<p>The chemical transport model (CTM) is often used for PM_2.5 forecasting. However, due to the uncertainty sources in modelling processes such as meteorology, emissions, and chemical reactions, the PM_2.5 forecasting of the CTM still has high uncertainty, especially during severe air pollution episodes with complex meteorological and chemical coupling. In this study, by using low-memory GPU, we developed a Model-Coupled Deep Learning method (MCDL method) by coupling a deep learning method with the CTM to improve the spatiotemporal forecasting of PM_2.5 for 3 days in China during a severe haze episode. This method assimilates surface-observed PM_2.5 into the CTM to produce an analysis field of PM_2.5. Then, an encoder-decoder-based deep learning algorithm (ConvGRU) was used to predict the bias between the PM_2.5 of the CTM and the analysis field, therefore improving the accuracy of the forecasting. The PM_2.5 of the CTM and meteorological variables of the WRF model were added as auxiliary variables for training. The results suggest that the MCDL method can significantly improve the PM_2.5 forecast of the CTM, where the root-mean-square error (RMSE) decreased by 24.5%. The MCDL method also performed better than the data-driven deep learning method with 20% lower RMSE. During this severe haze episode, the MCDL method effectively corrected the significant underestimation or overestimation of the CTM spatiotemporal forecast results and has better long-term spatiotemporal forecasting performance than the data-driven deep learning method. Besides, the MCDL method has the highest accuracy in forecasting pollution levels of PM_2.5 when daily concentrations exceed 75&#xa0;µg/m³.</p>