Research On Deep Learning Based Air Pollution Prediction System

In recent years,the air quality in China has improved significantly,but the situation of air pollution management remains grim,with high pollution weather still frequent in the autumn and winter seasons.The Ministry of Ecology and Environment has proposed an air quality improvement target of basically eliminating heavily polluted weather by2025,which places high demands on both monitoring and forecasting of air pollution.Most statistical models for prediction do not take into account the parameter quantity of the model,or only predict the concentration of some important pollutants such asPM_2.5orPM₁₀,and few models make holistic predictions on all six pollutants(PM_2.5,PM₁₀,₂,₂,and₃)that affect the air quality index.Furthermore,the existing atmospheric pollution monitoring system currently has sparse distribution of monitoring points or outdated data updates,which inevitably leads to lagging monitoring data.Therefore,in this paper,a model based on feature extraction and LSTM structure（FE-LSTM）and a discrete wavelet and convolution based auto encoder（DW-CAE）were designed for single-factor concentration prediction and prediction of concentrations for all six pollutants,respectively.The FE-LSTM model requires fewer training parameters,and the DW-CAE model have high prediction accuracy.In addition,an air quality monitoring and predicting system was designed and implemented to monitor real-time air pollution concentration levels in various regions.Multiple predictive models,including DW-CAE,were incorporated into the prediction module of the system to meet the requirements of environmental protection departments for air pollution prevention and control.The main research contents of this paper are as follows:（1）A model based on feature extraction and LSTM structure（FE-LSTM）was designed for single-variable concentration prediction.The FE-LSTM model was used forPM_2.5 concentration prediction on the BeijingPM_2.5 dataset and for concentration prediction of six major pollutants on the Yining air pollution dataset.The performance of the FE-LSTM model was compared with other models such as LSTM,bi-LSTM,and Auto-encoder.TakingPM_2.5 concentration prediction on two datasets as an example,the mean squared error（MSE）of FE-LSTM model decreased by 15.74%and 2.05%compared to Auto-encoder,while the parameter volume reduced by 27.86%compared to Auto-encoder.This indicates that the FE-LSTM model has slightly better prediction accuracy than the Auto-encoder model and requires fewer training parameters.（2）A discrete wavelet and convolution based auto encoder（DW-CAE）was proposed,which combines deep learning and signal processing techniques for multi-variable concentration prediction.The model uses discrete wavelet transformation to obtain the high and low frequency components of the target sequence.Then,a feature extraction module is designed to capture the correlations among variables,and the obtained feature matrix is input into an LSTM-based auto-encoder for prediction.The DW-CAE model was used to predictPM_2.5 concentrations in the BeijingPM_2.5 dataset,and to perform single-variable and multi-variable predictions in the Yining air pollution dataset.The predictive accuracy of the DW-CAE model was compared with that of LSTM,IMV-Full,DARNN,and other benchmark models.The results showed that the DW-CAE model outperformed other benchmark models in terms of accuracy,whether predicting individual or multiple pollution factors.In the overall prediction of six air pollutants,the R2 correlation coefficients between the predicted and true concentrations for each variable were all above 93%.This indicates the effectiveness of the DW-CAE model and its potential to provide technical and theoretical support for the prediction and control of overall air pollution.（3）An air quality monitoring and predicting system was designed and deployed in regions such as Yining City.The system provides real-time monitoring data and future pollution forecast data,and includes functionalities such as real-time monitoring,historical data query,comparative analysis,correlation analysis,pollution visualization,pollution tracing,pollution forecasting and station reports.The system presents real-time data and changes of atmospheric pollutants at each monitoring site through visualizations such as charts and GIS maps,which helps environmental protection department to analyze the air pollution situation in various regions,and formulate targeted countermeasures.