Early-Warning Detection of Grid Coherence Loss: Curvature-Based Precursor Detection Across Distinct Failure Modes in Power Systems

This paper presents and validates a curvature-based (second-derivative) early-warning detector for gradually-developing voltage instability precursors in power systems, built on standard PMU-class telemetry. The detector computes a rolling local trend in voltage magnitude, then tracks the acceleration of that trend, and fires when the acceleration crosses a threshold calibrated from each bus’s own steady-state behavior. It is validated against three genuine dynamic transitions across two IEEE-standard reference systems (IEEE 39-bus, Kundur two-area) — two local voltage-collapse events and one inter-area synchronism separation — achieving 100% detection across 20 seed-trials with zero false positives and positive lead time in every trial (1.4–9.1 s). The paper’s central contribution is narrower than “faster or more reliable early warning”: a comparison against four baseline detectors (fixed threshold, raw rate-of-change, moving-average deviation, simple first-derivative) shows the fixed threshold achieves substantially longer raw lead times, and all methods tie at zero false positives in this study’s engineered-collapse scenarios. What the acceleration-based construction demonstrates specifically is robustness across qualitatively distinct precursor geometries — including a decelerating-then-reversing precursor shape (Kundur) that both first-derivative baselines fail to detect at all (0/5 seeds), despite succeeding on both local-collapse cases. This identifies a structural blind spot in first-derivative detection methods that curvature-based measurement covers. Limitations are stated without qualification: this is a 3-transition, 20-seed-trial engineering validation, not a statistically powered study; near-miss testing (disturbances that approach but do not reach instability) is absent and flagged as the single most important prerequisite for any operational deployment claim; and validation is limited to two small reference systems, with extension to a ~2,000-bus synthetic grid identified as in-progress future work. This paper is the first empirical output of a broader multi-corridor research program (the Standard Coherence Fidelity Layer, or SCFL) but is written and evaluated to stand entirely on its own power-systems merits — any connection to that broader framework is explicitly separated out as a claim for companion work, not a precondition for evaluating the results reported here.