Improving Artificial Neural Network Models for Predicting Hourly Water Consumption

Abstract

Smart water meter devices are now widely installed in single family residences, allowing water consumption data to be collected at a high resolution from both the temporal and spatial perspectives. Such data allows improved prediction of future water consumption - an important task for water utilities as they manage the water supply. The dataset in this thesis consists of hourly water consumption data from the 9,045 single-family residences in Abbotsford, British Columbia from September 2012 to August 2013. This research focuses on predicting hourly water consumption by using improved artificial neural network (ANN) models and makes five main contributions. The first contribution is accurately predicting hourly water consumption at a finer spatial and temporal scale than previous work. The second contribution is gathering and studying a wide variety of datasets and related features for predicting future water consumption. In addition to water consumption data, daily weather information, demographic information, property information and date information during the same period of time are collected in the raw dataset. The third contribution is to systematically perform feature selection, an important step in building machine learning models but one that is absent from previous work on predicting water consumption. For different experiment criteria, customized feature sets assist the corresponding models to accurately predict the hourly usages. The fourth contribution is to improve prediction accuracy by building separate models for weekday and weekend prediction. Residents consume water in different patterns between weekdays and weekends. By tackling the predictions separately, better performance can be achieved with less complicated models. Lastly, this research investigates the performance of multi-hidden-layer ANN models versus single-hidden-layer models. Although, single-hidden-layer models are sufficient in theory, we show that multi-hidden-layer ANNs can lead to improved performance.