Evaluation of Artificial Neural Networks Performance Using Various Normalization Methods for Water Demand Forecasting

Water is one of the crucial resources that must be fed for domestic use as well as for agriculture and economic activities. Since Saudi Arabia is located in a desert region where the seawater desalination process is expensive, there is an urgent need to forecast efficiently the future of water demand. In this paper, the Artificial Neural Networks (ANNs), which are supervised machine learning algorithms inspired by the biological neurons in the human brain are used. ANNs performance is affected by the quality of the input dataset. In fact, time series data must be pre-processed to make it suitable for use in a machine learning framework such as ANNs. The task of converting raw data into a dataset is called Feature Engineering (FE). There are many techniques used for FE. The widely used one in the case of time-series is normalization. This study aims to evaluate ANNs performance using various normalization methods (min-max, z-score, decimal, median and Median Absolute Deviation (MAD)) for short-term water demand forecasting in Jeddah city, Saudi Arabia. Results show that the minimum-maximum normalization method produced the best overall performance in terms of root mean square error.