The insitu monitoring data of surrounding rock displacements reflect the changing of mechanical situation of a cavern. In order to overcome the excessive learning of ANN, a new method, ACA-LSSVM , is presented to forecast the nonlinear displacements of surrounding rock. An ant colony algorithm is used to choose parameters of support vector machine. It can escape from the blindness of manmade choice and enhances the efficiency and the capability of forecasting. The method can forecast in rolling the surrounding rock displacements on the basis of monitoring data, in order to discover abnormal situation in time, adjust the supporting schemes dynamically and ensure the stability of surrounding rock of the cavern. The engineering case studies indicate that it is scientific and there is an extensive prospect for this real time forecasting.