WindFix: Harnessing the power of self-supervised learning for versatile imputation of offshore wind speed time series

Abstract

The accurate prediction of offshore wind speed is crucial for effective wind energy management. Deep learning models, which utilize large-scale wind speed time series, have emerged as a prominent approach for this task. However, the availability of wind history and other oceanic data is often capriciously incomplete, with intermittent gaps in the time series due to transmission difficulties or measurement device malfunctions. This presents a significant challenge for both data preparation and model training. In this paper, a self-supervised framework, WindFix, is proposed for continuous missing-value imputation of offshore wind time series using meteorological features from both the self-spot and neighboring spots. This cutting-edge innovation integrates multiple masking techniques with enhanced transformer models, resulting in a highly efficient and versatile process that adapts seamlessly to various scenarios. The numerical results validated on wind dataset from Pacific Ocean show that the model can serve for different missing types with a mean squared error around 0.002 for the simplest case. The comprehensive analysis also provided in-depth explanations on the different patterns observed in various imputation types, attributing these to the allocation of attention weights on features.