Abstract
In Belgium's efforts to limit global warming, the share of renewable energy generation increased from 19.8% in 2022 to 29.8% in 2024. This rapid energy transition poses a significant risk to our grid stability due to the intrinsically high variability of renewable energy sources. As our electricity supply is becoming more variable, flexibility will have to be provided by consumers through demand response programs. Recently, Reinforcement Learning (RL) demonstrated superior performance for demand response compared to conventional control algorithms, including rule-based approaches and model predictive control (MPC), which are currently considered best practice by industry. However, adoption of RL remains low due to its (1) lack of robustness, (2) computationally intensive training, and (3) time-consuming digital twin modeling. These challenges increase the implementation costs of demand response programs, hindering their widespread adoption. In this study, we introduce Description-Based Controllers (DBCs) as a solution to these challenges. DBCs provide robust control, automatically generate digital twins based on system descriptions, and significantly reduce training time in RL, enabling an acceleration in demand response development. The effectiveness of this methodology will be demonstrated for industrial demand response of conventional heating, ventilation and air-conditioning (HVAC) systems using three case studies with increasing complexity.
Researcher(s)
Research team(s)
Project type(s)