Based on C-band weather radar and ground precipitation data from the Helan Mountain area in Yinchuan between 2017 to 2020, we evaluated the forecasting performances of 15 mainstream deep learning models used in recent years, including recurrent-based and recurrent-free models. The critical success index (CSI), probability of detection (POD), false alarm rate (FAR), mean square error (MSE), mean absolute error (MAE), and learned perceptual image patch similarity (LPIPS), were used to evaluate the forecasting abilities. The results showed that (1) recurrent-free models have significant parameter quantity and computing power advantages, especially the SimVP model. Among the recurrent-based models, PredRNN and PredRNN++ demonstrate good predictive capabilities for changes in echolocation and intensity, PredRNN++ performs better in predicting long sequences (1 h); (2) SimVP uses Inception to extract temporal features, which cannot capture the complex physical changes in radar echo images and fails to extract spatial–temporal correlations and accurately predict heavy rainfall areas effectively. Therefore, we constructed the SimVP-GMA model, replacing the temporal prediction module in SimVP and modifying the spatial encoder part. Compared with SimVP, the MSE and LPIPS indicators were improved by 0.55 and 0.0193, respectively. It can be seen from the forecast images that the forecast details have been significantly improved, especially in the forecasting of heavy rainfall weather.

Hail disaster prediction is a pivotal research topic within the field of meteorology, carrying substantial importance for the prevention and mitigation of disasters. The complex terrain and frequent occurrence of hail in Qinghai Province present challenges for traditional prediction methods, which are hindered by issues such as sample imbalance and geographic complexity, resulting in less than ideal outcomes. This study, based on ERA5 reanalysis data and ground-based observation records, proposes a LightGBM hail prediction model that integrates Bayesian optimization with a dual-output header (DOH) structure. First, the DOH structure is developed to separately optimize positive and negative samples, effectively mitigating the sample imbalance problem. Additionally, a Bayesian optimization strategy is applied to conduct global hyperparameter tuning, thereby improving model performance. Experimental results show that, compared to mainstream single and ensemble classification methods, the proposed model achieves superior accuracy, precision, and recall. In the test set, the model demonstrated a prediction accuracy of 0.97, a recall of 0.939, a precision of 0.966, a critical success index (CSI) of 0.909, and a false alarm rate (FAR) of only 0.015. This research provides an effective technical solution for hail prediction in the Qinghai region, offering significant practical value for enhancing early warning capabilities.