Determination of the Representative Time Horizons for Short-term Wind Power Prediction by Using Artificial Neural Networks

Abstract

Wind power is one of the major renewable energy sources, and this source has reached to compete with conventional energies. Wind speed has very complex variations during different time horizons. Prediction of wind speed shows some uncertainties depending on atmospheric parameters, such as temperature, pressure, solar irradiation, and relative humidity. Additionally, wind turbines could not generate electricity at all wind speed values that are less than cut-in and greater than cut-out. These conditions add new uncertainties to the prediction of this meteorological parameter. In this case, wind power prediction from generated electricity data will be better than direct wind speed or other meteorological parameters. Generally, in engineering applications wind power prediction is based on hourly mean, in other words, one hour time horizon data. This time horizon is accepted as a reference and representative but physically this horizon does not represent data homogeneity and should be changed depending on location and time of year. The main aim of this article is to determine the highest representative time horizon to predict wind power in a considered system. First, artificial neural networks methodology is applied to generated power with a 1.5-kW wind turbine by using 1 and 60 minutes average data. It is seen that 35 minutes time horizon gives the best representative scale for wind power prediction in April and 15 minutes in August, respectively. During these months, errors between measured and testing values are decreased to 9.72 and 3.62%, respectively, for the mentioned time horizons. Especially, it is seen that mean values that are evaluated from less than 8-10 minutes and greater than 40 minutes time horizon data give high errors. In other words, using very short time horizons of power data cause high prediction variations and, as a result of these deviations, wind power prediction shows high errors.