A Comprehensive Evaluation of Physics-Informed Long Short-Term Memory Model for Ocean Current Prediction

Abstract

In an era where renewable energy sources are essential for sustainable development and ensuring the quality of life for future generations, ocean energy has emerged as a significant research focus. This is due to its vast potential for energy production from currents, waves, and tides. A primary challenge in harnessing this energy lies in the inadequacy of accurate assessments using near-real-world ocean models. This study addresses this challenge through the implementation of Physics-Informed Long Short-Term Memory (PI-LSTM) models, which integrate physics-based equations with machine learning techniques. The dataset used in this study comprises 25,600 data points, with 100 time points for each of the 256 location points, ranging within [0,1) for time and [-1,1] for location. Simulation data was generated by solving Burger’s 1D equation, and a custom loss function was incorporated to embed Burger’s equation as a physics loss term to train the model. The study's primary objective was to predict the velocity component and evaluate the performance of PI-LSTM models against traditional LSTM models and current Physics-Informed Neural Network (PINN) models based on validation loss values. Results demonstrated that PI-LSTM models could predict the velocity component with a Mean Squared Error (MSE) as low as 6.7e-5, outperforming basic LSTM models in 55.56% of cases and PINN models in 77.78% of cases. Additionally, PI-LSTM models required only half the number of epochs in the best cases compared to basic LSTM models, indicating superior training efficiency. The findings suggest that PI-LSTM models not only achieve lower validation loss values but also demonstrate enhanced training efficiency by incorporating physical laws through custom loss functions, capturing the underlying dynamics of oceanic processes more accurately with reduced computational effort. In conclusion, PI LSTM models show considerable potential as a superior tool for modeling ocean energy systems, offering improved predictive performance and lower computational costs. This study underscores the viability of PI-LSTM models in advancing ocean energy research, providing a robust framework for more effective energy assessment and harnessing, and highlighting the promising impact of integrating physics-informed methodologies within machine learning models for the future of renewable energy