Wind-speed prediction model based on variational mode decomposition, temporal convolutional network, and sequential triplet loss

Abstract

Accurate wind-speed prediction is a persistent challenge in efforts to ensure the safe and stable operation of wind-power systems. Prediction lag is one of the greatest obstacles to accurate wind-speed prediction. This study established a novel wind prediction model based on the variation mode decomposition, temporal convolutional network, and sequential triplet loss. The proposed model primarily includes two parts: the variational mode decomposition part used for noise reduction and feature extraction, and the temporal convolutional network part used for short-term wind-speed prediction based on dilated casual convolutions and residual connections. In addition, to address the prediction lag problem, this study develops a new loss called sequential triplet loss for sequence prediction, which can push the target value away from the unideal point. To verify the prediction performance of the proposed model, several models were used for comparison. The experimental results reveal that the proposed model (1) has considerably better prediction performance than the comparison models, (2) can reduce the prediction lag, and (3) the variational mode decomposition can improve the prediction accuracy.