基于深度学习的城市PM2.5浓度时空分布预测及不确定性评估

Deep learning-based spatio-temporal prediction and uncertainty assessment of urban PM2.5 distribution

城市PM2.5浓度的时空分布预测旨在基于有限观测样本实现研究区域内PM2.5分布的全范围感知。理想的预测模型需同时保证结果的高精度与高可靠性。然而,现有研究大多以提升精度为唯一目的,忽视了由于数据质量与模型结构的各异所导致预测结果的不确定性,这极大限制了高精度预测结果的可靠性与可用潜力,从而难以有效辅助空气污染治理等实际应用。为此,本文提出一种耦合不确定性评估的PM2.5浓度时空分布预测模型。通过构建以图卷积和循环网络为主的预测模块,实现PM2.5浓度的高精度预测;同时,基于对抗学习策略与变分自编码思想构建不确定性量化模块,同步揭示预测结果的不确定性水平。深圳市实际数据实证表明,本文方法能有效兼顾PM2.5浓度预测结果的精度与可靠性,能为包括监测站点布局选址在内的环境治理工作提供科学决策支持。

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 results,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 autoencoder 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 results,providing scientific decision-making support for environmental management.