Wind direction prediction based on nonlinear autoregression and Elman neural networks for the wind turbine yaw system

In this article two wind direction prediction models based on nonlinear autoregression (NAR) and Elman neural network (ENN) were proposed, using modified error back propagation optimization algorithm and gradient descent optimization algorithm respectively, aiming to enhance prediction accuracy and reduce wind direction error. By establishing neural network wind direction prediction models, and the historical data of wind turbines are used for verification. The model has demonstrated a high prediction accuracy of wind direction determination (up to 1.79°), comparing with the results obtained by traditional instruments (exceeding 5°). In most cases, the NAR neural network has better performance, while the Elman neural network is slightly better in a few cases. Due to their good performance, both models can be further developed according to actual situations. The modeling in Matlab/Simulink confirmed that neural network-based strategies can accurately and quickly predict wind direction changes, which significantly improves the accuracy and performance of the yaw system. This not only reduces the yawing process time to a certain extent, but also allows to effectively reduce the number of yawing in time and extend the lifetime of wind turbine mechanical components, at the same time increasing the output power by more than 6 %.