| 引用本文: | 苏楠,章少辉,白美健,等.融合随机森林和SHAP方法的灌区用水调度经验分析[J].灌溉排水学报,0,():-. |
| sunan,zhangshaohui,baimeijian,et al.融合随机森林和SHAP方法的灌区用水调度经验分析[J].灌溉排水学报,0,():-. |
|
| 摘要: |
| 【目的】为定量表征灌区积累的丰富的用水调度经验极其重要，但迄今无法被定量表征，导致难以使其能够被被其他管理人员复制和应用。解决该问题的难点表现在，影响用水调度目标流量的原因要素或特征变量形成了一个复杂的非线性决策网络、且特征变量存在冗余相关。 【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，来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】即首先采用随机森林模型非线性回归算法，来构建用水调度目标流量与各特征变量之间的非线性映射，并用SHAP方法来优化该非线性映射的具体表征，以剔除不同类特征变量和考虑空间变异时同类特征变量之间的冗余相关，由此获得用水调度目标流量对各特征变量响应的重要程度及排序，提出适用于实际用水调度的特征变量组合；同时利用SHAP得到用水调度目标流量对各特征变量响应的正负方向。 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调度员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】本文所用方法在此过程中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。 |
| 关键词: 淠史杭灌区；用水调度；随机森林模型；SHAP方法；特征变量；非线性表征 |
| DOI： |
| 分类号:S |
| 基金项目:中国水科院专项项目(ID0145B052021) |
|
| Analysis of water regulation experience by combining random forest and SHAP method in irrigation district |
|
sunan, zhangshaohui, baimeijian, zhangbaozhong
|
|
China Institute of Water Resources and Hydropower Research
|
| Abstract: |
| 【Objective】The rich experience of water scheduling accumulated in irrigated areas is extremely important. Whereas, the experience cannot be quantitatively characterized so far, making it difficult to be replicated and applied by other managers. The difficulty of solving this problem is that the characteristic variables which affect the target flow rate of water dispatching constitute a complex nonlinear decision network, and simultaneously there is redundant correlation among the characteristic variables. In order to quantitatively characterize the irrigation area to accumulate rich experience in water management, so that it can be copied and applied by other managers.【Method】To solve this problem, based on the measured data samples of three typical years in the irrigation area of Waxi trunk Canal in Pishihang Irrigation Area, this paper fully considered the spatial variation of characteristic variables such as temperature, rainfall and soil moisture, then deeply integrated the interpretation method of random forest model and SHAP model. Furthermore, the nonlinear quantitative characterization between the target flow of water dispatching and each feature variable was constructed under a limited data sample. 【Result】In other words, firstly the random forest model nonlinear regression algorithm is used to construct the nonlinear mapping between water dispatching target flow and each characteristic variable. Secondly, using SHAP to optimize the representation of the nonlinear mapping, which aims to eliminate the redundant correlation between different class feature variables and class feature variables considering spatial variation. In addition, the importance and ranking of the response of water dispatching target flow to each characteristic variable can be obtained, and the combination of characteristic variables suitable for actual water use scheduling is proposed, at the same time, the positive and negative directions of the response of target flow to each characteristic variable are given. By using this method, the importance scores and changes of each characteristic variable in long time series and different typical year scenarios can be obtained. In addition to finding the combination of characteristic variables suitable for actual water dispatching, the indexes of characteristic variables that dispatchers mainly refer to in 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 can be obtained.【Conclusion】On the whole, the quantitative knowledge representation of historical experience of water dispatching in irrigated areas is realized, which provides scientific basis for rational prediction of different water dispatching flows in the future. In this process, the method used in this paper has realized the quantitative knowledge representation of the historical experience of water dispatching in irrigated areas and provided scientific basis for rational prediction of different water dispatching flows in the future.
Key words: Pishihang Irrigation District; Water dispatching; Random forest model; SHAP method; Characteristic variable; Nonlinear characterization |
| Key words: Pishihang Irrigation District; Water dispatching; Random forest model; SHAP method; Characteristic variable; Nonlinear characterization |
