Gaussian process for wind power forecasting: Effect of training and test data size

Abstract

Probability is used to estimate various quantities in physics, and the Gaussian process is a probabilistic method. Since wind power forecasting is uncertain (for example, uncertainty is introduced by measurement noise), we use a random variable to represent it. This random variable represents uncertain quantities, and the probabilities assigned to it constitute its probability distribution. In probability theory and statistics, a Gaussian process is a stochastic process (a collection of random variables). Therefore, research on the Gaussian process (GP) and its applications is still ongoing. However, significant alternatives have been identified. Additionally, the application of this method in wind power prediction is limited. In this work, we investigate the impact of the size of the training and test data on the quality of the forecast. We therefore propose a methodology to study this approach. The simulation results were compared with the persistence method, Artificial Neural Networks (ANNs) and GP with Uncertain Inputs. We concluded that forecasts are uncertain statements whose uncertainty is verifiable. It is possible to adjust the degree of uncertainty of a forecast in light of new information (newly observed correlations). Moreover, the more training data we add, the more accurate the forecasts become (in the context of our wind power data). Moreover, the measured performance of our method is independent of the size of the test data. This methodology enabled us to investigate the impact of the size of the training and test data on the prediction quality, which helps to address the integration problem of the wind energy conversion system.