WorksMathematics
Counterfactual Screening of Rare Secondary Overloads in Power Grids with a Stochastic Surrogate for Electromagnetic Transient Simulations
Abstract
Power-system operators continually ask what would happen if a component failed now. Single-phase-to-ground faults cleared by single-pole tripping and automatic reclosure usually have no lasting consequences. In rare cases, however, the transient excited during the open-phase interval drives distant lines or transformers into overload that persists after the faulted line is restored. We develop a counterfactual screening framework for these secondary overloads, combining a linear stochastic swing surrogate with cross-entropy importance sampling. The surrogate represents the open-phase corridor through an effective coupling factor and localized multiplicative noise, with both parameters learned from electromagnetic-transient (EMT) simulations rather than prescribed. Closed-form propagation of the mean and covariance permits deterministic calibration without Monte Carlo sampling. On the reduced WECC 240-bus system, a globally calibrated model substantially improves finite-time agreement with the EMT electromechanical state. The global models are calibrated at reclosure and tested on disjoint faulted lines under a longer open-phase interval. Line- and operating-state-specific models use additional reclosure data from the assessment lines and provide smaller further gains at the cost of more separate fits. The fitted coupling is lower than the phase-counting baseline, consistent with the direction anticipated by sequence-network analysis. Cross-entropy importance sampling directs computational effort toward contingencies associated with persistent overload and yields rare-event probability estimates together with a map linking initiating faults to the elements they endanger. By separating offline EMT calibration from repeated surrogate evaluation, the method supports probabilistic contingency screening and identifies where detailed high-fidelity assessment is most valuable.
Introduction
Most of the time, a transmission-system operator is watching a grid in which nothing alarming is happening. The job is nevertheless anticipatory: to know, before anything happens, which plausible events would put the system in danger. Steady-state security assessment answers this question for post-contingency equilibria and runs continuously in control rooms [2, 3]. It does not answer it for the transient that connects the prefault and post-fault states, even though that transient can drive other lines and transformers above their limits for seconds. This paper develops a tool for that missing, dynamic part of the question. The analysis is counterfactual. It is not a response to an ongoing emergency but an estimate, made in advance and refreshed as conditions change, of what would happen if a given fault occurred now.
Guided by the operating experience of the Israeli transmission system operator (NOGA), we focus on the most common and least threatening transmission fault: a single-phase-to-ground fault on an extra-high-voltage line, cleared by single-pole tripping and automatic reclosure [4, 5]. Opening only the faulted pole keeps the two healthy phases in service. After a dead time of a fraction of a second the pole recloses and, in the overwhelming majority of cases, the event passes without consequence. Occasionally, however, the electromechanical swing excited during the open-phase interval travels through the network and pushes a remote critical element, typically a transformer, into an overload that persists after the faulted line has been restored. We call such a consequence a secondary contingency. It has three features that shape everything that follows. It is triggered by a frequent, benign primary event. It is rare. And when it happens it is costly, since a sustained transformer overload can lead to damage, protective tripping, or the first step of a cascade.
Screening for secondary contingencies is therefore a rare-event problem posed under real-time constraints, and three requirements pull against each other. The first is fidelity. The open-phase interval is an unbalanced three-phase transient, and electromagnetic-transient (EMT) simulation is the industry-accepted description of it [6, 7]. Even a parallel EMT code needs about half a minute of computation per simulated second on the reduced 240-bus model of the Western Electricity Coordinating Council (WECC) system [8]. The second is rarity: resolving a probability of order to within ten percent by direct Monte Carlo requires of order trajectories. The third is extent and latency: every line is a possible initiating fault, and the whole assessment must be repeated as load, dispatch, and renewable output drift. Direct EMT Monte Carlo meets the first requirement and misses the other two by orders of magnitude.
In previous work [1] we addressed rarity and extent. Following the swing-equation tradition of transient-stability analysis [9, 10, 11], we represented the post-fault dynamics by a linear stochastic swing model in which the faulted corridor has a reduced coupling during the open-phase interval and carries multiplicative noise local-
faulted line has been restored. We call such a consequence a secondary contingency. It has three features that shape everything that follows. It is triggered by a frequent, benign primary event. It is rare. And when it happens it is costly, since a sustained transformer overload can lead to damage, protective tripping, or the first step of a cascade.
Screening for secondary contingencies is therefore a rare-event problem posed under real-time constraints, and three requirements pull against each other. The first is fidelity. The open-phase interval is an unbalanced three-phase transient, and electromagnetic-transient (EMT) simulation is the industry-accepted description of it [6, 7]. Even a parallel EMT code needs about half a minute of computation per simulated second on the reduced 240-bus model of the Western Electricity Coordinating Council (WECC) system [8]. The second is rarity: resolving a probability of order to within ten percent by direct Monte Carlo requires of order trajectories. The third is extent and latency: every line is a possible initiating fault, and the whole assessment must be repeated as load, dispatch, and renewable output drift. Direct EMT Monte Carlo meets the first requirement and misses the other two by orders of magnitude.
In previous work [1] we addressed rarity and extent. Following the swing-equation tradition of transient-stability analysis [9, 10, 11], we represented the post-fault dynamics by a linear stochastic swing model in which the faulted corridor has a reduced coupling during the open-phase interval and carries multiplicative noise localized at its endpoints. Combined with cross-entropy (CE) importance sampling [12, 13, 14], this model screens many initiating single-phase faults in parallel and concentrates computational effort on the rare scenarios that produce long overloads. The weak point of that construction was its physics. The reduced coupling and the noise amplitude were postulated, not derived from or tested against a model that operators trust. Whether such a light model can stand in for the EMT description of the open-phase transient was left open.
This paper identifies the surrogate’s effective parameters from ParaEMT reference trajectories on the reduced WECC 240-bus system [7]. Calibration matches the electromechanical state at reclosure; subsequent assessment tests the globally fitted model on disjoint faulted lines under a longer open-phase interval. A hierarchy of global, line-specific, and line–operating-state-specific fits examines the value of additional calibration information. The comparison uses the phase-angle and frequency variables retained by the swing model, with parameter identification based on its deterministic moment equations.
The calibrated surrogate is then used to estimate overload probabilities and construct a contingency-to-vulnerable-element map: for each initiating fault, the probability that a monitored line or transformer accumulates more than a prescribed time in overload. This is the operator-facing output of the proposed background screening workflow in Figure 1.

Figure 1. Logic of the paper. (a) The counterfactual question. A frequent and usually benign single-phase fault on line is cleared by single-pole tripping and reclosed after the open-phase interval . The electromechanical swing it excites can drive a remote critical element into persistent overload within the screening horizon . (b) The fast physics model. The open-phase coupling and the noise amplitude of a stochastic swing surrogate are learned offline from EMT (ParaEMT) trajectories by closed-form moment matching. The learned coupling is compared with phase-counting and sequence-network estimates. (c) Online, cross-entropy importance sampling over all initiating faults turns surrogate trajectories into the contingency-to-vulnerable-element map . The intended tool runs in the background and alerts the operator only when an entry exceeds a threshold. Blue labels mark elements carried over from Ref. [1], orange marks what is new here, and red marks secondary overload and the operator alert. The network and map are schematic; quantitative results are in Sec. IV.
Section II develops the model, calibration, and probability estimation; Section III specifies the experiments. Sections IV and V present the findings and their interpretation, with scope and limitations collected in § 2.4 A.
Methods
Table I collects the notation and time conventions used in the main text and appendices. The method separates representation from estimation: EMT simulations calibrate the reduced dynamics, and importance sampling estimates overload probabilities under those fixed dynamics. Calibration, state assessment, and screening use distinct time settings. The state variables, switching rule, and observation convention remain the same; the fitted physical parameters are held fixed when the screening sampler varies fault location and open-phase duration.
| Symbol | Meaning |
| Network graph, bus and line sets, and number of initiating lines; in screening. | |
| Faulted line and a noise-bearing endpoint; . | |
| Pre-fault operating state and its dimensionless uniform injection multiplier. | |
| Reduced state: frequency deviations and phase angles in the fixed reference. | |
| Inertia matrix and the canonical vector for endpoint . | |
| Nominal drift, open-phase drift, and equivalent Itô drift. | |
| Endpoint-local diffusion matrix and its independent Wiener driver. | |
| Open-phase coupling factor: during only. | |
| Common amplitude of endpoint-localized multiplicative noise; . | |
| Fitted parameter collection. | |
| Calibration-hierarchy model names, not inertia matrices. | |
| Instant the faulted pole opens; fault clearing is treated as instantaneous. | |
| Open-phase duration (auto-reclosing dead time). | |
| State-evaluation time: at reclosure for every fitted hierarchy level; after a open-phase interval for their common assessment. | |
| Upper limit of the overload integrals, fixed at in screening; distinct from the calibration target time. | |
| Observation convention, projected EMT state, and finite-time reference state-transfer map. | |
| Surrogate mean and state covariance. | |
| Observation-space neighborhood mask and its binary bus-selection entries. | |
| Covariance target formed across calibration operating states at fixed line. | |
| Endpoint-masked predicted covariance and its diagonally regularized version. | |
| Calibration objective, scalar regularizer, and objective weights. | |
| Matrix argument of the second-moment operator and an eigenvalue of its matrix representation. | |
| Zero- to positive-sequence reactance ratio in Appendix B. | |
| Accumulated line exposure, summed network exposure, and the prescribed strict-exceedance threshold. | |
| Global and line-specific probabilities under the nominal scenario law. | |
| Corresponding probabilities conditional on initiating fault . | |
| Sampled open-phase duration, nominal exponential mean, and shared proposal mean, all in seconds. | |
| Nominal line probability, proposal line probability, and proposal-parameter vector. | |
| Samples per adaptation iteration, number of iterations, and fresh final-batch size. | |
| CE elite fraction, previous-parameter retention factor, and elite count. | |
| Adaptation iteration, sample index, and selected elite index set. | |
| Binary elite-membership indicator and smallest selected elite score. | |
| Raw likelihood ratio, normalized diagnostic weight, and weighted event contribution. | |
| Probability estimate, estimated standard error, and relative sampling standard error. | |
| Relative finite-time state error; distinct from sampling uncertainty. | |
| Probability, expectation, and the exponential function. |
Table I. Principal notation and time conventions. Physical calibration parameters, scenario-law parameters, and sampling diagnostics have distinct roles. The same time symbol may take different values in different experiments.
Contingency and Reduced Dynamics
The modeled contingency is a single-phase-to-ground fault cleared by single-pole tripping and followed by successful automatic reclosure. The reduced description starts at pole opening and retains the open-phase and post-reclosure electromechanical response.
Fault inception and clearing. Protection opens the faulted pole of line . Fault inception and breaker operation occur over tens of milliseconds, below the electromechanical time scale retained by the surrogate. Clearing is therefore treated as instantaneous, and denotes the instant of pole opening. The reduction begins from the corresponding electromechanical state and does not resolve the preceding fault or breaker transient.
Open-phase interval . One phase is de-energized, but the other two remain in service, so the line continues to transmit power. The interval permits de-ionization of the secondary arc sustained by capacitive and inductive coupling from the energized phases; premature reclosure can restrike the fault. In protection terminology, is the auto-reclosing dead time [15]. We use open-phase duration to distinguish it from the evaluation and screening times. During this interval, the effective positive-sequence coupling changes from to , with .
Reclosure and post-reclosure interval . At , the open pole is reclosed and the nominal coupling is restored [4]. The machines do not instantaneously return to equilibrium: their ensuing swing can cause transient overloads on other elements. The transferred electromechanical state therefore matters even after the network configuration is restored.
Evaluation time and screening horizon . The surrogate state is compared with its EMT reference at the experiment-specific time ; overload duration is accumulated up to . State assessment at or after reclosure has . Calibration uses the boundary case , so the target is the state attained at reclosure. The main hierarchy comparison instead uses and , one second after reclosure. Screening integrates exposure through and does not fix every sampled duration at a calibration or test value. A full-trajectory illustration may also end at without making that endpoint the calibration time. Varying changes the observation time; varying changes the switching event. Table I records these roles. Throughout, fault clearing initiates the open-phase interval at ; denotes reclosure, not a second fault-clearing event.
Why a stochastic reduced model? The swing variables retain the phase and frequency response used by the screening calculation, but not the three-phase electrical and device states resolved by ParaEMT. Their influence on the retained variables is represented by an effective coupling and a linear stochastic correction, whose parameters are identified from the EMT response.
The two fitted terms address different features of the reduction. The coupling factor represents the change in transfer through the open-phase corridor. The diffusion matrices act only at its endpoints, so the model introduces the correction at the location of the disturbance and retains the network dynamics that transmit its effects elsewhere. Linearity in the state gives the closed moment equations used for calibration.
The stochastic correction represents unresolved variability within the prescribed operating-state family. Its empirical assessment combines the open-phase comparison in Figure 3 with the deterministic–stochastic comparison in the calibration hierarchy. Section IIB specifies the across-operating-state covariance target used to identify its amplitude.

Figure 3. ParaEMT–surrogate comparison during the open-phase interval. Phase-angle trajectories for the illustrative large-excursion 9#4 contingency over . Solid curves show ParaEMT; dashed curves show one realization of (2) with , not an ensemble mean or confidence band. The vertical markers identify pole opening and reclosure. Figure 11 continues this same example to 10 s and displays the corridor separation . The full-trajectory endpoint is not the 0.8-s calibration target or the 2.5-s hierarchy-test time.
For a faulted line and a pre-fault operating state , the surrogate evolves the electromechanical state through the open-phase and post-reclosure intervals of Sec. II A.
The reduced electromechanical state is
where and are the frequency-deviation and phase vectors over the bus set . The formulation preserves the reduced WEC network graph and transmission-line connectivity rather than Kron-reducing the system to generator internal nodes. Thus, the graph used for the neighborhoods in (C1)–(C2) is the graph used to construct the swing dynamics. Phases are absolute coordinates relative to a fixed reference bus, not deviations from the pre-fault phase vector; denotes the equilibrium state for operating condition in that reference convention.
The symbols and denote the coupling and common endpoint-noise amplitude used for a specified scenario. Their sharing across scenarios is determined by the calibration hierarchy in Sec. III A, under Comparison hierarchy, and Figure 2. They are global constants in M/M, line-specific values in M, and line–operating-state–specific values in M. The baseline M fixes , and M has . In every case, one amplitude is shared by the two endpoints of a given faulted line. We suppress these calibration indices in the dynamical equations; the scenario indices on and still identify the modified corridor and the endpoint where noise acts.

Figure 2. Calibration hierarchy and common comparison protocol. (A) fixes ; fits one global deterministic coupling and one global stochastic pair. fits 25 line-specific pairs and fits 1,150 line–operating-state-specific pairs. All fits use s. The specialized models additionally use reference data from the 25 assessment lines, which are excluded from the global fits. The arrows describe fitting specificity and effort, not measured runtime ratios. (B) The common protocol fixes the WECC model, observations, and assessment conditions. Figure 4 scores every model on 25 lines and 46 operating states at s. Those line identities are unseen by , but have separate 0.8-s calibration records for . The count identifies a reference family, not one pool split among all models. The sequence-informed deterministic reference only motivates the coupling correction; it is not implemented or tested in the hierarchy. The line and operating-state indices are and , respectively.
Across the open-phase and post-reclosure intervals, the dynamics are
where
is the endpoint set of the faulted line, equivalently its zero-radius graph neighborhood. During , the multiplicative forcing represents variability localized to the affected corridor. The equation is interpreted in the Stratonovich sense, as indicated by , and the are independent Wiener processes. The vector denotes the power-injection forcing for the pre-fault operating state. At reclosure, the stochastic forcing is removed and the nominal drift is restored; the attained electromechanical state supplies the initial condition for the post-reclosure interval. The linear description therefore yields a finite-time map composed across the switching event, rather than one time-invariant faulted-network propagator.
The contingency-dependent matrix is constructed from the swing dynamics by modifying the coupling of the affected line,
where is its nominal coupling. The modification applies only during the open-phase interval. Reclosure restores and , so the calibrated parameters refer to the same piecewise network configuration used in trajectory evaluation.
The parameter is the effective positive-sequence coupling factor of the open-phase corridor. Counting the two intact phases gives the physically transparent baseline , but not an exact sequence-network result. Negative- and zero-sequence contributions modify the positive-sequence transfer under a one-conductor-open condition, providing a physical basis for a coupling below the phase-counting value. The value used by the surrogate is identified from the electromechanical response rather than imposed by phase counting; its comparison with the sequence-network estimate is reported in Sec. IV B.
The noise acts on the phase coordinates of the two endpoints. Let be the th canonical vector in . For ,
Here is the inertia matrix of the retained swing model. Each diffusion matrix has a single endpoint contribution in the phase block. Since has units , the amplitude has units ; this is not a dimensionless perturbation. Because the multiplicative term uses phase coordinates, its specification is tied to the fixed phase reference introduced above. The same reference must be retained in the EMT projection, initial states, and surrogate evaluation; common-mode cancellation in the overload observable does not make the stochastic state equation reference-invariant.
The two parameters are identified jointly. The coupling controls the deterministic open-phase interaction, while controls localized variability and contributes to the equivalent Itô drift through . The baseline provides a warm start, not a constraint on the fit. The stochastic correction can consequently influence both the transferred mean state and its covariance, as made explicit in Sec. II B.
The sequence-network construction in Appendix B provides a physical motivation for the coupling correction. The hierarchy and the status of the sequence-informed control are specified in Sec. III A.
EMT Reference and Moment-Based Calibration
Shared observation space. ParaEMT provides the reference trajectories from which the reduced finite-time map is identified [7]. Its three-phase network and device model supplies the high-fidelity response, which is projected onto the observation space shared with the swing surrogate. Appendix A identifies the reference implementation and case inputs.
Let denote the faulted line and the pre-fault operating state. ParaEMT generates a high-dimensional trajectory
where is the evaluation time defined in Sec. II A. Its electromechanical observation is
where extracts the phase and frequency variables represented by the surrogate. The same observation convention is used when evaluating the reduced state, so the mismatch concerns corresponding physical variables.
The projection retains the electromechanical response through which the reference model’s fast electromagnetic and device dynamics influence the calibration target. The selected observations determine the dimension of the comparison.
The finite-time reference map is
where is the reduced electromechanical state associated with the pre-fault equilibrium. Fixing the initial state, switching convention, and observation time makes the map a direct comparison of the reference and reduced responses to the same scenario. Its approximation is the calibration objective. In the reported fit this target is evaluated at ; the subsequent assessment propagates the fitted model to its separately specified observation time.
Deterministic moment propagation. Linearity with multiplicative noise gives a closed deterministic system for the first two moments after conversion to the equivalent Itô form. For parameters , define the mean
and covariance
The equivalent Itô drift during the open-phase interval is
The second term is the noise-induced drift associated with the Stratonovich convention [16, 17, 18]. Consequently, changing affects the mean as well as the covariance; its role is not restricted to adding fluctuation around an unchanged deterministic trajectory. The mean satisfies
and the covariance satisfies
The moments at reclosure initialize their deterministic continuation under . In calibration, evaluation stops at reclosure; in the post-reclosure tests, the moments are continued to the specified . This retains the switching event explicitly and avoids Monte Carlo sampling when evaluating candidate parameters [16, 17].
The effective coupling and noise-induced drift jointly determine the mean-based metric in (15). This supplies the mechanism by which M can improve the finite-time map beyond the deterministic fit M: adjusts the corridor interaction, while the Stratonovich correction permits to modify the transferred mean. Their joint interpretation allows part of the effective correction to be represented by noise-induced drift rather than by alone. Appendix C 2 records the second-moment operator and the spectral diagnostic for the reduced stochastic dynamics.
Endpoint-local objective. The primary global fit uses only the two endpoints of the faulted line. The mask selects their observed phase and frequency components, and denotes the scatter of the projected EMT state across the training operating-state family at fixed fault , evaluated at for the training runs. The joint objective is
The first term matches the mean transferred state, the second matches its covariance target, and regularizes the predicted covariance. The weights are and ; the absolute log-determinant regularizer uses the diagonal shift . Their explicit definitions and mask construction are given in Appendix C 1.
ParaEMT supplies one deterministic trajectory for each , and is computed across the training operating states. The fitted therefore represents operating-point variability unresolved by the global reduced description. Only training states from the global calibration-line set enter this target; observations from the global testing lines are not used to identify the shared parameters of M or M. Models M and M instead receive additional calibration observations from those lines at the same 0.8-s reclosure target, as specified in Sec. III A. The later observations used to score Figure 4 are not fitting targets for any model. The restriction applies to parameter identification, not to the propagation of the disturbance: network-wide responses remain available for validation and screening.

Figure 4. Distribution of across the WECC 240-bus model hierarchy at and . All models are scored on the same 25 lines and 46 uniformly scaled operating states. Marker shapes identify lines; colors identify injection changes. were fitted on 25 disjoint calibration lines. instead uses 25 line-specific pairs fitted across the 46 operating states, and uses 1,150 line–state-specific pairs, obtained from additional reference runs on the assessment lines. Every fit uses and is held fixed here. Thus, the results measure duration and observation-time transfer after case-specific calibration, not transfer to unseen lines. Their smaller additional gains beyond require more separate fits. Unlike Figs. 9–10, this comparison includes injection scaling.
Validation metric. The principal validation metric compares the surrogate mean and the projected EMT state at the same evaluation time,
The numerator measures discrepancy in the observed electromechanical state; the denominator expresses that discrepancy relative to the reference state’s norm. Both the phase-angle and frequency components selected by enter this metric; it is not a phase-only error unless selects only phases. Within a given comparison, all hierarchy levels use the same variables, reference convention, switching duration, and observation time. This common scoring protocol does not imply identical fitting data: the global and scenario-specific models use the allocations in Sec. III A. All fitted coefficients are held fixed when the 0.8-s calibration setting is replaced by the longer-duration assessment setting.
In Figure 4, the finite-time state metric is evaluated at after reclosure at , with the coefficients fitted at the 0.8-s calibration target held fixed. Noise affects through the mean dynamics in Sec. II B. Covariance information enters parameter identification through (14); validation then scores the projected mean state.
Overload Observables and Conditional Risks
Let be the phase-angle vector of a sampled surrogate trajectory from (2). For a monitored line , the linearized transfer is compared with its allowable power-flow threshold . The line-specific overload indicator is
where equals one when the stated limit is exceeded and zero otherwise. Thus, is a cumulative duration: separated periods of overload contribute to the same integral. The monitored element need not be the faulted line. This distinction allows a local initiating contingency to be assessed through the persistent stress it produces elsewhere in the network.
The system-level indicator sums exposure over the monitored set,
where . Simultaneous violations on different elements contribute separately, so measures accumulated line-overload time rather than the duration for which at least one line is overloaded. It can consequently exceed . Neither indicator weights a violation by the amount of its excess above the threshold; their purpose is to distinguish persistent excursions from brief crossings that may have comparable peak phase separations.
Equation (16) uses a linearized-flow screening approximation. Common angular displacements cancel from the phase differences. Figure 11 and Appendix D 4 examine the corridor separations that determine its small-angle interpretation; the scope of the indicator is set out in Sec. V A.

Figure 11. Full illustrative transient and corridor-separation diagnostic. The large-excursion example uses s and is observed through 10 s. The in-figure label s denotes the endpoint of this diagnostic, not the calibration or hierarchy-test time. Panel (a) shows phase-angle trajectories through reclosure; panel (b) shows . Solid curves show ParaEMT; dashed curves show one surrogate realization with . Figure 3 displays the open-phase portion of the same run. This example illustrates oscillatory post-reclosure behavior but does not use the 1.5-s switching duration of Figure 4. Excursions beyond the shaded 0.5-rad diagnostic band limit the small-angle interpretation of (16); the band is not a protection setting.
For a prescribed overload-duration threshold , define
This is the probability that monitored line accumulates more than seconds above its flow limit. The corresponding global probability is
Here, the event concerns accumulated exposure over the monitored network, not the union of line-specific events at the same threshold. In particular, is not obtained by adding the . The strict comparison also defines as the probability of any positive overload, rather than the certain event associated with a nonnegative indicator.
The probability law specifies both the initiating scenario and the surrogate uncertainty conditional on it. A scenario comprises a faulted line and its open-phase duration ; the stochastic drivers then determine a realization of the electromechanical response. The nominal scenario law remains fixed when the sampling algorithm changes. The conditional risks defined below answer a different question from averaging over possible fault locations.
Conditional risk scores. For an initiating fault and threshold , define
The probability averages over the nominal open-phase duration and surrogate drivers conditional on . Its strict comparison follows (19). Sorting gives a ranking by the probability of exceeding a specified total line-overload time. Because includes the monitored set, the score can include exposure on the faulted element as well as secondary exposure elsewhere.
For an individual monitored line, define
The strict comparison is the convention in (18) and (E18). The collection
forms one row of the contingency-to-vulnerable-element map. Entries for identify secondary exposure.
Cross-Entropy Importance Sampling
As the overload threshold increases, most nominal samples may contribute zero to the event indicator. Importance sampling reallocates the trajectory budget without changing either the event or the calibrated physical model. The stochastic driving law is the same under the nominal and proposal simulations; only fault location and duration are resampled.
Direct Monte Carlo draws the faulted line and open-phase duration from the nominal joint law . CE adaptation instead fits a proposal to high-score contingencies and then freezes the proposal for a separate estimation batch. The WECC implementation retains both the learned categorical line probabilities and a shared duration parameter. With denoting the sampled open-phase duration, the estimator is
The ratio corrects both line selection and duration sampling. Positive proposal support on the nominal support, an unchanged stochastic driving law, and a fresh estimation batch preserve the nominal target. The joint law and likelihood ratio are specified in Appendix E [12, 13, 14]. The elite updates and Algorithm 1 are given in the same appendix. The duration parameter is an exponential mean, measured in seconds; the nominal mean remains fixed during proposal adaptation.
During probability estimation, the physical parameters remain fixed while CE adaptation changes the scenario proposal. Appendix E2 records the calibrated parameters and their use in the estimator. Each sampled scenario is evaluated with the specified dynamics and overload criterion before likelihood weighting is applied.
WECC Study
The experiments use the reduced WECC 240-bus system to assess state transfer across fault locations and operating conditions, then evaluate probabilistic screening with the calibrated global model.
Reference Case and Calibration Comparisons
The reduced WECC 240-bus transmission system includes conventional and renewable generation, loads, transmission corridors, and dynamic devices. The ParaEMT implementation supplies the network and dynamic-model inputs [8, 19]; the case description and one-line diagram appear in Appendix A1. The topology and dynamic parameters are fixed throughout the calibration and testing experiments. Only the specified fault, open-phase duration, observation time, and pre-fault operating condition change.
Global calibration data and target. The calibration set for M₁ and M₂ contains 25 faulted lines, each evaluated at 46 pre-fault operating states, giving 25 × 46 = 1,150 training scenarios. The operating states are generated by uniformly scaling the nominal active-power injection vector,
with the steady-state equilibrium recomputed for each scale. The nominal case is ; the family spans injection changes from −25% to +20% while retaining the electrical and dynamic-model parameters [20, 21].
Every calibration run, including the additional runs used for M₃ and M₄, uses s. The objective in (14) therefore matches the projected ParaEMT state at the end of the open-phase interval. This state supplies the initial condition for subsequent restored-network evolution. The endpoint-local mask selects the observations used for fitting; the covariance target is constructed from the 46 training operating states at the same 0.8-s snapshot. No later testing time enters the global calibration objective.
Line-disjoint testing of the global models. Let and denote the calibration and testing fault-line sets for M₁ and M₂. Each contains 25 distinct lines, and . The exact line identifiers and their training/test assignments are specified in the project repository [22]. The split is by faulted-line identity, not a random division of the 1,150 training scenarios. Testing uses a separate 25 × 46 family on the same prescribed operating scales. The global-model records are therefore partitioned as
For the global fits M₁ and M₂, only enters parameter identification. The fitted coefficients are then fixed when assessing the 25 previously unused faulted lines. The shared operating scales make this a test of transfer to different fault locations within the stated operating-state family, not a test on unseen networks or on an independently held-out range of injection scales.
Scenario-specific calibration. The reference collection is enlarged for M₃ and M₄ by including reclosure observations for the 25 lines in at all 46 operating states, again with s. These are additional calibration runs, not the s, s observations used for assessment. Thus, a line reserved for testing M₁ and M₂ can supply separate fitting observations for the scenario-specific models without entering the global fit.
Model M₃ identifies 25 parameter pairs , one for each assessment line, using its 46 calibration operating states jointly. Each pair is then shared across those operating states. Model M₄ identifies pairs , one for each line and operating state. The two specialized models use the same additional reference family but differ in parameter sharing. Their coefficients are fitted at the 0.8-s target and held fixed during the subsequent assessment.
Post-reclosure assessment. Figure 4 uses s and s for the 25 testing lines across the 46 scaled operating states. This is a deliberately demanding transfer test: the open-phase interval is longer than in calibration, and the observation is made one second after reclosure, while the response can still exhibit substantial oscillations. It measures how well the calibrated description reproduces the subsequent electromechanical state, rather than how closely it fits the training snapshot. These settings are not asserted to be a worst case over all faults or operating conditions. The full transient in Figure 11(a) illustrates oscillatory post-reclosure behavior under its own, shorter 0.5-s open-phase interval; it is not the trajectory used to define the timing of Figure 4.
Within every reference–surrogate pair, the initial operating state, faulted line, switching times, phase reference, and observation operator are matched. All five models are scored on the same network-wide observations, with no refitting to the assessment response. For M₁ and M₂, the assessment combines transfer to unseen lines with transfer in duration and observation time; for M₃ and M₄, it measures the latter transfer after case-specific calibration. Subsequent probability estimates use only M₂, with its global pair held fixed.
Separate temporal checks. Appendices D2 and D3 vary and , respectively, with unscaled injections and the global parameter pair held fixed. These experiments separate observation-time sensitivity from switching-duration sensitivity.
Comparison hierarchy. Figure 2(A) defines the five model levels. The baseline M₀ fixes and ; M₁ fits a shared deterministic coupling, and M₂ jointly fits the shared coupling and stochastic amplitude. Models M₃ and M₄ add line and operating-state dependence. The comparison tests increasing parameter specificity within this model family.
Panel (B) also shows the sequence-informed deterministic reference , . This reference motivates the coupling correction; it is neither implemented nor tested as part of the calibration hierarchy. The common protocol means matching all model comparisons within each stage; it does not mean using the same fault catalogue or time settings for training and testing.
Probabilistic Screening Setup
CEM and MC use the same calibrated surrogate, nominal contingency law, screening horizon s, and trajectory-evaluation interval s on the reduced WECC system [23]. The tested global thresholds are s in (19). The global parameters remain fixed during sampling as specified in (E19).
The nominal law selects among modeled faulted lines uniformly and uses an exponential open-phase duration with mean s; see (E2). The probabilities are conditional on a fault occurring within this line set. This 329-line screening catalogue is broader than either of the 25-line calibration and testing sets. Applying the global pair throughout that catalogue extrapolates it beyond the EMT-assessed line sets. The nominal mean 0.1 s specifies the probability law; it is distinct from the fixed 0.8-s calibration duration and 1.5-s hierarchy-test duration. CE adaptation uses 250 trajectories per iteration for 20 iterations, with the 25 largest scores forming the elite set. These 5,000 adaptation evaluations are separate from the final estimation batch and are included in the reported total CEM runtimes.
The duration-resolved assessment in Figs. 12a and 12b uses surrogate simulations. It varies the open-phase duration rather than marginalizing it under the nominal duration law. Its line-specific probabilities must therefore be distinguished from the durationmarginalized map in Figure 5. The event definitions themselves are those of § 1.1 C; they are not redefined for the WECC results.

Figure 5. Contingency-to-vulnerable-element map for the WECC 240-bus assessment at s over s. The x-axis shows critical transformer-connected monitored lines; the y-axis identifies the ten displayed initiating faults. The comma-separated axis label denotes the same fault identifier as Xn#m in the experiment records. The row order is not a scalar-risk ranking. Each cell represents , the probability of more than one second of accumulated overload on element conditional on fault , with open-phase duration and surrogate uncertainty marginalized under the specified nominal law. Entries with identify secondary exposure. The continuous colorbar includes labeled values through 0.015; its colors are not the duration-resolved warning classes.
Results
The results address three questions: what the surrogate retains during the open-phase interval, how much parameter flexibility is needed for state transfer, and what is learned by resolving overload risk across initiating faults and monitored elements. Sampling diagnostics are reported separately from model accuracy.
Open-Phase Trajectory Comparison
Figure 3 compares ParaEMT and one realization of the calibrated stochastic surrogate for the illustrative large-excursion 9#4 contingency during , with . Both calculations show an uneven phase response: many traces remain in a relatively narrow range, while others develop larger excursions before reclosure. The surrogate follows this overall separation of responses, with differences in fluctuation amplitudes. The comparison makes the retained response and the remaining trajectory discrepancies visible during the interval in which the modified coupling and stochastic forcing act.
The full trajectory and corridor-separation diagnostic in Figure 11 continue this example through reclosure. The hierarchy below provides the quantitative comparison of projected mean states.
Calibration Accuracy and Parameter Dependence
Joint optimization of the global stochastic model under the protocol in Sec. III A gives
At and , the fixed global pair gives errors of approximately 0.0968–0.15 across the 25 testing lines and 46 uniformly scaled operating states, with mean 0.098. The one-sided distribution places the mean near its lower end.
Figure 4 compares the hierarchy under this common assessment. The fixed deterministic baseline has errors of approximately 0.24–0.4, while fitting the deterministic coupling in brings most cases into the 0.17–0.29 range. The dominant reduction occurs from to . Because RSE scores the projected mean, this comparison shows the contribution of the noise-induced drift to state transfer within the tested hierarchy.
Models and produce narrower error distributions and smaller further reductions using the additional case-specific calibration records specified in Sec. III A. Substantial variation in their fitted parameters, reported in (D2) and Figure 8, therefore accompanies relatively modest accuracy gains. This is the main calibration trade-off: the global pair captures most of the improvement, while the specialized fits require additional reference observations and more independent parameter fits. The subsequent screening calculations retain on this basis.

Figure 8. Scenario-specific M parameters fitted at for 25 assessment lines and 46 operating states, relative to the global pair . Each of the 1,150 pairs is held fixed in the later hierarchy assessment. Maximum reported deviations are and . Ellipses summarize linewise parameter distributions, not confidence regions.
Coupling and temporal transfer. The fitted coupling lies below the phase-counting baseline, in the same direction as the representative sequence-network estimate . The residual difference distinguishes the line-level sequence calculation from the joint fit to the finite-time network response.
The unscaled-injection studies in Appendices D2 and D3 separate two temporal effects. After a fixed reclosure, discrepancies continue to increase even though the stochastic forcing has ceased: restoring the nominal operator does not erase the mismatch in the attained state. Varying the open-phase duration instead changes how long the modified dynamics act before that state is transferred. These trends complement the hierarchy comparison by showing how error depends on time as well as on the fitted parameterization.
Overload Probabilities and Sampling Efficiency
We evaluate the probabilities defined in (19) using the WECC setup in Sec. III B. Table II reports the probability estimates, final estimation-batch sizes, and total runtimes. Including the adaptation budget specified in Sec. III B, CEM uses 30,000–160,000 surrogate evaluations, compared with 125,000–300,000 for MC. Tables III and IV in Appendix E3 provide the sampling diagnostics and stage-by-stage costs.
Both estimators resolve the decrease in exceedance probability from approximately 0.056 at to 0.0077–0.0079 at and 0.0027–0.0029 at . The absolute differences between their point estimates range
| [s] | Method | Total [s] | ||
| 0 | CEM | |||
| MC | ||||
| 0.5 | CEM | |||
| MC | ||||
| 5.0 | CEM | |||
| MC | ||||
| 10.0 | CEM | |||
| MC |
Table 2. Probability estimates and total runtimes for CEM and Monte Carlo (MC) [24]. Here, is the final estimation-batch size. CEM total runtimes include both the 5,000 adaptation evaluations and final probability estimation.
from to . The estimates distinguish any positive overload from larger accumulated network exposure and reach the probability range.
At every reported threshold, CEM uses a smaller final batch and less total runtime than MC. The MC-to-CEM runtime ratios are 4.952, 4.438, 2.303, and 1.961 for increasing . The advantage at these budgets is therefore largest for the lowest threshold and narrows as the target event becomes rarer. At , CEM reaches an estimate of 0.002723 in 28,281.6 s including adaptation, compared with 0.002900 in 55,468.7 s for MC.
The contribution-variance analysis in Appendix § 2.3.2 favors CEM at and 5 s, and MC at 0 and 10 s. Adaptation selects the largest global scores, so its per-sample benefit varies with the final exceedance threshold.
Contingency-to-Vulnerable-Element Map
The conditional probabilities defined in § 1.1 C resolve the location of the consequences associated with each initiating fault. They are conditional on that fault and the specified operating state, rather than estimates of the fault’s occurrence frequency.
Figure 5 presents the conditional map at for ten initiating faults and the monitored elements shown. Rows are initiating faults, columns are monitored elements, and each cell is . The row order is not a descending ranking of the scalar scores . The figure instead resolves the fault–element pairs underlying the screening output.
The colorbar has labeled ticks at 0, 0.005, 0.010, and 0.015. Larger entries occur in different columns for different faults. The X2#227 contingency, for example, has a prominent secondary-risk entry on . A prominent entry for X112#155 instead lies on the faulted line itself. These contrasting cases show why a global score alone is insufficient: it does not identify the monitored element or distinguish self-exposure from secondary consequences.
Interpretation of the screening thresholds. The duration-resolved assessment in Figure 12b uses and probability boundaries 0.025 and 0.04: safe denotes probabilities below 0.025, warning values from 0.025 through 0.04, and emergency values above 0.04. The heatmap marginalizes open-phase duration under the nominal law, whereas the curves in Appendix § 2.12 condition on duration. Its continuous color scale is independent of those classes. § 2.4 A states the interpretation of these screening thresholds.
The duration-resolved curves supply complementary information: an element can have substantial exposure under a long open-phase interval while retaining a small duration-marginalized probability under the nominal law.
Computational Cost and Reuse of Calibration
Replacing repeated EMT simulations with reduced trajectories addresses the largest per-realization computational burden. A direct EMT ensemble requires a high-fidelity trajectory for each sampled contingency and open-phase duration. The reported mean surrogate evaluation time is approximately 0.19 s; the published WECC benchmark reports 29 s on one core and 28 s on eight cores for a 1-s ParaEMT simulation [8]. The benchmark settings are given in Table V. § 2.4 A states the scope of this computational comparison.
| Cores | Horizon [s] | EMT step [s] | Runtime [s] |
| 1 | 1 | 50 | 29 |
| 8 | 1 | 50 | 28 |
Table V. Published ParaEMT runtime on the reduced WECC 240-bus system [8]. These external benchmark settings differ from the present surrogate screening configuration.
Reference generation and parameter identification are offline costs. The screening ensemble reuses the calibrated model for every adaptation and estimation trajectory. This reuse is shared by CEM and MC; their sampling costs and precision are compared in Sec. IV C. The additional calibration records used by belong to the hierarchy assessment, not to the repeated screening calculation.
Discussion
The advance over the earlier screening construction is the identification of its effective parameters from EMT and the assessment of their transferred response. The WECC hierarchy supports a compact global parameterization, while the specialized fits show the value of additional case-specific information. The temporal studies further identify how switching duration and subsequent propagation affect the state mismatch.
Deterministic moment propagation separates parameter identification from trajectory sampling. The calibrated model can then be reused across counterfactual scenarios, with CE adaptation directing evaluations toward large accumulated exposure. The resulting fault–element map identifies where detailed high-fidelity follow-up is most valuable, connecting EMT reference information to probabilistic screening.
Scope and limitations
The assessment covers the reduced WECC network, the prescribed uniform injection-scale family, and the timing protocols in Sec. III A. The fitted coefficients are effective parameters for those conditions. The stochastic correction is an empirical representation of unresolved variability, rather than a white-noise model derived from the EMT equations. The hierarchy establishes its value within the tested family; distinguishing this mechanism from other reduced descriptions with comparable fitting flexibility requires additional comparisons. Applying the global pair to the broader 329-line screening catalogue extends beyond the EMT-assessed line sets. Changes in topology, device models, or operating conditions require further assessment of the calibration domain.
The metric evaluates the projected mean state at a specified time. It does not establish agreement of the complete waveform, the path-dependent overload duration, or the associated tail probability. Errors that change whether can affect a rare-event probability substantially, whereas state discrepancies away from the threshold may have little effect on that event. CEM and MC evaluate the same surrogate: their agreement assesses probability estimation under that model, not independent agreement with EMT. Increasing the sampling budget reduces statistical uncertainty without correcting model mismatch. EMT-informed overload, probability, and ranking comparisons under the same scenario law remain the corresponding physical-validation tasks. The state-error ordering of the hierarchy is therefore not an established ordering of probability accuracy. The second-moment spectral diagnostic likewise concerns the reduced stochastic dynamics.
The overload indicator uses a linearized power–angle relation and measures accumulated threshold exposure. The large corridor separations in Figure 11 limit its small-angle interpretation during that transient. Outside this regime, the indicator serves as a reduced-model severity measure for high-fidelity follow-up, rather than an exact AC power-flow or thermal-damage calculation. The duration threshold, probability-class boundaries, and angle-diagnostic band are illustrative; they are not NERC or WECC criteria or relay settings. In particular, safe denotes the low-probability screening class, not a guarantee of security. The resulting map prioritizes assessment rather than prescribing protection actions.
The reported probabilities also depend on the specified exponential open-phase-duration law with mean 0.1 s and on successful reclosure. This law is a duration model for the study, not an empirical distribution of protection outcomes. A failed reclosure requires a different switching sequence. Protection-specific branches and broader operating-state families are extensions of the present formulation.
The runtime comparison uses the reported sample budgets and their associated statistical uncertainties, rather than a matched-precision benchmark. Its contribution-variance advantage is event-dependent. Calibration fit counts describe fitting requirements, not measured wall-clock multipliers. The published ParaEMT timings provide external computational context: the model fidelity, numerical settings, and potential hardware differences preclude a controlled speedup claim or an EMT-ensemble runtime extrapolation from those values. The present screening implementation takes hours per threshold and remains a proof of concept. Complete update latency must include adaptation and final estimation at the required precision. Parallel trajectory evaluation and reuse of a CE proposal across nearby operating states remain options to test before operational deployment.
Data Availability
The project code and WECC calibration materials are available through the PowerGridREserach repository [22] and its calibration-data directory [25]. The model-specific calibration and assessment allocations are specified in Sec. III A. Historical Israel-case materials are stored separately. ParaEMT is available from its upstream repository [26]. and moves in the direction of the EMT-calibrated value . Relative to the baseline, this sequence correction accounts for approximately 71% of the observed reduction toward the fitted value. The residual difference, , shows that the line-level sequence equivalent alone does not determine the calibrated coupling.
Figure 7 examines the sensitivity of this result to the unavailable line-specific WECC zero-sequence parameters. Writing , neglecting resistance for this sensitivity calculation, and using gives

Figure 7. Sequence-network coupling versus , with . Shading marks the representative overhead-line range 2.0–4.6. The representative 230-kV ratio gives (complex-impedance estimate: 0.560); reproducing the fitted by this correction alone requires .
Representative overhead-line values –4.6 give –0.549. The representative value gives , consistent with the complex-impedance estimate 0.560 above. In contrast, reproducing through the line-level sequence correction alone would require , well outside the representative range considered here.
Local Calibration and Moment Diagnostics
Graph-Local Calibration
Calibration uses the electromechanical response at the endpoints of the faulted line, where the stochastic forcing in (2) acts [30]. The endpoint restriction applies to parameter identification, not to propagation through the network.
Let denote the structure-preserving network graph introduced in § 1.1 A. For a bus and a faulted line , define the graph distance from the bus to the line by
where denotes the shortest-path distance on . For a locality radius , define
The present analysis uses the endpoint-local case . The general neighborhood in (C2) then reduces to the endpoint set introduced in (3):
Only these two buses enter the calibration objective. Broader neighborhoods, including , are not used in the reported fits.
Let denote the corresponding diagonal bus mask,
When both phase and frequency variables are included in the observation operator, the same bus mask is applied to both components; the resulting observation-space mask is denoted by . In all results reported here, this mask is evaluated at .
ParaEMT produces one deterministic trajectory for each faulted line and pre-fault operating state . Consequently, the covariance target used in the calibration is not an ensemble covariance obtained from repeated EMT realizations. Instead, it measures the scatter of the projected finite-time EMT state across the training operating-state family at fixed faulted line. For the global fit, and all observations in this target use s. Let denote the 46 scaled training operating states specified in Sec. III A, and define
The corresponding sample covariance is
The target measures variation across training operating states, not numerical randomness in ParaEMT. It constrains the variability represented by in the global surrogate. The fitted parameter distributions are reported in Appendix D.1.
The objective is given in (14) in the main text. The calibration weights are fixed throughout the analysis as
The mean term in (14) constrains the transferred endpoint response, while the covariance term uses the operating-state scatter in (C6). The parameters are fitted jointly because also enters the mean dynamics. The regularization acts on the endpoint-masked predicted covariance
Because the mask introduces zero directions outside the endpoint subspace, the determinant is evaluated after adding the fixed diagonal regularization
The regularization term is then
The diagonal shift makes the covariance entering the determinant nonsingular. The absolute log-determinant penalizes departures from the reference determinant scale in either direction; it does not bound the condition number or control each covariance eigenvalue separately. This penalty supplements, rather than replaces, the operating-state covariance target in (C6).
The global fit uses only training-line endpoint observations at the 0.8-s reclosure snapshot. Models and additionally use 0.8-s observations from the assessment lines, with one pair per line or per line–operating-state combination, respectively. Those additional observations do not enter the globally shared fit. The coupled swing dynamics propagate the disturbance throughout . Network-wide comparison at then assesses all models at a duration and observation time not used for identification. The faulted lines themselves are held out only for the global models.
Mean-Square Stability Analysis
The second-moment diagnostic characterizes the reduced stochastic dynamics. Multiplicative noise can destabilize that moment even when the deterministic drift is stable [31]. During , the homogeneous operator, acting on a covariance-matrix argument , is
The term in (13) is a nonhomogeneous forcing term and therefore does not enter the spectral stability criterion.
The corresponding mean-square stability margin is defined by
Thus, implies exponential mean-square stability of the homogeneous open-phase stochastic dynamics, whereas indicates a noise-induced second-moment instability. For a scalar SDE , this criterion reduces to , illustrating why deterministic stability alone is insufficient in the presence of multiplicative noise [32].
We evaluate (C13) at the globally calibrated WECC parameters using the same inertia, damping, network-Laplacian, and fault-local noise structure employed by the surrogate. The calculation removes the global-phase mode before evaluating the reported physical spectrum. For the limiting tested contingency X15≠228, the rightmost physical second-moment eigenvalue is
which gives
The reported operator therefore has a positive margin at for the homogeneous second moment. Because the stochastic forcing is removed at , the subsequent interval is governed by the deterministic covariance operator
The post-reclosure covariance therefore evolves under the nominal deterministic operator.
Supporting Calibration and Transient Diagnostics
This appendix reports the parameter distributions, the two temporal sensitivity studies, and the full illustrative large-excursion trajectory. Their interpretation is given with the main calibration results in Sec. IV B.
Parameter Variation Across Scenarios
The scenario-specific fits use the additional reclosure observations and parameter-sharing allocation described in Sec. III A. They yield 1,150 parameter pairs,

Figure 9. Horizon-dependent for the 25 WECC testing contingencies at fixed and unscaled injections , with held fixed. The observation time varies from reclosure to . Dashed red curves are empirical guides, not universal bounds.
Across the 25 assessment lines and 46 operating states, these 0.8-s calibrated parameters span approximately
with mean values of approximately
Figure 8 displays the variation with fault identity and operating condition in the respective parameter units. The corresponding state-transfer errors and calibration-effort trade-off are analyzed in Sec. IV B.
Horizon Dependence of the Map Error
This experiment varies at fixed for the 25 testing lines. The injection vector is not rescaled: in (24). The initial operating condition and global pair remain fixed; the pair is not refitted as the observation time changes. In the notation below, the fixed operating state is suppressed. For the projected EMT state , the relative map error is
Figure 9 shows that line-dependent discrepancies continue to grow after reclosure, reaching approximately – at . These are nominal-injection results, whereas Figure 4 aggregates the 46 scaled operating states. The increasing spread indicates that the subsequent propagation retains a dependence on fault location.
Sensitivity to Open-Phase Duration
The complementary experiment fixes and varies from 0 to for the same 25 testing lines. The injections remain unscaled , and the global pair is not refitted. This study changes the switching duration, not the operating-state scale. For faulted line , the error is
Unlike (D4), this comparison changes the switching time rather than the observation horizon.
Figure 10 shows error approaching zero as and increasing with the open-phase duration. A longer interval extends the action of and the localized multiplicative forcing in (2), changing the state supplied to post-reclosure evolution. The widening spread shows that this sensitivity depends on fault location.

Figure 10. Open-phase-duration dependence of for the 25 WECC testing contingencies at fixed and unscaled injections (). The global pair is held fixed while varies from 0 to 1.5 s. Dashed red curves are empirical guides, not universal bounds.
The two temporal sweeps use the same random-seed configuration and agree at , where both report . The upper endpoint is also the largest displayed error over the tested duration range. The seed identifies the numerical run; the exact conditional mean in (12) remains seed-independent for fixed inputs.
Illustrative Transient and Large-Angle Diagnostic
The corresponding linearized no-overload region is written as
Because (16) depends on phase differences, uniform common-mode phase motion does not contribute directly to the screening metric. Nevertheless, the linearized relation assumes sufficiently small phase separation. We examine this assumption for the illustrative large-excursion 9#4 contingency in Figure 11. Its 0.5-s open-phase interval and 10-s diagnostic window are independent of the calibration and hierarchy-test timings. Panel (a) shows the absolute phase trajectories obtained from ParaEMT and the calibrated surrogate, while panel (b) reports the corresponding separation . The phase difference repeatedly exceeds the 0.5-rad diagnostic threshold, particularly during the large initial transient, and remains intermittently above this level during the subsequent response. For this illustrative case, the maximum separation is approximately 3.77 rad in ParaEMT and 4.24 rad in the surrogate. Thus, the large phase excursions in Figure 11 cannot be attributed solely to common-mode drift, and the small-angle approximation is not uniformly valid throughout this severe transient. These corridor-separation results give the diagnostic evidence for the indicator’s interpretation in § 2.4 A.
Sampling Laws, Estimators, and Statistical Records
We estimate the probabilities in (18)–(19) by combining Cross-Entropy Method (CEM) proposal adaptation with a separate importance-sampling estimation stage [12, 13, 14]. Both the learned line probabilities and the learned mean duration are retained for final estimation. Adaptation favors contingencies with large overload scores without changing the nominal probability law or refitting the physical surrogate parameters. Throughout this appendix, denotes the mean of the exponential proposal, in seconds, while s is the mean of the nominal duration law.
Sampling space and distributions. Let .
The contingency sampling space is
where identifies the faulted line and is a realization of the open-phase duration . The screening horizon is held fixed. The WECC implementation samples only faulted lines, assigning zero probability to under both the nominal and proposal laws. Thus, the estimates are conditional on a fault occurring within the modeled line set.
The nominal law uses uniform line selection and an exponential duration,
The adaptive proposal has categorical line probabilities and a single shared exponential mean,
with . The mean is shared by all faulted lines; no line-specific duration mean is fitted. Both and have units s. The nominal expected open-phase duration is therefore 0.1 s, whereas the proposal mean changes during adaptation. The operating-state scaling factor in (24) is a different quantity.
The categorical probabilities satisfy
and the adjustable parameters are
The required support condition is
where denotes the final estimation proposal defined below. Uniform initialization, positive duration parameters, and positive categorical smoothing preserve support during a finite adaptation stage.
Elite selection and parameter updates. The WECC adaptation uses
Here, is the number of samples per adaptation iteration, is the fixed iteration count, and is the fraction of the previous parameter value retained in each update. The symbol distinguishes this numerical smoothing parameter from the physical coupling . The initial proposal mean s differs from the nominal mean s. Updating the proposal does not change the expected duration under the nominal law.
At iteration , draw contingencies from , evaluate the corresponding trajectories through (2), and compute . The elite set contains the indices of the 25 largest scores,
The operator selects exactly 25 observations; ties do not enlarge the elite set. The empirical threshold is distinct from the event threshold used for final estimation.
Elite membership is represented by the binary indicator
Maximizing gives
The mean-duration update uses the pooled elite durations, consistent with the shared mean in (E3). The smoothed updates are
The categorical probabilities are renormalized after each update to remove roundoff. Even when a line has no elite observations, in exact arithmetic.
This is an unweighted elite-based CE proposal-learning rule. Unlike the nominal-density-weighted CE formulation for rare-event estimation, it does not include in the adaptation objective. Importance weights are applied in the final probability estimate. Adaptation is also independent of the chosen final threshold ; it favors large global scores rather than stopping when a prescribed event level is reached.
The stopping rule is a fixed computational budget,
No parameter-convergence tolerance or early-stopping test is used. The largest score, elite threshold, shared mean duration, and maximum categorical probability are recorded at every iteration.
Fresh-batch probability estimation. After the twentieth iteration, the complete learned proposal is frozen:
Thus, : the final batch retains both learned components. The probabilities belong to the nominal law and the initialization, not to the final categorical proposal. In general, . The superscript * denotes the frozen proposal, not a claim of global optimality or verified convergence.
Let denote the stochastic driving realization. Matching the strict threshold comparison in the probability-estimation code, define
for . The additional coordinate makes explicit that overload is not determined by alone. The stochastic drivers retain the same conditional law under the nominal and proposal simulations, so their likelihood factors cancel. At , strict exceedance selects any positive overload; the event would instead have probability one.
After adaptation, a fresh batch is drawn,
with fresh stochastic driving realizations. Conditional on the adaptation history, this estimation batch consists of independent draws from the frozen proposal; no adaptation observations are reused in the estimator.
(23) in the main text gives the estimate. The likelihood ratio corrects both changes of distribution:
The categorical factor cannot be omitted merely because the nominal fault law is uniform. The denominator must use the probabilities that actually generated the final line samples.
Provided that samples are generated from the density used in this ratio, (E6) and the fresh-batch construction imply , where denotes the adaptation history. The estimator is therefore unbiased for the specified surrogate probability without self-normalizing the weights [14]. This property does not remove model error relative to ParaEMT.
| Algorithm 1 N1Plus with elite-based categorical–exponential adaptation |
| Require: Surrogate score; nominal law (E2); screening horizon ; threshold ; final sample size . |
| 1: Set , , , and . |
| 2: Initialize and . |
| 3: for do |
| 4: Draw 250 contingencies from and fresh stochastic drivers. |
| 5: Evaluate (2) with and compute the global overload scores to . |
| 6: Select the 25 largest scores using (E8). |
| 7: Apply (E10)–(E11). |
| 8: Renormalize and record the adaptation diagnostics. |
| 9: end for |
| 10: Freeze and for every . |
| 11: Draw fresh line indices from and durations from the exponential law with mean , with fresh stochastic drivers. |
| 12: Compute the global and monitored-line overload indicators. |
| 13: Evaluate (23) using (E16). |
| 14: Report the estimate, sampling uncertainty where applicable, final proposal, and separate adaptation and estimation costs. |
Table 1 (PDF p. 21).
For weighted indicators in (23), the estimated standard error is
where includes the full joint likelihood ratio in (E16). This standard-error interpretation requires and finite variance. For finitely many line categories with , the condition ensures finite second moments of the weights under the mean-parameterized exponential proposal. This follows because the duration factor in decays only when . Small categorical probabilities can still produce large weights; a finite empirical standard error alone does not establish reliable tail sampling.
Algorithm 1 summarizes proposal adaptation and the fresh-batch estimation stage using the complete learned joint distribution.
The adaptation budget is 250 × 20 = 5,000 surrogate evaluations. Consequently, a run with fresh estimation samples requires evaluations; the supplied example therefore requires 40,000 evaluations in total. End-to-end computational comparisons must include both stages. The supplied scorer uses . When a sampled duration exceeds , re-closure lies beyond the screening window; the original duration must still be retained in the likelihood ratio.
Both stages use the learned fault-location and duration components. The categorical probabilities emphasize particular corridors, while the shared mean duration changes the frequency of long open-phase intervals. Because (E3) factorizes, it does not introduce line–duration dependence. Concentration of is an interpretable indicator of which corridors dominate the adapted elite set, not a change in their physical failure frequencies. Quantitative contributions to the nominal overload probability, and conditional criticality rankings, must be computed with the full joint likelihood ratio rather than identified with itself.
The same estimator applies to a line-specific event by replacing the global score with . A global-score proposal can be reused when the support condition holds, although its efficiency need not be uniform across monitored lines. Further justification and definitions for rare-event/instanton sampling are given in Sections VI.A–B of [1]; the connection to static contingency screening is discussed in [33].
Conditional Contingency-to-Element Estimates
Under the uniform nominal fault law , a fresh CEM batch gives
where and is the full joint likelihood ratio in (E16). Replacing the final indicator by gives the global conditional estimate. The normalization by distinguishes conditional risk from a fault’s contribution to the unconditional mixture over fault locations. The learned proposal masses are not substitutes for these weighted conditional probabilities.
Integration of EMT-Informed Calibration
For each sampled scenario , the dynamics in (2) are evaluated with and the global M parameters
These physical parameters remain fixed during CE adaptation and final estimation; they are not proposal parameters. The stochastic driving law, pre-fault operating-state prescription, screening horizon, and overload definitions are unchanged from the nominal assessment. Each trajectory is scored using (16) and (17), and its contribution is weighted with the complete likelihood ratio in (E16). The conditional normalization is given in (E18).
WECC Probability Diagnostics and Runtime Records
Assessment protocol. This appendix supplies the statistical diagnostics and timing breakdown behind the main-text comparison in Table II. The nominal law, fixed physical parameters, horizon s, evaluation interval s, and thresholds are those of Sec. IV C. MC samples directly from the nominal law. CEM uses the complete frozen proposal , and the likelihood ratio in (E16); its final batch is independent of the adaptation draws conditional on the fitted proposal.
The reported probability estimates use the global overload event at the four thresholds specified in the WECC study,
Probability estimates and sampling diagnostics. Table III reports the probability estimates and final-batch uncertainty and weight-concentration diagnostics. Table IV separately records the evaluation budgets and adaptation and estimation times. The main-text table combines only , method, , , and total runtime; the additional columns are retained here for reproducibility and the variance comparison below.
| [s] | Method | RelSE | |||||
| 0 | CEM | ||||||
| MC | |||||||
| 0.5 | CEM | ||||||
| MC | |||||||
| 5.0 | CEM | ||||||
| MC | |||||||
| 10.0 | CEM | ||||||
| MC |
Table III. Final-batch statistical reporting. Probability estimates, standard errors, relative standard errors, weight-based ESS, and maximum normalized weights are obtained using the WECC simulations.
| [s] | Method | Adaptation [s] | Estimation [s] | Total [s] | ||
| 0 | CEM | |||||
| MC | 0.000 | |||||
| 0.5 | CEM | |||||
| MC | 0.000 | |||||
| 5.0 | CEM | |||||
| MC | 0.000 | |||||
| 10.0 | CEM | |||||
| MC | 0.000 |
Table IV. Separately reported CEM and MC [24] evaluation budgets and stage runtimes. CEM totals include 5,000 adaptation evaluations, and total time is the sum of the supplied stage records.
For the reported estimator, define
where , , and . The categorical factor is retained in every importance weight.
For independent final samples with finite-variance weighted indicators, the estimated standard error and weight diagnostics are
The relative standard error RelSE measures sampling uncertainty in the probability estimate and is distinct from . The raw likelihood weight is ; its normalized diagnostic counterpart is , with . Normalized weights are used only for diagnostics; the probability estimate remains the ordinary importance-sampling estimate in (23). Weight-based ESS and maximum normalized weight describe weight concentration, not event-specific precision by themselves [34].
For independent, unweighted MC samples,
These relative standard errors require and . Normalized nonnegative weights also satisfy
The entries in Table III satisfy these identities.
Contribution-variance comparison. Comparing different sample budgets requires separating sampling precision from the number of evaluations. Since is the sample variance of an individual estimator contribution, define
Here, both sample sizes refer exclusively to the final estimation batches. For independent samples, this compares the estimated per-draw variances: values above one favor CEM, whereas values below one favor MC. It is therefore the relevant sample-based comparison of fixed- efficiency, rather than the ratio of the selected batch sizes. It does not account for proposal-adaptation cost.
For the results shown in Table III, the ratios are
The contribution variance is lower for CEM at and , and lower for MC at and . The CEM final batches are smaller, so their standard errors also reflect the different sample counts. Equation (E25) separates that sample-size effect from the variability of individual contributions.
Computational cost. The specified CE adaptation uses
Each displayed CEM run includes one adaptation. Reusing a proposal across thresholds incurs that adaptation cost only once. Complete CEM runtime is
This includes the pilot trajectories, sampling, elite selection, parameter updates, final trajectory evaluations, and weighting.
Table IV separates adaptation time from final estimation and records both final-batch and total evaluation counts. Adding its two time columns gives the totals summarized in Table II; no per-realization extrapolation is used. The comparison includes one adaptation per CEM run and pertains to the reported sampling budgets.
Surrogate runtime and published ParaEMT benchmark. The separately reported mean computation time for one surrogate trajectory is approximately . The complete stage timings are recorded separately in Table IV.
The published ParaEMT benchmark in [8] reports
These measurements correspond to a 1-s simulated horizon and a internal integration step on Windows, using a machine equipped with two Intel Xeon Platinum 8280 2.7-GHz processors and 512-GB RAM.
The 0.01-s trajectory-evaluation interval in the screening procedure and the published internal EMT integration step describe different numerical operations. Table IV supplies the measured stage costs for the surrogate ensemble; Sec. V A states the limits of comparisons with the external EMT benchmark.
Duration-Resolved Screening
Line-specific screening. Figure 12 displays line-specific screening quantities as the open-phase duration varies over a fixed 10-s observation window. The overload integrals use the screening horizon , as defined in Sec. II A. The top panel shows accumulated overload duration; the bottom panel shows the probability of with . This line-specific event is distinct from the global event . The faulted element and the monitored element need not coincide.

(a) (a) Accumulated overload versus . Green identifies faulted lines (); red identifies critical transformer-connected monitored lines (). The indicator measures overload duration, not peak flow.

(b) (b) Exceedance probability for s. Illustrative classes: safe (), warning (–), and emergency ().
Figure 12. Duration-resolved WECC screening: (a) accumulated line overload; (b) line-specific exceedance probability. Both panels vary while accumulating exposure through s. The s notation inside these original graphics denotes that total trajectory window, not a calibration target. The label in panel (b) denotes the exposure threshold s used throughout the text. These duration-resolved quantities are distinct from the global, duration-marginalized estimates in Table II.
These curves condition on duration. Their screening classes are defined in Sec. IV D, with the scope of their interpretation stated in Sec. V A.
Relation to the main-text screening thresholds. The duration-resolved assessment in Figure 12a identifies ten monitored transmission lines with for the reported open-phase durations ; these lines are distinct from the ten initiating faults displayed as heatmap rows. The associated duration-resolved probabilities in Figure 12b enter the illustrative emergency class for the reported longer-duration cases. These are duration-conditioned results for severe cases within the tested set. The fault-conditioned, duration-marginalized probabilities in Figure 5 instead lie below the first classification boundary and use an independent continuous color scale.
References
- [1]Almada, A., Pagnier, L., Goldshtein, I., Kazi, S. R., & al., . et . (2026). Real-time dynamic n-1 screening: Identifying high-risk lines and transformers after common faults. In arXiv preprint. arXiv:2602.12293
- [2]ENTSO-E. (2007). Operational handbook and system operation guidelines. <a href="https://doi.org/https://eepublicdownloads.entsoe.eu/clean-documents/pre2015/publications/ce/report_2007_2.pdf" rel="noopener noreferrer" target="_blank">https://doi.org/https://eepublicdownloads.entsoe.eu/clean-documents/pre2015/publications/ce/report_2007_2.pdf</a>
- [3]North American Electric Reliability Corporation (NERC), “Reliability standards for the bulk electric systems of north america.” https://www.nerc.com/pa/Stand/Pages/ReliabilityStandards.aspx, 2023. Accessed: 2026-04-27.nerc.com/pa/Stand/Pages/ReliabilityStandards.aspx
- [4]A. Montanari, M. Tavares, and C. Portela, “Adaptative single-phase autoreclosing based on secondary arc voltage harmonic signature,” parameters, vol. 6, p. 7, 2009.
- [5]Godoy, E., Celaya, A., Altuve, H. J., Fischer, N., & Guzmán, A. (2012). Tutorial on single-pole tripping and reclosing. 39th Annual Western Protective Relay Conference.
- [6]Van Cutsem, T., & Vournas, C. (2007). Voltage stability of electric power systems.
- [7]M. Xiong, B. Wang, D. Vaidhynathan, J. Maack, M. J. Reynolds, A. Hoke, K. Sun, D. Ramasubramanian, V. Verma, and J. Tan, “An open-source parallel emt simulation framework,” Electric Power Systems Research, vol. 235, p. 110734, 2024.DOI
- [8]M. Xiong, B. Wang, D. Vaidhynathan, J. Maack, M. J. Reynolds, A. Hoke, K. Sun, and J. Tan, “Paraemt: An open source, parallelizable, and hpc-compatible emt simulator for large-scale ibr-rich power grids,” IEEE Transactions on Power Delivery, vol. 39, no. 2, pp. 911–921, 2023.DOI
- [9]Wood, A. J., Wollenberg, B. F., & Sheblé, G. B. (2013). Power generation, operation, and control.
- [10]P. W. Sauer, M. A. Pai, and J. H. Chow, Power system dynamics and stability: with synchrophasor measurement and power system toolbox. John Wiley & Sons, 2017.
- [11]Pavella, M., Ernst, D., & Ruiz-Vega, D. (2012). Transient stability of power systems: a unified approach to assessment and control.
- [12]R. Y. Rubinstein, “Optimization of computer simulation models with rare events,” European Journal of Operational Research, vol. 99, pp. 89–112, 1997.DOI
- [13]P.-T. De Boer, D. P. Kroese, S. Mannor, and R. Y. Rubinstein, “A tutorial on the cross-entropy method,” Annals of Operations Research, 2005.DOI
- [14]A. B. Owen, Y. Maximov, and M. Chertkov, “Importance sampling the union of rare events with an application to power systems analysis,” Electronic Journal of Statistics, vol. 13, no. 1, pp. 231–254, 2019.arxiv.org/abs/1710.06965
- [15]Commission, I. E. (1984). Iec 60050-441-17-44: dead time (during auto-reclosing). IEC Electropedia.
- [16]C. Gardiner, Stochastic Methods: A Handbook for the Natural and Social Sciences. Berlin, Heidelberg: Springer, 4 ed., 2009.
- [17]Øksendal, B. (2003). Stochastic Differential Equations: An Introduction with Applications.
- [18]R. Mannella and P. V. E. McClintock, “Itô versus stratonovich: 30 years later,” Fluctuation and Noise Letters, vol. 11, no. 1, p. 1240010, 2012.DOI
- [19]Rockies, N. L. of . the . (n.d.). ParaEMT Public: 240-Bus WECC PSS/E Model. GitHub Repository. <a href="https://doi.org/https://github.com/" rel="noopener noreferrer" target="_blank">https://doi.org/https://github.com/</a>
- [20]P. Kundur and O. P. Malik, Power system stability and control, vol. 7. McGraw-hill New York, 1994.
- [21]Chiang, H.-D. (2011). Direct methods for stability analysis of electric power systems: theoretical foundation, BCU methodologies, and applications.
- [22]Almada, A. (n.d.). PowerGridREserach: Power-grid dynamic screening code. GitHub Repository. <a href="https://doi.org/https://github.com/AyrtonAlmada/PowerGridREserach" rel="noopener noreferrer" target="_blank">https://doi.org/https://github.com/AyrtonAlmada/PowerGridREserach</a>
- [23]J. E. Price and J. Goodin, “Reduced network modeling of wecc as a market design prototype,” in 2011 IEEE Power and Energy Society General Meeting, pp. 1–6, IEEE, 2011.DOI
- [24]Kroese, D. P., & Rubinstein, R. Y. (2012). Monte carlo methods. Wiley Interdisciplinary Reviews: Computational Statistics, 4, 48–58.
- [25]Almada, A. (2026). WECC240 calibration reference trajectories and provenance. GitHub Repository. <a href="https://doi.org/https://github.com/AyrtonAlmada/PowerGridREserach/tree/main/data/wecc240/calibration" rel="noopener noreferrer" target="_blank">https://doi.org/https://github.com/AyrtonAlmada/PowerGridREserach/tree/main/data/wecc240/calibration</a>
- [26]developers, P. (n.d.). ParaEMT: Public source-code repository. GitHub Repository. <a href="https://doi.org/https://github.com/NatLabRockies/ParaEMT_public" rel="noopener noreferrer" target="_blank">https://doi.org/https://github.com/NatLabRockies/ParaEMT_public</a>
- [30]G. A. Pavliotis, “Stochastic processes and applications,” Texts in applied mathematics, vol. 60, pp. 41–43, 2014.DOI
- [31]Arnold, L. (1974). Stochastic Differential Equations: Theory and Applications.
- [32]R. Khasminskii, Stochastic Stability of Differential Equations. Berlin, Heidelberg: Springer, 2 ed., 2012.DOI
- [33]M. Chertkov, F. Pan, and M. G. Stepanov, “Predicting failures in power grids: The case of static overloads,” IEEE Transactions on Smart Grid, 2010.arxiv.org/abs/1006.0671
- [34]V. Elvira, L. Martino, and C. P. Robert, “Rethinking the effective sample size,” International Statistical Review, vol. 90, no. 3, pp. 525–550, 2022. # EMT Reference, WECC Case, and Event Notation {#sec:A} This appendix identifies the reference implementation and case inputs used in {xref}`Sec. III` and summarizes the switching and observation settings using the time conventions in {xref}`Table I`. ## Reference Model and WECC Case {#sec:A.1} ParaEMT couples three-phase nodal network equations with synchronous-machine, excitation, governing, stabilizer, and converter models. Its differential-algebraic formulation is evaluated using trapezoidal integration and sparse linear algebra {citep}`[7, 27, 28]`. The observation operator in (7) selects the electromechanical variables compared with the surrogate; fast electrical and device states remain in the reference model. The experiments use the public reduced WECC 240-bus case, with conventional and renewable generation, loads, transmission corridors, and approximately 20% inverter-based-resource penetration {citep}`[8]`. The case is distributed in the `models/240bus_psse` directory of ParaEMT, with the `WECC240.raw` network description and `240bus.xls` model data {citep}`[19]`. All calibration levels use this same topology and set of dynamic-model parameters; only the specified contingency and operating condition change. {xref}`Figure 6` shows the network. ::: figure #fig:6 (6)  bbox=0.52,0.07,0.92,0.47 caption: One-line representation of the reduced WECC 240-bus calibration case, adapted from [29]. ::: ## Timeline Reference {#sec:A.2} The switching sequence is defined in {xref}`Sec. II A` and illustrated in {xref}`Fig. 1`. {xref}`Table I` collects the time conventions; the experiment-specific settings and calibration allocations are given in {xref}`Sec. III A`. The 9#4 illustration uses $T_{\mathrm{op}} = 0.5$ s and ends at 10 s. These settings remain distinct from the nominal screening-duration distribution. # Sequence-Network Interpretation of the Coupling {#sec:B} A quantitative sequence-network calculation provides a first physical check of the fitted coupling. Among the WECC testing contingencies, line X110#154 has positive-sequence reactance $X_1 = 0.02886$ p.u., equal to the median reactance of the tested line set, and is therefore selected as a representative case. For a one-conductor-open condition, the negative- and zero-sequence networks contribute through $$ Z_{20} = Z_2 \mathbin\Vert Z_0 = \frac{Z_2 Z_0}{Z_2 + Z_0}, $$ so that the effective positive-sequence transfer impedance is $$ Z_{\mathrm{op}} = Z_1 + (Z_2 \mathbin\Vert Z_0). $$ Using the standard transmission-line approximation $Z_2 \simeq Z_1$ together with representative overhead-line sequence ratios $R_0/R_1 = 13.03$ and $X_0/X_1 = 3.49$ gives $$ Z_1 \simeq 0.00202 + \mathrm{j}0.02886,\qquad Z_0 \simeq 0.02628 + \mathrm{j}0.10072, $$ and $$ Z_2 \mathbin\Vert Z_0 \simeq 0.00249 + \mathrm{j}0.02257. $$ Because $\alpha$ scales the line coupling in (4), the corresponding comparison is made at the susceptance level, $$DOI