Energy Management in Residential Dwellings: Forecasting Hot Water Demand with Hybrid Machine Learning Models

Optimal and accurate forecasting of hot water consumption is essential for optimizing grid energy usage and supporting sustainable energy management in the fight against global warming and fluctuating demand in a changing climate. This study hybridises and compares supervised learning algorithms, Support Vector Regression (SVR), and the statistical regression technique ARIMA, both integrated with Multi-Layer Perceptron (MLP). The analysis utilises one year and five months of domestic hot water consumption data from a residential building, categorized into daily seasonal consumption for winter and summer scenarios as endogenous inputs. SVR-MLP demonstrates superior accuracy in winter at 94%, effectively modeling irregularities and non-linearity dependencies in hot water demand, despite higher computational costs of 3 minutes and 35 seconds. In contrast, ARIMA-MLP performs optimally in summer, offering computationally efficient and precise forecasts under stable variable seasonal conditions with 98% accuracy. This study highlights the strengths and limitations of each model, offering valuable insights for optimizing domestic hot water and energy consumption forecasting based on seasonal variability.