A Weekday and Weekend Separation-Based Algorithm for Short-Term Load Forecasting

Abstract

This paper presents a methodology for forecasting hourly loads over seven days, using a separation of weekday and weekend data. The data used for the forecasting included the date, weather, and load data from January 1, 2021, with the forecast being conducted for four weeks from November 1 to 28, 2021. The methodology involved separating the data into weekdays and weekends, conducting separate forecasts for each, selecting features for use in each process, and optimizing the forecasting model's parameters through hyperparameter tuning. To select features, the Shapley Additional Explanations method was used to calculate feature importance, and Pearson correlation coefficients were used to measure linearity between features. This allowed for the selection of input features with high importance while avoiding those with high linearity with other features. The parameters of the forecasting model were optimized through Grid search, and the optimal combination of the XGBoost forecasting model's Learning Rate, n estimators, and Max depth was assessed. After all processes were completed, forecasting was performed, and the forecasting error was presented through Normalized Mean Absolute Error, Mean Absolute Percentage Error, and Normalized Root Mean Squared Error. The results obtained using the proposed algorithm were compared with those obtained without using it. In future studies, the authors plan to investigate forecasting for special dates, such as public holidays, and explore the use of forecasting algorithms that consider the probability of input data.