Deep neural networks for time series prediction with applications in ultra-short-term wind forecasting

The aim of this paper is to present deep neural network architectures and algorithms and explore their use in time series prediction. Existing and novel input variable selection algorithms and deep neural networks are applied for ultra-short-term wind prediction. Since gradient-based optimization starting from random initialization often appears to get stuck in poor solutions, recent research effort aimed at training methods for such deep networks is summarized. Shallow and deep neural networks coupled with two input variable selection algorithms are compared on a ultra-short-term wind prediction task. Initial results show that deep neural networks outperform shallow ones. Depth adds additional computational cost and input variable selection use reduces it. Keywords—Deep Neural Networks, Input Variable Selection, ultrashort-term, wind forecasting