Exploring the impact of feature selection on load profile analysis and energy demand forecasting through enhanced support vector regression

Abstract This paper investigates the impact of weather and calendar data features on load profile analysis and proposes an improved load profile forecasting approach using Support Vector Regression (SVR). A detailed load profile was constructed for a single-family house in Algiers, Algeria, based on an in-depth analysis of survey responses over three years. The SVR model, employing a built-in split method, has demonstrated the highest efficiency for short-term predictions, particularly for one-day forecasts. The initial results from the standard SVR model yielded a Mean Absolute Percentage Error ( MAPE ) of 39.50%, a Mean Squared Error ( MSE ) of 0.0461, and an R 2 score of 0.8679. Additionally, the study compared the performance of other machine learning models, including Random Forest Regressor (RFR), Gradient Boosting Regressor (GBR), and Artificial Neural Network (ANN) for one-day forecasting. The RFR achieved an MSE of 0.20088, MAPE of 90.45%, and an R 2 score of 0.4243; the GBR yielded an MSE of 0.13274, MAPE of 80.76%, and an R 2 score of 0.6196; while the ANN demonstrated an MSE of 0.0618, MAPE of 59.71%, and an R 2 score of 0.6407. Notably, the SVR model emerged as the superior performer across various forecast horizons, prompting further exploration to enhance its capabilities. In addition to the standard SVR method, this study introduces an enhanced SVR approach utilizing the Radial Basis Function (RBF) kernel and fine-tuning its parameters. This enhanced model achieved a significantly reduced MSE of 0.0419, MAPE of 18.89%, and an improved R 2 score of 0.8799 for one-day forecasts, surpassing the standard SVR model's performance. Fourier transform analysis was also applied to uncover underlying patterns in the consumption data, complementing the time-domain results from the SVR model. A grid search optimized hyperparameters, revealing that C = 5 and ε = 0.01 provided the best model performance. These findings offer practical implications for energy management, policy-making, and the development of smart grid technologies, contributing to the sustainability and efficiency of energy consumption in residential settings.