Vertical wind speed extrapolation model using long short‐term memory and particle swarm optimization

Abstract

Abstract Investigating the potential for deploying a wind farm requires accurate knowledge of the vertical profile of wind speed (WS), both temporal and with height. Commonly, WS estimations at different heights involves calculating site‐dependent parameters such as roughness length, atmospheric conditions, and wind shear coefficient. This study proposes a hybrid model to estimate WS at different heights based on measurements at lower heights using particle swarm optimization and long short‐term memory (PSO‐LSTM). The training procedure consists of two steps, namely the Levenberg–Marquardt (LM) and the PSO, to minimize the mean squared error. To perform vertical extrapolation, the PSO‐LSTM utilizes the measured WS at 10‐40 m heights to predict WS at 50 m. The predicted results at 50 m are then utilized along with the measured WS at 10–40 to extrapolate WS to 60 m. This process continues until the extrapolation of WS at 120 m is made. The PSO is compared with other optimization methods like genetic algorithm (GA), tuna swarm optimization (TSO), grey wolf optimization (GWO), sparrow search algorithm (SSA), and bald eagle search (BES) on improving the LSTM accuracy. The performance of the PSO‐LSTM is compared with the standard LSTM, feedforward neural network and the logarithmic law (LogLaw) based estimation. The extrapolated values obtained using these methods are compared with LiDAR system‐based measurements. The proposed method outperformed the other methods.

Keywords

• artificial neural networks
• logarithmic law
• long short‐term memory
• particle swarm optimization
• vertical extrapolation
• wind speed