Prediction of wind shear layer for dynamic soaring by using proper orthogonal decomposition and long short term memory network

Abstract

The wind shear layer is a naturally formed airflow that enables the albatross to soar for six days at almost no cost. The modeling and prediction of the wind shear layer can be very helpful for a long-endurance flight (dynamic soaring), but the existing studies usually ignore the turbulence structures of wind shear layers. In this paper, the wind shear layer on the leeward side of the ridge is simulated by a large eddy simulation (LES) method to analyze the turbulence structures. In the numerical simulation, the three-dimensional (3D) elevation data of the mountain is used as the topography at the bottom and the synthesized turbulent velocity is used as the inlet boundary. Because of the huge computational cost of 3D simulations, a data-driven predicting framework is also established to reduce the cost and maintain the prediction accuracy, which includes an offline training stage and an online forecasting stage. In the offline stage, the proper orthogonal decomposition (POD) is used to extract features from the LES data of wind velocity fields and the obtained POD coefficients are used to train the long short term memory (LSTM) networks. In the online stage, the future wind fields are predicted by the trained LSTM networks in the noisy and real-time environment. In conclusion, this paper analyzed the physical characteristics of the wind shear layer on the leeward side of the ridge and provided the accurate prediction for these characteristics.