关键词: TimeGAN, 重大天气, 小样本扩充, 光伏出力预测, 时间序列

Abstract: Accurately predicting photovoltaic (PV) power generation during major weather events is crucial for ensuring reliable energy supply and grid stability. However, the uncertainty of major weather events results in a relatively small amount of historical data stored by the photovoltaic station regarding such events. Due to a shortage of historical data, the accuracy of predicting PV power under major weather conditions is relatively low. In response to the above issues, this paper proposes a small amount of historical data augmentation prediction method based on Time series Generative Adversarial Network (TimeGAN), which captures the complex time dependence relationship between PV power and weather conditions. Based on the existing small amount of historical data of the photovoltaic station, realistic time series data is generated to simulate the process of major weather events, and then PV power prediction is carried out. Through experiments and comparative analysis, the results showed that the prediction results of using TimeGAN to expand the small sample were better fitted than those of GAN expanded small sample data. When 25% of the data was expanded, MAE decreased by 1.14, with a reduction rate of 21%, and RMSE decreased by 1.09, with a reduction rate of 18%; When the data was amplified by 50%, MAE decreased by 1.08, with a reduction rate of 23%, and RMSE decreased by 0.99, with a reduction rate of 20%. This shows that the accuracy of TimeGAN amplification data has been significantly improved when processing time series data.

Key words: TimeGAN, Major weather events, Small sample expansion, Photovoltaic output prediction, Time series