Lightning Warning Based on Extrapolation of Radar Composite Reflectivity Data

Lightning activity significantly impacts power system stability. In recent years, proactive lightning prediction and prevention have emerged as a novel approach to lightning protection. Addressing the demand for lightning activity early warning, this study constructs an improved CL-Radar deep learning model based on ConvLSTM to achieve radar data extrapolation and prediction. The model introduces a spatiotemporal memory flow mechanism and a data reverse input strategy to mitigate long-term feature loss and enhance generalization capability. Using Doppler radar data from Fujian Province (March–September 2024) with a temporal resolution of 6 minutes and a spatial resolution of 1 km, we constructed a dataset and trained the model. The results indicate that evaluation metrics stabilize after 80,000 iterations, achieving an average Probability of Detection (POD) of 0.862, a False Alarm Rate (FAR) of 0.209, and a Critical Success Index (CSI) of 0.702 at a 30 dBZ threshold. Furthermore, this study conducts a comparative analysis of lightning processes in the target region, verifying the spatial consistency between strong radar echo areas and lightning activity occurrence zones.