Short-term prediction of wind vector at multi-heights via deep learning techniques based on marine measurements from Light-Detection-and-Ranging device

Accurate short-term forecasts of offshore wind vectors are essential for exploring wind energy in marine environment. Prior studies largely emphasize wind-speed prediction at a single height, while short-term forecasting of wind vectors at multi-heights has been insufficiently investigated. Meanwhile, despite the fast development of deep learning (DL) techniques, there is a lack of research on comparing the prediction performance of varied DL models. Using long-term LiDAR measurements over 10 altitudes (22–179 m), we benchmark 4 DL models, i.e., long short term memory (LSTM), Convolutional LSTM (ConvLSTM), Transformer, and Informer, for short-term forecasts of wind vectors (i.e., speed and direction) with varied lead times. A vector-decomposition strategy is adopted to remove directional wraparound discontinuities. Results demonstrate that: for the single-height forecasting scenario, ConvLSTM and Informer yield the most accurate speed predictions (R2 = 0.996, coefficient of determination), while Informer best predicts direction (R2 = 0.916); for the scenario at multi-heights, ConvLSTM provides the most accurate and vertically stable results, reflecting its advantage in spatiotemporal feature extraction. Meanwhile, the vector-decomposition strategy is found to work well. Results also show that prediction errors tend to be larger under weak/non-stationary wind conditions, and they also grow with enlarged lead time (e.g., ConvLSTM wind-speed Root Mean Square Error increases from 0.72 m/s at 30 min to 1.25 m/s at 120 min). These results offer useful guidance on model selection: ConvLSTM for multi-height speed, Informer for direction at short leads, and both for operational forecasting under marine conditions.