融合随机森林和SHAP方法的灌区用水调度经验分析--以淠史杭灌区瓦西干渠灌域为例

Water Allocations in Irrigation Districts Determined by Random Forest Combined with SHAP Method

【目的】定量表征灌区积累丰富的用水调度经验,使其能够被其他管理人员复制和应用。【方法】本文基于淠史杭灌区瓦西干渠灌域的3个典型年实测数据样本,在充分考虑温度、降雨和土壤墒情等特征变量的空间变异基础上,通过融合随机森林模型和SHAP方法,构建有限数据样本下灌区用水调度目标流量与各特征变量之间的非线性定量表征。【结果】应用该方法可得到长时间序列以及不同典型年情境下各特征变量的重要性得分及变化情况,在找到适用于实际用水调度特征变量组合的同时,可分析得到不同调度情景下主要参考的特征变量指标;结合SHAP值正负情况分析,还可得到用水调度目标流量对各特征变量响应的正负方向。【结论】本文所用方法实现了灌区用水调度历史经验的定量知识化表征,为理性预测未来不同用水调度流量提供科学依据。

【Objective】Decades of management of various irrigation districts have accumulated valuable experiences,but how to make most of these experiences has not yet been appreciated.The purpose of this paper is to fill this gap,proposing a method to harness these experiences to improve water allocation calculation in irrigation districts.【Method】The analysis is based on data measured over a three-year period from an irrigation area at the Waxi branch in the Pishihang irrigation district.We consider spatial variation in variables such as temperature,rainfall and soil moisture in integrating the interpretation method of the random forest model with the SHAP model.Nonlinear quantitative relationship between the target water allocation and each variable was constructed using limited data samples.【Result】The model obtains the important scores and changes in each variable in a long time series for different typical years.In addition to the combination of characteristic variables suitable for actual water dispatching,the indexes of characteristic variables that dispatchers mainly refer to different dispatching scenarios can be analyzed and obtained.Combined with the analysis of SHAP value,the positive and negative directions of the response of water dispatching target flow to each characteristic variable is obtained.【Conclusion】The method proposed has realized quantitative representation of historical experience of water dispatching in irrigated areas and provided guidelines for rational prediction of different water dispatching flows in irrigation districts.