WATER DEMAND FORECASTING BY MEMORY-BASED LEARNING

Recently, a ''memory based'' approach towards various kinds of problem s has been proposed. The underlying principles of the memory based app roach are : (1) storing past examples in a memory. (2) searching ''nea r'' examples to a given input in a memory. In this paper, we apply the memory based approach to water demand forecasting, and present a hybr id method which consists of MBL (Memory Based Learning) and the conven tional multiregression. In the memory based method, the distance metri c is crucial. In our method, the local distance metric is defined from examples when an input data is given then the neighborhood of the inp ut is determined based on the distance metric. If there exist examples within the ''neighborhood'' of the input data, then the forecast is g iven by MBL. Otherwise, local multiregression is used. We applied this method to the daily water demand forecasting. The forecasting results by our method are approximately 5% better than those by the multiregr ession method. Especially, when there exist past examples in the neigh borhood, namely in the case where MBL is applicable, the results are a pproximately 10-30% better than those by the conventional multiregress ion.