Probabilistic water demand forecasting focussing on the impact of climate change and the quantification of uncertainties in the short- and mid-term

As could be seen in recent years, the impact of climate change is already detectable in water demand patterns and results in new challenges for the water supply sector. Demand peaks caused by changing climate conditions such as longer dry periods force water suppliers to a more efficient control and management of their assets and water resources to avert supply shortages. Especially demand peaks of multiple hours during the day or persisting demand peaks of several days and weeks threaten the supply demand-balance. By utilizing accurate forecasts of the expected water demand, suppliers are enabled to better prepare their assets for such extreme conditions.

To adapt to the consequences of changing hydro-climatic and demand conditions, this research proposes a water demand forecasting model to predict such extreme demand conditions caused by climate change for the short- to mid-term range. Here, a special emphasis is put on modelling the impact of weather variables on the water consumption caused by climate change. Those effects are complex, non-linear and multidimensional in nature and therefore challenging to model. Focusing on the practical usage, the forecasting model is appropriate for real-time application providing accurate forecasts coupled with a high interpretability. This allows the quantification of the ongoing effects of climate change and enables a better consideration of the underlying uncertainty.

Our case study uses real data on district level from two regions in West and Central Germany. To appropriately account for the practical need of varying forecast schemes, historical demand and weather data are used at quarter-hourly, hourly as well as daily resolution.

Multiple linear, non-linear and stacked models tailored to the forecasting purpose and the varying horizons are implemented with a clear focus on interpretability and forecasting accuracy. To model the underlying uncertainty, complete probablistic forecasts are proposed. Model assessment takes place by utilizing appropriate metrics as the MAE, CRPS or energy score.