A fault-aware, explainable proximal policy optimization based intelligent control framework with neuromorphic-inspired encoding for voltage and frequency stability

Abstract High penetration of inverter-interfaced generation, faults, and sudden load changes in power systems makes them more vulnerable to voltage and frequency deviations. Traditional controllers often struggle to adapt to varying disturbance types and severities in real time. Adaptive, data-driven control methods are required for responding dynamically to such disturbances. The proposed research work addresses this need by combining Proximal Policy Optimization (PPO)-based Reinforcement Learning (RL) with neuromorphic-inspired spike encoding, enabling temporally meaningful representation of system disturbances for interpretable and effective stability analysis in a simulation-based surrogate framework across different fault types. The controller adjusts its actions based on observed deviations in voltage and frequency. Instead of fixed control, the proposed framework enables dynamic fault-aware responses. For systematic fault investigation, a synthetic fault dataset is developed for the IEEE New England 39-bus system. It provides a controlled and flexible environment for reproducible experimentation and focused evaluation of the proposed strategy. The fault types covered are Single Line-to- Ground fault (SLG), Line-to-Line fault (LL), Line-to-Line-to Ground fault (LLG), and Three-Phase faults (3PH). Voltage deviations and frequency deviations capture the disturbances in the system. After the application of the PPO controller, the average voltage drop is reduced from ≈ 0.13 p. u. to ≈ 0.015 p. u., and the average frequency deviation is reduced from ≈ 0.069 p. u. to ≈ 0.010 p. u. A Deviation-Based Stability Score (DBSS) is defined as a data-driven metric derived from voltage and frequency deviations during fault scenarios. It reaches approximately (0.99), indicating improved stability performance after the application of the controller within the simulation-based framework. The other analysis includes recovery time estimation, surface plots for AI control effectiveness, surface plots for recovery time effectiveness vs. fault severity and AI control strength, correlation heatmap, comparison with baseline, droop, and PPO-based control strategies, convergence analysis, ablation study (raw vs. spike encoding), training with multiple seeds, plotting of smoothed training rewards, and PPO agent performance analysis across different observation configurations. The present PPO-based intelligent control framework provides transparent and interpretable decision-making unlike traditional black-box models through explicitly relating voltage and frequency deviations, spike visuals, surface-based, feature importance plots, and control actions revealing clear cause-effect controller behaviour, making the work explainable. It contributes to meeting Goal-7 of 17 Sustainable Development Goals (SDG). It aligns with meeting the objective of affordable and clean energy for all by reducing the risk of blackouts, effective utilization of energy, and improved reliability of the power grid. The present work provides a scalable, adaptive, explainable, and fault-aware control solution for next-generation and intelligent power systems.