Ensemble Deep and Machine Learning for Improving Short-Term Water Demand Forecast in Cities

Abstract

Intelligent water distribution systems are essential to ensure water availability for all city users amidst increasing water scarcity driven by urbanization and climate change. In this study, our goal was to enhance the forecasting accuracy of a water utility using a Bayesian Moving Averaging (BMA) ensemble model. This model combines Random Forest (RF), Extreme Gradient Boosting (XGB), and Multilayer Perceptron (MLP) algorithms. We utilized daily monitoring data from the water utility spanning from 2018 to 2023 to predict short-term water demand, incorporating weather features. Our findings demonstrate that the proposed ensemble model achieved the best performance, yielding the lowest mean absolute percentage error (MAPE) of 15.99% and the highest R squared (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>) value of 0.98 on the testing set. Additionally, the RF model exhibited better results compared to the other single models. These outcomes underscore the potential of the ensemble approach in short-term water demand forecasting, surpassing more traditional and state-of-the-art methods.