A Model Free Control Based on Machine Learning for Energy Converters in an Array

Abstract

This paper introduces a machine learning based control strategy for energy converter arrays designed to work under realistic conditions where the optimal control parameter can not be obtained analytically. The control strategy neither relies on a mathematical model, nor does it need a priori information about the energy medium. Therefore several identical energy converters are arranged so that they are affected simultaneously by the energy medium. Each device uses a different control strategy, of which at least one has to be the machine learning approach presented in this paper. During operation all energy converters record the absorbed power and control output; the machine learning device gets the data from the converter with the highest power absorption and so learns the best performing control strategy for each situation. Consequently, the overall network has a better overall performance than each individual strategy. This concept is evaluated for wave energy converters (WECs) with numerical simulations and experiments with physical scale models in a wave tank. In the first of two numerical simulations, the learnable WEC works in an array with four WECs applying a constant damping factor. In the second simulation, two learnable WECs were learning with each other. It showed that in the first test the WEC was able to absorb as much as the best constant damping WEC, while in the second run it could absorb even slightly more. During the physical model test, the ANN showed its ability to select the better of two possible damping coefficients based on real world input data.

Digital Twin for the Prediction of Extreme Loads on a Wave Energy Conversion System

“…For the estimate of important hydrodynamic coefficients, an ANN was created to predict the reflection coefficient of a submerged breakwater, and it was successfully validated based on experimental data [43]. Model-free ANN methods were developed and applied to control of wave energy arrays by [44,45], and several other approaches of using model-free reinforcement learning techniques in wave energy systems have recently been published [46][47][48]. Even if these works have not used the concept of DTs, they share the common idea of modeling or optimizing wave energy systems using innovative and model-free machine learning techniques.…”

Advances and Challenges in Wave Energy Park Optimization—A Review

“…In the former experiment, the surface elevation was tracked by videogrammic measurements, and the experimentally measured interaction factor was presented for 1-2 floats moving in heave and surge (Nader et al, 2017). In the latter, both a linear damping and an advanced control algorithm based on machine learning and artificial neural networks were used as PTO systems (Thomas et al, 2018b). The collaborative control algorithm required no previous knowledge of the incident waves but still increased the energy absorption through communication between the WECs (Thomas et al, 2018a), and the performance of the WECs with linear damping was compared for three array layouts by Giassi et al (2019b).…”

Towards Real-Time Reinforcement Learning Control of a Wave Energy Converter

“…After training, the identified model is coupled with standard strategies used for the control of WECs, e.g., resistive or damping control in [17], reactive or impedance-matching in [18] and latching control in [19,20]. On the one hand, some studies have proposed the use of neural networks to find the optimal parameters for impedance-matching control on a time-averaged basis [21], thus being readily applicable to the centralised control of multiple WECs [22]. On the other hand, other works have focused on real-time control [19,20,23], exploiting the capability of neural networks to handle the predicted wave elevation over a future time horizon, similar to MPC.…”