air pollutants using artificial intelligence is based on the historical observation data,using machine learning,deep learning and other methods to analyze the temporal variation pattern of the data,in order to provide a reasonable prediction of the subsequent instantaneous values.The current research on air pollutants prediction for China is mainly focused on the eastern region with high economic development and population density,but few researches on the northwest region,meanwhile these researches mainly focus on just two aspects,AQI（Air Quality Index）which describes the overall condition of environmental air pollution on one hand,and fine particulate matter（PM2.5）which triggers haze on the other hand.However,the prediction results of AQI values only represent the overall pollution level of the air,and the prediction studies of PM2.5values ignore other categories of air pollutants.Although these studies have achieved some results,they are not applicable to Jinchang City,where air pollution characteristics are significant.According to the national standard,the calculation of AQI takes into account six basic air pollutants,which are"sulfur dioxide（SO₂）,nitrogen dioxide（NO₂）,carbon monoxide（CO）,ozone（O₃）,respirable particulate matter（PM10）and fine particulate matter（PM2.5）".In order to investigate the real air pollutants in Jinchang city more precisely,this thesis refines the air pollutants into the six items mentioned above,and uses deep learning techniques to carry out research on the prediction of the concentration values of the six air pollutants in different time dimensions.Three types of data were prepared to predict the six air pollutants:historical monitored data of air pollutant concentrations,meteorological observed data of Jinchang urban area and monitored data of major industrial exhaust emission points in the city for the correspondent periods.The main work is as follows:（1）Characterization of air pollutant concentrations in Jinchang City.In my thesis,firstly,the statistics and analysis of the six air pollutant concentrations in Jinchang were made.By calculating the extreme values and coefficients of variation of each pollutant,it was found that the dispersion of PM2.5,PM10 and SO₂was relatively large;by calculating the autocorrelation coefficients,it was found that the values of each pollutant in previous moments had a significant influence on the following moments,and the degree of influence was varying for different pollutants,among which the correlation of SO₂was the lowest.At the same time,it was found that the concentrations of O₃,CO and NO₂have a 24h periodical variation pattern;by calculating the Pearson correlation coefficients among the six air pollutants,it was found that each pollutant has a certain level of positive or negative correlation with each other;by comparing the concentrations of PM2.5 and PM10 in three adjacent cities with those of Jinchang,it was found that the corresponding pollutants in different cities also had a higher correlation with each other.Therefore,some of the air pollution data from neighboring cities can be used as input variables for the model to supplement the predictions.These studies demonstrate that the concentrations of each pollutant do not varies in isolation,but are correlated with each other,and it is feasible to predict the future moment values from the historical moment data of each pollutant.（2）Design of deep learning models and comparative experiments.In order to improve the prediction of air pollutants concentrations,a Sequence-to-Sequence model based on GRU network was designed for the experiments.The experiments were conducted by feeding three categories data in a given historical range to the model,then extracting characteristic information in the encoding stage,and subsequently using the characteristic information in the decoding stage to sequentially predict the output air pollutant concentration values of a certain air pollutant from the1st to the nth hour in the future.For the purpose of enabling the model working in a relatively optimized state,four sets of comparison experiments were conducted:the first set of experiments compared the differences between using LSTM and GRU networks during the coding and decoding stages of the sequence-to-sequence model;the second set of experiments compared the effects of different GRU network layers in the model on the prediction performance;the third set of experiments compared the effects of different loss functions on the fitting performance during the model training;the fourth set of experiments compared the effects of different range of historical input data on the prediction performance.The conclusions of these comparison experiments were used as the basis for the optimum parameters setting of the Deep Model.（3）Research on the prediction of pollutants and results analysis.The experimental results show that the observed concentration values of PM2.5 and PM10pollutants have very similar characteristics and variation patterns,so the performance of the model for predicting these two pollutants is almost the identical,and the prediction of their variation patterns and concentration levels have high accuracy in the near future.The concentration values of SO₂are affected by local emissions and meteorological conditions,with no obvious variation pattern,so the model can just predict the concentration trends and levels in the next 1～2 h.The concentration values of NO₂,O₃and CO have a 24-h periodic cycle,so the deep learning model has relatively excellent prediction performance for them,which can effectively predict the concentration values in the following 12 h.In summary,the method proposed in this paper has high prediction accuracy and practical utility in predicting the concentration values of six air pollutants for the next few hours in Jinchang City.