Short-Term Load Forecasting using Machine Learning for Decarbonisation and Cost-Reduction in a Water Network-based Power System

For a water company to power their essential operations such as water and sewage treatment and supplying clean water to consumers, integrating Renewable Energy Systems (RESs) is a cleaner and more sustainable alternative to using grid-electricity. However, due to the intermittent nature of RESs, it becomes a challenge to match the supply of clean energy with the energy demand. Demand-side Energy Management approaches, such as load-forecasting, have been previously proposed but not yet applied to water networks. In recent years, Machine Learning (ML) models such as Neural Networks have become popular for load-forecasting applications, however, have not been applied for forecasting energy demand of a water network, keeping external factors such as the water demand into consideration. This study aims at conducting Short-term Load Forecasting (STLF) for a water treatment site in the northeast of the UK using two ML models namely, Nonlinear Autoregressive with Exogenous input(s) (NARX), and Long Short-Term Memory (LSTM). Both models were designed and trained using historical values of energy consumption, total water outflow, and other time-series variables for comparison. An average of multiple forecast tests showed that the LSTM (MAPE 7.95%) performs more accurately than the NARX network (MAPE 40.61%). The study showed that using ML models, a water company can, to a large extent, accurately forecast their energy demand for efficient Demand-side Energy Management, potentially leading to carbon and energy cost-savings.