Machine Learning-Based Electricity Load Forecasting for Smart Grid Energy Management: A Comparative Study Using Austrian Demand Data

Accurate electricity load forecasting is essential for efficient and reliable smart grid energy management, particularly as power systems integrate growing shares of variable renewable generation. This study develops and compares four machine learning models Linear Regression, Support Vector Regression (SVR), Random Forest, and Extreme Gradient Boosting (XGBoost) for short-term electricity demand forecasting. Four years of hourly demand, solar generation, and wind generation data for Austria were obtained from the ENTSO-E Transparency Platform (35,000 hourly observations) and enriched with temporal features (hour, day of week, month, year, and weekend indicator). Models were trained on a chronologically ordered 80:20 split and evaluated at three forecast horizons a 500-hour rolling window, four selected 2018 timestamps, and a full 24-hour daily profile using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination (R²). XGBoost achieved the strongest overall performance (MAE = 195.29 MW, RMSE = 243.71 MW, MAPE = 2.38%, R² = 0.9617), followed closely by Random Forest (MAPE = 3.20%, R² = 0.9178), while Linear Regression and SVR produced substantially higher errors (MAPE of 15.90% and 14.98%, respectively) and weaker explanatory power. These results confirm that tree-based ensemble methods capture the nonlinear, temporally dependent structure of electricity demand far more effectively than linear or kernel-based approaches, and can be deployed without deep learning infrastructure. The proposed framework offers a practical, scalable basis for load forecasting, demand response, and renewable energy integration in smart grid operations.