基于多因子最近邻抽样回归模型的径流相似性预报

针对传统径流预报精度不高、预见期不足的问题，提出基于降雨、径流相似性的径流预报方法，采用大数据挖掘在历史降雨产流过程中搜索相似过程，预测后期径流最可能的过程线。为了延长径流预报预见期，实时接入降雨预报信息，提出3种径流滚动预报方式，实现了7 d预见期的径流逐日滚动预报；针对流域在涨退水等不同阶段的产汇流特性，建立可根据实时水雨情自适应切换的降雨、径流输入模式，进一步提高径流预报精度。该研究成果在大渡河的应用表明预报效果达到预期:3 d预见期的纳什系数大于0.9，平均相对误差小于10%；7 d预见期的纳什系数大于0.8，平均相对误差小于15%。

Runoff similarity forecast based on multi-factor nearest neighbor bootstrapping regressive model

Focusing on the low accuracy and insufficient foreseen period of traditional runoff forecast, this study proposed a runoff forecast method based on the similarity of rainfall and runoff. Data mining was used to search for the similar historical rainfall and runoff process, and the most likely runoff hydrograph in the later period was predicted. To prolong the runoff foreseen period to seven days, the real-time rainfall forecast information was inputted into the model and three rolling forecast schemes were proposed. The forecast models could be adaptively switched according to real-time rainfall conditions to further improve the forecast accuracy. The application in Dadu River showed that the Nash coefficients of forecasting the third day and seventh day were greater than 0. 9 and 0. 8, and the average relative errors were less than 10% and 15%, respectively. The research is of great significance to improve the forecast accuracy, extend the foreseen period, and promote the management and operation level of the reservoir group.