A Hybrid Backpropagation Network - Statistical Model for Water Demand Time-series Forecasting

This paper investigates a range of statistical, neural network and hybrid approaches for making one-step-ahead forecasts of a monthly water demand time-series on the basis of 108 historical data points. A uni-variate approach, using solely the water demand time-series, is taken to construct two stand-alone forecasting models: a backpropagation network and a statistical model. A bi-variate approach, in terms of noise ltering, data pre-processing based on the underlying components of the water demand series and use of an innuential external parameter, is taken to build a hybrid statistical and backpropagation network model. We benchmark the results of the models against the forecasting method currently used at Melbourne Water. It is found that when there is strong prior knowledge about the problem domain, neural networks can be used to improve traditional statistical forecasts even when faced with a relatively small amount of historical data.