This paper introduces a system designed to improve the accuracy of short term water demand forecasts by combining a proven mathematical prediction model with a knowledge base of information relating to non-cyclic abnormal demand occurrences. In part one of the paper, a prototype system is described that comprises a mathematical prediction module and a simplified knowledge base of FORTRAN rules. The results derived from the testing of the prototype show that the methodology is capable of providing significant improvements in prediction accuracy when the normal cyclic demand pattern is disrupted. Part two of the paper describes the steps taken towards implementing the elements that make up a full combined forecasting system in the light of the knowledge gained from building and testing the prototype.