다중 계절성 전력 부하 예측을 위한 하이브리드 MSTL-SARIMAX 모델의 성능 분석

초록

Predicting power loads in university buildings is challenging due to the complex weekly and annual seasonality driven by academic calendars and seasonal shifts. In this study, we develop hybrid Multiple Seasonal-Trend decomposition using locally estimated scatterplot smoothing (LOESS) and Seasonal Autoregressive Integrated Moving Average with eXogenous variables (MSTL-SARIMAX) model for effective decomposition and interpretable forecasting. The model uses MSTL to separate the time series into trend-residual and seasonal components, which then serve as exogenous variables for the SARIMAX model in predicting the non-seasonal elements. Using actual university data, the proposed model outperformed standard SARIMAX and Fourier–ARIMA models. Although hyperparameter-tuned machine learning models, such as eXtreme Gradient Boosting (XGBoost) and Light Gradient-Boosting Machine (LightGBM), achieved the highest accuracy, our hybrid model offers a distinct advantage in terms of interpretability. By decomposing each component, it explains the basis of its predictions, making it highly valuable when causal analysis is as crucial as accuracy.

키워드