Abstract:With the continuous increase in the demand for refined meteorological forecasting, high-resolution numerical models make progress in the forecasting capability of extreme precipitation events. However, limited by factors such as errors in the initial field of the model, parameterisation uncertainties, and terrain, there are still systematic bias problems in precipitation forecasting. The convective-scale numerical model in Guangxi (GX(R1)) faces the technical bottleneck of urgently needing to improve the accuracy in short-term strong precipitation forecasting. In response to this issue, this study proposes a deep learning method based on multi-source data fusion, the MResUNet model. This model achieves performance breakthroughs through three technical improvements: first, the squeeze-and-excitation (SE) module is introduced to construct a channel attention mechanism, dynamically adjust feature weights, and suppress noise interference in the model output; second, the atrous spatial pyramid pooling (ASPP) module is integrated to fuse multi-scale features and improve the positioning accuracy of precipitation areas; third, a multi-modal data fusion framework is constructed to integrate the advantages of convective-scale numerical forecasting products in Guangxi, radar echo data, ground observation data, and the ResUNet++ model. The weighted loss function is optimised, and different weights are assigned to different precipitation intensities according to the measured rainfall data at stations, significantly enhancing the model’s sensitivity to extreme precipitation events. To verify the effectiveness of the MResUNet model, four sets of control experiments are designed: Scheme 1 retains the original output of the GX(R1) model as the benchmark; Scheme 2 uses the input containing only numerical model products (X1, X2); Scheme 3 uses only radar echo data (X3); Scheme 4 fuses the model products and radar echo data and implements a multi-source data fusion strategy. The experimental results show that all MResUNet schemes are significantly better than the original model output, and the TS (Threat Score) scores exhibit positive skill characteristics. In particular, for Scheme 4, under the thresholds of strong precipitation of ≥20 mm/h and ≥30 mm/h, the TS scores increase by 9.77% and 8.98%, respectively. It shows significant advantages in reducing the false alarm rate. The hit rate at the precipitation level of ≥30 mm is higher than that of Scheme 3 and much higher than that of Scheme 2, with an increase of 5.45% and 177%, respectively, and the prediction bias is the closest to the ideal value of 1. Further, the extreme precipitation event of “Dragon Boat Water” in Guangxi in 2023 and typical typhoon precipitation processes are selected for case verification. The analysis shows that the MResUNet effectively solves the problem of dispersed prediction of the strong precipitation centre in the GX(R1) model and demonstrates excellent prediction capabilities in both the magnitude of precipitation intensity and the location of precipitation areas. This study proves that by deeply integrating multi-source observational data and numerical model products and optimising the architecture of the deep learning model, the forecasting accuracy of different precipitation magnitudes can be significantly improved. In particular, the scheme based on multi-source data fusion shows obvious advantages in strong precipitation forecasting. The research results provide a new technical approach to solve the technical problems in short-term precipitation forecasting.
