Stacking integration algorithm based on CNN-BiLSTM-Attention with XGBoost for short-term electricity load forecasting
Shucheng Luo, Baoshi Wang, Qingzhong Gao, Y. Wang, Xinfu Pang
Abstract
Improving the accuracy of electric load forecasting is critical for grid stability, industrial production, and residents' daily lives. Traditional short-term load forecasting methods often struggle to fully capture the long-term dependencies and deep-seated features in unknown datasets, thus limiting their generalization ability. In this paper, we propose an algorithm for short-term power load forecasting based on the stacking integration algorithm of Convolutional Neural Network-Bidirectional Long Short-Term Neural Network-Attention Mechanism (CNN-BiLSTM-Attention) with Extreme Gradient Tree (XGBoost). First, an adaptive hierarchical clustering algorithm (AHC) selects a dataset with similar day characteristics. Then, combined with influencing factors, the Stacking integrated algorithm based on CNN-BiLSTM-Attention and XGBoost is employed for forecasting short-term load data. Finally, the integrated algorithm model was applied to the multi-feature load dataset in the Quanzhou area from 2016 to 2018. Comparative analysis showed that MAPE could be reduced by 5.88–69.40 % in the four selected typical days compared to the comparative algorithm, significantly improving load forecasting accuracy.