Deep learning-based spatio-temporal prediction and uncertainty assess-ment of urban PM2.5 distribution
The goal of predicting PM2.5 concentration is to achieve a comprehensive perception of the PM2.5 distribution in the study area based on limited observations.Ideal prediction models are required to ensure both high accuracy and high reliability of the results.However,most of the existing studies prioritize the efforts to improve accuracy,which ignores the uncertainty of results caused by data and model.This greatly limits the reliability and potential availability of high-precision prediction re-sults,making it difficult to assist practical applications such as air pollution control effectively.To overcome this problem,this paper proposes a PM2.5 concentration spatiotemporal distribution prediction model with coupled uncertainty assessment.The prediction module,mainly based on graph convolutional and recurrent networks,achieves high-precision prediction of PM2.5 concentration.Meanwhile,the uncertainty quantification module based on adversarial learning strategies and variational autoen-coder is constructed to synchronously reveal the uncertainty level of the prediction results.Extensive evaluations of real-world dataset show that the proposed model can effectively balance the accuracy and reliability of PM2.5 concentration prediction re-sults,providing scientific decision-making support for environmental management.