Scalable and Efficient Support Vector Regression Model in Power Systems

Support Vector Regression (SVR) is a crucial tool for modeling complex relationships in power systems and is widely applied to tasks like demand and load forecasting. SVR, rooted in statistical learning theory and the principle of structural risk minimization, operates by mapping input space to a high-dimensional feature space and constructing an optimal separating hyperplane therein to achieve non-linear regression. In power systems, SVR effectively captures complex non-linear relationships among multidimensional variables such as load, weather, and economic factors, thus excelling in areas like short-term load forecasting and renewable energy generation prediction. However, the high computational cost of SVR presents significant limitations for its practical applications in real-world scenarios, especially in the era of smart grids with their everincreasing data volumes. Traditional SVR algorithms have a time complexity of $\mathcal{O}\left(n^{3}\right)$, where n is the number of training samples. This implies that as data scale increases, training time rises dramatically, becoming impractical for large-scale realtime applications. To address this problem, we propose two scalable acceleration algorithms, ALM-SVR and ADMM-SVR, to improve the efficiency of SVR. Both algorithms have linear time complexity and thus are able to be scalable to large-size datasets. We further discuss their extensions and provide theoretical proofs of their boundedness and convergence. Experimental results on benchmark and real power system datasets show that our methods significantly outperform conventional SVR solvers in terms of running time and are therefore suitable for large-scale applications in power systems.