|
Abstract
|
Accurate and timely prediction of photovoltaic (PV) power generation is crucial for ensuring the stability of power systems. To address the challenge of limited PV power data on rainy days, this paper proposes a method using a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) for data augmentation. The introduction of a progressive gradient penalty strategy mitigates the issue of gradient vanishing in generative adversarial networks. To forecast PV power generation, a Bidirectional Long Short-Term Memory neural network (BiLSTM) is employed. Additionally, an adaptive hyperparameter optimization method based on the Sparrow Search Algorithm (LSSA-BiLSTM) is introduced, enhancing the model's predictive performance. The optimization ability of the Sparrow Search Algorithm is further improved by incorporating a chaos mapping strategy. The developed model is implemented on a 1MW PV power station and compared with four baseline prediction models (BiLSTM, LSTM, GRU, SVM). Results indicate that, under uniform initial parameters across all models, utilizing LSSA for hyperparameter optimization yields the best predictive outcomes. Employing WGAN-GP for proportional augmentation of rainy-day datasets results in a 15.3% reduction in Mean Absolute Error (MAE) and a 6.15% reduction in Root Mean Square Error (RMSE). This suggests that WGAN-GP effectively learns the features and distribution patterns of PV power data, significantly improving the accuracy of predictions on rainy days.
|