Unlocking few-shot wind speed prediction through a novel end-to-end transfer learning paradigm based on decomposition and gating information fusion

Abstract

Accurate wind speed prediction is one of the key technologies for achieving intelligent and sustainable development in the engineering field. In the field of wind speed prediction, we are confronted with a challenging few-shot prediction problem. Specifically, due to the fact that some wind turbines in wind farms are newly established or there is data loss during the data collection process, these turbines only contain a small amount of wind speed data. This scarcity of data poses great difficulties for the prediction work, and traditional prediction methods often fail to achieve the desired prediction accuracy. In order to overcome the above difficulties, we propose an novel prediction paradigm of end-to-end transfer learning based on data decomposition and gated information fusion. We use the Fourier transform to find the source domain similar to the target domain to achieve feature alignment. Then, we pre-train the model on the source domain and transfer this model to the target domain, thus solving the problem of low prediction accuracy when directly predicting the target domain. In the first step, the data is decomposed and denoised by using the Variational Mode Decomposition. According to the sample entropy, the decomposed data is reorganized into three frequency components. Each component is input as an independent channel into the end-to-end prediction model. Firstly, the features of each channel are expanded to a high-dimensional space through the Multilayer Perceptron. Then, the gating mechanism is utilized to mix the features of the three channels into the features of one channel, thus achieving information fusion. Finally, the prediction result of the end-to-end model is output through the Gated Recurrent Unit. In the second step, the model pre-trained on the source domain is transferred to the small-sample target domain. The Dynamic Time Warping and cosine similarity are used to quantify the similarity of each channel between the two domains. The parameters of the channels with high similarity are locked, and at the same time, the parameters of other channels are fine-tuned to output the final prediction result. In addition, multiple sets of comparative experiments conducted using the wind speed data from wind farms in Queensland, Australia, have demonstrated the superiority of this prediction paradigm. Our strategy outperforms various baseline models in all three sets of data. Moreover, ablation experiments have proven the effectiveness of each component in this framework in improving prediction accuracy, opening up a new path for solving the difficult problem of few-shot prediction in practical engineering.