Ultra-short-term load forecasting based on the combination of EEMD and Autoformer multi-model

Abstract

Ensuring the secure and stable operation of the power grid heavily relies on accurate and efficient load forecasting. To advance this endeavor, this study presents an ultra-short-term load forecasting methodology that merges the Ensemble Empirical Mode Decomposition (EEMD) technique with the Autoformer multi-model approach. Firstly, a comprehensive input feature matrix is crafted by selecting load data, historical weather data, and date information, which are meticulously preprocessed before analysis. Subsequently, the EEMD algorithm is enlisted to break down historical load data into distinct frequency components. Each frequency component, combined with weather data, undergoes individualized training and prediction within a separate model. The Autoformer model is harnessed for predicting lower frequency components, while the XGBoost model is employed for higher frequency components. In the final stage, the prediction outputs from each model are amalgamated and reconstructed to yield the ultimate load prediction. To expedite computation, a CPU/GPU heterogeneous collaborative parallel computing strategy is employed, enhancing the model's speed. The proposed approach is validated through real historical data sourced from a specific geographical area. The findings affirm its superiority over traditional models in terms of accuracy. The model showcases high-quality load forecasting capabilities, thereby establishing itself as a promising tool for ensuring the secure and stable operation of power grids.