A Reinforcement Learning Framework for Scalable and Cost-Efficient Energy Management in Smart Grids

Abstract

Smart grids integrate renewable energy sources and enable dynamic demand responses to transform energy management. The complexity of managing multiple agent systems with different devices presents challenges in terms of scalability, computational efficiency and real-time adaptability. This paper introduces the new framework MARL-SG (Multi-Agent Reinforcement Learning for Smart Grids), which aims to optimize energy consumption across devices while maintaining grid stability, reducing costs and satisfying users. With MARL-SG, training is centralized, and execution is decentralized to ensure scaling, and advanced technologies such as Heuristic masking ensure the allocation of computational resources to critical tasks. According to the experimental results, the MARL-SG reduces energy costs during peak and off-peak hours, achieves almost perfect grid stability and provides reliable and cost-effective energy distribution. The framework will enable modern smart grids to manage energy more intelligently.