PINN-based mesoscale wind field reconstruction of radar measured convective storm in Nanchang

Radar echo data, which include surface meteorological information, are frequently required for various types of weather predictions. However, the spatial resolution of meteorological radar data is typically on the order of kilometers. Additionally, various environmental factors can interfere with radar data acquisition, leading to instability and frequent data gaps. To solve this problem, the physics informed neural network (PINN) is utilized to reconstruct the radar measured convective storm in Nanchang, so as to supplement the missing radar echo data. Navier-Stokes equation is encoded in PINN as the prior physical knowledge. The reconstruction errors are also statistically analyzed. What's more, the evolutionary mesoscale features of the convective storm are briefly described. The field survey is conducted to evaluate the wind disaster losses in Nanchang. Results show that the severe convective weather in Nanchang on March 30, 2024, was a typical squall line event. The reconstruction method based on PINN effectively compensates for missing radar wind data. The reconstruction errors approximately follow a Weibull distribution, with most significant errors occurring near the boundaries of the missing data regions, which has a limited impact on analyzing mesoscale features of convective storm.