Single-Point Thunderstorm Forecasting Based on Second-Order Moist Potential Vorticity and Deep Learning

Thunderstorms are the most frequent type of severe convective weather, which pose great threats to buildings, power transmission, communication facilities, and air transportation. Their analysis and forecasting have long been challenges in meteorological operations. Currently, deep learning-based lightning forecasting has a short valid period, mostly relying on satellite imagery, radar echoes, and lightning location data, focusing on very-short-range forecasting. The longest valid period does not exceed 6 h, and the forecasting accuracy is not high. Based on the physical quantities of the ECMWF numerical prediction model and the actual observations of single-point thunderstorms, this paper constructs a single-point thunderstorm forecasting model with a long validity period (>6 h). The study integrates multi-dimensional parameters such as thermal, dynamic, water vapor, and stratification instability, introduces the second-order moist potential vorticity S as a comprehensive predictor, systematically compares the forecasting performance of eight models, such as 1D PreRNN and ConvLSTM, and verifies the actual operational capability of the model through independent cases. The results show that the 1D PreRNN model has the best overall performance in all periods, which can effectively capture the temporal evolution characteristics of meteorological physical quantities and still has stable generalization performance under unbalanced samples. The model performs well in the 1st, 2nd, and 4th periods, and especially still has significant operational reference value in the 4th period with the longest forecasting validity period; only the 3rd period is weakly affected by the small number of samples. The effect of second-order moist potential vorticity has significant time-dependent differences. Its overall improvement effect is limited in short-term forecasting, but it can provide key disturbance signals in the 4th period with the longest forecasting validity period, and the model forecasting performance drops significantly after removal. The original binary cross-entropy loss is most suitable for the unbalanced sample scenario in this study, and weighted losses are prone to overcorrection. The method in this paper can achieve stable and reliable single-point thunderstorm forecasting for more than 6 h, and can provide long-term fixed-point meteorological support for key scenarios such as aerospace and new energy stations.