Adaptive Fault-Resilient and Self-Healing Energy Management for IoT-Enabled Microgrids via Deep Q-Learning with Real-Time Edge Control

This research introduces a fault-tolerant energy management system for IoT microgrids. It uses Deep Q-Learning (DQL) with real-time edge control. The system was trained and tested using the GridSTAGE dataset. The dataset contains detailed data and labeled fault events like islanding, DER dropout, and load imbalance. Tests showed the DQL agent had a 98.93% fault detection accuracy and a load satisfaction rate above 96.47%. Test accuracy was 98.81%, with false-positive rates below 2%. Inference latency stayed under 100 ms, which meets IEEE 2030.7 standards for edge-based microgrid controllers. The model showed it could adapt to different operating situations, as seen in confusion analyses. It modified the control actions according to the availability of DER to continue the flow of energy when there were problems. DQL system was also more flexible to the changes of time and nonlinear fault propagation as compared to LSTM-based and rule-driven systems. Exploration and learning parameters were studied and led to finding stable learning areas. It can be run on embedded control platforms and its fast decision-making makes it suitable. There was also fault-conscious dispatch and DER switching that reduced the recovery time. It can be used in changing grid conditions because it has high F1-scores (gt;0.94). Such a system minimizes latency issues, better islanding determination, and better utilization of resources. It is a non-adaptive control system that is scaled up. Future directions will consider the multi-agent systems, adversarial fault modeling, and safety-constrained reinforcement learning in order to enhance resilience in significant microgrid applications.