The high proportion of power electronics in new power systems introduces the "dimensionality disaster" challenge in system modeling and analysis. Among existing model order reduction methods, theoretical analysis approaches are highly targeted but suffer from insufficient universality. The state variable selection principle of the singular perturbation method is tied to linearized models, making it difficult to accurately capture comprehensive dynamic characteristics. To address this, the state variable selection principle is improved by leveraging data-driven techniques to obtain full dynamic characteristics. The least absolute shrinkage and selection operator (LASSO) is employed for state variable selection, supplemented with critical state variables based on physical information. Perturbation methods are then applied to achieve system dimensionality reduction. Finally, the approach is tested in a typical grid-connected converter system, demonstrating that the reduced-order system can reflect the small-disturbance stability characteristics of the original system, validating the method′s effectiveness.