Very short-term wind power forecasting with neural networks and adaptive Bayesian learning

Abstract

This article presents an adaptive very short-term wind power prediction scheme that uses an artificial neural network as predictor along with adaptive Bayesian learning and Gaussian process approximation. A set of recent wind speed measurements samples composes the predictor’s inputs. The predictor’s parameters are adaptively optimized so that, at a given time t, its outputs approximate the future values of the generated electrical power. An evaluation of this prediction scheme was conducted for two tests cases; the predictor was set to simultaneously estimate the values of the wind power for the following prediction horizons: 5 min, 10 min and 15 min for test case n°1 and for the test case n°2, the prediction horizons were 10 min, 20 min and 30 min . The neural predictor performs better than the persistent model for both test cases. Moreover, the Bayesian framework also permits to predict, for a specified level of probability, the interval within which the generated power should be observed.