To implement computer control of a distribution network, it is desirable to have some knowledge of the demand variations a number of hours ahead, at a number of points of the network. A prediction model which facilitates fast computation and requires modest storage and little updating is desirable. Such a model based on a potential-function algorithm has been developed, and its performance with various load data has been analysed and compared with a selected number of other prediction methods. The accuracy of prediction can be increased at the expense of storage space and computing effort.