Conformal Prediction Intervals For Water Demand Forecasting

In a world with accelerating climate change, rapid population increase and urbanization, urban water systems are under a growing stress. Precise short- and medium-term water demand forecasting are needed to optimize water supply operations. While machine learning methods are commonly used for this task, most studies rely on point predictions which lack a robust characterization of prediction errors. This undermines decision making under uncertainty and related applications. In this work, we employ real data to demonstrate the advantages of probabilistic water demand forecasting up to a week ahead. In particular, we explore the benefits of conformal predictions, a set of novel techniques providing distribution-free prediction intervals. Conformal predictions are model agnostic and may guarantee the validity of the prediction intervals under some assumptions. We apply the conformal prediction framework on several ML models, including tree-based methods, deep neural network models and classical time series analysis. We compare these conformalized approaches against traditional probabilistic methods such as quantile regression and Monte-Carlo dropout.