Auto-optimization of Energy Generation for Wave Energy Converters with Active Learning

TL;DR

This paper introduces an auto-optimization control framework for wave energy converters that actively learns and adapts to changing ocean conditions to maximize energy output, outperforming traditional control methods.

Contribution

The paper proposes a novel dual control-based auto-optimization framework with active learning for WECs, effectively handling unknown and dynamic ocean conditions.

Findings

Outperforms model predictive control, extremum seeking, and Bang-Bang control in simulations.

Effectively learns and adapts to unknown regular and irregular wave conditions.

Demonstrates robustness and improved energy generation in diverse ocean scenarios.

Abstract

This paper presents an auto-optimization control framework for wave energy converters (WECs) to maximize energy generation under unknown and changing ocean conditions. The proposed control framework consists of two levels. The high-level controller operating at a longer time scale aims to maximize the average energy generation over several wave periods. The generated Power Take-Off (PTO) profile as the reference for the low-level physical system to follow. The new auto-optimization process leverages the parameterization of the non-stationary operation condition in WECs, establishing the relationship between the average energy generation and the key design parameters of the PTO force subject to the unknown wave parameters. The high-level controller is designed based on the concept of Dual Control for Exploration and Exploitation (DCEE) to quickly learn the unknown wave parameters by actively probing the ocean condition, while generating the optimal PTO profile. During this process, the uncertainty of the estimated wave condition is quantified and embedded in the optimization cost function to enable active learning. Simulation results under unknown regular and irregular waves demonstrate the effectiveness and robustness of this novel auto-optimization WEC systems with active learning, outperforming model predictive control, extremum seeking and classic Bang-Bang control approaches.