The excavation and construction of underground engineering is a dynamically adjusting process of system. The paper starts with the index of surrounding rock displacement which can reflect both observability and controllability of underground engineer system. The nonlinear machine learning tool - support vector machine (SVM) which based on statistic learning theory is utilized to construct the time series model. Because penalty factor and kernel parameter of SVM affect the predicting accuracy evidently, and SVM has not provided the selection method, the parameters are optimized by global optimization arithmetic - particle swarm optimization. Based on the PSO-SVM evolutionary predictive model, appending the up-to-date monitoring information, multi-step extrapolating forecast model of surrounding rock displacement is constructed, and according to control criteria, the supporting scheme is adjusted, realizing the predictive control for underground engineer. An engineer sample is studied, the result states that the PSO-SVM model is feasible. The proposed predictive control method provides new approach for underground construction.