基于Stacking集成算法的抛石护岸水毁破坏预测研究

Stacking Integration Algorithm-Based Water Damage Prediction in Riprap Revetment

抛石护岸在顶冲等极端情况下易发生水毁破坏,给人民的生命财产带来威胁。通过水槽试验获取496组样本数据,利用互信息(MI)筛选出6个关键特征属性,并采用支持向量机(SVR)、广义回归神经网络(GRNN)和随机森林(RF)等机器学习算法构建多个预测模型。然后,将这些模型作为基学习器,结合BP神经网络(BPNN)作为元学习器,采用Stacking集成学习方法构建抛石护岸破坏程度预测模型。最后,通过决定系数(R^(2))、均方根误差(R_(RMSE))及平均绝对误差(M_(MAE))等评价指标对模型性能进行评估。结果表明,Stacking模型在抛石护岸破坏高度、长度、范围上的平均R^(2)为0.98、RRMSE为0.02、M_(MAE)为0.03,相较于单一模型(SVR、GRNN、RF),Stacking模型的R_(RMSE)、M_(MAE)皆为最小,R2最高。在抛石护岸水毁破坏程度的预测中,融合的Stacking模型展现出更高的准确性与稳定性。

Riprap revetments are prone to water-induced deterioration during severe circumstances,hence presenting risks to both human life and property.The present work employs flume experiments in order to gather a dataset consisting of 496 samples.In this study,the selection of six essential feature attributes is performed using mutual information(MI).Subsequently,several machine learning methods,including support vector regression(SVR),generalized regression neural network(GRNN),and random forest(RF),are utilized to construct several prediction models.The aforementioned models perform as foundational learners,with a back-propagation neural network(BPNN)operating as a meta-learner.The construction of a prediction model for assessing the degree of damage to riprap revetments is achieved using the Stacking ensemble learning approach.The evaluation of model performance involves the utilization of metrics,such as the coefficient of determination(R^(2)),root mean square error(R_(RMSE)),and mean absolute error(M_(MAE)).The findings indicate that the Stacking model produces an average R^(2)value of 0.98,root mean square error(R_(RMSE))of 0.02,and mean absolute error(M_(MAE))of 0.03 when predicting the height,length,and range of revetment damage.When comparing the performance of the Stacking model to individual models such as support vector regression(SVR),Generalized regression neural network(GRNN),and random forest(RF),it is observed that the Stacking model achieves the lowest root mean squared error(R_(RMSE))and mean absolute error(M_(MAE)),while also achieving the greatest r-squared(R^(2))value.The Stacking model,when integrated,demonstrates enhanced precision and consistency in forecasting the magnitude of water-induced deterioration in riprap revetments.