Overcoming Data Scarcity in Wind Power Forecasting: A Deep Learning Approach With Bidirectional Generative Adversarial Network and Neighborhood Search PSO Algorithm

The precision and stability of wind power prediction (WPP) are critical for the grid-connected operation of wind farms. However, the insufficient availability of historical data poses challenges for traditional deep learning prediction models to accurately forecast for new-built wind farms (NWF) under the background of a substantial increase in wind power installed capacity worldwide. Hence, there is practical scientific significance in exploring high-precision prediction methods within the domain of NWF WPP. To address the challenge of few sample in WPP, a novel data-enhanced WPP method is proposed, which integrates BiGAN (BiGAN) module, self-attention mechanism (SAM) and neighborhood search particle swarm optimization (NSPSO). Within the data enhancement module, BiGAN is proposed to mitigate convergence difficulties and gradient instability encountered during the training of traditional GANs, thereby fostering closer alignment between the generated distribution and the real distribution. During the prediction stage, SAM is designed to obtain a new input matrix for weight allocation before BiGRU, enhancing its sensitivity to critical input information. Furthermore, to prevent SAM-BiGRU from succumbing to local optima, the Dense layer is optimized by the NSPSO algorithm to improve the prediction accuracy. Extensive experimental results in two scenarios demonstrate that the proposed approach surpasses other advanced methods to a certain extent, achieving one-step-ahead prediction accuracy rates of 0.9775 and 0.9810, respectively. This study provides novel ideas to the field of WPP and demonstrates the potential of the proposed model to improve the accuracy of wind farms in power prediction, especially for those with limited historical data.