کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

A Comparative Study of GAM, Bagging regression trees and Random Forest for Short-Term Electrical Load Forecasting

نویسندگان :

Javad Hajiannezhad¹ Habib Rajabi Mashhadi² Maryam Baradaran Naseri³ Mohammadreza Mohammadhasani⁴

1- دانشگاه فردوسی مشهد 2- دانشگاه فردوسی مشهد 3- دانشگاه صنعتی شریف 4- دانشگاه فردوسی مشهد

کلمات کلیدی :

Short-term load forecasting،Generalized Additive Model (GAM)،Random Forest،Bagging،Machine learning،Electricity demand prediction

چکیده :

hort-term electrical load forecasting plays a critical role in power system operation, planning, and reliability. Accurate hourly demand prediction helps utilities reduce operational costs and improve energy management. In this paper, a comparative study of three regression-based machine learning models—Generalized Additive Model (GAM), Bagging regression trees (BRT), and Random Forest (RF)—is presented for short-term electrical load forecasting. The models are evaluated using historical load data combined with weather variables and temporal features. The dataset includes hourly electricity consumption along with meteorological variables such as temperature, humidity, wind speed, and calendar-related indicators. Data preprocessing, feature engineering, and hyperparameter tuning are performed before model training. Model performance is assessed using Mean Squared Error (MSE) and the coefficient of determination (R²) on both training and test datasets. In addition, the impact of feature selection on forecasting accuracy is investigated. The results show that all three models achieve acceptable forecasting performance, with R² values above 0.93 on the test data. Among them, the Random Forest model provides the highest accuracy, benefiting from variance reduction and random feature selection. The GAM model demonstrates strong interpretability and captures nonlinear relationships effectively, while the Bagging approach offers a balance between accuracy and model stability. The findings confirm that ensemble tree-based models, particularly Random Forest, are well-suited for short-term load forecasting when sufficient historical and weather data are available.