Deep Learning-Driven Control Optimization for Wave Energy Converters via Real-time Sea State Prediction

Wave Energy Converters (WECs) are devices designed to harness the vast and untapped energy potential of ocean waves to generate electricity. According to studies, wave energy has an estimated global potential of 2 TW, making it a critical contributor to future sustainable energy solutions. However, efficient wave energy extraction requires advanced control strategies due to the nonlinear and stochastic nature of ocean waves.

Current control strategies for WECs primarily include passive control, model predictive control (MPC), and reinforcement learning-based approaches. MPC has been widely adopted due to its ability to optimize control forces while maintaining system constraints. Reinforcement learning (RL)-based control strategies have emerged as an alternative, leveraging self-learning algorithms to optimize control laws. Some other control approaches are also widely adopted in the WEC domain, such as PID, latching control, reactive control, etc.

A major limitation of all these control approaches is their dependence on the incoming wave signal, either through direct measurement or estimation methods. In practical conditions, measuring ocean waves in real-time is highly challenging due to the rapidly changing ocean dynamics. Our approach addresses this problem by estimating the incoming ocean waves using only the physical measurements of the WEC device—such as position, velocity and PTO force. These parameters are easier to measure and free from the complexities associated with direct ocean wave measurement. By leveraging these readily available WEC signals, our method makes real-time control more practical, robust, and adaptive to diverse ocean conditions.

The core of our method involves transforming the above-mentioned WEC motion data into time-frequency images using Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT). These representations capture the spectral characteristics of the WEC’s dynamic response to incoming waves, providing an image-based dataset for training a convolutional neural network (CNN) based regression model. Unlike direct time-series processing, CNNs can effectively extract spatial and frequency-domain patterns from these images, improving the robustness of Tp estimation.

To train our models, we generate a diverse dataset of WEC responses using simulations of a WaveBot WEC device under different sea conditions. To construct the dataset, we simulate WEC responses under a range of sea conditions using a JONSWAP wave spectrum. The generated time-frequency images are labeled with corresponding Tp and Hm values. A ResNet-18 CNN architecture is employed, replacing the standard classification head with a fully connected regression layer that outputs a single scalar value (Tp). The model is trained using Mean Squared Error (MSE) loss and optimized using Adam, ensuring fine-grained continuous Tp estimation.

By shifting the focus to CNN-based feature extraction from time-frequency images, our approach eliminates the need for explicit wave forecasting and complex numerical models. This method enhances control adaptability by dynamically adjusting PTO control settings to improve power extraction and operational reliability.