Identification of simplified control strategies in multisource optimisation for electric demand flexibility in building energy management

The building sector significantly contributes to current energy challenges due to its high consumption and carbon footprint. The record high energy prices and increasing integration of intermittent renewable resources necessitates a paradigm shift in energy management. A critical component of this shift is the synergy between energy supply and demand, where demand-side response (DSR) plays a crucial role by providing grid flexibility. This communication focuses on enhancing electric demand flexibility in tertiary buildings as part of its carbon response. Specifically, it involves optimising electrical flows to identify energy-efficient strategies that utilize local resources effectively.

Previous research highlights the widespread use of advanced control methods such as model predictive control (MPC) in DSR for buildings. Despite their effectiveness, these methods face practical challenges due to their complexity and limited compatibility with building energy management systems (BEMS). Simplified control strategies for BEMS, which do not require prior dynamic modelling nor disturbance prediction, are gaining traction. Data-driven control strategies using machine learning techniques – such as reinforcement learning, supervised or unsupervised learning – have proven effective for optimising demand flexibility in buildings.

The aim of this work is to identify energy management strategies that can be easily integrated into BEMS through rule-based control. Instead of relying solely on expert knowledge, this study uses a learning process to identify simplified control actions based on simulation results from a previously developed MPC environment. The MPC calculates optimal trajectories for the chosen case study and provides a diverse range of strategies to account for disturbance uncertainties (such as weather, occupant behaviour, and energy usage). The simulation-generated data are then analysed to identify patterns for robust and simplified control schemes. The rule-based control, guided by machine learning, does not seek to provide an optimal solution but rather learns from optimal scenarios under varying input uncertainties.

The case study involves a tertiary building equipped with local photovoltaic production and electric vehicles. Machine learning-derived rules will be compared to those based on expert knowledge, and their deviation from the optimal solution provided by a perfect feedback MPC will be assessed. The robustness of the identified strategies will be evaluated against the optimal baseline using key performance indicators such as cost, electricity imports and exports, self-consumption rate and peak load shifting.

The findings will inform experimental applications in BEMS on the university campus, contributing to policy on efficient energy management practices.