Ultra Short Term Load Power Prediction Considering Feature Recombination and BiGRU-Attention-XGBoost Model

The ultra-short-term load forecasting, as a basic component of power system forecasting, is of great significance for production scheduling planning. However, the power load, being nonlinear, time-varying and uncertain in nature, necessitates the improvement of to prediction accuracy by fully exploiting its potential characteristics and conducting separate prediction. In this paper, a combined forecasting model of BiGRU-Attention-XGBoost based on ALIF-SE is proposed. The historical load series are decomposed into periodic series, fluctuation series and trend series based on ALIF-SE. An attention mechanism is employed to improve the BIGRU model. It is then combined with XGBoost model to construct a power load forecasting model based on time-varying weight combination. The experimental results demonstrate a significant improvement in prediction accuracy of the input model after ALIF-SE treatment. Moreover, the combined model exhibits good prediction effect for both working days and holidays, with the majority of the prediction errors below 5%. The mobility of the proposed method is verified by comparing the results of experiments with different load data sets. The experimental results demonstrate the effectiveness, accuracy, and feasibility of the proposed model.