Machine Learning for Planning Water Supply Infrastructure in the Face of Climate Change

The evaluation of risks and assessment of developments in drinking water supply under the influence of climate change require reliable predictions of future per capita drinking water demand and supply over more extended periods. The consumption of drinking water and the water supply are strongly dependent on various environmental factors, such as temperature, precipitation, and the number of summers and hot days, which are directly influenced by climate change. There is currently no established mathematical model that can reliably predict based on the limited available measurement data and sociological survey data. Therefore, the declared goal of our research project is to develop a machine learning (ML) strategy that enables the generation of ML models capable of reliably predicting water demand and supply in the context of climate change.

One of the special features of this challenge for ML strategies is the relatively small number of training samples per individual local water supplier. This is unavoidable, as many local water suppliers only record consumption and supply monthly and often do not continuously monitor their own databases. This scarcity of local data requires ML models specialized for data-poor applications.

The relatively large number of local training datasets for this task makes the automation of the training process particularly attractive. For some years now, the concept of automated ML (AutoML) has enjoyed increasing popularity. AutoML focuses on the automatic selection of models, selection of model components, hyperparameter tuning, and model evaluation. We intend to develop an AutoML strategy for data-poor applications that focuses on predicting water demand and supply as functions of climate parameters. Using the ML models produced by AutoML, it would be possible to make individual and ensemble predictions of drinking water consumption and supply for each local water supplier. Data from water suppliers throughout southern Germany will be used to train the models. The result of this work would enable each local water supplier to implement an early warning system, with the aim of defining specific measures for better management of water resources in a timely manner.