Abstract: Based on the collection of samples of tunneling projects, the mapping relation of the project factors including dimension, submergence and surrounding rock quality to the maximal deformation of surrounding rock is regressed using the least squares support vector machines and a forecast model of maximal deformation of surrounding rocks is built. The kernel function parameters and penalty factors are optimized according to the learning samples. Compared with testing samples, the forecast has a maximal relative error of 14.4%, which indicates that the method is effective.