A hybrid architecture with bidirectional gating mechanism for spatiotemporal air quality prediction

Abstract

Abstract Accurate air quality prediction plays a crucial role in supporting various socio-economic activities, including agriculture, transportation, and disaster prevention. While traditional numerical air quality prediction methods and single-model deep learning approaches struggle to capture complex spatiotemporal dependencies, this study proposes a CNN-Transformer-LSTM model with a Bidirectional Gating Mechanism, termed BG-Hybrid, which dynamically balances local spatial features (CNN) and global temporal dependencies (Transformer) through the bidirectional gating core. The BG-Hybrid model enables adaptive feature fusion by projecting pooled features from both the two branches into sigmoid-activated weights ( $$\alpha , \beta$$ ), where $$\beta = 1 - \alpha$$ , ensuring information conservation. In terms of predictive performance, the model outperforms baseline models in error reduction and surpasses conventional hybrid architectures. This performance advantage is coupled with robust generalizability across diverse datasets, suggesting potential application in real-world scenarios.