Volume 60

Dynamic Identification and Prediction of a Marine Parallel IRGT/SOFC Hybrid System Using LSTM Network Jiale Wen, Shengying Xiao, Xicong Mi, Catalina Spataru, Yiwu Weng, Xiaojing Lv

Abstract

In response to the frequent faults and low efficiency caused by response delay and power imbalance in dual-unit intercooled reheat gas turbine/solid oxide fuel cell (IRGT/SOFC) hybrid systems under multiple operating scenarios, this study proposes a multi-input multi-output (MIMO) long short-term memory (LSTM) neural network for dynamic identification and prediction. By integrating physics-based modeling with data-driven training, the proposed framework captures the transient behaviors of a marine parallel IRGT-SOFC system across varying fuel scenarios. Sensitivity analyses on sampling rate, hidden layer configuration, and learning rate demonstrate that a single-layer LSTM with 50 neurons and 0.01 s sampling interval provides the optimal balance between prediction accuracy and computational cost. The trained model achieved R² values exceeding 0.998 across all key outputs, with MAPE consistently below 0.79% except T_FCavg, effectively predicting power output, SOFC temperature, and overall system efficiency. The proposed model effectively captures the transient response patterns of dual-unit systems under varying fuel input, laying a high-precision modeling foundation for predictive power control of next-generation marine propulsion systems in complex operational scenarios.

Keywords All-electric propulsion system; Dual unit IRGT/SOFC; MIMO-LSTM; Real-time prediction