Accurate estimation and prediction of wind speed are helpful for power dispatching and optimizing control strategies. However, in newly-built wind farms, there is often a problem of insufficient historical data, which brings difficulties to precise modeling.

Although there are many ways to expand historical data, the statistical characteristics of wind speed data patterns (such as distribution features and quantiles) are constantly changing, which further hinders the modeling process under the condition of missing historical data. As shown in Figure 1, the distribution characteristics and quantile characteristics of the measured data change.

To address the time-varying statistical feature issues encountered in few-shot modeling, the research team innovatively constructed an incremental transfer learning multi-task prediction method based on temporal-frequency convolution interaction, as shown in Figure 2. This method introduces a temporal-frequency convolutional interactive neural network, integrates it into a parallel framework that utilizes circular convolution, and adds gated recurrent units.

The experimental results using the measured data show that, compared with the classic LSTM algorithm, the prediction error of the proposed method has been significantly reduced. Furthermore, the fine-tuning of the incremental transfer learning model indicates that the prediction effect can be further improved over a longer period of time. Some of the comparison results are shown in Figure 3.

The relevant research results are titled “Incremental transfer learning based on temporal-frequency convolution interaction for multi-task prediction of wind speed and wind power”. Published in Neural Networks on July 5, 2025.