A Two-stage Short-term Load Forecasting Method Based on Comprehensive Similarity Day Selection and CEEMDAN-XGBoost Error Correction

Abstract

The precise short-term load forecasting (STLF) represents a pivotal technology in maintaining the equilibrium between energy supply and demand, as well as enhancing the effectiveness of power system operations. And two key steps to accurately realize STLF process are the utilization of relevant historical data from similar days and the revision of forecasting errors. However, the current STLF methods solely rely on weather similarity as a parameter for identifying similar days, leading to imprecise outcomes when predicting loads that are less susceptible to weather-related variables. Furthermore, the existing error revising methods are insufficient in feature extraction of high frequency fluctuation error series, which limits the further improvement of accuracy. To this end, this paper proposed a two-stage STLF approach. Firstly, the concept of comprehensive similar day is constructed. Based on numerical weather prediction (NWP) and load sequential characteristics, the comprehensive similar day of the time period to be forecasted (TPF) are selected by morphological similar distance (MSD). Secondly, the load data of comprehensive similarity days will be input into long short-term memory (LSTM) as extra information to get preliminary load forecasting results. Finally, the error correction model based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and XGBoost is proposed for a further enhancement in accuracy. The validity of the proposed STLF method is verified by using the data set from Ausgrid in Sydney. Compared to directly using LSTM for prediction, the proposed method achieved 2.78% and 6.16% improvement in the accuracy of hour-ahead and day-ahead load forecasting.