Meteorological data assimilation (DA) improves the forecasting capabilities of numerical weather prediction (NWP) models by using observations. Meanwhile, the development of deep learning (DL) technology has enabled the generation of reasonably accurate “future observations.” Among all these “future observations,” weather radar echo extrapolation (WREE) data derived from DL can be utilized to predict the development of convective systems, which is valuable for weather forecasts. Based on a DL model, which combines the residual convolution long short‐term memory (rcLSTM) network with the innovative framework of a generative adversarial network (GAN) named GAN‐rcLSTM, this study extrapolates the S‐band radar echo mosaics at 3‐km altitude from seven radars in Guangdong and assesses the performance of the GAN‐rcLSTM model. After confirming the good performance of the DL model, DA experiments are designed to explore the possibility of assimilating WREE data and assess its impact on the NWP. The results show that assimilating WREE data or radar echo observations (REOs) data can lead to a reduction in the reflectivity errors for both the analysis and forecast fields, which consequently improves the reflectivity distribution on analyses and forecasts. Additionally, assimilating WREE data not only has positive impacts on the location and intensity of precipitation but also improves the precipitation critical success index (CSI), which is similar to REOs assimilation results. Therefore, assimilating WREE data can achieve similar positive impacts compared to assimilating REOs data, offering a novel approach to improve the short‐term forecasts from NWP.

Data assimilation,Radar echo extrapolation,Deep learning