Machine Learning-Based State of Charge Estimation for the Energy Management System of a 3 GT Hybrid Battery-Solar Electric Boat

Abstract

Small-scale electric boats face highly restricted energy margins due to limited battery capacities, whereas a single fishing cycle typically lasts up to eight hours. In operational contexts, the state of charge (SoC) acts as the primary decision variable in the Energy Management System (EMS), making the overall system accuracy heavily dependent on the precision of SoC estimation. For LiFePO₄ batteries, the highly flat open-circuit voltage (OCV) curve in the middle SoC range causes conventional estimation methods to lose their sensitivity. This study proposes an integrated simulation model for a 3 GT electric fishing boat, combining a 48 V 200 Ah LiFePO₄ battery pack using an OCV-resistance-RC approach, an 860 Wp photovoltaic array, and a 10 kW Brushless DC (BLDC) motor. This modeling generates an eight-hour mission dataset at a one-minute resolution, comprising 60 voyage variations equivalent to 28,800 samples. Random Forest and Long Short-Term Memory (LSTM) algorithms are evaluated using data leakage-free feature engineering that relies solely on measured quantities from the battery management system, excluding cumulative current integration. Based on 5,052 test samples, the LSTM method achieves superior accuracy with an RMSE of 0.587 and an R² of 0.9970, outperforming Random Forest, which yields an RMSE of 0.696 and an R² of 0.9958. However, the LSTM computation demands approximately 33 times longer training time. Energy audit results demonstrate that the photovoltaic array supplies 29.4% of the total mission energy requirement, with 8 out of 60 voyages reaching the battery protection threshold. This confirms that accurate SoC estimation is an absolute operational requirement to ensure the electrical reliability of fishing boats.