Data-driven deep learning model for short-term wind power prediction assisted with WGAN-GP data preprocessing

Accurate data-driven wind power prediction plays an increasingly key role in ensuring the stable operation of wind farms (WFs). However, the historical data of WFs are often limited and noisy, which can significantly impact prediction performance. To ensure the accurate wind power prediction, this paper presents a novel phased model that includes sequence preprocessing, sequence decomposition-reconstruction, and hybrid prediction model. The purpose of sequence preprocessing is to enhance the quality of historical data by combining random sample consensus and isolation forest algorithms for noise screening and a missing value imputation method considering the randomness of the generator in Wasserstein generative adversarial network with gradient penalty (WGAN-GP) model is proposed. To reduce the complexity of the sequence and increase the computational speed, a novel combination method of improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and sample entropy (SE) is proposed, resulting in a high-frequency subsequence and a low-frequency subsequence. Based on temporal convolutional network (TCN), a novel hybrid deep learning model is proposed to handle complex time series. By combining TCN with transformer (TR) or bidirectional gated recurrent unit (BiGRU), the models become more suitable for wind power prediction, and they are employed for different subsequences considering the frequency characteristics. The proposed framework is validated in comparison with 18 other models, the results confirm the efficacy and superiority of the proposed framework, achieving up to 73.50% prediction error (RMSE) reduction.