The two-stage stochastic program with proba-bilistic constraints is studied by assuming normally distributed uncertain parameters. At stage I, the decision, represented by discrete variables, is made such that normal operations can be applied at stage II with high probability. At stage II, the uncertainty is realized and the optimal decision, represented by continuous variables, is determined accordingly. Different from the conventional two-stage decision process, recovery operations are introduced at stage II if normal operations are infeasible. This two-stage decision process can be approximately modeled by a scenario-based mixed-integer linear program (MILP). Two issues of this scheme are concerned. First, with more samples, the solution of the scenario-based MILP is closer to the original problem, but the computational demands increase significantly. Second, the feasibility of such a solution under unseen scenarios cannot be guaranteed. To overcome these two limitations, an improved Benders decomposition method is first proposed to calculate the globally optimal solution of the scenario-based MILP more efficiently than existing algorithms. Second, the probabilistic feasibility of a stage I solution from the improved Benders decomposition is rigorously evaluated through a chance constrained program with linear decision rule. A refinery optimization problem is solved through the proposed scheme to verify the computational efficiency and probabilistic feasibility.