A Multi-objective Approach to Safe Reinforcement Learning for Smart Grids
2025 · Balancing energy costs and safety with multi-objective reinforcement learning.
Building energy management systems commonly employ Reinforcement Learning (RL) controllers to overcome the dependence on accurate forecasts of key quantities by optimal controllers like Model Predictive Control (MPC). The challenge for RL approaches lies in informing the controller about realworld safety constraints. Existing work simplistically treats safety violations as a penalty to be added to the scalar reward term. In this work, we instead cast the problem as the true Multi-Objective Reinforcement Learning (MORL) task that it actually is by making the reward vector-valued and explore the performance of different MORL algorithms on a smart household scenario. Our results suggest that MORL algorithms can effectively learn the Pareto front of cost and safety constraints and outperform model-based MPC controllers, while allowing flexible selection of operating points without the need for retraining.