Multi-Model Short-Term Load Forecasting Based on VMD-XGBoost-BiGRU

To enhance the predictive accuracy of ultra-short-term power load, this study introduces a hybrid forecasting model that integrates VMD (Variational Mode Decomposition, VMD) with a diverse set of machine learning techniques. Initially, VMD is employed for modal decomposition, coupled with sample entropy to assess component volatility, thereby facilitating data reconstruction into stationary and non-stationary components. For the stationary components, prediction is conducted using XGBoost (eXtreme Gradient Boosting, XGBoost). Meanwhile, non-stationary components are forecasted utilizing a BiGRU (Bidirectional Gated Recurrent Unit, BiGRU) neural network equipped with an attention mechanism. Through empirical case analysis, this approach demonstrates a reduction in the MAPE (Mean Absolute Percentage Error, MAPE) values by 28.27% and 21.66% respectively, compared to the basic GRU (Gated Recurrent Unit, GRU) and BiGRU models. This validation underscores the efficacy and precision of the proposed model in addressing non-stationary ultra-short-term load forecasting, thereby significantly enhancing the predictive performance of the model.(Abstract)