Shunting scheduling problems at railway stations can be regarded as a sort of resource constrained project scheduling problem (RCPSP). Unlike the normal RCPSPs, however, shunting scheduling problems require that the number of work which consists of a project has to be dynamically changed in the process of solving the problem and some of the work has to be performed at a prescribed timing, for example. We propose an efficient algorithm for shunting scheduling problems combining probabilistic local search and PERT (Program Evoluation and Review Technique). Local search and PERT are combined so that the candidates for answers in the local search process are evaluated by PERT. This enables us to reduce the search space of the local search to a great extent and makes the algorithm work quite fast. We have confirmed the effectiveness of our algorithm through several experiments using actual train schedule data.