Improving PM2.5 simulations using LSTM: a study on spatiotemporal generalization

Abstract

Deep learning models, particularly long short-term memory (LSTM) networks, have shown strong potential for bias correction in air quality simulations. Although LSTM-based models have demonstrated success in improving PM_2.5 simulations during specific periods such as winter months or heavy pollution events, their ability to generalize across varying temporal intervals and geographical locations remains underexplored. This study systematically investigates the spatiotemporal generalization capabilities of LSTM-based (LSTM_Local, LSTM_Regional, LSTM_Regional-idx, LSTM_Regional-sub) models across 12 monthly intervals and 34 monitoring sites in Beijing. Benchmark comparisons with alternative deep learning models (RNN, GRU, BPNN, CNN) demonstrate the superiority of LSTM-based model in improving predictive performance. Seasonal analysis reveals that the LSTM_Local and LSTM_Regional-sub model achieved modest gains during summer half-year (May to October) but achieved significant improvements during winter half-year (November to April), with average RMSE reductions of −5.89 % (November), −17.40 % (December), −6.37 % (January), 0.36 % (February), −5.87 % (March) and −1.57 % (April). Spatially, urban sites show moderate gains, but suburban sites exhibit greater improvement, with average RMSE of −5.14 % (Center), −2.26 % (South-East), −7.05 % (North-East), −8.39 % (South-West) and −12.37 % (North-West). These findings highlight the robust spatiotemporal generalization of LSTM-based models and support their applicability for real-time bias correction, long-term forecasting, and air quality dataset enhancement at fine spatial and temporal resolutions.