This paper presents a procedure for a stochastic dynamic programming model conserving the performance of stochastic dynamic programming but requiring rather less computation time than the classical procedure. The approach is based upon the considerations that forecasting beyond a certain time has little effect and that this effect can be estimated approximately. It means that the forecasting horizon can be limited to less time periods than those usual in dynamic programming. An application of the method is presented for a single multipurpose reservoir. The results for autoregressive flow (lag 1) show that the forecasting horizon can be restricted to four weeks for some time periods such that real-time operating problems can be solved with a micro-computer.