Customized Neural Network training to predict the highly imbalanced data of domestic hot water usage

Abstract

Despite space heating and cooling, the energy use for hot water production has not changed significantly over time and accounts for a big share in modern, well-insulated buildings. The main challenge of hot water generation lies in the highly stochastic nature of the domestic hot water (DHW) demand. Prediction of DHW demand can significantly help to a more efficient operational strategy in water heating systems. However, the time-series data of hot water demand is very sparse and imbalanced, including many zero demands, which makes it challenging to be predicted properly by Machine Learning methods. This study uses data recorded from a single-family building in South Africa and aims to understand how the customizations of a neural network for learning imbalanced datasets can affect the prediction of hot water demand. Four different customizations (Random over-sampling, Random under-sampling, Weight Relevance-based Combination Strategy, Synthetic Minority Over-sampling Technique for Regression) are compared with the baseline model to predict the hot water demand data. The performance of 9 different simulations is compared and the challenges are outlined. The over-sampling technique shows promising results for practical implementation by over-predicting high peaks by up to 20%, which will guarantee enough hot water production at peak usage.