As the mainstream technology of modern weather forecast, numerical weather prediction (NWP) has been developing in the direction of refinement in recent years, yet the prediction error is still unavoidable. Therefore, it is of great significance to improve the accuracy of numerical weather forecast by revising the results. A traditional method of prediction correction, i.e., the Anomaly Numeral-correction with Observations (ANO), is used to correct the forecast based on statistics of historical data. Results indicate that this method has a good effect. As an emerging method, deep learning has been gradually applied to the field of meteorology in recent years, and has achieved significant results in precipitation prediction and cloud image recognition. Domestic scholars in China used CU-Net, a deep learning model to correct the deviations of the model grid point forecast data of 2 m temperature, 2 m relative humidity and 10 m wind respectively from the European Centre for Medium-Range Weather Forecast (ECMWF), which significantly improved the forecast compared with the ANO method. Based on the above tests, this paper uses dense convolutional structure network model to improve the CU-Net model and forms a new deviation correction model for NWP, which is named as Dense-CUnet, and further develops a deviation correction model named Fuse-CUnet to integrates multiple meteorological elements from NWP and topographic features. Deviation correction tests and comparative analysis of these different models have been carried out. Root mean square error (RMSE) and mean absolute error (MAE) are used as the scoring metrics. By comparing with the original prediction results of ECMWF and the results revised by the ANO and CU-Net methods, it is found that the dense-convolution structure network model Dense-CUnet can be used to effectively modify the positive effect. Moreover, the Fuse-CUnet model that integrates multiple elements can greatly improve the revision effect.

The rapid development of economy and culture in Beijing-Tianjin-Hebei region has a higher requirement for instantaneous strong wind forecasts. Correctly estimating and predicting instantaneous strong winds on the ground level in winter, especially accurate high-resolution grid-point forecasting of gusts under complex terrain condition, is of great significance for improving the service for major Winter Olympics events, the safe operation in the capital and surrounding cities, and disaster prevention and mitigation capabilities. This study establishes a relationship between the gust coefficient and wind speed, wind direction and terrain height based on long-term series of observation data in Beijing-Tianjin-Hebei. Combined with objective statistical analysis method, gust observation data fusion technology and grid point deviation correction technology, an objective gust forecast method is developed, which not only retains the model physical parameters and local climate characteristics, but also utilizes the grid point deviation correction technology. The results of batch verification and case analysis during the Winter Olympics show that the average absolute errors in the Zhangjiakou competition area and the Yanqing competition area are below 2.3 m/s and 3.0 m/s, respectively. The forecast score of gust wind speed above level 8 in the Yanqing competition area is above 0.5. It solves the bottleneck problem of large gust prediction errors and meets the on-site service requirements of major Winter Olympic activities.

Thunderstorm gusts are a common and hazardous type of severe convective weather, characterized by a small spatial scale, short duration, and significant destructive power. They often lead to severe disasters, highlighting the critical importance of their accurate forecasting. Previous studies have explored the environmental factors and spatiotemporal distribution characteristics of thunderstorm gusts, highlighting the need for improved forecasting methods. In recent years, artificial intelligence techniques have shown promise in enhancing the accuracy of thunderstorm gust forecasting, with various machine learning algorithms and models having been developed. This paper proposes a multiscale feature fusion module called Thunderstorm Gusts Block (TG-Block) and a deep learning model named Thunderstorm Gusts net (TG-net) based on the Attention U-net and TG-TransUnet models, and employs interpretable methods such as Integrated Gradient, Deep Learning Importance Features, and Shapley Additive exPlanations to validate the model’s practical relevance and reliability. The analysis of feature importance underscores the model’s ability to capture key thermodynamic and multiscale weather characteristic information for thunderstorm gust nowcasting. It is, however, worth emphasizing that these conclusions are only based on a limited number of thunderstorm gust examples, and the evaluation results may be affected by specific weather types and sample sizes. Nonetheless, TG-net has been put into real-time operation at the Institute of Urban Meteorology, and we will continue to rigorously validate its performance and make any necessary optimizations and enhancements based on feedback to ensure the robustness and stability of the model.

Accurate and fine-scale short-term precipitation forecasting is crucial for disaster prevention, mitigation, and socioeconomic development. Currently, the direct precipitation forecasts of numerical weather prediction often face great challenges and correction methods are still needed to further improve the forecast accuracy. By utilizing the 500-m resolution fusion precipitation data from the Rapid-refresh Integrated Seamless Ensemble (RISE) system in the Beijing–Tianjin–Hebei (BTH) region, this study proposes a new Segmented Classification and Regression machine learning model based on the extreme gradient boosting (XGBoost) algorithm, termed SCR-XGBoost, which can be applied to correct hourly precipitation forecasts in areas with a dense network of weather stations at lead times of 4–6 h. The performance of the model is evaluated according to six metrics: the accuracy (AC), mean absolute error (MAE), root mean square error (RMSE), correlation coefficient (CC), threat score (TS), and bias score (BS). The results indicate that, although the XGBoost algorithm is almost ineffective for directly forecasting precipitation, the SCR-XGBoost model can significantly improve the forecast performance compared with the original RISE forecast, and the segmented correction method for torrential rainfall (≥ 20 mm h⁻¹) outperforms other precipitation grades, which can effectively alleviate the problem of false alarms in the RISE system for heavy rainfall and above (≥ 10 mm h⁻¹) . The optimization rates after applying the SCR-XGBoost model correction in precipitation forecasts can be improved by 6.49%–23.21% in terms of RMSE and MAE reduction, and the CC and AC can be greatly improved by 35.38%–84.39%. Therefore, the SCR-XGBoost algorithm, which introduces precipitation grade classification and multi-layer piecewise machine learning corrections, can significantly improve the 4–6-h precipitation forecast skill, especially for heavy rainfall. The results of this study not only provide new insights for machine learning-based precipitation forecasting, but also help improve rainfall forecasts and the level of disaster prevention and reduction in the BTH region.