摘要
针对机器学习模型的过拟合问题和非线性隐式型模型结构,基于贝叶斯正则化对某拱坝径向位移建立神经网络(NN)和支持向量机(SVM)模型,对比分析均方差(M_(MSE))、复相关系数、最大绝对误差、预测M增比等评价指标,并通过部分依赖图(PDP)来挖掘机器学习模型的因果解释能力。研究结果表明,贝叶斯正则化能显著提高机器学习模型的预测和解释能力,NN和SVM的预测M_(MSE)分别降至64.67%、3.85%,过拟合程度分别由70.99、9778.36降至41.22、17.15;基于PDP从机器学习模型中分离出的因果分量与传统多元线性回归模型(MLR)相近。采用贝叶斯正则化构建的SVM模型的解释能力与MLR模型最接近,预测性能亦较优。
Aiming at the over-fitting problem and nonlinear implicit model structure of the machine learning model,neural network(NN)and support vector machine(SVM)models were established for the radial displacement of an arch dam based on the Bayesian regularization.The evaluation indexes of the mean square error(MMSE),complex correlation coefficient,maximum absolute error and prediction MMSE increase ratio were counted to evaluate the prediction performance.Partial dependency graph(PDP)was used to mine the causal interpretation ability of machine learning models.Research results show that the Bayesian regularization can significantly improve the prediction and interpretation ability of the machine learning model.The prediction MMSEs of the NN and SVM are reduced to 64.67%and 3.85%,and the over-fitting degrees are reduced from 70.99 and 9778.36 to 41.22 and 17.15,respectively.Based on the PDP,causal components of the hydraulic,temperature and time effect can be effectively separated from the machine learning model,and the separation result is similar to that of the traditional multiple linear regression(MLR)model.The Bayesian regularization-optimized SVM model has the closest interpretation ability to the MLR model and also has a better prediction performance.
作者
隋旭鹏
朱圣辉
王少伟
徐丛
庄钧惠
SUI Xu-peng;ZHU Sheng-hui;WANG Shao-wei;XU Cong;ZHUANG Jun-hui(School of Environmental and Safety Engineering,Changzhou University,Changzhou 213164,China;Jiangsu Jiangyin Key Water Conservancy Project Construction Management Office,Wuxi 214431,China)
出处
《水电能源科学》
北大核心
2022年第9期120-124,共5页
Water Resources and Power
基金
国家自然科学基金项目(51709021)
中国博士后科学基金项目(2020M670387)
中国水利水电科学研究院流域水循环模拟与调控国家重点实验室开放研究基金(IWHR-SKL-KF202002)
中国水利水电科学研究院水利部水工程建设与安全重点实验室开放研究基金(202009)。
关键词
混凝土坝
位移监控模型
机器学习
贝叶斯正则化
性能提升
concrete dam
displacement monitoring model
machine learning
Bayesian regularization
performance improvement