On the Calibration of Metocean Time Series Using Machine Learning

Abstract

Nowadays, numerical model data is one of the primary inputs to all metocean studies, whether for deep-water locations or coastal applications. This paper presents the use of machine learning to calibrate long term metocean time series of wind and wave parameters obtained from numerical models against measurement records, usually covering shorter periods. We present the added value of machine learning compared to standard calibration methods to improve data used as primary input to both operability studies and engineering design studies. Time series of wind and wave parameters obtained from global numerical hindcast data sets are compared to oceanographic buoy measurements. We investigate the improvement brought by machine learning methods on the quality of the calibrated populations for the bulk of the distributions, but also the agreement between the calibrated data and the measurements for extreme events, not only for peak values but also for storm profiles. We evaluate the reliability of the method by comparing the results over different periods at 1 location and with varying length of training, validation and test sets.