Hyperparameter Optimization of Multi-Target Support Vector Regression with Sigmoid Particle Swarm Optimization-based Acceleration Coefficients for Electricity Consumption Prediction

Abstract

The predictive accuracy of Support Vector Regression (SVR) in electricity forecasting is often constrained by hyperparameter tuning. To address the documented poor performance of SVR on the multi-target Tetouan City Power Consumption dataset, an advanced metaheuristic, Particle Swarm Optimization with Sigmoid-Based Acceleration Coefficients (PSO-SBAC), is employed to optimize SVR's regularization constant ( ) and the RBF kernel coefficient ( ) hyperparameters. The superior stability of PSO-SBAC was first confirmed on ten mathematical benchmark functions, where it achieved near-perfect success rates, significantly outperforming the less consistent standard PSO on complex multimodal functions. When applied to the forecasting task, the proposed SVR-PSO-SBAC model demonstrated superior generalization capabilities. While the baseline model showed robustness on high-frequency 10-minute data, the proposed SVR-PSO-SBAC achieved the best performance on Zone 1 in the 1-hour interval, recording the lowest test Root Mean Squared Error (RMSE) of 34769.5. Furthermore, the model excelled in handling complex zonal characteristics, achieving a massive 27.4% reduction in error compared to the baseline in Zone 3 (10-minute interval) with an RMSE of 4878.1. The findings resolve a documented performance anomaly, concluding that SVR's perceived limitations are a function of suboptimal tuning, which can be overcome with a robust and adaptive optimization strategy.