Assessing wind-wave forces on smooth and rough horizontal cylinders by neural networks

This work investigates the use of feed-forward neural networks (FNNs) to predict wave-induced hydrodynamic forces on smooth and covered by barnacles horizontal cylindrical structures. The study is based on a dataset of 4867 field-recorded sea states obtained from small-scale field experiments conducted at the Marine Energy Laboratory (MEL) of Reggio Calabria, a site well suited for scaled testing under wind-wave-only conditions. Two neural network models were developed. The first predicts the 95th percentile of the horizontal and vertical forces, while the second reconstructs the probability density function (PDF) of the force time series by estimating their standard deviation. Both models showed high accuracy compared to the measured force data, with the percentile prediction network achieving an average MAE = 0.83 N/m, MSE = 1.29 N2[jls-end-space/]/m2 and (Formula presented) and the standard deviation network achieving an average MAE = 0.34 N/m, MSE = 0.22 N2[jls-end-space/]/m2 and (Formula presented) for smooth and rough cylinders. The second approach also enables applications in structural monitoring. For comparison, a multiple linear regression (MLR) model was also implemented. While the MLR provided acceptable results, the FNNs provided better prediction accuracy, especially for nonlinear loading conditions caused by marine growth. Although a Rayleigh distribution was assumed for the extraction of the extreme values from the reconstructed PDFs, this approximation is recognized as a simplification. The results highlight the potential of neural networks to support the design and condition-based monitoring of offshore structures affected by changing surface conditions.