The reservoir management problem for a hydrothermal power system is well suited to modeling via dynamic programing. In this paper we describe a dual approach which we term constructive dynamic programming (CDP) which has been successfully applied to optimize releases in a stochastic two-reservoir model of the New Zealand power system. That model ignores serial correlations of inflows, though, and hence assumes that current inflow observations do not have any impact on future release decisions. Tests show, however, that better decision rules can be produced by accounting for inflow correlation. Hence we have developed an extension to the standard CDP to explicitly deal with serial correlation of reservoir inflows, and we report on those extensions also. ¿ 1999 American Geophysical Union