樽海鞘算法优化支持向量机的RC柱抗侧移承载力预测

Salp Swarm Algorithm-optimized Support Vector Machines For Lateral Strength Prediction of RC Columns

现有钢筋混凝土(RC)柱抗侧移承载力预测模型缺乏泛化性能,延性柱抗弯承载力的预测模型不能用于非延性柱的抗剪承载力,反之亦然。机器学习(ML)方法能够解决这一问题,但由于无法自动剔除冗余和不相关特征,使得ML模型复杂度高且容易过拟合。为此,提出一种樽海鞘算法优化支持向量机(SSALS-SVM)方法,基于给定的数据集,SSALS-SVM能利用樽海鞘优化算法(SSA)自动剔除冗余和不相关的特征,筛选最具代表性且各特征之间相关性弱的特征子集形成最优特征组合,同时对控制模型非线性拟合能力的超参数进行优化。优化后的模型既能识别出影响延性和非延性RC柱抗侧移承载力的设计变量,又能反映最优特征组合与抗侧移承载力间的非线性映射关系。为了验证SSALS-SVM方法的泛化性能,基于248个RC柱抗侧移承载力试验数据,分别与现有的RC柱抗侧移承载力预测模型进行对比,结果表明,SSALS-SVM比现有预测模型的泛化性能最高提升了83%。

The existing methods of predicting the lateral strength of RC columns lack generalization performance,where methods of predicting the flexural strength of ductile columns cannot be used to predict the shear strength of non-ductile columns,and vice versa.While current machine learning methods can solve this problem,they cannot automatically remove a large number of redundant and irrelevant features from the dataset,which in turn increases the complexity of the ML model and leads to overfitting.To this end,this paper proposes a new method called salp swarm algorithm-optimized least squares support vector machines(SSALS-SVM),which can remedy the aforementioned problems.Based on a given data set,SSALS-SVM can adopt the salp swarm algorithm(SSA)to automatically eliminate redundant and irrelevant features and select the most representative feature subset with weak correlation among features to form an optimal feature combination,while the hyperparameters governing the nonlinear fitting ability of LS-SVM are also optimized.In this way,the optimized prediction model can not only identify the design variables that influence the lateral strength of ductile and non-ductile columns,but also reflect the nonlinear mapping relationship between the optimal feature combination and lateral strength of RC columns.The generalization performance of proposed SSALS-SVM for predicting the lateral strength of RC columns is verified by comparing with existing prediction models based on 248 experimental data of RC columns.Numerical results show that the generalization performance of proposed SSALS-SVM can be enhanced up to 83%higher than that of existing prediction models.