Reliable short-term predictions of the wind power production are critical for both wind farm operations and power system management, where the time scales can vary in the order of several seconds, minutes, hours and days. This comprehensive study mainly aims to quantitatively evaluate and compare the performances of different Box & Jenkins models and backpropagation (BP) neural networks in forecasting the wind power production one-hour ahead. The data employed is the hourly power outputs of an N.E.G. Micon 900-kilowatt wind turbine, which is installed to the east of Valley City, North Dakota. For each type of Box & Jenkins models tested, the model parameters are estimated to determine the corresponding optimal models. For BP network models, different input layer sizes, hidden layer sizes, and learning rates are examined. The evaluation metrics are mean absolute error and root mean squared error. Besides, the persistence model is also employed for purpose of comparison. The results show that in general both best performing Box & Jenkins and BP models can provide better forecasts than the persistence model, while the difference between the Box & Jenkins and BP models is actually insignificant.
Generation of Statistical Scenarios of Short-term Wind Power Production
