Autonomous Cybersecurity for Power Grids: A GAN andReinforcement Learning Framework for Edge Deployment

The digitalization of electrical substations has created critical vulnerabilities in national power grids. Existing Intrusion Detection Systems (IDS) are often too computationally intensive and costly for widespread deployment in resource-constrained Operational Technology (OT) environments. This paper introduces a novel, autonomous cyber-physical defense system designed for scalability and affordability. The system features a hybrid Transformer-LSTM detection engine, adversarially trained using a Generative Adversarial Network (GAN) to ensure resilience against zero-day attacks. Going beyond passive detection, a Reinforcement Learning (RL) agent is implemented to execute real-time, optimal mitigation strategies that balance security with operational continuity. A key contribution of this work is a novel software optimization pipeline, which enables this complex AI architecture to be deployed efficiently on low-cost Raspberry Pi 5 hardware. Experimental results on a challenging, class-balanced test set demonstrate that our system achieves high detection accuracy (>96%) and performs real-time inference within the strict resource limits of commodity edge hardware, presenting a practical and scalable solution for securing critical infrastructure.