Abstract: Uncertainties exist in the geological parameters, prediction model, and input data of the curtain grouting cement predictions; the traditional point prediction suffers considerable errors and lacks a capability of quantifying the uncertainty level. This study describes a Bootstrap method and an improved grey wolf support vector machine (Bootstrap-IGWO-SVM), and develops an interval prediction model of curtain grouting cement with quantification of its uncertainty. First, a new data set is created from the initial training set sampling using the Bootstrap method; the grey wolf method is used to optimize the penalty factor C, RBF kernel function variance g, and loss factor p so as to improve prediction accuracy. Then, we improve the grey wolf method using the nonlinear convergence factor, dynamic weight factor, probabilistic chaos map, and Levy flight, so that the imbalance between local and global searchings can be eliminated. Finally, we estimate the system error and random error using the IGWO-SVM method and the random forest method respectively, and sum up them as the total error, thereby obtaining a cement interval prediction through building a normal distribution model. Thus, the uncertainty of prediction model is quantified. The results show the prediction accuracy of this improved IGWO-SVM is RMSE = 85.3, R2 = 0.53, and MAE = 45.6, a significant improvement on the GWO-SVM’s values of 96.6, 0.40 and 48.5. It is also superior in prediction accuracy to the back propagation neural network (BPNN) or extreme learning machine (ELM). At the confidence level of 99%, its prediction interval coverage probability (PICP), mean prediction interval width (MPIW), and coverage width-based criterion (CWC) are 98.7%, 363.6 kg/m, and 363.6 kg/m respectively.

Key words: grouting cement prediction, improved grey wolf optimization algorithm, support vector machine, interval prediction, curtain grouting