Ensemble interpolation of missing wind turbine nacelle wind speed data in wind farms based on robust particle swarm optimized generalized regression neural network

The integrity of wind turbine nacelle wind speed data is of great value for wind farm maintenance and wind power prediction. However, for many reasons the nacelle wind speed data are missed from one time to another. It is difficult to design an appropriate interpolation model to accurately fill in missing wind speed data because the wind evolutions are highly nonlinear and non-stationary. In this paper, an innovative statistical approach, Robust Particle Swarm Optimized Generalized Regression Neural Network (RPSO-GRNN) algorithm is proposed to achieve the high-accuracy regeneration of missing nacelle wind speed data. Firstly, the Dynamic Time Warping (DTW) method, Pearson’s Correlation Coefficients (PCC) method, and Nearest Neighbor (NN) method are applied to evaluate the similarity of wind speed data between the wind turbine that contains missing wind speed data and other available turbines, constructing three candidate member models based on GRNN. Secondly, the RPSO algorithm is applied to optimize the GRNN’s parameters. Lastly, two superior ones of the three candidate member models are selected to construct an entropy weight-based ensemble estimation model. The experimental results with the dataset from a large wind farm in the Midwest region of the United States show that: (a) DTW is superior to the PCC method and the NN method in dealing with the nonlinear similarity of wind speed data; (b) The RPSO algorithm yields more practical and accurate structure and parameters of GRNN; (c) The ensemble model with entropy weight has a sound theoretical basis, achieving best estimation and stability.