Data-driven, Long-term Forecasting of Water Demand in the Face of Climate Change

Sufficient availability of drinking water and corresponding long-term planning are essential prerequisites for a sustainable future. This requires reliable long-term forecasts of future water demand. Hourly and daily demand forecasts using machine learning (ML) are well established, provided that sufficient data are available. Nevertheless, there are substantial challenges. First, many local water utilities only have monthly demand data. Second, the system is transient because of climate change and social, legal, and economic changes. Third, future weather and climate conditions, as well as the aforementioned processes of change, are uncertain. Overall, this leads to highly volatile and uncertain scenarios with limited data, a significant challenge for modeling and ML methods. Nevertheless, these methods should be broadly applicable in different climate and economic regions, allow reliable predictions over decades, and be manageable for experts in planning offices. 

This project aims to improve long-term forecasts of water demand by addressing the following four research questions: Which ML models for data-poor problems best describe water demand, and can model selection be automated? What explanatory variables are needed, and how are they distributed in the future? How can we address the varying explanatory power of data in transient problems? How can we achieve reasonable uncertainty intervals for risk assessments?

To answer these questions, we will develop, combine, and evaluate ML models explicitly designed for data-poor situations, and automate their selection. This will include selecting explanatory variables and studying their probability distributions. We will work on two time scales: the short term (local weather) and the long term (climate). For the short-term scale, we will use statistical weather generators. In contrast, for the long-term scale, we will use long-term weather forecasts from the German Weather Service (DWD) under different climate scenarios. Because technical, societal, or economic changes and their effects on water demand are generally difficult to predict and model, they must be treated as exogenous or fixed variables. They can affect the validity of data collected under current conditions. Therefore, we will develop multi-fidelity approaches that can learn from shorter time series in larger spatial areas.

For this project, we will build on prior work in polynomial chaos and Gaussian process regression. The study area for development and testing is southern Germany, with a range of climatic and economic regions and close to 100 local water suppliers. All methods will be made open-source to promote transparency in demand forecasting and thus make improved forecasting and decision support publicly available.