准确预测隧道洞口段沉降变形是保障安全进洞的重点工作,如何解决输入层维度高的问题以及准确描述机器学习预测模型的性能意义重大.因此,将主成分分析法(PCA)、优化算法和支持向量回归机(SVR)相结合,提出6个基于PCA和优化算法的SVR组合预测模型.首先,通过PCA筛选影响拱顶沉降的主要因素;其次,采用遗传算法(GA)、粒子群算法(PSO)和灰狼算法(GWO),对SVR的惩罚因子和核参数进行寻优;最后,将组合预测模型应用到温州市石鼻头隧道,采用相关性系数(R)、均方根误差(RMSE)和平均绝对误差(MAE)对预测模型的性能进行比较评价,并构建模型查询表.结果表明:组合预测模型均有较高精度,R≥0.9870,RMSE≤6.7924 mm,MAE≤3.4937 mm;PCA降维后,GA优化后的SVR预测模型的预测效率提高了65%,PSO和GWO优化后的SVR预测模型减少了输入层维度,但需要更大的k值,降低了预测效率,PCA-GWO-SVR模型尤为明显;PCA-PSO-SVR预测模型的鲁棒性更强.
Precise prediction of settlement and deformation at tunnel portal section is a priority in ensuring safe en-try into tunnel,and it is of great importance to solve the problem of high dimensionality of input layer and accurately describe the performance of machine learning prediction model.Therefore,the principal component analysis(PCA),optimization algorithm and support vector regressor(SVR)are combined to generate 6 combination prediction models based on PCA,optimization algorithm and SVR.First,the main factors that affect crown settlement are identified by using PCA.Next,the optimization algorithms including genetic algorithm(GA),particle swarm optimization(PSO)and grey wolf optimizer(GWO)are used to optimize the penalty factors and kernel parameters of SVR.At last,the combination prediction model is applied to the Shibitou Tunnel in Wenzhou,the comparative evaluation of the per-formance of the prediction model is conducted by using the correlation coefficient(R),root mean square error(RMSE)and mean absolute error(MAE),and the model query table is generated.The results indicate:the combina-tion prediction models have high precision,with R≥0.9870,RMSE≤6.7924 mm and MAE≤3.4937 mm.After dimen-sionality reduction by PCA,the GA optimization allows SVR prediction model to increase its prediction efficiency by 65%,while PSO and GWO optimization allows SVR prediction model to decrease the dimensionality of its input layer but increase its k value,which reduces the prediction efficiency,particularly so in the case of PCA-GWO-SVR.PCA-PSO-SVR prediction model has a better robustness.