Deep belief network based deterministic and probabilistic wind speed forecasting approach. (15th November 2016)

• Record Type: Journal Article
• Authors: Wang, H.Z.; Wang, G.B.; Li, G.Q.; Peng, J.C.; Liu, Y.T.
• Is Part Of: Applied energy. Volume 182(2016)
• Page Start: 80
• Page End: 93
• Publication Date: 2016-11-15
• Subjects: Deep belief network -- Wind speed forecast -- Deep learning -- Quantile regression -- Wavelet transform; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2016.08.108
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: For the first time, deep belief network is designed for wind speed forecast (WSF). The nonlinear features in wind speed are used to improve forecast accuracy. The uncertainties of wind speed are evaluated using quantile regression. The competitive performance and high-stability of the proposed method were proved. Abstract: With the rapid growth of wind power penetration into modern power grids, wind speed forecasting (WSF) plays an increasingly significant role in the planning and operation of electric power and energy systems. However, the wind speed time series always exhibits nonlinear and non-stationary features, making it very difficult to be predicted accurately. Recognizing this challenge, a novel deep learning based approach is proposed for deterministic and probabilistic WSF. The approach is a hybrid of wavelet transform (WT), deep belief network (DBN) and spine quantile regression (QR). WT is employed to decompose raw wind speed data into different frequency series with better behaviors. The nonlinear features and invariant structures of each frequency are completely extracted by layer-wise pre-training based DBN. Then, the uncertainties in wind speed are statistically synthesized via the QR method. Case studies using real wind farm data from China and Australia have been presented. The comparative results demonstrate that the high-level nonlinear and non-stationary feature in the wind speed series can be learned better, and competitive performance can thus be obtained. It is therefore convinced that the proposed method has a high potential for practical applications in electric power and energy systems.