基于特征工程的强降水物理要素提取及分析

Using Feature Engineering to Extract Important Physical Parameters for Heavy Rain

摘要: 灾害性强降水事件对人民生命和财产安全具有重大影响，其精准预报是防灾减灾的重要内容，但因其突发性强、变化快和强度大的特点，常用数值模式的预报准确率很低。近年来随着人工智能技术快速发展，融合物理特征和机器学习的强降水预报成为热点研究。本文利用特征工程提取强降水预报关键物理因子，发现随着预报精细度提升，模式直接预报降水的贡献越来越小，而相关物理特征更具参考价值。对6小时以上累计降水，中高层的动力维持作用、整层抬升和整层含水量具有较大贡献；而对于3小时以内降水预报，近地面层的动力辐合、不稳定能量特征、及影响降水效率的云微物理因子重要性更加凸显，这与降水系统的多尺度特征和模式可预报性相对应。通过特征工程精简后的物理特征可有效去除冗余提升机器学习模型效率，为提升强降水的预报准确率提供支撑。

Abstract: Hazardous heavy rain events have significant impacts on people’s lives and property safety. It’s important to make accurate prediction to reduce the loss. However, due to its high intensity and rapid change, the accuracy of operational numerical models is very low. In recent years, with the rapid development of artificial intelligence technology, more and more studies devote on integrating physical understanding and machine learning to improve heavy rain forecast. In this study, the key physical factors of heavy precipitation prediction are extracted by feature engineering methods. It is found that the finer the forecast is, the contribution of forecasted rain is smaller, but the physical features are more important. For the above 6 h rainfall, the dynamics of the middle and high levels, the uplift of the whole layer and the water content of the whole layer contribute greatly; for forecast less than 3 hours, the importance of dynamic uplift at low level, the unstable energy and some microphysical factors are prominent, which is in according to the multi-scale characteristics of the atmosphere. The physical features simplified by feature engineering can effectively remove redundancy and improve the efficiency of machine learning model. The results would help to improve the prediction of heavy rain events.