Abstract: A new method (i.e. a heuristic prediction method) is presented for long time horizon optimal control problems. The key features of this method are that a heuristic term is added to its cost function for each subproblem, and a prediction method is adopted in the high level coordination resulting in reduced amount of computation for the high level problem. For strictly convex problems, it is shown that the decomposed two-level problem is equivalent to the original one. Finally, an approach for finding the heuristic terms is presented for two types of optimal control problems, and a parallel algorithm is given for the heuristic prediction method. In comparison with P. B. Luh's method (Chang, 1990), the heuristic prediction method can reduce the amount of computation for the high level problem and. needs much less high level iterations. Moreover, it is easy to choose the initial values of the coordination variables. Numerical results show its feasibility and its noticeable potential for applications in real-time control problems.