Offshore floating photovoltaics (OFPVs) show great potential for future renewable energy generation by complementing existing wind energy. To maximize power output, it is essential to accurately predict the response of these structures under wave-induced forces.

Although finite-element fluid-structure interaction (FE-FSI) models can simulate the response of very large floating structures (VLFSs) with high accuracy, they are computationally expensive and time-intensive. To address this challenge, this thesis develops and optimizes three surrogate deep learning models: a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM) network, and a hybrid CNN-LSTM model. An iterative trial-and-error approach was employed to fine-tune the hyperparameters and improve model performance.

The CNN model accurately captured spatial features but struggled with generalization across varying sea states. The LSTM model captured temporal patterns well, but had lower accuracy and longer training times. The hybrid CNN-LSTM model combined the strengths of both approaches, achieving a low weighted absolute percentage error (WAPE) of 0.98 %. All models performed well under typical wave conditions but exhibited reduced accuracy in extreme scenarios, particularly for low-amplitude tilts. A sensitivity analysis indicated that these errors were likely caused by insufficient representation of extreme sea states in the training data.

The hybrid CNN-LSTM model was selected as the model that most accurately predicts the response of VLFSs due to its balanced performance, minimal prediction time, and generalization capability across typical operating conditions. While the model shows high reliability for typical sea states, its performance declines for extreme conditions, such as extremely large waves or very small tilts. Despite reduced reliability in extreme conditions, the model provides a promising tool for real-time OFPV response prediction and preliminary studies.