Vertical extrapolation of Advanced Scatterometer (ASCAT) ocean surface winds using machine-learning techniques

Abstract

Abstract. The increasing demand for wind energy offshore requires more hub-height-relevant wind information, while larger wind turbine sizes require measurements at greater heights. In situ measurements are harder to acquire at higher atmospheric levels; meanwhile the emergence of machine-learning applications has led to several studies demonstrating the improvement in accuracy for vertical wind extrapolation over conventional power-law and logarithmic-profile methods. Satellite wind retrievals supply multiple daily wind observations offshore, however only at 10 m height. The goal of this study is to develop and validate novel machine-learning methods using satellite wind observations and near-surface atmospheric measurements to extrapolate wind speeds to higher heights. A machine-learning model is trained on 12 years of collocated offshore wind measurements from a meteorological mast (FINO3) and space-borne wind observations from the Advanced Scatterometer (ASCAT). The model is extended vertically to predict the FINO3 vertical wind profile. Horizontally, it is validated against the NORwegian hindcast Archive (NORA3) mesoscale model reanalysis data. In both cases the model slightly over-predicts the wind speed with differences of 0.25 and 0.40 m s−1, respectively. An important feature in the model-training process is the air–sea temperature difference; thus satellite sea surface temperature observations were included in the horizontal extension of the model, resulting in 0.20 m s−1 differences with NORA3. A limiting factor when training machine-learning models with satellite observations is the small finite number of daily samples at discrete times; this can skew the training process to higher-/lower-wind-speed predictions depending on the average wind speed at the satellite observational times. Nonetheless, results shown in this proof-of-concept study demonstrate the limited applicability of using machine-learning techniques to extrapolate long-term satellite wind observations when enough samples are available.