Comparison of Water Consumption Forecasting Methods Including Artificial Neural Networks

Multiple linear regression, time series methods and artificial neural networks were investigated to determine their value in forecasting per capita water consumption for the Northern Adelaide Plains, S.A. Over fourteen years of monthly data was used to calibrate and test models, including variables such as month of the year, rainfall, evaporation, number of wet days per month, and real marginal price of water. Methods of model preparation, artificial neural networks and factors affecting urban water demand are discussed. The study has shown that artificial neural networks have modelling capabilities superior to regression and time series methods. Seasonal patterns have been predicted equally well by both simple and statistically sophisticated methods. However, neural networks demonstrated the ability to provide improved forecasts with fewer variables.