基于SMOTE算法和逻辑回归模型算法的江苏短时强降水潜势预报

短时强降水是导致城市内涝和山洪、滑坡等灾害的重要原因，其突发性强，局地性明显，预报难度大，是强对流天气业务预报的重点和难点之一。本研究利用2011-2018年的江苏省国家气象观测的逐小时降水资料对江苏省短时强降水时空分布特征进行分析，江苏短时强降水频次分布为典型的南多北少，主要降水出现在早晨04-10时和午后15-19时，前半夜出现降水的概率则较低。基于ERA5再分析资料，选取了对于短时强降水有较强判断能力的气象要素，合成少数类过取样算法（SMOTE）和逻辑回归（LR）方法，构建短时强降水的预报模型，利用2019年的，欧洲中期天气预报中心（ECMWF）预报产品基于该模型进行短时强降水的确定性预报和概率预报，并使用同期实况数据进行系统性检验和天气过程检验。结果表明该模型总体性能较好，对短时强降水出现与否有较好的判别能力和预报指示意义。未来24h以内的6h时效预报，TS评分在0.23以上，未来60h以内的6h时效预报TS评分均在0.2以上，但也存在着一定程度的空报和漏报。基于SMOTE+LR短时强降水预报模型对短时强降水的潜势预报具有较好的指示作用，对气象防灾减灾具有重要意义。

The potential forecast for short-term heavy precipitation in Jiangsu Province base on SMOTE and Logistic Regression Combination Algorithm

Short-term heavy precipitation is a type of strong convective weather, which is prone to cause geological disasters such as urban waterlogging and mountain floods, landslides, etc., which is sudden, strong, local, and difficult to predict. It is one of the key and difficult businesses with strong convective weather forecasts. This study uses the hourly precipitation data of Jiangsu Province from 2011 to 2018 to analyze the spatiotemporal distribution features of short-term heavy precipitation in Jiangsu Province. The frequency distribution of short-term heavy precipitation in Jiangsu is more south than north, and the major occurrence period is at 04-10 am and 15-19 pm, the probability of short-term heavy precipitation in the first half of the night is lowest. Based on the ERA5 reanalysis data, meteorological features with strong judgment ability for short-term heavy precipitation were selected, and a small number of oversampling algorithms (SMOTE) and logistic regression (LR) methods were synthesized to construct a short-term heavy precipitation prediction model. Based on this model, the forecast outputs of the European Center for Medium-Term Weather Forecast (ECMWF) perform deterministic and probabilistic forecasts of short-term heavy precipitation based on this model, and use the real-time data of same period for systematic verification and weather process verification. The results show that the overall performance of the model is better, and it has a better ability to discriminate the presence or absence of short-term heavy precipitation. The forecast within 24h lead time will have a TS score above 0.23, and the forecast within 60h lead time will have a TS score above 0.2, but there will be some false warnings and omission. The SMOTE + LR short-term heavy precipitation forecasting model has a good indication for the potential forecast of short-term heavy precipitation, and is of great significance for meteorological disaster prevention and reduction.