Adaptive hot water production based on Supervised Learning

A major challenge in the common approach of hot water generation in residential houses lies in the highly stochastic nature of domestic hot water (DHW) demand. Learning hot water use behavior enables water heating systems to continuously adapt to the stochastic demand and reduce energy consumption. This paper aims to understand how machine learning (ML) can predict the stochastic hot water use behavior, and to investigate the potential reduction in energy use by an adapting hot water system. Different ML models are implemented on a data set of 6 residential houses, and their average performance is compared. Ten different models were evaluated, including four single models (Random Forest, Multi-Layer Perceptron, Long-Short Term Memory Neural Network, and LASSO regression), four Sequential Multi-Task models combining classification and regression models, and two Parallel Multi-Task models based on Random Forest and Multi-Layer Perceptron. Dynamic simulation of a smart hot water supply system, which adapts to the predicted demand, shows that adaptive hot water production can provide significant energy use reduction