Abstract
To address the challenge of rotor speed drift, large overshoot, and even DC microgrid instability caused by constant generator rotor speed inaccurate prediction in marine IRGT/SOFC all-electric propulsion systems, this study proposes a physics-informed Long Short-Term Memory framework to enhance transient prediction accuracy and robustness. This method integrates rotor nonlinear differential equations into the loss function, enforcing physical characteristic consistency while leveraging the LSTM memory capability to capture long-term nonlinear dependencies. The dataset from the mechanism and data-driven hybrid model which is validated in 2.5% error by experiment is used for training.
Results demonstrate that it reduces 55% Root Mean Square Error for key power parameters of system output power, with RMSE dropping from 107 to 52 and 183 to 83, respectively. For generator rotor speed, the RMSE improved by 50%, from 2.08 to 1.05, with an R² increase from 0.926 to 0.998, which mitigate rotor speed drift and overshoot, ensuring accurate and reliable predictions. Under dynamic conditions such as sudden load decreases, the PI-LSTM framework maintains stability and limits the prediction error for generator rotor speed to a maximum of 5 rpm, while the traditional LSTM suffers deviations as large as 30 rpm (1%). This work introduces a novel technical approach for next-generation marine hybrid propulsion systems, enabling highly accurate dynamic predictions and enhancing the reliability of marine microgrid operations.