Graph Convolutional Recurrent Neural Networks for Water Demand Forecasting

Short-term forecasting of water demand is a crucial process for managing efficiently water supply systems. This paper proposes to develop a novel graph convolutional recurrent neural network (GCRNN) to predict time series of water demand related to some water supply systems or district metering areas that belong to the same geographical area. The aim is to build a graph-based model able to capture the dependence among the different water demand time series both in spatial and in temporal terms. This model is built on a set of different graphs, and its performance is compared to two methods, including a state-of-the-art deep long short-term memory (LSTM) neural network and a traditional seasonal autoregressive moving average model. Additionally, the forecasting model is tested in a condition when a sensor has a malfunction. The results show the ability of the GCRNN to produce accurate and reliable forecasting, especially when based on graph built while accounting for both time-series correlation and spatial criteria. The GCRNN consistently outperforms the LSTM during the fault test, showing its ability to generate a robust prediction for days after a sensor malfunction, given the GCRNN's ability to benefit from the other time series of the graph.