Interpretable regional meteorological feature extraction enhances deep learning for extended 120-h PM2.5 forecasting

While deep learning models perform well in short-term PM2.5 forecasting, their performance tends to decay significantly with increasing forecast spans. This study proposes a "Domain-to-Point" approach to enhance deep learning models for extended 120-h PM2.5 forecasting. We used convolutional autoencoders (ConvAEs) to extract regional meteorological features (RMFs) from both surrounding and future meteorological data provided by an advanced weather forecasting system. The RMFs and temporal indicators were used as predictor variables in a time-attention-based bidirectional long short-term memory neural network to forecast the next 24- to 120-h PM2.5. For four megacities in China, the ConvAEs compressed the high-dimensional meteorological data into a lower-dimensional representation, while preserving most of the relevant information (R-2 > 0.75). Using the RMFs as inputs substantially mitigated the performance decay of the PM2.5 forecasting models over extended forecast spans, enabling accurate longer-term forecasting, with R-2 improvements of 14.7-79.2% for the 120-h forecasts. This study further revealed that errors in the meteorological forecasts are the primary cause of performance decay, suggesting that improving the accuracy of meteorological forecasts could be an effective approach for enhancing PM2.5 forecasting performance. With Beijing as a case study, model interpretation using the Shapley-additive-explanations method demonstrated the crucial role played by RMFs in accurately forecasting a typical air pollution episode and identifying the main meteorological factors affecting pollution. The proposed "Domain-to-Point" approach enhances the capability of deep learning models for longer-term PM2.5 forecasting, thus facilitating the development of preventative measures to safeguard public health against air pollution.