10th International Congress on Information and Communication Technology

Authors - Angel Peredo, Hector Lugo, Christian Narcia-Macias, Jose Espinoza, Daniel Masamba, Adan Gandarilla, Erik Enriquez, Dong-Chul Kim
Abstract - This paper explores the under-examined potential of offline reinforcement learning algorithms in the context of Smart Grids. While online methods, such as Proximal Policy Optimization (PPO), have been extensively studied, offline methods, which inherently avoid real-time interactions, may offer practical safety benefits in scenarios like power grid management, where suboptimal policies could lead to severe consequences. To investigate this, we conducted experiments in Grid2Op environments with varying grid complexity, including differences in size and topology. Our results suggest that offline algorithms can achieve comparable or superior performance to online methods, particularly as grid complexity increases. Additionally, we observed that the diversity of training data plays a crucial role, with data collected through environment sampling yielding better results than data generated by trained models. These findings underscore the value of further exploring offline approaches in safety-critical applications.