基于注意力常微分方程的PM_(2.5)浓度预测及其可解释性分析

An Attentional Ordinary Differential Equation for Predicting PM_(2.5) Concentration and Its Interpretability Analysis

可解释的准确预测PM_(2.5)浓度变化可以有助于人类规避暴露风险,对人类健康风险评估和政策实施具有重要意义。目前已有PM_(2.5)浓度预测模型过多专注于提升模型预测精度,但忽略了模型的可解释性,造成模型可复用性和可信任度较差。鉴于此,本文提出了一种兼顾模型预测精度与模型可解释性的注意力时空常微分方程模型(Attentional SpatioTemporal Ordinary Differential Equation,ASTODE)用于PM_(2.5)浓度预测任务。具体而言,本文将神经常微分方程集成至PM_(2.5)浓度预测任务中,以提升预测模型的可解释性。此外,针对传统神经常微分方程难以挖掘PM_(2.5)浓度数据中空间依赖关系的挑战,本文提出了一种新颖时空导数网络将传统神经常微分方程扩展到了时空常微分方程。针对传统神经常微分方程难以挖掘PM_(2.5)浓度数据中长期依赖关系的挑战,本文设计了一种时空注意力机制去融合多个时间节点的隐藏状态。本文采用真实的PM_(2.5)浓度数据集对提出的ASTODE模型进行了验证。实验结果表明,ASTODE模型不仅在预测精度上优于或逼近于存在的6个基线方法,并且在可视化的视角下具有良好的可解释性。

Accurate and explainable prediction of PM_(2.5) concentration can help humans avoid exposure risks to air pollution,which is of great significance for human health risk assessment and policy implementation.Currently,the existing PM_(2.5) concentration prediction models focus on improving the model prediction accuracy without considering model interpretability,resulting in poor model reusability and trustworthiness.Therefore,this paper proposes an Attentional Spatiotemporal Ordinary Differential Equation(ASTODE)model for PM_(2.5) concentration prediction tasks considering both prediction accuracy and model interpretability.Specifically,this paper integrates the Neural Ordinary Differential Equation(NODE)into the PM_(2.5) concentration prediction task to improve the interpretability of the prediction model.In addition,to address the challenge of traditional NODE in mining spatial dependencies in PM_(2.5) concentration data,this paper proposes a novel spatiotemporal derivative network that extends the traditional NODE to spatiotemporal ordinary differential equations.To address the challenges of traditional NODE in mining long-term dependencies in PM_(2.5) concentration data,this paper proposes a spatiotemporal attention mechanism to fuse hidden states of multiple time nodes.In the experimental section,the proposed ASTODE model is validated using a real PM_(2.5) concentration dataset.This paper quantifies the prediction errors of the ASTODE model in both temporal and spatial dimensions.By comparing with six existing baseline methods,our proposed ASTODE model obtains a similar or higher prediction accuracy.This paper also analyzes the interpretability of our proposed ASTODE model from a visualization perspective,demonstrating that the proposed ASTODE model balances the prediction accuracy and interpretability to some extent.