基于机器学习的变电站地下水埋深智能预报预警模型

Intelligent prediction and early warning model for substation groundwater depth based on machine learning

摘要

【目的】设计一套准确且高效的地下水位动态预测模型对于变电站排水系统的智能监控与预警系统的有效应用以及确保变电站安全稳定运行至关重要。【方法】聚焦于220kV园区变电站的试点研究项目，针对极致梯度提升树(XGBoost)、随机森林(RF)和长短期记忆模型(LSTM)三种机器学习模型进行了综合评估，重点分析了三种模型在预测暴雨情景下地下水埋深动态方面的性能表现。模型的训练数据来源于经过校准和验证的地下水流数值模型，并结合多种暴雨情景下的地下水埋深动态预测结果作为基准参考值。为了较好地评估这些模型的预测准确性和可靠性，采用了Nash-Sutcliffe效率系数(NSE)、均方根误差(RMSE)、Pearson相关系数和平均绝对误差(AE)作为评价指标。【结果】研究结果显示，XGBoost、RF和LSTM三种模型均能在时间尺度上模拟出与基准结果相近的地下水埋深动态，NSE、RMSE和Pearson相关系数分别达到0.999 8、0.003 1 m和0.999 9,但在空间上表现差异大，RF模型模拟的AE小于0.01 m, XGBoost模型模拟的AE小于0.26 m, LSTM模型给出的AE小于0.12 m。使用20%网格点的模型数据进行机器学习训练输入，RF模型的性能依然表现最佳，同时模型训练和预测的时间效率提升了5倍。【结论】基于机器学习模型构建的地下水埋深动态预测模型表现良好，在排水系统的智能监控与预警系统中具有良好的应用前景。

Abstract

[Objective] Designing an accurate and efficient groundwater depth dynamics prediction model is crucial for the effective application of intelligent monitoring and early warning systems for substation drainage systems and for ensuring the safe and stable operation of substations. [Methods] Focusing on the pilot study project of the 220kV substation in the industrial park, a comprehensive evaluation was conducted on three machine learning models: Extreme Gradient Boosting(XGBoost), Random Forest(RF), and Long Short-Term Memory(LSTM). The performance of these models in predicting groundwater depth dynamics under heavy rainfall scenarios was analyzed in detail. The training data for the models were derived from a calibrated and validated groundwater flow numerical model, using prediction result of groundwater depth dynamics under various rainfall scenarios as benchmark reference values. To thoroughly assess the prediction accuracy and reliability of these models, the Nash-Sutcliffe Efficiency Coefficient(NSE), Root Mean Square Error(RMSE), Pearson Correlation Coefficient, and Mean Absolute Error(AE) were used as evaluation indicators. [Results] The research result showed that XGBoost, RF, and LSTM models could simulate groundwater depth dynamics consistent with the benchmark result over the time scale, with NSE, RMSE, and Pearson correlation coefficients reaching 0.999 8, 0.003 1 m, and 0.999 9, respectively. However, the spatial performance varied significantly. The AE simulated by the RF model was less than 0.01 m, the AE simulated by the XGBoost model was less than 0.26 m, and the AE given by the LSTM model was less than 0.12 m. When using model data from 20% of the grid points for machine learning training, the RF model still showed the best performance, and the time efficiency of model training and prediction improved by 5 times. [Conclusion] The groundwater depth dynamics prediction model based on machine learning models demonstrates excellent performance and shows promising application prospects in the intelligent monitoring and early warning systems for drainage systems.

关键词

变电站 / 地下水埋深预警系统 / 机器学习 / 智能监控 / 数值模拟

Key words

substation / groundwater depth early warning system / machine learning / intelligent monitoring / numerical simulation

基金资助

清华大学-宁夏银川水联网数字治水联合研究院专项统筹重点项目(SKL-IOW-2023TC2307)

“十四五” 国家重点研发项目(2023YFC3708903)

广东揭阳惠来园区输交电工程项目(0352004823220005)