Improving Water Demand Forecasting Models Understanding with XAI Methods

Abstract

Understanding the main drivers of water demand predictive models is important for model applications and usability. The ever-increasing computational capacity has enabled multiple complex models to excel at predicting demand values. However, the understanding of how the model works and what predictors are responsible for the model’s outputs can be improved with the use of eXplainable artificial intelligence (XAI) methods. We present a comprehensive and systematic XAI implementation to enhance the understanding of common predictive models, such as tree-based ensembles and artificial neural networks and compare them with a simple yet highly explainable multiple linear regression model. By using a monthly residential demand dataset, we focused on discovering the positive and negative effects of multiple features, including socioeconomic and weather determinants. XAI methods include the SHapley additive exPlanations values, the local interpretable model-agnostic explanations values, and partial dependence plots. Results show that property size variables, like lot and building areas, contribute to above-average predictions of water demand. Additionally, weather-related predictors, such as mean temperature and humidity, significantly enhance the models’ predictability. XAI methods also show that an approximate constant monthly demand can be expected from houses when temperature values do not significantly vary. Finally, an uncertainty analysis highlights tree-based models as the most robust when working with noisy predictors. XAI methods may assist water management strategies by providing important details on behind-the-scenes complex predictive models that enhance the models’ usability in the water sector.