Session: 09-02-03: Wave Energy - Design and Performance Analysis 2

Paper Number: 105123

105123 - A Preliminary Study of Learning a Wave Energy Converter System Using Physics-Informed Neural Network Method 

Physics-informed neural network (PINN) is a new type of neural network method that can be used to solve the physical problem by providing a given data-set in a machine learning process with embedded physics information directly described by differential equations. As a result, the PINN method does not require collecting data in advance, and one can trains the model using information directly obtained from the differential equations and constraints. Because of its powerful graphics processing unit (GPU) capability and modeling flexibility, the PINN method has gained significant interest in many research areas. The method has been applied to solve the Navier-Stokes equations for CFD simulations, including lid driven and dam breaking problems, and the governing equation of motion for mass spring damper systems, which is often used to represent the common wave energy converter (WEC) systems. Neural networks contain a set of linear transformations with non-linearities from the physical problem embedded implicitly. When using PINN method for WEC application, this suggest that those nonlinear forcing terms, including the convention integral terms and viscous drag terms, and those from complex array effects interaction, can be directly linearize in the system of equations, and the method can be beneficial for WEC control application. However, the neural networks method is very computationally demanding. Nevertheless, solving the linearized system of equations can be effectively speed up through the use of GPU.

In this study, a simple heaving sphere is simulated using a PINN method, which is developed using the MATLAB Deep Learning toolbox. The PINN method trains the model by enforcing the output of the response of the WEC, which is a function of time, to fulfill the governing Cummins equation and the initial conditions. The objective of this research is to investigate the use of PINN method for the analysis of wave energy converter. A series of PINN simulations is performed, including a decay test and a set of wave conditions, and the results are compared to those obtained from WEC-Sim. The study is focused on numerical benchmarking, in particular the collection points resolution study, overall accuracy of PINN method and the potential GPU acceleration. Finally, a discussion on potential advanced techniques used in the PINN, including adaptive learning and the neural tangent kernel algorithm to balance the scales of the loss terms and improve the model accuracy and stability, and the effectiveness of GPU acceleration will be included.

Presenting Author: Bo-Chen Chen National Yang Ming Chiao Tung University

Presenting Author Biography: Mr. Bo-Chen Chen is a Master's Student at National Yang Ming Chiao Tung University in Taiwan

Authors:

Bo-Chen Chen National Yang Ming Chiao Tung University
Yi-Hsiang Yu National Yang Ming Chiao Tung University, Department of Civil Engineering