betaFeedback

WorksEngineering

A Hybrid Systems Approach to Parameterizing the PERC1 Model for Data Centers Using Synchrophasor Measurements

Abstract

The Power Electronic Reconnecting and Ceasing (PERC1) load model represents the timed ceasing and reconnection of large power-electronic loads, including data centers. Voltage thresholds and timers govern its response, with a ramp on reconnection, making it a hybrid dynamical system. This paper formulates PERC1 as a hybrid automaton and uses that structure to identify its parameters from synchrophasor records of grid disturbances. We propose a four-stage identification procedure: mode-sequence identification, estimation of logic parameters from event-time constraints, and joint Gauss-Newton refinement using hybrid trajectory sensitivities and saltation matrices. The resulting salted Fisher information is then used to analyze parameter identifiability. Two structural results are established. The detection and actuation delays are identifiable only through their sum unless a second dip interrupts the reconnection ramp, and a single disturbance leaves one unidentifiable direction in each threshold-delay pair, which two disturbances with different voltage slopes resolve. A case study using synthetic 60 frames/s synchrophasor records demonstrates recovery of the ceasing and reconnection dynamics and identifies poorly observable parameters.

Introduction

Data centers used 4.4% of United States electricity in 2023, and projections for 2028 range from 6.7% to 12% [1]. On July 10, 2024, a permanent 230 kV fault and three reclosing attempts sent about 1,500 MW of data center load to on-site backup supply. Most of it did not return for hours [2]. On July 22, 2026, a single 230 kV line fault in the Dominion zone of PJM removed about 3,800 MW in two waves one minute apart [3]. In both events, the customers’ own protection and controls initiated the disconnection. NERC’s incident review therefore recommends that interconnection requirements include dynamic response models of large loads [2].

The Power Electronic Reconnecting and Ceasing (PERC1) model was introduced in late 2025 to address this modeling gap [4], [5], and NERC recommends its use for representing the ceasing and reconnection of any type of data center [6]. Derived from the EPRI aggregated electric-vehicle charger model [7], it includes voltage-dependent load characteristics, current limits, washout and lead–lag blocks, and logic governing load ceasing and reconnection. The model reduces the connected load fraction when voltage remains below a specified threshold for a prescribed duration and ramps up following voltage recovery.

The PERC1 parameters are generally unknown to system planners. Although uninterruptible power supply (UPS) vendors provide parameter ranges [6], these may not represent the response of individual facilities. To the best of our knowledge, no method has been proposed for jointly estimating PERC1’s continuous and logic parameters from synchrophasor records while also analyzing their identifiability. Phasor measurement units (PMUs) at the point of interconnection provide disturbance records that could support such estimation. The challenge is to estimate the model parameters from these records and determine which parameters can be uniquely identified.

Measurement-based approaches have been widely studied for load model parameter estimation [8], [9], including PMU-based estimation [10]. Gradient-based methods rely on smooth dependence of model outputs on the parameters. In PERC1, however, changes in the logic parameters alter when the load ceases and reconnection begins. Consequently, a least-squares objective evaluated using fixed-step simulations can be piecewise constant with respect to these parameters, limiting the effectiveness of gradient-based optimization. For the WECC composite load model, which also includes load tripping and reconnection, derivative-free optimization has been used with replayed or closed-loop simulations [11]. However, such approaches do not establish which parameters can be identified from the available measurements, as illustrated by the difficulties in estimating trip thresholds reported in [11].

Hybrid dynamical systems offer a natural framework for addressing this problem [12]. Hiskens and Pai Hiskens and Pai [13] showed that the trajectory sensitivities of such systems jump at every transition and used them to fit models to disturbance recordings [8]. The saltation matrix describes the jump [14], and the salted Fisher information framework of [15] uses hybrid trajectory sensitivities to analyze parameter identifiability and establishes a hybrid persistence-of-excitation condition for local identifiability.

This paper develops a hybrid dynamical systems approach for estimating PERC1 parameters from synchrophasor measurements. We formulate PERC1 as a hybrid automaton and propose a four-stage identification procedure based on event-time constraints and hybrid trajectory sensitivities. Using salted Fisher information [15], we establish conditions for parameter identifiability and show how multiple disturbances can resolve ambiguities between voltage thresholds and timing parameters.

This paper proceeds as follows. Section II writes PERC1 as a hybrid automaton, and Section III builds the identification on it. Section IV tests the procedure on synthetic PMU records of a data center.

PERC1 as a Hybrid Automaton

In what follows, parameter and variable notation follows the vendor documentation [4]. Quantities are in per unit on the model base MBase=MWinit/Lfm\mathrm{MBase} = \mathrm{MW}_{\mathrm{init}}/L_{\mathrm{fm}}, where MWinit\mathrm{MW}_{\mathrm{init}} and Mvarinit\mathrm{Mvar}_{\mathrm{init}} are the pre-disturbance active and reactive power and LfmL_{\mathrm{fm}} is a loading factor [4]. The inputs are the terminal voltage magnitude VtermV_{\mathrm{term}} and the frequency deviation Δω\Delta\omega. The block diagram in [4] has eight states: the filtered voltage VfiltV_{\mathrm{filt}} and frequency wfiltw_{\mathrm{filt}}, the washout outputs dP\mathrm{dP} and dQ\mathrm{dQ}, the lead–lag outputs PctrlP_{\mathrm{ctrl}} and QctrlQ_{\mathrm{ctrl}}, and the currents IpI_p and IqI_q. We represent the washout and lead–lag blocks with the states wpw_p, wqw_q, ℓp\ell_p, ℓq\ell_q. Each washout is written in the partial-fraction form sK/(1+sT)=(K/T)[1−1/(1+sT)]sK/(1+sT)=(K/T)[1-1/(1+sT)], so the measured voltage is never differentiated, and each lead–lag as a feedthrough plus a low-pass state, (1+sTa)/(1+sTb)=Ta/Tb+(1−Ta/Tb)/(1+sTb)(1+sT_a)/(1+sT_b)=T_a/T_b+(1-T_a/T_b)/(1+sT_b). The corresponding continuous dynamics are

TvV˙filt=Vterm−Vfilt,Tfw˙filt=σ(Vterm)Δω−wfilt,(1)T_v\dot{V}_{\mathrm{filt}}=V_{\mathrm{term}}-V_{\mathrm{filt}},\qquad T_f\dot{w}_{\mathrm{filt}}=\sigma(V_{\mathrm{term}})\Delta\omega-w_{\mathrm{filt}}, \tag*{(1)}
Tvpw˙p=Vterm−wp,Tvqw˙q=Vterm−wq,(2)T_{vp}\dot{w}_p=V_{\mathrm{term}}-w_p,\qquad T_{vq}\dot{w}_q=V_{\mathrm{term}}-w_q, \tag*{(2)}
Tbpℓ˙p=up−ℓp,Tbqℓ˙q=uq−ℓq,(3)T_{bp}\dot{\ell}_p=u_p-\ell_p,\qquad T_{bq}\dot{\ell}_q=u_q-\ell_q, \tag*{(3)}

with outputs

dP=KvpTvp(Vterm−wp),Pctrl=ℓp+TapTbp(up−ℓp),(4)\mathrm{dP}=\frac{K_{vp}}{T_{vp}}(V_{\mathrm{term}}-w_p),\qquad P_{\mathrm{ctrl}}=\ell_p+\frac{T_{ap}}{T_{bp}}(u_p-\ell_p), \tag*{(4)}
dQ=KvqTvq(Vterm−wq),Qctrl=ℓq+TaqTbq(uq−ℓq),(5)\mathrm{dQ}=\frac{K_{vq}}{T_{vq}}(V_{\mathrm{term}}-w_q),\qquad Q_{\mathrm{ctrl}}=\ell_q+\frac{T_{aq}}{T_{bq}}(u_q-\ell_q), \tag*{(5)}

and lead–lag inputs uq=Mvarinit/MBase+dQu_q=\mathrm{Mvar}_{\mathrm{init}}/\mathrm{MBase}+\mathrm{dQ} and up=MWinit/MBase+Kdroopdb(wfilt)+dPu_p=\mathrm{MW}_{\mathrm{init}}/\mathrm{MBase}+K_{\mathrm{droop}}\mathrm{db}(w_{\mathrm{filt}})+\mathrm{dP}, where db\mathrm{db} is the deadband [dbfl,dbfh][\mathrm{db}_{\mathrm{fl}},\mathrm{db}_{\mathrm{fh}}] and the gate σ(Vterm)\sigma(V_{\mathrm{term}}) in (1) is 1 while Vterm≥0.7V_{\mathrm{term}}\geq0.7 pu and 0 otherwise. A voltage step thus produces a request pulse KvpΔVterm/TvpK_{vp}\Delta V_{\mathrm{term}}/T_{vp} in (4). The requests are converted to currents with the filtered voltage, limited, scaled by the connected fraction FracOn\mathrm{FracOn}, and delayed:

Ipv=sat⁡[Ipmin⁡,Ipmax⁡](Pctrl(Vfilt/Vinit)nP/Vfilt),(6)I_{pv}=\operatorname{sat}_{[I_{p\min},I_{p\max}]}\left(P_{\mathrm{ctrl}}\left(V_{\mathrm{filt}}/V_{\mathrm{init}}\right)^{n_P}/V_{\mathrm{filt}}\right), \tag*{(6)}
Iqv=sat⁡[Iqmin⁡,Iqmax⁡](Qctrl(Vfilt/Vinit)nQ/Vfilt),(7)I_{qv}=\operatorname{sat}_{[I_{q\min},I_{q\max}]}\left(Q_{\mathrm{ctrl}}\left(V_{\mathrm{filt}}/V_{\mathrm{init}}\right)^{n_Q}/V_{\mathrm{filt}}\right), \tag*{(7)}
TtI˙p=FracOn Ipv−Ip,P=VtermIp,(8)T_t\dot{I}_p=\mathrm{FracOn}\,I_{pv}-I_p,\qquad P=V_{\mathrm{term}}I_p, \tag*{(8)}
TtI˙q=FracOn Iqv−Iq,Q=VtermIq,(9)T_t\dot{I}_q=\mathrm{FracOn}\,I_{qv}-I_q,\qquad Q=V_{\mathrm{term}}I_q, \tag*{(9)}

where PP and QQ are PMWP_{\mathrm{MW}} and QMvarQ_{\mathrm{Mvar}} on the base MBase\mathrm{MBase} [4].

Ceasing and reconnection logic

The PERC1 logic is represented as a finite-state machine with modes Monitor, Cease, and Ramp [4]. In Monitor or Ramp, if VfiltV_{\mathrm{filt}} stays below VceaseV_{\mathrm{cease}} for TceaseT_{\mathrm{cease}} seconds, a delay timer starts, and a recovery of VfiltV_{\mathrm{filt}} does not cancel it; after TdelayT_{\mathrm{delay}} seconds, FracOn\mathrm{FracOn} drops to 1−Fcease1-F_{\mathrm{cease}} and the mode becomes Cease. In Cease mode, once VfiltV_{\mathrm{filt}} has stayed above Vrecon≥VceaseV_{\mathrm{recon}}\geq V_{\mathrm{cease}} for TreconT_{\mathrm{recon}} seconds, the time TStartRampT_{\mathrm{StartRamp}} is recorded and FracOn\mathrm{FracOn} ramps linearly over TrampT_{\mathrm{ramp}} seconds to 1−Fcease+FreconFcease1-F_{\mathrm{cease}}+F_{\mathrm{recon}}F_{\mathrm{cease}}. The mode then returns to Monitor. Subsequent ceasing events drop FracOn\mathrm{FracOn} to the same 1−Fcease1-F_{\mathrm{cease}}, so the share Fcease(1−Frecon)F_{\mathrm{cease}}(1-F_{\mathrm{recon}}) leaves at the first cease and stays off, while FreconFceaseF_{\mathrm{recon}}F_{\mathrm{cease}} cycles with every event. A

dip during the ramp goes through the same detector: FracOn\mathrm{FracOn} keeps ramping during the TceaseT_{\mathrm{cease}} dwell and freezes while the delay timer runs, then drops to 1−Fcease1-F_{\mathrm{cease}}.

We encode this as two timed automata [16] that share the event cease (Figure 1). The cease detector D\mathcal{D} has clock τc\tau_c and arms only while R\mathcal{R} is in Monitor or Ramp. The mode automaton R\mathcal{R} has clock τr\tau_r and operates over Monitor, Cease, and Ramp. The transition from timing to armed resets τc\tau_c. Entry to Cease and every crossing of VreconV_{\mathrm{recon}} reset τr\tau_r, so τr=Trecon\tau_r=T_{\mathrm{recon}} with Vfilt>VreconV_{\mathrm{filt}}>V_{\mathrm{recon}} indicates that the voltage has remained above the threshold for the required duration.

Timed automata of the PERC1 logic

Figure 1. Timed automata of the PERC1 logic. The arrows represent edges EE, their transition conditions define guards GG, and the arrowheads at update states define resets RR. D\mathcal{D} reads VfiltV_{\mathrm{filt}} from the plant (1)–(9) and emits cease, which resets FracOn:=1−Fcease\mathrm{FracOn}:=1-F_{\mathrm{cease}} on both edges into Cease; the edge Ramp→\toMonitor resets FracOn:=1−Fcease+FreconFcease\mathrm{FracOn}:=1-F_{\mathrm{cease}}+F_{\mathrm{recon}}F_{\mathrm{cease}}. In Ramp, dFracOn/dt=FreconFcease/Tramp\mathrm{dFracOn}/\mathrm{d}t=F_{\mathrm{recon}}F_{\mathrm{cease}}/T_{\mathrm{ramp}} while D\mathcal{D} is not in delay, and FracOn\mathrm{FracOn} is constant otherwise. Clocks are reset to zero on entry to timing, delay, Cease, and Ramp, and τr\tau_r is also reset by the self-loop of Cease.

The hybrid automaton

Let x=(Vfilt,wfilt,wp,wq,ℓp,ℓq,Ip,Iq,FracOn,τc,τr)∈R11x=(V_{\mathrm{filt}},w_{\mathrm{filt}},w_p,w_q,\ell_p,\ell_q,I_p,I_q,\mathrm{FracOn},\tau_c,\tau_r)\in\mathbb{R}^{11} and q=(d,r)∈Q=QD×QRq=(d,r)\in\mathcal{Q}=\mathcal{Q}_{\mathcal{D}}\times\mathcal{Q}_{\mathcal{R}}. The PERC1 model is the hybrid automaton H=(Q,R11,f,E,G,R)\mathcal{H}=(Q,\mathbb{R}^{11},f,E,G,R) in the sense of [12]. In mode qq, the state flows according to x˙=fq(x,u;θ)\dot{x}=f_q(x,u;\theta), given by (1)–(9) with

dFracOndt∈{0,FreconFceaseTramp},τ˙c,τ˙r∈{0,1}.\frac{\mathrm{dFracOn}}{\mathrm{d}t}\in\left\{0,\frac{F_{\mathrm{recon}}F_{\mathrm{cease}}}{T_{\mathrm{ramp}}}\right\},\qquad\dot{\tau}_c,\dot{\tau}_r\in\{0,1\}.

The edges EE, guards GG, and resets RR are shown in Figure 1. The parameter vector is divided into θ=(θc,θd)\theta=(\theta_c,\theta_d). The continuous parameters θc\theta_c include gains, time constants, exponents, and current limits that govern the system dynamics. The logic parameters θd\theta_d (VceaseV_{\mathrm{cease}}, TceaseT_{\mathrm{cease}}, TdelayT_{\mathrm{delay}}, VreconV_{\mathrm{recon}}, TreconT_{\mathrm{recon}}, TrampT_{\mathrm{ramp}}, FceaseF_{\mathrm{cease}}, FreconF_{\mathrm{recon}}) appear only in the guards, resets, and ramp rate. Thus, the output y=(P,Q)y=(P,Q) depends on θd\theta_d through the transition times and the values of FracOn\mathrm{FracOn}.

For a data center, 1−Fcease1-F_{\mathrm{cease}} represents the load that rides through the disturbance (cooling, auxiliaries, and servers on static UPS units), FceaseF_{\mathrm{cease}} is the share transferred to backup supply, FreconFceaseF_{\mathrm{recon}}F_{\mathrm{cease}} is the share that automatically returns to grid supply, and Fcease(1−Frecon)F_{\mathrm{cease}}(1-F_{\mathrm{recon}}) is the share that requires manual reconnection [2].

Identification from Synchrophasor Data

Problem statement

A PMU at the point of interconnection reports the positive-sequence voltage magnitude, frequency, active and reactive power at times tk=k/fst_k = k/f_s, with fs∈{30,60}f_s \in\{30,60\} frames/s [17]: the inputs VtermV_{\mathrm{term}}, Δω\Delta\omega and the outputs y=(P,Q)\mathbf{y} = (P,Q) of the model. We drive the model with the measured inputs, interpolated between frames, so the grid beyond the point of interconnection need not be modeled, as in the play-in approach of [5], [8]. A phasor estimate is a windowed average of the underlying signal. We therefore apply the same window MW\mathcal{M}_{W} to the model outputs before comparing them with the records yk\mathbf{y}_k, and we discard samples within one window of a voltage step. The estimate minimizes, over all available records,

J(θ)=12∑k∥Wk1/2rk(θ)∥2+12∥Σ0−1/2(θ−θ0)∥2,(10)J(\theta) = \frac{1}{2}\sum_k \lVert W_k^{1/2}r_k(\theta)\rVert^2 + \frac{1}{2}\lVert\Sigma_0^{-1/2}(\theta-\theta_0)\rVert^2, \tag*{(10)}

where rk=MWy(tk;θ)−ykr_k = \mathcal{M}_{W}\mathbf{y}(t_k;\theta)-\mathbf{y}_k, WkW_k weights by the inverse noise variances, and the second term is a weak prior centered at the initialization θ0\theta_0 of Stages 1–2 below. It moves the estimate only along directions the data do not inform.

Hybrid trajectory sensitivities

Between transitions the sensitivity S=∂x/∂θS = \partial x/\partial\theta satisfies the variational equation S˙=(fq)xS+(fq)θ\dot{S} = (f_q)_xS+(f_q)_\theta. At a transition at time tet_e defined by g(x(te−),θ)=0g(x(t_e^-),\theta)=0 with reset x(te+)=R(x(te−),θ)x(t_e^+) = R(x(t_e^-),\theta), the event time depends on the parameters through [15 Prop. 3.4]

∂te∂θ=−gxS(te−)+gθgxf−,(11)\frac{\partial t_e}{\partial\theta}=-\frac{g_xS(t_e^-)+g_\theta}{g_xf^-}, \tag*{(11)}

and the sensitivity jumps as [13], [15 Prop. 3.2]

S(te+)=ΞS(te−)+Rθ−(Rxf−−f+)gθgxf−,(12)S(t_e^+) = \Xi S(t_e^-) + R_\theta- (R_xf^- - f^+)\frac{g_\theta}{g_xf^-}, \tag*{(12)}

where f∓f^\mp are the vector fields before and after the transition and Ξ=Rx−(Rxf−−f+)gx/(gxf−)\Xi= R_x-(R_xf^- - f^+)g_x/(g_xf^-) is the saltation matrix [14]. The last term of (12) is the explicit dependence of the guard on the parameters, carried by gθg_\theta in (11). For PERC1, it and RθR_\theta are the only ways θd\theta_d enters the output sensitivity. For a clock guard g=τ−Tg=\tau-T, ∂te/∂T=1\partial t_e/\partial T=1. For a voltage guard g=±(Vfilt−V∙)g=\pm(V_{\mathrm{filt}}-V_\bullet), where V∙∈{Vcease,Vrecon}V_\bullet\in\{V_{\mathrm{cease}},V_{\mathrm{recon}}\}, ∂te/∂V∙=1/V˙filt(te)\partial t_e/\partial V_\bullet=1/\dot{V}_{\mathrm{filt}}(t_e). Thus, the slope of the filtered voltage at the threshold crossing determines how threshold changes affect the transition time. At a cease from Monitor, the only nonzero rows of Rxf−−f+R_xf^- - f^+ are those of IpI_p and IqI_q, with values FceaseIpv/TtF_{\mathrm{cease}}I_{pv}/T_t and FceaseIqv/TtF_{\mathrm{cease}}I_{qv}/T_t respectively. Consequently, each timing parameter produces a pulse in ∂P/∂θ\partial P/\partial\theta of area VtermFceaseIpv ∂te/∂θV_{\mathrm{term}}F_{\mathrm{cease}}I_{pv}\,\partial t_e/\partial\theta that decays with TtT_t (Figure 2(a)). We integrate SS along with the state and apply (12) at each crossing, which we locate by interpolation within the step.

Sensitivities and level sets of the cease time

Figure 2. (a) Normalized sensitivities θj∂P/∂θj\theta_j \partial P/\partial\theta_j in event A to VceaseV_{\mathrm{cease}}, TceaseT_{\mathrm{cease}}, and FceaseF_{\mathrm{cease}}, from (12). The timing parameters act through a pulse that decays with TtT_t; the fraction acts through a level change. (b) Level sets of the cease time in the (Vcease,Tcease)(V_{\mathrm{cease}},T_{\mathrm{cease}}) plane for the two events.

Four-stage procedure

Stage 1: Mode sequence. Following [18], we first identify the mode sequence. For a single sag, the mode sequence is Monitor→\toCease→\toRamp→\toMonitor. The connected fraction FracOn follows the template FracOnT(t;tc,tr,te,a,b)\mathrm{FracOn}_{T}(t;t_c,t_r,t_e,a,b): 1 before tct_c, aa until trt_r, linear until tet_e, and bb afterwards. The template is fitted to the ratio of the measured active power to the active power predicted by the model with FracOn =1=1 (Algorithm 1, lines 1–3). The levels give Fcease=1−aF_{\mathrm{cease}}=1-a and Frecon=(b−a)/FceaseF_{\mathrm{recon}}=(b-a)/F_{\mathrm{cease}}, while the times give Tramp=te−trT_{\mathrm{ramp}}=t_e-t_r, with trt_r corresponding to the vendor’s TstartRampT_{\mathrm{startRamp}}.

Algorithm 1 Four-stage identification of PERC1
Require: Records r=1,…,Rr=1,\ldots,R of VtermV_{\mathrm{term}}, Δω\Delta\omega, yk=(Pk,Qk)\mathbf{y}_k=(P_k,Q_k); vendor θcnom\theta_c^{\mathrm{nom}}, initial θ0\theta_0, prior Σ0\Sigma_0
1: Stage 1, each record rr: ρ^k←Pk/MWP(tk;θcnom,FracOn≡1)\hat{\rho}_k\leftarrow P_k/\mathcal{M}_{W}P(t_k;\theta_c^{\mathrm{nom}},\mathrm{FracOn}\equiv1)
2: Fit FracOnT\mathrm{FracOn}_{T} to {ρk(r)}\{\rho_k^{(r)}\} on the frame grid by least squares, linear in (a,b)(a,b), to obtain (tc(r),tr(r),te(r),a(r),b(r))(t_c^{(r)},t_r^{(r)},t_e^{(r)},a^{(r)},b^{(r)})
3: Fcease←1−aˉF_{\mathrm{cease}}\leftarrow1-\bar{a}, Fˉrecon←(bˉ−aˉ)/Fcease\bar{F}_{\mathrm{recon}}\leftarrow(\bar{b}-\bar{a})/F_{\mathrm{cease}}, Tramp←tˉe−tˉrT_{\mathrm{ramp}}\leftarrow\bar{t}_e-\bar{t}_r (means over rr)
4: Stage 2: t↓(r)(V)t_{\downarrow}^{(r)}(V), t↑(r)(V)t_{\uparrow}^{(r)}(V) (after tc(r)t_c^{(r)}) of VtermV_{\mathrm{term}} filtered with TvT_v
5: Vcease←arg⁡min⁡V[max⁡r−min⁡r]{tc(r)−t↓(r)(V)}V_{\mathrm{cease}}\leftarrow\arg\min_V[\max_r-\min_r]\{t_c^{(r)}-t_{\downarrow}^{(r)}(V)\} over VV with nonnegative delays (vendor default if R=1R=1); Tcease←T_{\mathrm{cease}}\leftarrow mean of tc(r)−t↓(r)(Vcease)t_c^{(r)}-t_{\downarrow}^{(r)}(V_{\mathrm{cease}}); Tdelay←0T_{\mathrm{delay}}\leftarrow0
6: estimate Vrecon,TreconV_{\mathrm{recon}},T_{\mathrm{recon}} from tˉr−tˉ↑(r)(V)\bar{t}_r-\bar{t}_{\uparrow}^{(r)}(V); θ←θ0:=(θc0,θd)\theta\leftarrow\theta_0:=(\theta_c^0,\theta_d)
Stage 3 (λ←10−2\lambda\leftarrow10^{-2})
7: repeat
8: for each rr, integrate x˙=fq\dot{x}=f_q, S˙=(fq)xS+(fq)θ\dot{S}=(f_q)_xS+(f_q)_\theta, applying (11)–(12) at each guard crossing (located by interpolation)
9: rk←MWy(tk;θ)−ykr_k\leftarrow\mathcal{M}_{W}y(t_k;\theta)-y_k, Jk←MW∂y(tk)/∂θJ_k\leftarrow\mathcal{M}_{W}\partial y(t_k)/\partial\theta; H←∑kJkTWkJk+Σ0−1H\leftarrow\sum_kJ_k^{\mathsf T}W_kJ_k+\Sigma_0^{-1}, g←∑kJkTWkrk+Σ0−1(θ−θ0)g\leftarrow\sum_kJ_k^{\mathsf T}W_kr_k+\Sigma_0^{-1}(\theta-\theta_0)
10: solve (H+λdiag⁡H)δ=−g(H+\lambda\operatorname{diag}H)\delta=-g; if J(θ+δ)<J(θ)J(\theta+\delta)<J(\theta) in (10): θ←θ+δ\theta\leftarrow\theta+\delta, λ←λ/3\lambda\leftarrow\lambda/3; else λ←4λ\lambda\leftarrow4\lambda and re-solve
11: until the relative decrease of JJ is below 10−610^{-6} or no step is accepted
12: Stage 4: F←∑kJkTWkJk\mathcal{F}\leftarrow\sum_kJ_k^{\mathsf T}W_kJ_k at θ^\hat{\theta}; rank and null directions from the singular value decomposition of W1/2JW^{1/2}J; if F≻0\mathcal{F}\succ0, σj←[F−1]jj\sigma_j\leftarrow\sqrt{[\mathcal{F}^{-1}]_{jj}}

Table 1.

Stage 2: guard parameters from event-time constraints. Let t↓(V)t_{\downarrow}(V) be the first time the filtered voltage of a record drops below VV and t↑(V)t_{\uparrow}(V) the first time after the cease that it rises above VV. A cease observed at tct_c and a ramp start observed at trt_r impose

tc=t↓(Vcease)+Tcease+Tdelay,t_c=t_{\downarrow}(V_{\mathrm{cease}})+T_{\mathrm{cease}}+T_{\mathrm{delay}},
tr=t↑(Vrecon)+Trecon,(13)t_r=t_{\uparrow}(V_{\mathrm{recon}})+T_{\mathrm{recon}}, \tag*{(13)}

together with the dwell conditions that VfiltV_{\mathrm{filt}} stays below VceaseV_{\mathrm{cease}} on [tc,tc+Tcease][t_c,t_c+T_{\mathrm{cease}}] and above VreconV_{\mathrm{recon}} on [tr,tr+Trecon][t_r,t_r+T_{\mathrm{recon}}]. Each record contributes one equation per observed transition and an inequality for each dip below a candidate VceaseV_{\mathrm{cease}} without a cease. With two or more records, we select the threshold for which the delays implied by (13) are consistent across the records.

Stage 3: joint refinement. Starting from the estimates obtained in Stages 1–2, we use a Levenberg–Marquardt iteration to minimize (10) over (θc,θd)(\theta_c,\theta_d), with the Jacobian computed using (12) (Algorithm 1, lines 7–11).

Stage 4: identifiability. The Gauss–Newton matrix F=∑kJkTWkJk\mathcal{F}=\sum_kJ_k^{\mathsf T}W_kJ_k is the salted Fisher information of [15] for sampled outputs. Its rank and small singular values indicate which parameters can be identified from the records. When F\mathcal{F} is positive definite, the square root of the Cramér–Rao bound, [F−1]jj\sqrt{[\mathcal{F}^{-1}]_{jj}}, provides a lower bound on the standard deviation of unbiased estimates of θj\theta_j.

Two structural results

Proposition 1. For a record in which every cease is detected while R\mathcal{R} is in Monitor, the output of HH depends on (Tcease,Tdelay)(T_{\mathrm{cease}},T_{\mathrm{delay}}) only through Tcease+TdelayT_{\mathrm{cease}}+T_{\mathrm{delay}}. A cease detected while R\mathcal{R} is in Ramp separates the two.

Proof. In Monitor, FracOn ≡1\equiv1 whatever the mode of D\mathcal{D}, so the detector affects the plant only through the event cease, emitted at t↓(Vcease)+Tcease+Tdelayt_{\downarrow}(V_{\mathrm{cease}})+T_{\mathrm{cease}}+T_{\mathrm{delay}}. The delay mode has no exit guard on VfiltV_{\mathrm{filt}}, so two parameter vectors with equal sum generate identical executions. In Ramp, dFracOn/dt=FreconFcease/Tramp\mathrm{dFracOn}/\mathrm{d}t=F_{\mathrm{recon}}F_{\mathrm{cease}}/T_{\mathrm{ramp}} while D\mathcal{D} is timing and 0 while it is in delay, so the level at which FracOn freezes, 1−Fcease+FreconFcease(t↓+Tcease−tr)/Tramp1-F_{\mathrm{cease}}+F_{\mathrm{recon}}F_{\mathrm{cease}}(t_{\downarrow}+T_{\mathrm{cease}}-t_r)/T_{\mathrm{ramp}}, depends on TceaseT_{\mathrm{cease}} alone.

For initialization, we set Tdelay=0T_{\mathrm{delay}}=0 and assign the total delay to TceaseT_{\mathrm{cease}}; a second dip during ramp permits separate identification.

Proposition 2. For a record with a single cease at tct_c and a single ramp start at trt_r, the Fisher information restricted to (Vcease,Tcease)(V_{\mathrm{cease}},T_{\mathrm{cease}}) has rank one with null direction (1,−1/V˙filt(t↓))(1,-1/\dot{V}_{\mathrm{filt}}(t_{\downarrow})), and restricted to (Vrecon,Trecon)(V_{\mathrm{recon}},T_{\mathrm{recon}}) has rank one with null direction (1,−1/V˙filt(t↑))(1,-1/\dot{V}_{\mathrm{filt}}(t_{\uparrow})). Records of two disturbances whose filtered voltages cross the thresholds with different slopes make both blocks full rank.

Proof. By (13), the output depends on (Vcease,Tcease)(V_{\mathrm{cease}},T_{\mathrm{cease}}) only through tct_c (the re-arm edges are unobservable without a second cease), so by (11) the Jacobian columns satisfy ∂y/∂Tcease δTcease+∂y/∂Vcease δVcease=0\partial y/\partial T_{\mathrm{cease}}\,\delta T_{\mathrm{cease}}+\partial y/\partial V_{\mathrm{cease}}\,\delta V_{\mathrm{cease}}=0 whenever δTcease=−δVcease/V˙filt(t↓)\delta T_{\mathrm{cease}}=-\delta V_{\mathrm{cease}}/\dot{V}_{\mathrm{filt}}(t_{\downarrow}); likewise for (Vrecon,Trecon)(V_{\mathrm{recon}},T_{\mathrm{recon}}) through trt_r. Two records give two null directions that are independent when the slopes differ, so their intersection is trivial.

In the language of [15], one record fails hybrid persistence of excitation along these directions, and a second record with a different crossing slope restores it.

Case Study

We consider a data center represented by PERC1 with the true parameters listed in Table I. Following a sustained voltage sag below 0.70 pu for 50 ms, 60% of the load transfers to backup. Of the disconnected load, 75% reconnects after the filtered voltage remains above 0.88 pu for 0.5 s, with a 2 s ramp. The active power is nearly constant-power (nP=0.1n_P=0.1), with a current limit Ipmax⁡=1I_{p\max}=1. The frequency path is disabled (Kdroop=0K_{\mathrm{droop}}=0).

Two disturbances are considered. Event A represents a nearby fault, with voltage dropping to 0.45 pu for nine cycles, recovering to 0.85 pu at clearing, and returning to 1 pu with a 0.3 s time constant. Event B represents a remote disturbance, with voltage falling to 0.60 pu in 150 ms, holding for 60 ms, rising to 0.90 pu in 250 ms, and returning to 1 pu with a 0.5 s time constant. The model is simulated with a 1 ms step. Voltage and power are averaged over two cycles and sampled at 60 frames/s, with noise standard deviations of 0.002 and 0.003 pu, respectively.

Stage 1 identifies the cease times within approximately one PMU frame and estimates Fcease=0.595F_{\mathrm{cease}}=0.595, Frecon=0.748F_{\mathrm{recon}}=0.748, and Tramp=1.975T_{\mathrm{ramp}}=1.975 s. Stage 2 estimates the threshold–delay pairs from the observed transition times. As shown in Figure 2(b), the intersection of the two level sets gives (Vcease,Tcease)=(0.690,0.064)(V_{\mathrm{cease}},T_{\mathrm{cease}})=(0.690,0.064), compared with the true values (0.70,0.05)(0.70,0.05). The reconnection parameters are estimated as (Vrecon,Trecon)=(0.918,0.431)(V_{\mathrm{recon}},T_{\mathrm{recon}})=(0.918,0.431).

Stage 3 reduces the cost (10) from 4369 at initialization to 736 after three iterations and 725 after forty, compared with 751 at the true parameters and a noise floor of 710. The identified model (Figure 3) reproduces the cease times within 0.4 ms and the ramp start and end within 3 ms, below one PMU frame. Table I reports the parameter estimates and Crámer–Rao standard deviations from the two-event Fisher information evaluated at the true parameters.

Synthetic PMU records for events A and B and cessation of event A

Figure 3. Synthetic PMU records of (a) a nearby fault, (b) a remote slow sag, and (c) the cessation of event A. Top: voltage and active power, measured and modeled at the Stage 1–2 initialization and after Stage 3; crosses mark masked samples within one phasor window of a voltage step. Bottom: FracOn, initial, true, and identified; dotted lines mark the true transitions.

The logic parameters are recovered within two standard deviations, except TreconT_{\mathrm{recon}}, whose 64 ms error is seven times its bound. Over 20 noise realizations, the root-mean-square errors are 0.006 pu and 4 ms for (Vcease,Tcease)(V_{\mathrm{cease}},T_{\mathrm{cease}}), 0.042 pu and 70 ms for (Vrecon,Trecon)(V_{\mathrm{recon}},T_{\mathrm{recon}}), 8 ms for TrampT_{\mathrm{ramp}}, and 0.003 and 0.001 for FceaseF_{\mathrm{cease}} and FreconF_{\mathrm{recon}}. The spread of (Vrecon,Trecon)(V_{\mathrm{recon}},T_{\mathrm{recon}}) exceeds the bound because input noise shifts the voltage threshold crossings, whereas the bound assumes exact voltage measurements. For TrampT_{\mathrm{ramp}}, FceaseF_{\mathrm{cease}}, FreconF_{\mathrm{recon}}, and Ipmax⁡I_{p\max}, the spread is within 10% of the bound, consistent with [15].

The current limit and voltage exponents are recovered, while TvpT_{vp}, TapT_{ap}, and TbpT_{bp} remain poorly identifiable. During deep voltage sags, current limiting masks their effects, while phasor-window averaging smooths the short-duration responses needed to identify these time constants.

The single-event results in Table I confirm Proposition 2. For Event A alone, the Fisher information has rank 15 of

ParameterTrueStages 1–2Stage 3, A+BCRB σ\sigmaStage 3, A only
KvpK_{vp}0.0200.0100.0170.0150.027
TvpT_{vp}0.1000.1000.1150.440.107
KvqK_{vq}0.0100.0050.0110.00180.012
TvqT_{vq}0.1000.1000.1220.0250.127
TapT_{ap}0.0500.0500.0000.260.000
TbpT_{bp}0.2000.1000.1150.440.107
nPn_P0.1000.0000.0960.0290.146
nQn_Q1.0001.0000.9910.0351.067
Ipmax⁡I_{p\max}1.0001.0001.0020.00301.002
TvT_{v}0.0200.0200.0210.00760.023
VceaseV_{\mathrm{cease}}0.7000.6900.6940.00830.520
TceaseT_{\mathrm{cease}}0.0500.0640.0470.00450.017
VreconV_{\mathrm{recon}}0.8800.9180.9050.0130.599
TreconT_{\mathrm{recon}}0.5000.4310.4360.00880.578
TrampT_{\mathrm{ramp}}2.0001.9752.0020.00692.007
FceaseF_{\mathrm{cease}}0.6000.5950.6010.00140.603
FreconF_{\mathrm{recon}}0.7500.7480.7500.000690.751

Table 1 (PDF p. 4). Identification results (per unit, seconds)

17, with null directions (Vcease,Tcease)∝(0.997,0.079)(V_{\mathrm{cease}},T_{\mathrm{cease}}) \propto(0.997,0.079) and (Vrecon,Trecon)∝(0.522,−0.853)(V_{\mathrm{recon}},T_{\mathrm{recon}}) \propto(0.522,-0.853), consistent with the crossing slopes of −12.3-12.3 and 0.550.55 pu/s, respectively. Starting from the vendor default thresholds of 0.50.5 and 0.60.6 pu, the single-event estimates converge to (Vcease,Tcease)=(0.52,0.017)(V_{\mathrm{cease}},T_{\mathrm{cease}})=(0.52,0.017) and (Vrecon,Trecon)=(0.60,0.58)(V_{\mathrm{recon}},T_{\mathrm{recon}})=(0.60,0.58). These estimates fit Event A as well as the true model but may predict incorrect transition times under other disturbances. Combining both events resolves the ambiguity, increasing the Jacobian’s smallest singular value from 10−1310^{-13} to 1.51.5.

Proposition 1 is also verified by setting Tdelay=50T_{\mathrm{delay}}=50 ms. For Event A, the Jacobian columns corresponding to TceaseT_{\mathrm{cease}} and TdelayT_{\mathrm{delay}} are identical, so the delays cannot be identified separately. Introducing a second voltage dip to 0.550.55 pu during the ramp makes the corresponding Jacobian block full rank, with singular values of 1662 and 27. The bound on TdelayT_{\mathrm{delay}} is 31 ms, approximately two PMU frames. The 60 frames/s reporting rate limits the information available for estimating short detection delays, so the millisecond-scale detection delays reported by several vendors call for a PMU or fault recorder streaming at a higher rate.

Conclusion

PERC1 is a hybrid system. Treating it as one turns its parameterization from curve fitting with a nonsmooth cost into a structured problem involving mode sequence identification, estimation of logic parameters from transition times, joint parameter estimation using hybrid trajectory sensitivities, and analysis of which parameters can be uniquely identified from the measurements using the Fisher information matrix. The hybrid formulation also makes explicit two facts that the block diagram hides. The detection and actuation delays cannot be identified separately unless a dip interrupts the reconnection ramp, while a single disturbance cannot uniquely identify the threshold-delay pairs. Two synchrophasor records of different voltage profiles recover the logic parameters of a data center load within the PMU frame resolution. The washout and lead–lag time constants remain poorly identifiable because their effects are largely masked by current limiting during deep voltage sags.

References

  1. [1]A. Shehabi et al., “2024 United States data center energy usage report,” Lawrence Berkeley Nat. Lab., Tech. Rep. LBNL-2001637, Dec. 2024.
  2. [2]NERC, “Incident review: Considering simultaneous voltage-sensitive load reductions,” NERC, Atlanta, GA, USA, Tech. Rep., Jan. 2025.
  3. [3]M. Wharton, “Dominium load transfer event,” Presented at the PJM Operating Committee, Audubon, PA, USA, Aug. 6, 2026.
  4. [4]PowerWorld Corporation, “Load characteristic model: PERC1, Simulator 24 documentation,” Chappaquoin, IL, USA, 2026. [Online]. Available: https://www.powerworld.com/WebHelp/Content/TransientModels_HTML/Load%20Characteristic%20PERC1.htm
  5. [5]J. Weber, “Power electronic reconnecting and load model,” Presented at the NERC Load Modeling Task Force webinar, Nov. 18, 2025. [Online]. Available: https://www.powerworld.com/files/Power-Electronic-Reconnecting-and-Ceasing-PERC1-Load-Model.pdf
  6. [6]NERC, “Data center load modeling technical reference,” NERC, Atlanta, GA, USA, Tech. Rep., Jul. 30, 2026.
  7. [7]L. Sundares h and P. Mitra, “A positive sequence model for aggregated representation of electric vehicle chargers: EPCL model description,” EPRI, Palo Alto, CA, USA, Tech. Rep. Project 1-116982, Nov. 2022.
  8. [8]I. A. Hiskens, “Nonlinear dynamic model evaluation from disturbance measurements,” IEEE Trans. Power Syst., vol. 16, no. 4, pp. 702–710, Nov. 2001.
  9. [9]B.-K. Choi et al., “Measurement-based dynamic load models: Derivation, comparison, and validation,” IEEE Trans. Power Syst., vol. 21, no. 3, pp. 1276–1283, Aug. 2006.DOI
  10. [10]A. Rouhani and A. Abur, “Real-time dynamic parameter estimation for an exponential dynamic load model,” IEEE Trans. Smart Grid, vol. 7, no. 3, pp. 1530–1536, May 2016.DOI
  11. [11]J. H. Woo, V. D. Paduani, Z. Yang, Y. Ma, Q. Xiao, and N. Lu, “Reinsertion-aware parameter estimation of the WECC composite load model via closed-loop equivalent-network co-estimation,” preprint, arXiv:2610.34497, 2026.arxiv.org/abs/2610.34497
  12. [12]J. Lygeros, K. H. Johansson, S. N. Simić, J. Zhang, and S. S. Sastry, “Dynamical properties of hybrid automata,” IEEE Trans. Autom. Control, vol. 48, no. 1, pp. 2–17, Jan. 2003.DOI
  13. [13]I. A. Hiskens and M. A. Pai, “Trajectory sensitivity analysis of hybrid systems,” IEEE Trans. Circuits Syst. I, Fundam. Theory Appl., vol. 47, no. 2, pp. 204–220, Feb. 2000.DOI
  14. [14]N. J. Kong, J. J. Payne, J. Zhu, and A. M. Johnson, “Saltation matrices: The essential tool for linearizing hybrid dynamical systems,” Proc. IEEE, vol. 112, no. 6, pp. 585–608, Jun. 2024.DOI
  15. [15]B. G. Odunlami, M. Netto, and H. Lin, “Salted Fisher information for hybrid systems,” preprint, arXiv:2603.29862, 2026.arxiv.org/abs/2603.29862
  16. [16]R. Alur and D. L. Dill, “A theory of timed automata,” Theor. Comput. Sci., vol. 126, no. 2, pp. 183–235, 1994.
  17. [17]IEC/IEEE, “Measuring relays and protection equipment—Part 118-1: Synchrophasor for power systems—Measurements,” IEC/IEEE 60255-118-1:2018, 2018.DOI
  18. [18]S. Paoletti, A. L. Juloski, G. Ferrari-Trecate, and R. Vidal, “Identification of hybrid systems: A tutorial,” Eur. J. Control, vol. 13, no. 2–3, pp. 242–260, 2007.DOI

Paper details

vlnc:W_am3wuzdd
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents