基于可解释性分析的大坝变形监控模型对比研究

Comparison of monitoring model for dam deformation based on interpretability analysis

近年来,经典统计模型和机器学习模型在大坝安全监控领域并行发展,然而前者的“预测能力”和后者的“可解释性”通常存在一定局限,且关于量化多重因素对大坝监测量影响程度的对比研究相对较少。基于闽江支流上GTX重力坝的水平位移和垂直位移原型监测数据,分别采用多元线性回归(MLR)、偏最小二乘回归(PLS)、随机森林算法(RF)建立兼顾预测能力和解释能力的大坝变形监控模型;同时,针对每种模型开展特征重要性分析,探究不同因素对大坝变形的影响程度。研究结果表明:3种模型中随机森林模型的拟合能力最佳,偏最小二乘回归模型的预测能力最佳;3种模型提供的可解释性基本符合实际规律,且特征重要性排序规律定性一致,水压分量和温度分量对该坝体位移影响显著,时效分量所占比例最低。研究成果可为后续开展大坝安全监控模型优选提供参考。

In recent years,classical statistical models and machine learning models have parallelly developed in dam safety monitoring field.However,the predictive ability of the former and the interpretability of the latter usually have certain limitations,and there are relatively few comparative studies on the impact of quantitative multiple factors on dam monitoring measured data.Based on the prototype monitoring data of horizontal displacement and vertical displacement of GTX gravity dam on the tributary of Minjiang River,this paper used multiple linear regression(MLR),partial least squares regression(PLS)and random forest algorithm(RF)to establish different dam deformation monitoring models that takes both predictive ability and interpretability into account.At the same time,the feature importance analysis was carried out for each model to explore the influence of different factors on dam deformation.The results showed that the random forest model had the best fitting ability and the partial least squares regression model had the best prediction ability among the three models.The interpretability provided by the three models was basically in line with the actual law,and the order of feature importance was consistent:the water pressure component and the temperature component had a significant impact on the displacement of the dam body,and the proportion of the aging component was the lowest.The research results can provide reference for the subsequent optimal selection of dam safety monitoring model.