Experimental Insight and data-driven modeling for predicting cyclic lateral response of monopiles supporting offshore wind turbines

Monopiles support nearly 80% of global offshore wind towers, with demand rising as offshore wind energy grows. Predicting stiff monopile rotation under millions of low-amplitude lateral load cycles is challenging. Conventional methods like the p–y approach, developed for long flexible piles, usually fail to capture such behavior. This study addresses the existing gap by creating precise and interpretable machine learning (ML) models based on high-quality experimental observations to forecast monopile head rotation under a range of cyclic lateral loading scenarios, tentatively simulating the offshore environment. Using a wide-ranging experimental dataset of 7800 rotation measurements from controlled 1 g model tests, ten predictive models were developed, including multilinear regression (MLR), nonlinear regression (NLR), and machine learning (ML) algorithms such as artificial neural networks (ANN), recurrent neural networks (RNN), decision trees (DT), random forests (RF), and support vector regression (SVR) approaches. The effects of four key input variables—number of cycles ( N ), loading frequency ( fr ), load amplitude ( α ), and directionality ( β )—were evaluated through SHAP values, perturbation, and permutation sensitivity methods. The robustness and generalizability of the proposed models were confirmed through a k-fold cross-validation technique. The RF demonstrated the best balance of accuracy and generalization as indicated by R 2 and MSE indices, followed by DT, SVR with radial basis function, and ANN . β appeared as the most significant input variable for the predictive models. Closed-form equations derived from shallow ANN models offer reliable rotation predictions. These findings support ML as a practical and interpretable estimation tool for the performance of monopiles.