Learning-Based Robust Control for Offshore Wave Energy Converters
Mr Shuo Shi and Professor Ron Patton, University of Hull
Wave energy has the potential to contribute up to 2.1 TW power, particularly considering suitable high energy density locations worldwide. However, the high Levelized Cost of Energy (LCoE) of wave power impedes the development of commercial applications of Wave Energy Converter (WEC) devices due to the low energy conversion efficiency and high maintenance requirements. Optimal control of WEC can maximise the overall efficiency and guarantee the safety as well as reliability of WEC, hence reducing the LCoE. This PhD research focuses on Data-driven and Robust learning control method for wave energy converters to maximize the power take-off from waves. Using this learning control algorithm the Hull University Control & Intelligent Systems research team has participated in the International Wave Energy Control Competition - WECCCOM, organised by a consortium of institutions in R&D of marine renewable energy systems. A novel wave excitation force prediction and estimation method has been developed due to the requirement for implementing a power efficiency maximisation control of wave energy converters (WECs). The Hull team is ranked among the top 3 competitors in the competition.
(1) Short-term Wave Forecasting using Gaussian Process for Optimal Control of WECs
The Gaussian Process (GP) model for short-term wave forecasting is developed and shows better or comparative performance with Neural Network (NN) and Autoregressive (AR) modelling methods. The GP strategy is not only capable of forecasting the mean value of wave elevations but also provides the uncertainty of forecasting, which is beneficial to the safe operation, robust and optimal control of the WEC device.
(2) Robust Data-driven Estimation of Wave Excitation Force for WECs
A data-driven technique has been developed to estimate the wave excitation force (WEF), which uses a robust Bayesian filter in WEC hydrodynamic system which is described by non-parametric Gaussian process (GP) models. This modern way of incorporating the first principle model into a probabilistic framework is more robust than calculating estimates of a parametric function representation. Unlike sample-based non-linear Kalman filters, the means and covariances of joint probabilities can be directly computed based on analytic moment matching.
(3) Learning-based prediction-less resonating controller for WECs
We developed a data-efficient learning approach for the complex-conjugate control of a wave energy point absorber. Particularly, the Bayesian Optimization algorithm is adopted form aximizing the extracted energy from sea waves subject to physical constraints. Thealgorithm learns the optimal coefficients of the causal controller. Less than 20 function evaluations are required to converge towards the optimal performance of each sea state. Additionally, this model-free controller can adapt to variations in the real sea state and be insensitive and robust to the WEC modeling bias.