Incorporating Weather Forecasts into Short-Term Water Demand Prediction using Probabilistic Deep Learning with Long Short-Term Memory Networks

Abstract

<p>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 due to climate change. Especially demand peaks of multiple hours during the day or persisting demand peaks of several days caused by prolongued dry periods and more heat days throughout the summer force water suppliers to more efficiently control and manage their resources. Being able to take proactive and informed decisions through reliable short-term probabilistic forecasts is therefore crucial in this context.</p> <p>This research proposes two probabilistic deep learning architectures based on long short-term memory (LSTM) networks to forecast hourly water demand up to 10 days in advance. Both models processes different temporal sequences of data, including past observations of water demand and regressors as well as future regressors with different time lengths. The models encode long-term historic information of the water demand and features, including historic meteorological information, and simultaneously incorporate short-term future information on calender- and weather features using statistically optimized point forecasts (DWD MOSMIX) of the latter. Through implementing the models in an autoregressive manner, the output is fed back into itself at each step and predictions are made conditioned on the previous one to account for correct path dependency between consecutive hours. This way the model produces multi-step-ahead forecasts of variable length by using future information together with the historic context.</p> <p>In a case study of central Germany, the performance of the proposed deep learning models was compared to a Lasso estimated high-dimensional time series model and a conventional AR(p) model. Results indicate the potential of the proposed approach of using weather forecasts in short-term water demand prediction especially for lead times larger than 24 hours.</p>