基于深度学习和综合评价算法的城市极端降水灾害预警预测

Urban Extreme Precipitation Disaster Early Warning and Prediction Based on Deep Learning and Comprehensive Evaluation Algorithms

摘要: 针对数值模式在短临尺度局地极端降水预报中的漏报风险，以及固定阈值方法难以刻画多气象要素非线性耦合的问题，本文提出一种融合多尺度特征与特征词元化Transformer (FT-Transformer)的极端降水事件风险评估模型。该模型通过特征词元化与自注意力机制自动捕捉降水强度、累积量与温湿压等背景场的高阶交互，构建分类与危害指数回归相结合的多任务学习框架，利用回归任务正则化缓解极端样本稀缺带来的过拟合问题，并采用保序回归(Isotonic Regression)校准输出概率。基于郑州区域2015~2021年共25,465个降水事件的实验表明：模型在2021年测试集(3808个事件，含郑州“7∙20”特大暴雨)上取得F1 = 0.9573、AUC = 0.9998、高危事件召回率98.25%，显著优于随机森林(F1 = 0.9148)和逻辑回归(F1 = 0.9422)等基线方法；概率校准使平均绝对误差下降约35%。对“7∙20”特大暴雨的回顾性分析显示，模型能准确识别该极端事件并给出合理的风险刻度，验证了方法的有效性。

Abstract: Aiming at the missed-alarm risk of numerical weather prediction (NWP) models in short-term local extreme precipitation forecasting and the difficulty of fixed-threshold methods in characterizing nonlinear coupling among multiple meteorological factors, this paper proposes an extreme precipitation event risk assessment model integrating multi-scale features with Feature-Tokenizer Transformer (FT-Transformer). The model automatically captures high-order interactions among precipitation intensity, accumulation, and background fields (temperature, humidity, pressure) through feature tokenization and self-attention mechanism. A multi-task learning framework combining classification and hazard index regression is constructed, where the regression task serves as regularization to alleviate overfitting caused by extreme sample scarcity, and Isotonic Regression is employed to calibrate output probabilities. Experiments on 25,465 precipitation events from Zhengzhou area during 2015~2021 demonstrate that the model achieves F1 = 0.9573, AUC = 0.9998, and high-risk recall of 98.25% on the 2021 test set (3,808 events including the “July 20th” extreme rainstorm), significantly outperforming Random Forest (F1 = 0.9148) and Logistic Regression (F1 = 0.9422). Probability calibration reduces the mean absolute error by approximately 35%. Retrospective analysis of the “July 20th” extreme rainstorm shows that the model accurately identifies this extreme event and provides reasonable risk scores, validating the effectiveness of the proposed method.