| 余红玲,王晓玲,任炳昱,郑鸣蔚,吴国华,朱开渲.土石坝渗流性态分析的IAO-XGBoost集成学习模型与 预测结果解释[J].水利学报,2023,54(10):1195-1209 |
| 土石坝渗流性态分析的IAO-XGBoost集成学习模型与 预测结果解释 |
| IAO-XGBoost ensemble learning model for seepage behavior analysis of earth-rock dam and interpretation of prediction results |
| 投稿时间：2023-04-22 |
| DOI：10.13243/j.cnki.slxb.20230240 |
| 中文关键词: 土石坝 渗流性态分析 XGBoost 可解释性 SHAP理论 改进的天鹰优化算法 |
| 英文关键词: earth-rock dam seepage behavior analysis XGBoost interpretability SHAP theory improved Aquila optimization algorithm |
| 基金项目:国家自然科学基金雅砻江联合基金项目(U1965207) |
|
| 摘要点击次数: 1439 |
| 全文下载次数: 1151 |
| 中文摘要: |
| 针对现有土石坝渗流数值模拟方法计算效率较低、难以实时分析大坝渗流性态，而现有基于机器学习算法建立的代理模型又存在模型可解释性较差的问题，提出土石坝渗流性态分析的IAO-XGBoost集成学习模型，并基于Shapley加性解释(SHapley Additive exPlanation，SHAP)理论对预测结果进行解释。在采用多地质体自动建模方法和CFD技术对大坝渗流场进行计算分析的基础上，基于改进的天鹰(Improved Aquila Optimization，IAO)算法优化极限梯度提升(eXtreme Gradient Boosting，XGBoost)集成学习算法中的n_estimators、max_depth和learning_rate等超参数，进而建立基于IAO-XGBoost集成学习算法的大坝渗流性态指标预测模型，以揭示上下游水位和坝基地层渗透系数等输入特征变量与渗流性态指标模拟值间的复杂非线性映射关系。进一步地，将IAO-XGBoost集成学习算法与可解释机器学习框架SHAP理论相结合，挖掘影响大坝渗流性态指标预测结果的关键特征，并解释特征变量对渗流性态指标预测的影响。案例研究表明，IAO-XGBoost具有较高的预测精度，相比于IAO-GBDT、IAO-RF、IAO-DT和IAO-SVR算法，其预测精度分别提高了0.52%、11.64%、37.21%和25.07%;且相比于IAO-XGBoost、IAO-GBDT和IAO-RF算法的特征重要性分析方法，SHAP理论具有更强的模型可解释性，提高了预测结果的可信度。 |
| 英文摘要: |
| In view of the problems of low computational efficiency and difficulty in real-time analysis of dam seepage behavior in the existing seepage numerical simulation methods of earth-rock dam，and the problems of poor interpretability of the existing surrogate model based on machine learning algorithm，an IAO-XGBoost ensemble learning model for seepage behavior analysis of earth-rock dam is proposed，and the predicted results are explained based on the SHapley Additive exPlanation (SHAP) theory.On the basis of using multi-geological body automatic modeling method and CFD technology to calculate and analyze the seepage field of the dam，the hyper-parameters such as n_estimators，max_depth and learning_rate of the eXtreme Gradient Boosting (XGBoost) ensemble learning algorithm were optimized by the Improved Aquila Optimization (IAO) algorithm.Then a prediction model of seepage behavior index of dam based on IAO-XGBoost ensemble learning algorithm was established to reveal the complex nonlinear mapping relationship between the input characteristic variables such as the upstream and downstream water level and permeability coefficients of dam foundation and the simulated value of seepage behavior index.Furthermore，the IAO-XGBoost ensemble learning algorithm was combined with the explainable machine learning framework SHAP theory to excavate the key features affecting the prediction results of the dam seepage behavior index，and explain the influence of the characteristic variables on the prediction results.The case study shows that the IAO-XGBoost ensemble learning algorithm has high prediction accuracy.Compared with IAO-GBDT，IAO-RF，IAO-DT and IAO-SVR algorithms，the prediction accuracy of IAO-XGBoost ensemble learning algorithm increases by 0.52%，11.64%，37.21% and 25.07%，respectively.Compared with the feature importance analysis methods of IAO-XGBoost，IAO-GBDT and IAO-RF algorithms，SHAP theory has stronger model interpretability and improve the reliability of the prediction results. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |