Research On Air Quality Prediction Method Based On Multi-source Data Fusio

With the acceleration of industrialization and urbanization,severe air pollution has become one of the global environmental problems.Various air pollutants,such as PM_2.5,PM₁₀,SO₂,not only cause significant damage to the ecological environment but also seriously affect people’s health and quality of life.To obtain timely information on air quality,many air quality monitoring stations have been established in various regions to collect real-time air quality data and predict future air quality conditions.These efforts aim to assist governments and the public in taking measures to mitigate the impact of air pollution on public health and the environment.With further research,various statistical and deep learning models have been successfully applied to air quality prediction.However,the concentration of various air pollutants is affected by various factors,such as meteorological conditions and spatial geographic information,which to some extent,increase the complexity of air quality prediction.Additionally,the correlation between various pollutants and the volatility of pollutant indicator data also increases the difficulty of accurately predicting air quality.To address the above problems,this thesis incorporates different air pollutants,spatial information and meteorological conditions into the design of air quality prediction models,designing a statistical prediction model and a deep learning prediction model to improve air quality prediction accuracy.Specifically,the main work is as follows:（1）Aiming at the correlation of air quality data among different monitoring stations and the volatility of air quality data,a method is proposed for predicting air quality that incorporates spatial information through truncated singular value decomposition and autoregressive integrated moving average（ARIMA）.This method first utilizes truncated singular value decomposition to compress and encode air quality data from multiple monitoring stations,remove redundant data,and extract potential spatially related features.Then,to eliminate the influence of air quality data volatility on predictions,empirical mode decomposition is used to stabilize the data.Finally,the traditional single-sequence ARIMA model is extended to a matrix-sequence prediction model and combined with empirical mode decomposition and truncated singular value decomposition to construct a multi-site air quality prediction model called SE-ARIMA.To validate the effectiveness of this model,ablation and comparative experiments are conducted on a real-world dataset and the results show that SE-ARIMA has a certain advantage in predicting air quality accuracy.（2）For the characteristic that air quality has a certain correlation with spatial information and meteorological factors,an air quality prediction method based on the graph convolutional networks and gated recurrent unit is proposed.The method utilizes the graph convolutional networks to extract spatial features and gated recurrent units to extract temporal features,constructing the GWGCN-GRU model for prediction.Weight optimization matrices and gating mechanisms are incorporated into the model to better capture spatial features,further enhancing prediction accuracy and precision.Comparison and ablation experiments are conducted on a real-world dataset.The results show the effectiveness of the GWGCN-GRU model for improving the accuracy of air quality prediction.