Hybrid medium-voltage AC/DC (MV AC/DC) distribution networks provide enhanced efficiency and controllability while enabling compact integration of DC native resources and loads. However, converter-dominated operation increases vulnerability to rapid D...
Hybrid medium-voltage AC/DC (MV AC/DC) distribution networks provide enhanced efficiency and controllability while enabling compact integration of DC native resources and loads. However, converter-dominated operation increases vulnerability to rapid DC fault current escalation and to dynamic stability degradation. Existing approaches typically rely on hardware-intensive solutions such as full-bridge submodules or DC circuit breakers, or employ control strategies that lack formal constraint guarantees during severe transients. This dissertation proposes and validates a coordinated, software-centric control framework for half-bridge modular multilevel converters (HB-MMCs) that enhances DC fault ride-through (FRT) while limiting peak DC fault currents. By managing fault energy, the proposed framework relaxes the speed requirements
for DC protection, potentially enabling the use of cost-effective mechanical circuit breakers instead of ultra-fast hybrid or solid-state devices.
The first contribution is a safety-constrained model predictive control (MPC) layer for real-time DC fault current limiting and submodule energy management. The MPC explicitly enforces hard constraints (envelope bounds) on circulating current magnitude, submodule energy deviation, and AC-side power exchange. A momentum-accelerated projected-gradient solver with a fixed iteration budget, 120 iterations during normal operation, scaled to 420 during fault conditions, enables bounded, deterministic runtime execution at the converter control rate. This preserves effective control action during rapid transient and recovery.
The second contribution is a reinforcement learning (RL) supervisory layer based on Twin Delayed Deep Deterministic Policy Gradient (TD3), which adapts selected MPC meta-parameters, objective weights and envelope limits, to accommodate diverse operating conditions and model uncertainty, while retaining hard-constraint enforcement in the MPC inner-loop. The RL supervisor is trained via curriculum learning over representative MVDC contingencies and deployed with bounded action spaces to maintain stability of the hierarchical architecture.
A high-fidelity electromagnetic transient (EMT) test bench of a point-to-point MVDC link was used to validate the proposed framework. Across the tested fault resistance range (0.1–20 Ω), peak fault currents were reduced by 7.46% to 24.33% relative to PI-based vector control (VC). The TD3 supervisor achieved an additional 18.9% peak current suppression relative to fixed-parameter MPC under severe fault conditions. These results demonstrate that software-enabled FRT offers a viable alternative to hardware-based fault current limiting, providing a verifiable pathway towards adaptive resilience across diverse fault severities and operating conditions, and positioning hybrid MV AC/DC distribution systems to meet the reliability and resilience demands of future converter dominated distribution networks.