基于可解释机器学习的黄河源区径流分期组合预报

Combined Forecasting of Streamflow in the Source Region of theYellow River Based on Interpretable Machine Learning

黄河源区是黄河流域重要的产流区和我国重要的清洁能源基地,提高黄河源区径流预报准确率可为流域水资源科学调配和水风光清洁能源高效利用提供重要支撑。以黄河源区唐乃亥和玛曲水文站为研究对象,基于不同月份径流组分的差异,考虑积雪覆盖率及融雪水当量变化,构建了中长期径流分期组合机器学习预报模型及其可解释性分析框架。研究结果表明:1)年内的径流预报时段可划分为融雪影响期(3—6月)和非融雪主导(以降雨和地下水补给为主)期(7月—次年2月);2)与传统不分期模型相比,唐乃亥站和玛曲站分期组合预报模型的纳什效率系数分别达0.897、0.835,确定系数(R2)分别达0.897、0.839,均方根误差分别降低了10%、17%,提高了径流预报准确率,通过分位数映射校正,唐乃亥站和玛曲站预报模型的R2分别进一步提升至0.926和0.850;3)基于SHAP机器学习可解释性分析框架,辨识了预报因子对径流预报结果的贡献程度,由高到低依次为降水、前一个月流量、蒸发、气温、相对湿度、融雪水当量等,发现了不同预报因子之间交互作用散点分布具有拖尾式或阶跃式的特征。

The source region of the Yellow River is an important runoff⁃producing area of the Yellow River Basin and an essential clean ener⁃gy base in China.Improving the accuracy of streamflow forecasting in the source region of the Yellow River will provide significant support for the scientific allocation of water resources in the basin and the efficient production of wind and solar clean energy.This article took Tangnaihai and Maqu hydrological stations in the source region of the Yellow River as research objects.Based on the differences in streamflow compo⁃nents in different months,considering the changes in snow coverage and snowmelt water equivalent,a medium and long⁃term streamflow stag⁃ing combined machine learning forecasting model and its interpretable analysis framework were built.The research results show that a)the streamflow forecasting period within the year can be divided into a snowmelt⁃affected period(March to June)and a non⁃snowmelt⁃dominated period(mainly precipitation and groundwater recharge)(July to February of the following year);b)Compared with the traditional non⁃stag⁃ing model,the Nash efficiency coefficients of the staging combined forecasting model reach 0.897 and 0.835 respectively,and the coefficient of determination(R2)reaches 0.897 and 0.839,with a reduction in root mean square error by 10%and 17%,improving the accuracy of streamflow forecasting at Tangnaihai and Maqu stations.Through quantile mapping correction,the R2 of the forecasting models at Tangnaihai and Maqu stations is further improved to 0.926 and 0.850;c)Based on the interpretability analysis framework of SHAP machine learning,the contribution degree of forecasting factors to the runoff forecast results is identified,from high to low,as precipitation,previous month's discharge,evaporation,temperature,relative humidity and snowmelt water equivalent.It is found that the scatter distribution of the interac⁃tion between different forecasting factors has the characteristics of trailing or stepping.