Data assimilation (DA) integrates the latest observational data into initial background field, producing an optimal analysis field in continuous time and space within the realm of numerical model forecasting to improve forecasting. However, its computational efficiency remains a concern due to its relatively slow processing speed. Alternatively, deep learning (DL) methods train assimilation models using observational or forecast data as labels, but its effectiveness is often constrained by the accuracy of these labels. In this paper, we proposed a novel approach that combined three-dimensional variational assimilation (3D-Var) with DL method, which integrated the Hadamard attention mechanism and Transformer modules into the Unet framework (HT-Unet) to assimilate sulfur dioxide (SO2) and improve the forecast skill. This hybrid method not only significantly improved the forecast accuracy but also accelerated computational processes. Three forecast experiments were conducted with background filed, 3D-Var, and HT-Unet analysis fields to assess forecast improvements in SO2 concentration. The results showed that the correlation with HT-Unet analysis field was nearly 50% higher than the background field, which was comparable to those from 3D-Var analysis field. The RMSEs of the SO2 concentration in the HT-Unet and 3D-Var DA experiments were reduced by 2.59 µg/m³ (29.50%), and 2.60 µg/m³ (29.61%). Additionally, the HT-Unet method achieved computational efficiency far surpassing traditional methods, which was 34 times faster than traditional approaches for assimilating a single analysis. The new method demonstrated the potential to replace the traditional 3D-Var method, overcoming the high-cost limitations of conventional variational assimilation and significantly enhancing air pollution forecasting accuracy.