Posters

Abstract

The hub-height of commercial wind turbine has increased significantly over the last decade due to advancement in the technology. Commonly employed wind masts measure wind data upto 80 m, which is lower than hub-height of turbine installed in commercial wind farms. There is a need for extrapolation of wind data at hub-height for accurate prediction of wind power before deployment. Commonly used methods for extrapolation of wind speed are the Power Law and the Logarithmic Law, and their performance varies with climatic conditions as reported in literature. The aim of this study is to utilize Nonlinear Auto-Regressive with Exogenous Inputs (NARX) Recurrent Neural Network for extrapolation of wind speed, and to compare results with existing methods. The wind data having temporal resolution of ten minutes was measured according to standards at two sites for a duration of two years. Two sites representing power class 1 and 2 were selected with mean speed of 5.09 m/s and 7.42 m/s at 80 m height. The training of NARX was done by using the wind speed at 80 m as the output of the network while wind speed at three heights (20 m, 40 m, and 60 m), temperature at 5 m and 78 m, pressure at ground level, and relative humidity at 5 m. The NARX network was trained using three algorithms; Bayesian Regularization (BR), Levenberg-Marquardt (LM) and Scaled Conjugate Gradient (SCG). A parametric study was performed to select the optimum number of time delays and neurons. Performance of the network was evaluated by using relative Mean Biased Error (rMBE), relative Mean Absolute Error (rMAE) and relative Root Mean Squared Error (rRMSE) and results were compared with the Power Law and the Logarithmic Law. The optimum time delays and neurons were found to be 2 and 1, respectively. The error metrics for both sites indicated that BR algorithm performed best by giving lowest rMBE (-0.056% and 0.157% for power class 1 and 2, respectively), rMAE (1.451% and 0.883%), and rRMSE (2.140% and 1.452%). The rMBE for LM training algorithm was 0.076% and -0.117%, and the rMAE was 1.554% and 1.095% and the rRMSE for LM training algorithm was 2.261% and 2.126%. When SCG training algorithm was used, the rMBE was -0.364% and -0.190%, the rMAE was 2.368% and 1.707%, while the rRMSE was 4.814% and 2.761%. The rMBE for Power Law was 3.323% and 0.995%; the rMAE was 7.123% and 3.085%; and rRMSE was12.44% and 5.385%. The rMBE for Logarithmic Law was -0.212% and -1.058%, rMAE was 6.455% and 3.307%, and rRMSE was 10.413% and 5.170%, respectively. In comparison to the Power Law and the Logarithmic Law, NARX provides better performance at 80m height for both sites. These results highlight the application of NARX for extrapolating wind speed at higher hub heights and conducting onshore wind resource assessment. .