Very short term pitch angle optimization in wind turbines: A machine learning approach

Abstract

This paper proposes a pitch angle forecasting model based on the k-nearest neighbor classification. Air temperature, atmosphere pressure, wind direction, wind speed, rotor speed and wind power parameters were represented as a 6-dimensional attribute tuple in the forecasting model. Euclidean, Manhattan and Minkowski distance metrics for measuring the proximity between training and test tuples, mean absolute, mean absolute percentage, and normalized root mean square error metrics for measuring the forecasting accuracy were embedded into the forecasting model. The k-nearest neighbor classifier with Manhattan distance metric for k=1 achieved MAE, MAPE and NRMSE as 0.001°, 0.245% and 0.324%, respectively as the best forecasting accuracy. However, as the worst forecasting accuracy, MAE, MAPE and NRMSE were achieved as 0.015°, 3.236% and 2.613%, respectively for Minkowski distance metric and k=10.