Vertical Wind Profile Estimation Using Hybrid Convolutional Neural Networks and Bidirectional Long Short-Term Memory

Determining the potential for developing a wind farm requires accurate knowledge of the vertical profile of wind speed (WS) at present and in the future time domain. In general, WS estimation at different heights requires the site-dependent parameters such as wind shear coefficient, atmospheric conditions, and roughness length. This paper examines the use of the Convolutional Neural Networks and Bidirectional Long Short-Term Memory (CNN-BLSTM) to estimate WS at different heights based on measurements at lower heights. To perform vertical extrapolation, the CNN-BLSTM uses measured WS at 10–40 m heights to predict WS at 50 m. The predicted WS at 50 m are further used along with the measured WS at 10–40 to extrapolate WS at 60 m. This procedure is repeated until the extrapolation of WS at 180 m is achieved. The performance of the CNN-BLSTM is compared with the multi-layer perceptron (MLP), support vector machine for regression (SVM), and the standard wind shear exponent (WSE)-based estimation. The extrapolated values obtained using these methods are compared with Light Detection and Ranging (LiDAR) reference system-based measurements. The coefficient of determination (R²) between extrapolated and actual WS values at height 180 m corresponding to CNN-BLSTM, MLP, SVM, and WSE methods are found to be 69.13%, 64.75%, 61.96%, and 55.96%, respectively.