摘要
传统的机器学习模型主要围绕如何提升模型预测精度进行研究,从而忽略了预测结果的可解释性.本研究基于LightGBM (Light Gradient Boosting Machine)建立了预测掺气设施空腔积水深度的黑箱模型,与常用的机器学习模型如RF (Random Forest)、SVM (Support Vector Machine)及XGBoost (Extreme Gradient Boosting)对比结果表明LightGBM拥有较高的预测精度.进一步通过贝叶斯优化技术对LightGBM的四个超参数进行优化,较大程度上提升了模型的R^(2)(决定系数)得分.应用SHAP (Shapley Additive Explanation)事后解释方法对LightGBM的预测结果进行全局解释和局部解释.全局解释结果表明:流速、水舌冲击角、坎高及流量是影响空腔积水形成的主要因素,特征交互解释可以用来解释特征之间的复杂非线性关系,局部解释则可以显示单个样本各特征的影响大小.研究建立的基于LightGBM-SHAP的空腔积水深度可解释性机器学习模型在掺气设施体型优化及模型试验方案优化方面有很好的应用前景.
Traditional machine learning models mainly focus on improving the model's prediction accuracy,thus ignoring the interpretability of the prediction results.In this study,a black box model based on LightGBM(Light Gradient Boosting Machine)was established to predict the depth of water accumulation in the cavity of aeration facilities.Compared with commonly used machine learning models such as RF(Random Forest),SVM(Support Vector Machine),and XGBoost(Extreme Gradient Boosting),LightGBM has higher prediction accuracy.The four hyper-parameters of LightGBM are optimized by Bayesian optimization technology,which greatly improves the R^(2)(determinant coefficient)score of the model.The SHAP(Shapley Additive Explanation)post-interpretation method is applied to explain the prediction results of LightGBM globally and locally.The global interpretation indicates that the flow velocity,the impact angle of the water tongue,the height of the ridge,and the flow rate are the main factors affecting the formation of cavity water accumulation.The feature interaction interpretation can be used to explain the complex nonlinear relationship between the features,and the local interpretation can show the influence of each feature of a single sample.The interpretable machine learning model of cavity water depth based on LightGBM-SHAP has a good application prospect in the shape optimization of aeration facilities and the optimization of the model experiment scheme.
作者
李珊珊
孙朝阳
李国栋
LI Shanshan;SUN Zhaoyang;LI Guodong(State Key Laboratory of Eco-hydraulics in North Arid Region,Xi'an University of Technology,Xi'an 710048,Shaanxi,China)
出处
《力学季刊》
CAS
CSCD
北大核心
2024年第2期442-453,共12页
Chinese Quarterly of Mechanics
基金
国家自然科学基金(52079107,52309105)
陕西省科技厅自然科学基础研究计划(2023-JC-QN-0395)
陕西省教育厅一般专项科研计划(22JK0470)。
