Short-term predictions of multiple wind turbine power outputs based on deep neural networks with transfer learning

This paper proposes a novel deep and transfer learning (DETL) framework, which enables a more efficient development of data-driven wind power prediction models for a group of wind turbines. In DETL, a transfer learning scheme is developed to boost computations in modeling wind power generation processes from a data-driven perspective and derive latent features for conducting power predictions. To perform the transfer learning, a new data organization scheme, which separates a batch of wind turbine datasets into a source domain and multiple target domains, is adopted. Based on the source domain, the DETL attempts to extract homogeneous characteristics of multiple wind turbine system dynamics via developing a base Auto-encoder (AE), whose architecture is adaptively determined. Next, the DETL aims to specify heterogeneous characteristics among individual wind turbine system dynamics via learning target domains, which converts the base AE model into multiple customized AE models. Finally, the customized AE model representing system dynamics of each wind turbine is extended to conduct multi-step wind power predictions by additionally incorporating temporal features and prediction targets. Field data collected from 50 wind turbines in commercial wind farms are utilized to verify the proposed DETL. Computational experiments validate that the DETL outperforms conventional training methods on developing a batch of prediction models with a higher prediction accuracy and faster training speed.