Deep reinforcement learning adaptive under-frequency load shedding for frequency control under extreme events

The operational principles of the conventional Under-Frequency Load Shedding Schemes (UFLS) are increasingly challenged by the modern grid’s rapidly evolving load characteristics and significant integration of Converter Interfaced Generation (CIG). These factors undermine the UFLS’s static approach which assumes predictable load and generation, thereby increasing power mismatch risks. Adaptive UFLS (AUFLS), which uses real-time data to compute the shedding scheme, has been proposed as a solution; however, most approaches are model-based and struggle to capture complex and hard-tomodel system relationships. Additionally, these methods often require intensive computation due to the solving of system equations. Data-driven techniques, while having the potential to capture these complex non-linearities, are often implemented in isolation, yet they can be unreliable when faced with scenarios beyond their training domain—sometimes leading to detrimental actions during extreme events. This paper implements an AUFLS that integrates the conventional UFLS scheme with data-driven techniques through Reinforcement Learning (RL) as a supplementary controller, leveraging the strengths of both approaches. The conventional UFLS is retained as a last resort and remains operational. Simulation results on a modified IEEE 39-bus network demonstrate that the proposed AUFLS can reduce the required load shedding while maintaining stable system frequency, all with minimal computational overhead.