Topology-Adversarial Cognitive Graph Digital-Twin Reinforcement Learning for Resilient Power-Grid Control
Keywords:
Cyber-physical security, deep reinforcement learning, digital twin, epistemic uncertainty, false data injection, graph attention networks, power-grid control, safe reinforcement learning, world modelsAbstract
Abstract— The accelerating penetration of renewable generation, weather-driven demand, and cyber-physical threats is pushing modern power grids toward regimes in which purely model-based or reactive controllers are increasingly brittle. This paper proposes TAC-GDT-RL, a framework that couples four mechanisms in a single closed loop: (i) a multi-task graph-attention (GATv2) digital twin predicting nodal voltage, component-failure probability, and operational risk from an eleven-dimensional physics- and weather-aware node representation; (ii) Monte-Carlo-dropout inference giving the twin calibrated epistemic uncertainty; (iii) an uncertainty-triggered topology-dreaming module that synthesizes adversarial contingencies and fine-tunes the twin through prioritized replay; and (iv) a Proximal-Policy-Optimization controller trained under curriculum learning and protected by a rule-based safety shield. On the IEEE 14-bus network driven by one year of hourly weather across five contingency families, a seven-way ablation isolates each mechanism using paired non-parametric tests and rank-biserial effect sizes. The complete method attains the highest mean control reward (33.65 versus 11.13-22.14), significantly exceeding the fixed-topology, cyber-unaware, shield-free, and plain-dynamic variants, and reduces the voltage-violation rate to 0.31. The twin reaches a failure-detection AUC of 0.982 and a voltage MAE of 0.026 p.u. Results are reported transparently, including mechanisms not individually significant at the present compute budget.