Research on the Assimilation Algorithm of NO2 Combining Deep Learning and Three-Dimensional Variational Method

In the context of the escalating concern over air pollution issues, data assimilation within numerical weather prediction holds substantial significance in enhancing the precision of initial forecast values. In view of the deficiencies of existing assimilation methods, this paper proposes a three-dimensional deep learning assimilation algorithm that integrates the idea of variational assimilation. The WRF-Chem model is utilized to generate the background field, and the NO2observational data are obtained from observational stations and preprocessed. Drawing on the principle of the three-dimensional variational assimilation method, its cost function is introduced into deep learning as a loss function to participate in the training to construct the 3DVar-Unet architecture. Three sets of experiments, namely the Control group, the 3DVar-Lbfgs group, and the 3DVar-Unet group, are designed to test the assimilation effect. The results show that the RMSE of 3DVar-Unet has dropped from 24.67 to 12.18 (an improvement of 50.6%), and at the same time, the computing time has decreased from 427s to 11s (a reduction of 97.4%). This algorithm effectively improves the assimilation accuracy and speed, providing a new approach for solving related problems such as air pollution forecasting.