Short-Term Power Load Forecasting in FGSM-Bi-LSTM Networks Based on Empirical Wavelet Transform

Abstract

In this paper, a prediction model based on Empirical Wavelet Transform (EWT) for FGSM-Bi-LSTM network is proposed to address the short-term power load forecasting problem. The model performs noise reduction on the data by combining time windows with EWT. Additionally, the model stability is enhanced by introducing the Fast Gradient Sign Method (FGSM) to generate adversarial samples. Finally, case experiments based on real-world power station load data are conducted. The results demonstrate that compared with prediction models such as LSTM, ARIMA, XGBoost, QR-GRU, and Transformer, the proposed FGSM-Bi-LSTM model reduces RMSE by 85.22%, improves MAE by 62.60%, and increases r by 9.83%. This demonstrates the strong generalization ability of the model to be extended to other time series prediction tasks.

Keywords

• Bi-LSTM
• EWT
• deep neural networks
• FGSM
• short-term load forecast
• smart grid