Abstract:To improve the accuracy of precipitation nowcasting, this paper proposes an adversarial neural network model named CastNet that combines deep neural networks. This model utilizes the recurrent neur-al network to capture the spatio-temporal features of radar echo data, employs the adversarial neural network to simulate the generation and dissipation changes of cloud clusters, and then integrates the optical flow constraint into the neural network to guide the model training. This accelerates the learning process of the neural network and enhances the spatio-temporal consistency of the model, effectively solving the problem of forecast ambiguity and significantly improving the accuracy of precipitation intensity and location. Tests are conducted on 9 major precipitation processes in Guangxi and its surrounding areas from May to October 2023. The results show that under various precipita-tion intensities (≥0.1, ≥2, ≥7, ≥15, ≥25, ≥40 mm·h-1), the average TS scores of SWAN 2.0 are 0.458, 0.270, 0.085, 0.034, 0.014 and 0.003， respectively; the average TS scores of SWAN 3.0 are 0.452, 0.402, 0.225, 0.129, 0.085 and 0.048， respectively; and the average TS scores of the CastNet model are 0.439, 0.397, 0.225, 0.139, 0.104 and 0.073， respectively. It can be seen clearly that the TS scores by the CastNet are higher than those of SWAN 2.0 and SWAN 3.0 under high-intensity precipita-tion of ≥7 mm·h-1 and above, except for few data points that are flat. In addition, as the forecast lead time extends, the relative advantage of CastNet becomes more obvious.