Mechanism-Preserving Adaptive Calibration of Transient Multi-Infeed Short-Circuit Ratio for Safety Threshold Generation in Receiving-End HVDC Systems

Multi-infeed LCC-HVDC receiving systems need a transient-strength index that is both physically interpretable and adaptive to operating-mode, fault and reactive-compensation changes. This paper develops a corrected transient multi-infeed short-circuit ratio (CTMSCR) that keeps the short-circuit-ratio denominator intact while calibrating it against electromagnetic-transient evidence. The denominator contribution of each infeed is decomposed into an impedance-coupling term and a disturbance-intensity term, and each path carries a coupling-calibration coefficient defined as the ratio of the true transient peak sensitivity to the quasi-steady sensitivity that the classical denominator assumes. The first result of this study is negative and delimits the method: a single coefficient per infeed-to-bus path, held constant across scenarios, is not sufficient, reaching a cross-validated Spearman coefficient of only 0.326 against 0.322 for the uncalibrated denominator. Making the coefficient scenario-conditioned, in the same way that the safety threshold is already conditioned in the scenario group, removes the limitation. Coefficients are estimated by a regularized objective anchored on the classical values, and the transient-overvoltage limit together with a declared confidence level is converted into bus-specific safety boundaries. A three-infeed receiving-end benchmark with 360 bus-level samples is used for validation; a total of 126 samples reach or exceed the 1.30 p.u. limit. All reported figures are obtained by a nested protocol in which coefficients, links and thresholds are estimated in training cases only and applied to held-out cases. On that basis the calibrated index raises the Spearman coefficient from 0.322 to 0.970, the classification accuracy from 0.675 to 0.914, and the Matthews correlation coefficient from 0.184 to 0.789, while the missed-alarm rate falls from 0.673 to 0.153. Logistic-regression, support-vector, principal-component and Bayesian baselines trained on the same features reach cross-validated Spearman coefficients between 0.11 and 0.26, because they cannot extrapolate to operating scenarios absent from their training set, whereas the mechanism-preserving denominator can.