An online transfer learning model for wind turbine power prediction based on spatial feature construction and system-wide update

Accurate prediction of wind turbine power is important for the safe operation of wind farms. However, most of the previous online transfer learning methods are partially updated and time-consuming. Here we propose a novel system-wide update online transfer learning model to overcome these shortcomings. To improve the multi-source data fusion accuracy, a new time trend quantification method is applied to expand the data source, a convolutional neural network multi-source data fusion method is proposed to reduce the dimension of data, and a Hilbert spatial feature construction method is used to construct spatial information of data. To achieve system-wide update and rapid prediction, we have deleted the weight unit of traditional method and added two data buffers. The results show that: (1) the proposed multi-source data processing method has the smallest mapping errors, which mean absolute error for all wind turbines is less than 32.1; (2) the proposed online transfer learning model has the highest prediction accuracy, which is higher than 92.5%.

Automated summary

A novel system-wide update online transfer learning model is proposed in this study to accurately predict wind turbine power. It applies various methods such as time trend quantification, convolutional neural network multi-source data fusion, and Hilbert spatial feature construction to improve data accuracy and reduce dimension. The model achieved high prediction accuracy of over 92.5%.