Scaling a Deep Learning-Based Control for a BESS to Various Microgrid Sizing Configurations Through Transfer Learning

The increased penetration of renewable energy sources (RES) into power grids introduces significant challenges due to their inherent variability, uncertainty and the resulting reduction in grid inertia. Traditional model-based control strategies often struggle to accurately capture the grid’s rapid dynamics and stochastic behavior. Consequently, there is a growing need for fast, model-free control approaches that leverage the increasing availability of real-time data to ensure a balance between supply and demand while minimizing operating costs and CO<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> emissions. In response, this study presents a hybrid control framework combining classical droop control with a predictive data-based control in a battery energy storage system (BESS). Leveraging historical power measurements, the proposed approach employs a deep learning model initially trained on one islanded microgrid and subsequently adapted to another one by using transfer learning techniques, avoiding the need for complete retraining. The framework is evaluated on an optimized islanded microgrid designed using mathematical programming to balance economic costs and emissions. In this cost-effective and environmentally friendly setup, the proposed framework further reduces costs and emissions while effectively managing power reserves from a BESS amidst complexities introduced by RES integration.