Physics-Informed Gaussian Processes for Wave Loading Prediction

Abstract

The quantification of wave loading is an important step within the estimation of fatigue accrual within offshore structures. The direct measurement of wave loads can be both challenging and expensive, often requiring the installation of bespoke systems if at all possible. The estimation of wave loads based on data from other sensors, more commonly found on offshore structures (e.g. wave radars) is therefore highly desirable. This paper presents an experimental study of a monopile structure within a wave tank, instrumented with accelerometers, strain gauges, a force collar, wave gauges and a velocimeter subject to a range of wave conditions. Here the dataset is used to construct models which combine data-based Gaussian process NARX models with linear wave theory. The novel model structures presented rely on only wave gauge data as an input and achieve improved predictive performance over purely data-based approaches across a range of wave states.