Short-Term Power Load Forecasting Based on LightGBM-VMD-SE and Inception-BiGRU-Attention

As the new power system evolves, there is a growing requirement for higher levels of accuracy in short-term power load forecasting within the electricity spot market. A short-term power load forecasting method based on LightGBM-VMD-SE and Inception-BiGRU-Attention is proposed. Firstly, the LightGBM is employed in determining the significance of different factors that affect power load. This enables the selection of highly crucial factors while minimizing the impact of less relevant features on prediction accuracy. Next, to address the effects of non-stationary in power load time series, the variational mode decomposition is utilized to break down the historical load sequence into several stationary intrinsic mode functions. Then, the complexity of each intrinsic mode function is examined using sample entropy, and the computational complexity is reduced by reconstructing the high-frequency and low-frequency components. Ultimately, an Inception-BiGRU-Attention model incorporating convolutional kernels of varying scales is developed to predict the high-frequency and low-frequency components individually, and predictions for both components are then merged to generate the final forecasted value. The effectiveness of the proposed method is verified using real load data from Hubei province, China. Experimental results demonstrate that compared to the VMD-SE-CNN-BiGRU-Attention model, the VMD-SE-Inception-LSTM-Attention model, the VMD-SE-Inception-BiGRU model and the Inception-BiGRU-Attention model, the proposed model enhances the prediction accuracy by 8.0%, 9.1%, 11.9%, and 13.3% respectively, thereby confirming the effectiveness in shor-term power load forecasting.