An intelligent power control based on deep neural-networks and regularized actor-critic learning for marine systems

Renewable generation units along with storage systems are being integrated to shipboard microgrids (SHMGs) for enhanced operational efficiency and cost-effective benefits. However, the Direct Current (DC) voltage profiles of SHMGs are affected by sudden changes in load, intermittent renewable inputs, and unmodeled dynamics which necessitate advanced control strategies to address them. This paper focuses on designing an adaptive data-driven SHMGs voltage regulation framework with hybrid energy storage units (HESUs) to preserve robust stability across a wide range of operational situations. The proposed data-driven voltage regulator is developed in two stages: i) an ultra-local model control (ULMC) is designed to stabilize the voltage outcomes of SHMG using the input/output (I/O) data. In this stage, a regularized actor-critic (RAC) using deep neural networks is adopted to adaptively adjust coefficients of ULMC, and ii) in the next stage, the un-modeled phenomena and disturbances included in the marine power system are approximated by non-integer extended state observer (NIESO). In this approach, deep neural networks of RAC learning are trained in such a way that obtain optimal policy using a reward function which is defined based on the regulation requirements of SHMG. The proposed control framework provides this possibility for the system to have a robust estimation and compensation to tackle with the largely unknown nonlinear disturbances and unmodeled dynamics. When coefficients embedded in the ULMC controller are adjusted by RAC learning, it can stabilize the voltage outcomes of the marine power system. Hardware-in-the-Loop (HiL) simulation using OPAL-RT have been employed as the powerful real-time testbed to evaluate the feasibility and applicability of proposed voltage compensator under realistic operations of SHMG. The robustness of proposed ULMC was evaluated using several fault and disturbance scenarios such as load spikes and fluctuations in the energy source as well as partial system failures. The HiL results confirm the voltage stability, and disturbance rejection capability of the proposed NIESO based on ULMC (designed by RAC learning) framework surpasses traditional model-based controllers and intelligent controllers by a significant margin. It was demonstrated that the method not only enables rapid response during transient but ensures sustained operational suitability and long-term efficiency during maritime microgrid operations.