Robust deep neural network for wind speed prediction

Wind speed prediction is a basic requirement of wind energy generation with large generation capacity for large-scale wind power penetration. The intermittency and stochastic quality of wind speed leads to a big challenge for high penetration of wind power in electricity systems due to error-prone wind speed prediction methods. There are many artificial neural network (ANN) approaches proposed in the recent literature in order to tackle this problem. However, hand engineering features and applying shallow architectures can lead to poor prediction performance in these methodologies. Deep neural networks are ANNs with great generalization capability that can automatically extract meaningful features from the data with as less prior knowledge as possible. In this paper we propose a stacked auto-encoder (SAE) neural network for ultra-short-term and short-term wind speed prediction. To the best of our knowledge this is the first paper that applies deep learning on wind speed prediction of a wind site. Moreover, a rough regression layer is applied at the top of this model in order to deal with uncertainty factors existing in the wind speed data. Experimental results show significant improvement compared to other ANN methods that applied shallow architectures in the literature and the traditional SAE.