Probabilistic Forecasting of Hourly Water Demand

Timeseries forecasting holds a prominent position in the domain of urban water systems. Most forecasting approaches are designed to provide single-point deterministic forecasts, neglecting the uncertainty in model predictions. In this work, we propose a methodological framework, able to provide probabilistic predictions over lead times of operational interest, by combining machine learning (ML) methods with multivariate statistics (i.e., copulas). The idea is that ML methods can be used to provide deterministic forecasts, and copulas can be used to quantify the predictive uncertainty of the forecasts. We showcase the effectiveness of proposed framework using hourly water demand data from a real-world case study.