Energy management of domestic hot water systems with model predictive control and demand forecast based on machine learning

Abstract

• Predictive control can improve performance of domestic hot water systems. • Model predictive controllers with imperfect or perfect forecasts perform similarly. • Decreasing accuracy increases the amount of temperature constraints violations. • Optimal control shows controllers should produce hot water shortly before peaks. • Inability to prepare for peak demand is the main impact of forecasting inaccuracies. While domestic hot water (DHW) production represents approximately 17% of the energy consumption of the Canadian residential sector, studies aiming to reduce this proportion are scarce. Model predictive control (MPC) can improve the performance of DHW systems by relying on demand forecasting models to operate the system. However, DHW predictions are often inaccurate for single-family or small systems, which makes the viability of MPC uncertain in these cases. To validate the applicability of MPCs relying on machine-learning forecasting models, we assessed their performance through the simulation of 40 different DHW profiles, based on measurements in the same number of single-family residential units. For the sake of comparison, four controllers are developed and tested in this paper: a rule-based controller (current situation), an optimal controller, an MPC with perfect demand forecast, and an MPC with real forecast from machine-learning models (i.e., MPC with neural networks). Furthermore, we considered 100 combinations of relative importance allocated to energy saving versus to the respect of temperature constraints in the controllers. We found that MPCs with perfect and imperfect forecasts could reach a similar performance in most cases, validating the applicability of the machine-learning models. The MPC relying on a machine-learning forecasting model yielded energy savings from 4% to 8% when compared with traditional control. The ability of the MPC to keep the stored DHW temperature closer to its minimal temperature constraint reduced heat losses and caused these savings. In comparison with optimal control, the inability to act sufficiently in advance of imperfect-forecast MPCs was identified as the main impact of forecasting inaccuracies and resulted in more temperature constraint violations.