Predictive Modelling and Optimisation of Power Generation for the M4 Wave Energy Converter: A Deep Learning Approach

Abstract

This paper investigates the use of a nonlinear autoregression neural network for wave field predictions, and its implementation into a power-take off passive loading control system which tunes the damping coefficient for a wave energy converter. The wave energy converter considered in this study is a part of a multi-institutional demonstrator project which has seen the deployment of a moored multimodal multibody (M4) attenuator wave energy converter in King George Sound in Albany, Western Australia. The device consists of a 1-2-1 float configuration and is approximately 20 meters in length. The developed neural network was used to predict wave elevations and energy spectrums for 10-second and 20-second ahead of time intervals. Findings of this study show that the neural network was able to accurately predict up to 10 s intervals (where RMSE = 1.32E-02), however the accuracy of predictions fell for 20 s predictions (where RMSE = 5.20E-02). A linear numerical model of the prototype M4 device was used to find the optimal PTO damping coefficient for the observed wave fields at King George Sound. This allowed for optimisation of mean absorbed power for a generated 3-hour JONSWAP unidirectional timeseries using variable damping coefficients. Here, the power output was able to be increased by 106% for a significant wave height of 0.63 m and peak period of 3 s and resulted in an overall increase in capture width ratio across the 3-hour wave dataset.