Wave Record Gap-Filling Using a Low-Rank Tensor Completion Model

Abstract

Abstract The offshore wind farm industry has identified further refinement of marine operations as necessary to realize the lower strike prices seen in recent subsidy auctions. This requires extending working times by taking advantage of weather windows, even when operating in more remote sites. Key to this will be increasing the accuracy of forecasts and live metocean data from site for effective and safe operation scheduling. In recent years, statistical models or deep learning-based models, by learning spatial and temporal patterns in observations have shown their potential to support or even partly replace numerical weather modelling to provide accurate forecasts. The present paper implements a machine learning approach utilizing a nonconvex low-rank tensor completion considering truncated nuclear norm algorithm (LRTC-TNN) which can characterize the spatial and temporal dependencies rooted in the data to fill gaps in measurement data from wave buoys. The performance is assessed by manually masking valid data thereby introducing three types of missing entries. The proposed method is shown to successfully fill datasets missing up to 20% of the data with R2 values exceeding 0.9. The results are also compared against other intuitive methods including linear interpolation, cubic spline interpolation, and mean historical. Compared to these, the present method is shown to have more reasonable trends. Finally, sensitivity considering the ratio of missing data, the type of missing data, and two hyper-parameters of the algorithm were compared to characterize their impacts on the results and advise potential improvements. This work demonstrates multivariate time series gap-filling with arbitrary missing ratio and sensor availability that can be extended and deployed to sensor networks for possible forecasting applications.