A Spatiotemporal Multimodal Framework for Air Pollution Prediction Based on Bayesian Optimization—Evidence from Sichuan, China

In regions characterized by complex terrain and diverse pollution sources, high-precision air pollution prediction remains challenging due to nonlinear spatiotemporal coupling and the difficulty of modeling local pollutant agglomeration. To address these issues, this study proposes a CNN–LSTM–Transformer multimodal prediction framework integrated with Bayesian Optimization. First, the Local Moran’s Index (LMI) is introduced as a spatial perception feature and concatenated with pollutant concentration sequences before being input into the CNN module. This design enhances the model’s ability to identify local pollutant clustering and spatial heterogeneity. Second, the LSTM architecture adopts a dual-channel structure: the main channel employs bidirectional LSTM to extract temporal dependencies, while the auxiliary channel uses unidirectional LSTM to capture evolutionary trends. A Transformer with a multi-head attention mechanism is then introduced to perform global modeling. Bayesian Optimization is employed to automatically adjust key hyperparameters, thereby improving the model’s stability and convergence efficiency. Empirical results based on atmospheric pollution monitoring data from Sichuan Province during 2021–2024 demonstrate that the proposed model outperforms various mainstream methods in predicting six pollutants in Chengdu. For instance, the MAE for PM2.5 decreased by 14.9–22.1%, while the coefficient of determination (R2) remained stable between 87% and 89%. The accuracy decay rate across four-day forecasts was controlled within 12.4%. Furthermore, in PM2.5 generalization prediction tasks across four other cities—Yibin, Zigong, Nanchong, and Mianyang—the model exhibited superior stability and robustness, achieving an average R2 of 87.4%. These findings highlight the model’s long-term stability and regional generalization capability, offering reliable technical support for air pollution prediction and control strategies in Sichuan Province and potentially beyond.