Irish
Medina,
Master’s
candidate
David
R.
Cheriton
School
of
Computer
Science
Smart water meters have been installed across Abbotsford, British Columbia, Canada, to measure the water consumption of households in the area. Using this water consumption data, we develop machine learning and deep learning models to predict daily water consumption for existing multi-family residences. We also present a new methodology for predicting the water consumption of new housing developments.
This thesis contains three main contributions: First, we build machine learning models which include a feature engineering and feature selection step to predict daily water consumption for existing multi-family residences in the city of Abbotsford. This is motivated by the recent development direction towards denser living spaces in urban areas. We present the steps of the model building process and obtain models which achieve accurate performance. Second, we present a new methodology for building machine learning models to predict daily water consumption for new multi-family housing developments at the dissemination area level. Currently, the models used in the industry are simple baseline models which can lead to an overestimation of predicted water consumption for new developments, leading to costly and unnecessary investments in infrastructure. Using this new methodology, we obtain a machine learning model which achieves a 32.35% improvement over our best baseline model, which we consider a significant improvement. Third, we investigate the use of deep learning models, such as recurrent neural networks and convolutional neural networks, to predict daily water consumption for multi-family residences.
In our case, the main advantage of deep learning models over traditional machine learning techniques is the capability of deep learning models to learn data representations, allowing us to omit the feature engineering and feature selection steps and thereby allowing water utilities to save valuable time and resources. The deep learning models we build achieve comparable performance to traditional machine learning techniques.