Incremental transfer learning based on temporal-frequency convolution interaction for multi-task prediction of wind speed and wind power.

Abstract

In the domain of wind energy, predicting wind speed and power is a challenging and important task, yet they are closely intertwined. However, the temporal heterogeneity of wind speed and power time series, coupled with their time-varying statistical properties, limits the accuracy of the prediction model. To address this, an incremental transfer learning framework was proposed to perform multi-task prediction of wind speed and wind power simultaneously. This approach refines historical models with new data, capturing the statistical characteristics (e.g., distribution) of evolving data patterns to enhance prediction precision. Obtaining repetitive patterns in wind data, particularly for new wind farms with limited historical data, is inherently difficult. While model migration from other farms is possible, commercial confidentiality makes this less feasible, so our study focuses on modeling dynamic behavior from a limited number of samples, which also reduces the computational resource requirements. A temporal-frequency convolutional interactive neural network was introduced, which is then integrated into a parallel framework utilizing circular convolution. To capture the data's temporal dependencies, the classic gated recurrent unit was complement outside the parallel structure. The experiments using actual wind speed and power data from the largest wind farm in Northern China, testing for intervals of 15 min, 30 min, 45 min, and 1 h. The proposed model achieved the highest prediction accuracy, demonstrating its potential for practical engineering application. Furthermore, with the refinement of incremental transfer learning, the model's predictive performance is further enhanced over extended periods.