Data-Driven Modeling of Hydraulic-to-Electric Energy Conversion for Underwater Thermal Vehicles

Accurate modeling of hydraulic-to-electric energy conversion is critical for improving energy utilization and operational efficiency in underwater thermal vehicles. Traditional physics-based models, though interpretable and effective under fixed conditions, often rely on idealized assumptions and fixed parameters that limit their adaptability to dynamic operating environments. To address these limitations, this study explores data-driven modeling techniques for real-time efficiency prediction. Four regression models are investigated-linear regression (LR), backpropagation (BP) neural network, support vector regression (SVR), and long short-term memory (LSTM) network-using experimental data collected from a custom-built thermal energy harvesting platform. The SVR model demonstrates the best tradeoff between accuracy and computational feasibility, achieving an R2 of 0.978 and low prediction error on both internal and external test sets. While LSTM achieves the highest accuracy overall, its computational complexity limits its applicability on embedded hardware. These results suggest that well-regularized machine learning models such as SVR offer a promising and deployable alternative to conventional analytical models, particularly in embedded underwater systems where both accuracy and real-time performance are required.