An improved Soft Actor-Critic strategy for optimal energy management

The transition from the current electrical grid to a smart, sustainable, efficient, and flexible electrical grid requires detecting future capabilities in order to have a system that can monitor, predict, learn, and make decisions on local energy consumption and production in real-time. A microgrid with these characteristics will allow the integration of distributed renewable energy systems efficiently, reducing the demand on power plants. The use of reinforcement learning can help find creative ways to keep the grid balanced; reschedule energy consumption through incentives; make predictions of demand and available energy at the grid scale, and assess the complexity of making these decisions. This work proposes using the novel Soft Actor-Critic (SAC) Deep Reinforcement Learning technique to manage electrical microgrids efficiently. SAC uses an entropybased objective function that allows it to overcome the problem of convergence brittleness by encouraging exploration without assigning a high probability of occurrence to any part of the range of actions. Results show the benefits of the proposed technique for the coordinated energy management of the microgrid.