Spatiotemporal Causal Decoupling Model for Air Quality Forecasting

TL;DR

AirCade addresses the challenge of accurate air quality forecasting by explicitly modeling the causal relationships between AQI and meteorological features. It introduces a spatiotemporal learning module with Domain Knowledge Prompt embeddings, a causal decoupling transformer (CADE) with DK-MSA-based Cade and Cadi layers, and a causal intervention mechanism to account for uncertainty in future weather. On the KnowAir dataset, AirCade achieves over 20% relative improvement over state-of-the-art baselines, underscoring the value of explicit causal decoupling and robustness to weather uncertainty. The work advances spatiotemporal air quality forecasting by integrating domain knowledge, causal reasoning, and robust attention mechanisms into a Transformer-based framework.

Abstract

Due to the profound impact of air pollution on human health, livelihoods, and economic development, air quality forecasting is of paramount significance. Initially, we employ the causal graph method to scrutinize the constraints of existing research in comprehensively modeling the causal relationships between the air quality index (AQI) and meteorological features. In order to enhance prediction accuracy, we introduce a novel air quality forecasting model, AirCade, which incorporates a causal decoupling approach. AirCade leverages a spatiotemporal module in conjunction with knowledge embedding techniques to capture the internal dynamics of AQI. Subsequently, a causal decoupling module is proposed to disentangle synchronous causality from past AQI and meteorological features, followed by the dissemination of acquired knowledge to future time steps to enhance performance. Additionally, we introduce a causal intervention mechanism to explicitly represent the uncertainty of future meteorological features, thereby bolstering the model's robustness. Our evaluation of AirCade on an open-source air quality dataset demonstrates over 20\% relative improvement over state-of-the-art models.

Figures (6)

• Figure 1: Three different causal structures with our proposed perspective. $\mathbf{X}$ and $\mathbf{Y}$ represent the past and future AQI respectively. $\mathbf{\hat{Z}}$ and $\mathbf{\widetilde{Z}}$ represent the past and future weather conditions respectively.
• Figure 2: Details of the proposed model (AirCade), which consists of spatial and temporal transformers. Each Transformer layer contains a causal decoupling module that explicitly models the relationship between AQI and weather.
• Figure 3: Domain knowledge multi-head self-attention mechanism (DK-MSA). Adaptive adjacency matrices coefficients and gated mechanism are incorporated into the self-attention function, alongside masking of coefficients to simulate environmental intervention.
• Figure 4: Ablation results on KnowAir dataset.
• Figure 5: Hyperparameter sensitivity results on KnowAir dataset.