An adaptive fusion deep learning model integrating DenseNet and LSTM with spatiotemporal attention for urban air quality prediction

This study proposed an attention-enhanced Densely Connected Convolutional Network (DenseNet)–Long Short-Term Memory (LSTM) spatiotemporal fusion model for short-term urban air quality prediction. The model employed DenseNet to extract multi-station spatial features and LSTM to encode historical temporal dependencies. A spatiotemporal attention fusion module was introduced to dynamically adjust the contribution of spatial features at different forecasting horizons. The experiments adopted a chronological data split, and Min–Max normalization was performed exclusively using statistics derived from the training set to prevent future information leakage. Hourly air quality data collected in Shanghai from July 1, 2021, to December 1, 2023, were used for model evaluation. For Particulate Matter with aerodynamic diameter ≤ 2.5 μm (PM2.5) prediction, the model achieved a Root Mean Square Error (RMSE) of 6.35 μg/m 3 , a Mean Absolute Error (MAE) of 4.12 μg/m 3 , a Mean Absolute Percentage Error (MAPE) of 19.64%, a Nash–Sutcliffe Efficiency (NSE) of 0.91, and a Kling–Gupta Efficiency (KGE) of 0.89. The Wilcoxon signed-rank test demonstrated that the complete model significantly reduced prediction errors compared with baseline models, including LSTM, Gated Recurrent Unit (GRU), Transformer, and Informer. The model also maintained stable predictive performance across six pollutants, including PM2.5, Particulate Matter with aerodynamic diameter ≤ 10 μm (PM10), Sulfur Dioxide (SO2), Carbon Monoxide (CO), Nitrogen Dioxide (NO2), and Ozone (O3). These results indicated that the proposed architecture had strong potential for urban air quality early warning applications.