Overshoot-Oriented Frequency Security Assessment and Structural Auditing Through a Traceable BPA Automation Workflow

The increasing penetration of renewable generation and power-electronic-interfaced devices changes the sources of inertia, frequency-control resources, and structural sensitivity of power-system frequency responses. This paper develops a traceable automation workflow for overshoot-oriented frequency security assessment, addressing opaque label provenance, the limited reproducibility of batch transient simulations, and overoptimistic conclusions from random data splits. The workflow links rule-based scenario construction, BPA batch execution, frequency-trajectory parsing, quality gates, safety-prioritized model selection, and structural auditing. The maximum system frequency is denoted by f_max. For the overshoot-focused task, f_max is the primary continuous severity label, while the binary security label follows the prescribed frequency band. Using 202 aligned records with 44 features, classification and regression tasks are jointly evaluated under a scenario-family holdout protocol. Extra Trees is selected as the primary model because it attains a family-holdout false-secure rate (FSR) of zero, an accuracy of 0.946860, and an f_max mean absolute error (MAE) of 0.615192 Hz. A separate Random Forest robustness probe achieves a mean accuracy of 0.996721 and a maximum FSR of zero across ten repeated splits. An independent 81-record template package identifies four representative structural templates while retaining 27 heterogeneous records outside the template narrative. Four additional line-9 scenarios were automatically generated and executed; all exhibit frequency overshoot with f_max between 72.1621 and 72.1828 Hz. Because the corresponding BPA summaries also indicate rotor-angle instability, these scenarios are used only to audit the frequency-response layer and the traceability of the workflow, rather than to claim complete transient security. The contribution of this work is a reproducible BPA-based experimental evidence chain, not a universal frequency-stability mechanism, external-grid generalization, or a claim of deep learning superiority.