Abstract
Since wind flow is a nonlinear phenomenon, it is generally difficult to predict how the wind at a certain point will change at the next moment. However, since the wind speed fluctuation near the ground surface appears as a part of the turbulence phenomenon in the atmospheric boundary layer, it is not a completely random but has "some feature" accompanying the passage of the turbulent structure. In this study, we tried to predict wind speed fluctuation up to 10 seconds ahead by learning the "feature" of wind speed fluctuation using LSTM (Long Short-Term Memory) which is one of Deep Learning. In addition, considering the nature of the turbulent flow in the ground surface layer, we examined the change in accuracy of prediction which depends on input conditions of LSTM. Although the accuracy of prediction decreases as the lead-time is longer, it has been confirmed that appropriate setting of learning time length and adding the vertical wind speed to the input condition contributes to improving prediction accuracy of the wind speed.
