A Two-Stage Calendar-Aware Random Forest Framework for Robust Short-Term Load Forecasting in Heterogeneous Power Systems

Accurate short-term load forecasting (STLF) plays a critical role in modern power system operation, enabling reliable unit commitment, economic dispatch, and efficient integration of renewable energy resources. However, forecasting performance often deteriorates under irregular calendar conditions, such as public holidays and transitional days, when electricity consumption patterns deviate from typical weekday behavior. To address this challenge, this study proposes a calendar-aware two-stage forecasting framework that employs Random Forest (RF) as both a data-driven classifier and a localized regression model. In the first stage, an RF classifier segments historical load data into homogeneous groups using calendar-based variables, including month of year, day of week, holidays, and bridging holidays. In the second stage, RF regression models are trained within each classified segment to capture localized load dynamics. A hierarchical fallback mechanism is further incorporated to ensure forecasting stability under rare-event data scarcity. The proposed framework is validated using four national electricity datasets with heterogeneous climatic and demand characteristics, including Thailand (half-hourly intervals from EGAT), France and Norway (hourly data from ENTSO-E), and Singapore (half-hourly data from EMA). Experimental results demonstrate that the proposed RF-RF framework consistently outperforms conventional approaches such as Multiple Linear Regression (MLR), Support Vector Regression (SVR), and rule-based segmentation. The model achieves Mean Absolute Percentage Errors (MAPE) of 4.03%, 3.31%, 2.35%, and 2.52% for Thailand, France, Norway, and Singapore, respectively, with the most significant improvements observed during holidays and bridging-holiday periods. These results highlight the importance of calendar-aware segmentation and demonstrate that aligning ensemble-based classification with ensemble regression provides a robust and generalizable STLF solution for heterogeneous power systems.