WTCGRU: Short-term wind speed prediction based on window feature importance sampling

In short-term wind speed prediction, different features in the time window have different degrees of influence on the prediction results. However, the importance of different features in the time window is mostly treated equally, which results in the deficiency of responding promptly to the time-varying characteristics of wind speed and treating differently the time window feature weights that affect the prediction accuracy. Hence, a novel model called WTCGRU is proposed to improve the prediction accuracy by fully considering the influence of different features in the time window during wind speed prediction. In WTCGRU, the weights of different features in the time window are sampled through an adaptive genetic approach, which is called importance sampling, and fed into a hybrid model composed of GRU and TCN, to predict more precisely the short-term wind speed; This step is executed repeatedly until WTCGRU's prediction accuracy is higher than a predefined threshold. Finally, theoretical experiments are carried out on the wind speed dataset from Szeged of Hungary, and it is shown that the WTCGRU has outperformed the mainstream short-term wind speed prediction model TCN, BILSTM, CNN-LSTM, and ensemble-CNN.

I. Introduction

The overexploitation of traditional energy sources and the increasing demand for energy in today's society have made renewable energy more important than ever before. As a clean and renewable energy source, wind energy has received a lot of attention worldwide. But the efficient use of wind resources is highly dependent on the accurate prediction of wind speeds[1], which makes it increasingly necessary to accurately predict wind speeds, but due to the highly non-linear, intermittent and random nature of wind speeds, accurate prediction of wind speeds seem difficult. To address the challenges posed by the complexity of wind speed for accurate prediction, extensive research has been conducted on wind speed prediction models, typical include the following four categories: physics-based models, statistical-based models, machine learning or data mining-based models and decomposition prediction model[2].