Author(s): Gregor Johnen; Andre Niemann; Alexander Hutwalker; Christoph Donner
Linked Author(s):
Keywords: No Keywords
Abstract: As could be seen in recent years, ensuring the water supply-demand balance is a topic of increasing concern to supply companies facing the threat of increased demand scenarios resulting from long-term effects of climate change. Demand peaks of multiple hours during the day or persisting peaks over several days -- caused by prolonged dry periods and more heat days throughout the summer -- force water suppliers to manage their resources more efficiently. Reliable short-term probabilistic forecasts play a crucial role in enabling initiative-taking and informed decision-making in this context. This research proposes two probabilistic Long Short-Term Memory (LSTM) architectures to forecast week-ahead hourly water demand. These models encode long-term historic information of water demand and features, including meteorological data, while incorporating short-term future information on calendar- and weather-based features using pseudo-forecast meteorological data as driving forces. Additionally, we introduce intrinsically distributional neural networks with a probability layer that outputs parameters of a chosen density function, enabling the generation of distributional multi-step-ahead forecasts of variable length by combining future information with historical context. The approach is validated through a case study using demand data from a German water supply company. Results demonstrate that the proposed LSTM architectures outperform state-of-the-art models, including AR (p) and conventional LSTM models, in terms of forecast accuracy and reliability. The integration of pseudo-forecast meteorological data significantly enhances predictive performance by capturing the interplay between future weather and water demand. Moreover, since probabilistic forecasting is an essential component of risk management, the results have significant implications for the management of drinking water distributions in the water supply sector, particularly when the distribution's bias is not too pronounced.
Year: 2024