| 引用本文: | 郭 宁,向 瑶,左岗岗.基于决策树的机井以电折水系数预测通用模型研究[J].灌溉排水学报,2026,45(9):96-104. |
| Guo Ning,Xiang Yao,Zuo Ganggang.基于决策树的机井以电折水系数预测通用模型研究[J].灌溉排水学报,2026,45(9):96-104. |
|
| 摘要: |
| 【目的】农业灌溉机井数量多且分散，实现精准计量需要较高的建设与维护成本，亟须探索低成本、高效率用水计量方式。【方法】通过系统收集不同区域“以电折水”系数及其影响因素，构建了涵盖井龄、机井深度、地下水埋深、水泵型号、灌溉类型、泵龄等指标的多源数据集。利用极端梯度提升（Extreme Gradient Boosting，XGBoost）算法对缺失数据进行链式插补，并构建了基于XGBoost的系数预测模型，同时引入随机森林（Random Forest，RF）和LightGBM模型作为对比模型。通过OPTUNA自动化超参数优化框架结合K折分层交叉验证训练模型，进一步对不同模型的预测性能进行统计显著性分析，并通过引入误差棒的置换重要性方法定量解析关键驱动因子。【结果】①XGBoost链式插补法能够有效捕捉变量间的非线性交互关系，生成的插补数据具有高度的一致性；②XGBoost模型在测试集上的纳什效率系数（NSE）高达0.997，显著优于LightGBM和随机森林模型；③统计检验结果表明，XGBoost的预测精度在统计学上显著优于LightGBM，且在高扬程等极端条件下表现更强的稳定性；④水泵实际抽水扬程、水泵额定扬程和水泵额定流量是影响“以电折水”系数的重要因素。【结论】提出的“插补-优化-预测”方法体系可有效解决数据缺失下的建模难题，验证XGBoost模型在处理非线性灌溉参数中的优势。明确了以“实际抽水扬程”为核心驱动因子，为无计量设施地区的农业用水估算提供了通用的解决方案。 |
| 关键词: 以电折水；机井灌溉；XGBoost模型；缺失数据处理；机器学习 |
| DOI：10.13522/j.cnki.ggps.2025348 |
| 分类号: |
| 基金项目: |
|
| A low-cost method for estimating electricity-to-water conversion coefficients for well-irrigated areas without metering facilities |
|
Guo Ning, Xiang Yao, Zuo Ganggang
|
|
1. Dizhang Subdistrict Office, Xixian New Area, Xi’an 712035, China; 2. State Key Laboratory of Water Engineering
Ecology and Environment in Arid Area, Xi’an University of Technology, Xi’an 710048, China
|
| Abstract: |
| 【Objective】In well-irrigated districts, accurately metering water pumped from individual boreholes is challenging because of their dispersed distribution and high costs of meter installation and maintenance. This paper proposes a low-cost and efficient method for estimating groundwater pumped for irrigation in areas without metering facilities.【Method】Measured electricity-to-water conversion coefficients and factors that affect the conversion coefficients were collected across multiple regions. These factors included well age and depth, groundwater table depth, pump model and age, and irrigation type. Missing data were imputed using an XGBoost-based chained imputation method. Random Forest and LightGBM were used as benchmark model for predicting the conversion coefficient, and an optimized XGBoost model was developed by integrating the Optuna automated hyperparameter optimization framework with stratified K-fold cross-validation. Statistical significance tests were conducted to compare model performance; permutation importance with error bars was used to quantitatively identify the key factors that influence the electricity- to-water conversion coefficient.【Result】①The XGBoost chained imputation method effectively captured nonlinear interactions among variables while maintaining high physical consistency of the imputed data. ②The XGBoost model achieved a Nash-Sutcliffe efficiency (NSE) of up to 0.987 on the test set, significantly outperforming the LightGBM and Random Forest models. ③Statistical tests showed that XGBoost significantly outperformed LightGBM and was more robust than Random Forest under extreme conditions, such as high pumping lift. ④Actual pumping lift, rated pump lift, and rated flow rate were the most influential factors affecting the electricity-to-water conversion coefficient.【Conclusion】The proposed imputation-optimization-prediction framework effectively addresses the challenges associated with missing data and demonstrates the ability of XGBoost to capture nonlinear relationships among irrigation-related variables. Our results show that the actual pumping lift is the most important factor for data collection; they provide a generalizable, low-cost approach for estimating agricultural groundwater use for irrigation in areas without metering facilities. |
| Key words: electricity-to-water conversion; irrigation by motor-pumped wells; XGBoost model; missing data handling; machine learning |
