Research On Air Quality Analysis And Prediction Technology Based On Multi-source Data

With the development of economy,the expansion of industrial scale and the rapid increase of energy consumption in China,many cities are facing air pollution problems by different degrees.Severely air-polluted environment causes much inconvenience to people’s daily life.Therefore,grasping changing rules and predicting the trend of air quality are becoming important.Based on the public data from the air quality monitoring station and the mobile-data collected by the self-made miniaturized air pollution monitoring equipment,this paper conducts a temporal-spatial analysis of the air quality in Nanjing,and also uses machine learning and neural network algorithms to predict the PM_2.5 concentration on the time series.The main work of this paper is concluded as follows:1.For the difficulty of air quality evaluation,the index method and fuzzy mathematics method are introduced.According to the experiment results,the two evaluation methods are quite different,and the results obtained by the fuzzy mathematics method are better than the index method.The index method only takes the largest air quality sub-index in the pollutant project as the final evaluation result,while the fuzzy mathematical method fully considers the impact of each pollutant on the air quality evaluation result.So in real life,we can flexibly selected the two methods to make the results more reasonable and accurate.2.For the low granularity of data collected by the air quality monitoring station,after analyzing the pollutants concentration change from different time aspects（year,month,hour）,this paper proposes a moving way to collect pollutants concentration data in the Sipailou campus of Southeast University with surrounding areas by the miniaturized air pollution monitoring equipment.In order to study the distribution characteristics of pollutants from a spatial perspective,Kriging interpolation method was used.The results show that the concentration of each pollutant has certain changes in time and space,and is greatly affected by human activities and meteorological factors.3.For the complicated calculation of traditional prediction methods,this paper introduces new methods based on machine learning and neural network algorithms to predict PM_2.5 concentration.To optimize the prediction model paramrters,simulated annealing algorithms is proposed.The experiment results show that the SVR,BP and LSTM model can predict the trendency of PM_2.5concentration,but the prediction result has a lag;according to the evaluation indicators RMSE and MAE,the prediction error of each model is relatively small:RMSE and MAE are about 13μg/m³and 9μg/m³respectively,but the accuracy of the LSTM model is still the highest among the three models.4.For the lagging problem of prediction results by machine learning and neural network algorithms,a fusion prediction model based on wavelet decomposition and reconstruction theory is proposed.This model disperses the trend and fluctuation characteristics of PM_2.5 concentration sequence into different component sequences.A component model is constructed for each component sequence according to the SVR,BP and LSTM algorithms,and all the component models added up make the final prediction result.The experiment results show that the improvement is quite obvious after incorporating the wavelet transform for each algorithm model,because the hysteresis no longer occurs and the accuracy of all the three prediction algorithms is increased by about 50%.