A novel discriminated deep learning ensemble paradigm based on joint feature contribution for wind speed forecasting

Abstract

Accurate short-term wind speed forecasting is essentially important for wind turbines operation and wind power generation. Decomposition-ensemble learning method has been widely explored in wind speed prediction. On the basis of this method, this research proposes a new modeling paradigm integrating decomposition, prediction and ensemble, comprehensively improves the original architecture of decomposition-ensemble learning method. A novel discriminated learning neural network is constructed to create contrapuntal input interfaces according to the contribution of joint feature components to the original sequence. On this basis, the parallel operation of prediction models with different complex structures is realized, so as to enhance the operation efficiency of the model. Finally, a turbine controlling strategy is designed in the light of the output of the model to ensure the safety and stability of the turbine controlling system. Three data sets of wind farms in Xinjiang, China are used for empirical test. The evaluation indexes and statistical test results demonstrate that the proposed model has superior approximation to actual values and stronger adaptability. The root mean squared errors of the proposed model on the three data sets are 0.63288, 0.44041 and 0.55258 respectively, which are superior than those baseline models, and the proposed model makes an average reduction of 75.45%, 48.41% and 44.62% on the root mean square error compared with other models in published literatures. To sum up, the proposed model can provide a valuable reference for wind power prediction.