Prédiction de la consommation des chauffe-eau électriques pour la commande prédictive des micro-grids

Buildings are significant energy consumers within the European Union, accounting for approximately 40% of total energy usage, with domestic hot water systems contributing around 15% of this demand. In addition to their consumption, buildings also hold potential as energy producers, positioning them as active nodes within modern energy micro-grids. To optimise energy costs and reduce peak demand, model predictive control (MPC) combined with local energy generation and storage has emerged as an effective strategy, enhancing both system efficiency and on-site energy utilisation. A critical factor in the successful implementation of MPC is the availability of accurate energy consumption prediction models. This thesis focuses on predicting the energy consumption of domestic electric water boilers (EWBs) with thermal storage. The study begins with a comprehensive review of state-of-the-art energy prediction models, identifying key methodologies, and evaluating their applicability to EWB systems. A statistical analysis of experimental data from two real-world residential buildings is then conducted to characterise consumption patterns and inform model development. Three endogenous prediction models, autoregressive integrated moving average (ARIMA), support vector regression (SVR), and artificial neural networks (multilayer perceptron (MLP)) are developed and evaluated. Their performance is compared in terms of accuracy, computational efficiency, and robustness. To leverage their complementary predictive strengths, a hybrid approach is proposed. To further improve prediction accuracy, exogenous variables, including time of day, day of week, and hot water usage are integrated into the hybrid model. The final hybrid model demonstrates improved performance in both prediction and forecasting, making it an optimal framework for MPC applications in micro-grids.