PhD Defesne - Alaa Selim
Electrical and Computer Engineering Department Title: Learning-Based Optimization and Control of Active Distribution Systems: Voltage Security, Cybersecurity, and Outage Mitigation Abstract: Active Distribution Networks (ADNs) with high penetrations of inverter-interfaced Distributed Energy Resources require coordinated voltage control, outage resilience, and cyber-physical security. In this thesis the operational challenge is formulated as a constrained decision process whose state combines nodal voltages, power injections, load forecasts, and threat indicators, while the action space unifies inverter set-points, network-switch commands, and protection settings. Within this unified framework a distributed Volt–VAR optimizer—implemented with the actor–learner architecture on Ray RLlib—is trained across Monte-Carlo solar- and load-scenarios to sustain ±5% voltage compliance with sub-second convergence. The same decision-process formulation is extended with probabilistic storm forecasts, enabling a hybrid model-predictive/deep-reinforcement learning strategy that pre-positions storage and schedules switch operations, thereby reducing expected customer downtime by more than 60% in hurricane simulations. Building directly on the Volt–VAR control security, a two-stage cyber-defense pipeline safeguards the controller against data manipulation. First, a ResNet classifier augmented by explainable-AI examines streaming phasor measurements and large-language analysis of control logs to flag stealthy perturbations of inverter Volt–VAR curves in real time. Second, a Bayesian-optimized Stackelberg formulation leverages the same decision variables to prescribe counteractions and topology adjustments that restore voltage bounds under worst-case false-data injections, thus closing the loop between detection and mitigative control. The resulting secure and resilient framework is finally extended to islanded operation. Reduced-order dynamic models of grid-forming inverters, combined with safe deep-reinforcement learning, map admissible proportional–integral gains and co-optimize real and reactive power commands under explicit voltage, frequency, and power-sharing constraints. This agent admits only safe actions, enabling reliable black-start of multiple microgrids across uncertain load-pickup profiles and completing an end-to-end control architecture for tomorrow's inverter-dominated distribution systems.