水利学报2026，Vol.57Issue(4)：637-650,14.DOI:10.3724/j.slxb.20250414
基于分期自相关机器学习方法的月径流模拟与驱动机制分析
Monthly streamflow simulation and driving mechanism analysis using periodization-based autocorrelation and machine learning methods
摘要
Abstract
To improve monthly runoff prediction accuracy,this study proposes a staged autocorrelation machine learning method that incorporates the dynamic periodicity of runoff autocorrelation.Autocorrelation functions were first used to identify hydrological periods and optimal lag orders.Model combinations were then optimized using the TOPSIS method,and SHAP analysis was applied to interpret key driving factors.Applied to the Yingluoxia Station in the Heihe River Basin,China,the method yielded the following results:(1)The year was divided into three distinct periods:dry period(October-February),transition period(March-June),and high-flow period(July-September),with optimal lag orders of 2,3,and 5 months,respectively;(2)The optimal model(AP-RF-LSTM-RF)achieved an NSE of 0.91,increasing the TOPSIS score by 68.2％and improving extreme runoff prediction by 67.6％;(3)SHAP results showed that dominant factors varied by period:temperature and 2-month-lagged runoff in the dry period,precipitation and 3-month-lagged runoff in the transition period,and both precipitation and temperature in the high-flow period.The proposed method effectively enhances model interpretability and prediction accuracy by aligning machine learning with hydrological processes.关键词
月径流预测/自相关性/分时期/可解释机器学习/黑河莺落峡Key words
monthly runoff prediction/autocorrelation/staged modeling/interpretable machine learning/Yin-gluoxia of the Heihe River Basin分类
天文与地球科学引用本文复制引用
朱富达,周佳灵,张帆,解玉磊,肖月寒..基于分期自相关机器学习方法的月径流模拟与驱动机制分析[J].水利学报,2026,57(4):637-650,14.基金项目
国家自然科学基金项目(52409003) （52409003）
中国科协青年人才托举工程项目(YESS20240294) （YESS20240294）
北京林业大学大学生创新创业训练计划项目(202510022029) （202510022029）
