Zhige Yuan

The triple active bridge (TAB) enables three-port bidirectional transfer with high-frequency isolation, making it highly suitable for electric aircraft demanding efficient and lightweight power systems. However, TAB control is challenged by inherent port coupling, where disturbances at one port propagate throughout the system, and meanwhile current stress and efficiency constraints further complicate the control design. Recent advances in deep reinforcement learning (DRL) enable adaptive, model-free control but face sim-to-real limitations. To address these challenges, this paper proposes a physics-informed online-trained DRL (OT-DRL) framework for TAB converters: physical constraints are embedded into the reward function and action space, while real-time hardware data is integrated into the replay buffer to bridge the sim-to-real gap. The proposed online-trained DRL approach improves learning efficiency, transient dynamics, and ensures zero voltage switching under all conditions. Experiments demonstrate 11.3% higher episode efficiency, 2.6% greater final reward, and significant enhancements in dynamic response, 15% current stress reduction, and overall efficiency across various operating scenarios¹.

The integration of reconfigurable battery storage (RBS) systems into dc shipboard microgrids (SMGs) holds considerable potential by improving cell-level flexibility and system adaptability compared to conventional battery storage systems. This article proposes an online-trained deep reinforcement learning (DRL) control framework for an RBS system interfaced with a bidirectional dual-active-bridge (DAB) converter, aiming to manage its dynamic operation while enabling adaptive current regulation and cell balancing. Considering the nonlinear dynamics and real-time uncertainties inherent in practical SMGs, an automated online training framework is developed to bridge the sim-to-real gap that often hinders the generalization of offline-trained DRL methods. To this end, the proposed online training framework is implemented on a dSPACE real-time platform, enabling the controller to continuously interact with the physical system and refine its policy with real-world experiences. Experimental results validate that the proposed online-trained DRL approach achieves convergence 20.1% faster, in terms of training episodes, compared to simulation-based training. Comparative studies demonstrate that the proposed control strategy achieves superior dynamic performance, reducing settling time by 43.6% compared to the simulation-trained controller and by 53.7% compared to the benchmark predictive controller. It also enables faster charge-discharge transitions, completing mode switches within just 9.1 ms. Furthermore, the approach improves battery balancing and transient current handling, thereby improving its suitability for SMG applications.

Traditional current reference generators (CRGs) used in low-voltage ride-through (LVRT) control schemes often optimize a single objective, such as minimizing power oscillations, at the expense of other performance metrics such as total harmonic distortion (THD). This article proposes a many-objective optimized current reference generation strategy (MaO-OCRGS) that yields a set of tradeoff solutions, each defined by proposing two decision variables and their associated power quality performance metrics, i.e., power oscillations ($\Delta p$, $\Delta q$), $\mathit {THD}$, and the unbalance index ($\mathit {UI}$). A hybrid physics-based and data-driven (HPD) approach is proposed to construct the performance objective functions of a many-objective optimization (MaOO) problem, where the power oscillations are theoretically derived, and an artificial neural network (ANN) is proposed to model the relationships between the decision variables and current-related metrics ($\mathit {THD}$ and $\mathit {UI}$). A nondominated sorting genetic algorithm III (NSGA-III) is employed to solve the MaOO problem, yielding a set of optimal decision variables. A subsequent postoptimization selection then identifies feasible solutions that prioritize one performance metric while keeping the others within acceptable limits, thereby avoiding undesirable tradeoffs and resulting in three optimized CRGs. Experimental validation demonstrates that the optimized CRGs achieved through the proposed MaO-OCRGS effectively enhance one performance objective without unacceptable degradation in others, as the other metrics are maintained within their predefined limits, thus offering flexibility while enhancing the overall LVRT performance under unbalanced voltage conditions.