betaFeedback

WorksEngineering

Learn Locally, Extend Globally: Koopman Eigenfunctions and the Basin of Attraction

Abstract

Koopman eigenfunctions associated with the Jacobian eigenvalues of a stable equilibrium are Lyapunov functions defined exactly on its basin of attraction. We turn this into a computational procedure for power systems. Eigenfunctions estimated by extended dynamic mode decomposition (EDMD) are written as functions of the state, made to vanish at the equilibrium by centering the lifted variable cos⁡δ\cosδ, and verified on a grid with the model: two Lyapunov candidates from the EDMD spectrum certify up to 66% of the basin of a single-machine infinite-bus system from data in the linear region. The Koopman semigroup property then extends the EDMD-Lie eigenfunction to the whole basin, recovered at the grid resolution with its isostables and isochrons.

Introduction

The basin of attraction of a stable equilibrium determines which post-fault states of a power system return to synchronous operation, and direct methods for transient stability assessment rest on estimates of it [1]. The classical route is a Lyapunov function, constructed algorithmically [2] and then either enlarged by integrating a certified set backward in time [3] or checked on samples with Lipschitz margins [4]. Koopman eigenfunctions tied to the Jacobian eigenvalues of the equilibrium [5] exist on the whole basin, uniquely under nonresonance [6]. They are Lyapunov functions there [7], with the isostables as level sets [8].

Extended dynamic mode decomposition (EDMD) estimates eigenfunctions from data snapshots [9], [10], but the estimate is only as good as the dictionary [11], and its accuracy fades with the distance from the data. Some authors therefore build the candidate eigenfunctions from the model: Bierwart and Mauroy validate it afterward by sum-of-squares programs or an adaptive grid [12], Umathe and Vaidya [13] work with the eigenfunctions of the saddles, and Römer and Breitenhuber fit a rational ansatz to the Koopman generator equation through a non-convex program [14]. Others keep the eigenfunction as learned: Deka and Dimarogonas sharpen it with Laplace averages before learning a Lyapunov function from it [15], and Matavalam et al. [16] apply it directly for transient stability assessment.

This letter certifies the learned eigenfunction itself. The left eigenvectors of the EDMD matrix give an eigenfunction in closed form, so it can be checked against the model at every grid point with an explicit tolerance. Once certified, it extends to the whole basin of a single-machine infinite-bus system. An EDMD eigenfunction with discrete eigenvalue μk≠1\mu_k \ne1 vanishes at the fixed point whenever the lifted fixed point is invariant under the EDMD matrix. We show that centering cos⁡δ\cos\delta ensures that invariance for any data set and removes the spurious eigenvalue of [11]. Two Lyapunov candidates from the EDMD spectrum, certified on a grid with the model, give inner basin estimates for five dictionaries. The semigroup property extends the eigenfunction from a small certified seed to the whole basin; the extension is finite exactly on the basin and yields the isostables and isochrons [8]. The extension accepts any certified positively invariant seed, including a ball from the linearization. Seeding it with the eigenfunction learned from the analytical dictionary of [11], tailored to the model, keeps the object that serves state estimation. It also keeps the dictionary at six observables, since the seed need not be large.

This work is supported in part by Eaton Research Laboratories and in part by the National Science Foundation under Grant No. 2328241. M. Netto, A. Kale, and B. G. Odunlami are with the Department of Electrical and Computer Engineering, New Jersey Institute of Technology, Newark, NJ 07102, USA. A. Bernstein is with Eaton Research Laboratories, Golden, CO 80401, USA.

Preliminaries

Consider x˙=f(x)\dot{x}=f(x) on X⊆Rn\mathcal{X}\subseteq\mathbb{R}^{n} with flow StS_t. The Koopman semigroup Kt\mathcal{K}_t acts on observables g:X→Cg:\mathcal{X}\to\mathbb{C} by composition, Ktg=g∘St\mathcal{K}_t g=g\circ S_t [5]. A Koopman eigenfunction with eigenvalue λ∈C\lambda\in\mathbb{C} is an observable ϕ:D→C\phi:\mathcal{D}\to\mathbb{C}, defined on a positively invariant open set D⊆X\mathcal{D}\subseteq\mathcal{X} and continuously differentiable there, that satisfies

ϕ(Stx)=eλtϕ(x),(1)\phi(S_t x)=e^{\lambda t}\phi(x), \tag*{(1)}

(semigroup form)

Lfϕ:=∇ϕ⋅f=λϕ,(2)\mathcal{L}_f\phi:=\nabla\phi\cdot f=\lambda\phi, \tag*{(2)}

(generator form) for all x∈Dx\in\mathcal{D} and t>0t>0, where ∇\nabla denotes the vector differential operator and Lf\mathcal{L}_f denotes the Lie derivative, the semigroup generator. Eqs. Eqs. (1)–(2) state the eigenfunction property in two equivalent forms, the semigroup form in [5] and the generator form in [7]. Let x⋆x^\star be a hyperbolic stable equilibrium (no Jacobian eigenvalue on the imaginary axis) with basin B(x⋆)\mathcal{B}(x^\star) and Jacobian eigenvalues λ1,…,λn\lambda_1,\ldots,\lambda_n with left eigenvectors ℓ1T,…,ℓnT\ell_1^\mathsf{T},\ldots,\ell_n^\mathsf{T}. Under non-resonance conditions, the λk\lambda_k, k=1,…,nk=1,\ldots,n, are Koopman eigenvalues. The associated eigenfunctions ϕk\phi_k, called principal, are C1C^1 on B(x⋆)\mathcal{B}(x^\star), with D=B(x⋆)\mathcal{D} = \mathcal{B}(x^\star), satisfy ∇ϕk(x⋆)=ℓk\nabla\phi_k(x^\star) = \ell_k, and vanish only at x⋆x^\star [6, 7]. Let x0x_0 denote xx at t=0t = 0 and λk=σk+jνk\lambda_k = \sigma_k + j\nu_k. Along a trajectory x(t)=Stx0x(t) = S_t x_0, the semigroup form (1) yields

ϕk(x(t))=eλktϕk(x0),\phi_k(x(t)) = e^{\lambda_k t}\phi_k(x_0),
∣ϕk(x(t))∣=∣eλkt∣ ∣ϕk(x0)∣=∣eσkt∣ ∣ejνkt∣ ∣ϕk(x0)∣=eRe⁡(λk)t∣ϕk(x0)∣\begin{aligned} |\phi_k(x(t))| &= |e^{\lambda_k t}|\,|\phi_k(x_0)| = |e^{\sigma_k t}|\,|e^{j\nu_k t}|\,|\phi_k(x_0)| \\ &= e^{\operatorname{Re}(\lambda_k)t}|\phi_k(x_0)| \end{aligned}

since ∣ejνkt∣=1|e^{j\nu_k t}| = 1, and differentiating in tt,

ddt∣ϕk(x(t))∣=Re⁡(λk)eRe⁡(λk)t∣ϕk(x0)∣=Re⁡(λk)∣ϕk(x(t))∣<0.(3)\frac{d}{dt}|\phi_k(x(t))| = \operatorname{Re}(\lambda_k)e^{\operatorname{Re}(\lambda_k)t}|\phi_k(x_0)| = \operatorname{Re}(\lambda_k)|\phi_k(x(t))| < 0. \tag*{(3)}

The inequality holds away from x⋆x^\star because Re⁡(λk)<0\operatorname{Re}(\lambda_k) < 0 at a hyperbolic stable equilibrium and ϕk\phi_k vanishes only at x⋆x^\star. Hence, ∣ϕk∣|\phi_k| is a Lyapunov function on the basin. Products ϕkϕl\phi_k\phi_l are again eigenfunctions, with eigenvalue λk+λl\lambda_k + \lambda_l, and ϕ1∗\phi_1^\ast is an eigenfunction with eigenvalue λ1∗\lambda_1^\ast because ff is real. Thus, ∣ϕ1∣2=ϕ1ϕ1∗|\phi_1|^2 = \phi_1\phi_1^\ast is a real eigenfunction with eigenvalue λ1+λ1∗=2Re⁡(λ1)\lambda_1 + \lambda_1^\ast= 2\operatorname{Re}(\lambda_1). The basin is the largest set on which the principal eigenfunctions are C1C^1 [7]. Level sets of ∣ϕ1∣|\phi_1| are the isostables and level sets of arg⁡ϕ1\arg\phi_1 are the isochrons [8].

Extended dynamic mode decomposition

Given snapshot pairs {xk−1,xk}\{x_{k-1},x_k\}, k=1,…,mk = 1,\ldots,m, sampled with period Δt\Delta t, and a vector of qq observables g(x)g(x), EDMD forms G0=[g(x0) ⋯ g(xm−1)]G_0=[g(x_0)\ \cdots\ g(x_{m-1})], G1=[g(x1) ⋯ g(xm)]G_1=[g(x_1)\ \cdots\ g(x_m)] and

K=G1G0†∈Rq×q,(4)K = G_1G_0^\dagger\in\mathbb{R}^{q\times q}, \tag*{(4)}

with †\dagger the Moore–Penrose pseudoinverse [9]. If μk\mu_k and ξkT\xi_k^\mathsf{T} are the eigenvalues and left eigenvectors of KK, then λ^k=ln⁡(μk)/Δt\hat{\lambda}_k = \ln(\mu_k)/\Delta t approximates a Koopman eigenvalue and

ϕ^k(x)=ξkTg(x)=∑iξk,igi(x)(5)\hat{\phi}_k(x) = \xi_k^\mathsf{T}g(x) = \sum_i \xi_{k,i}g_i(x) \tag*{(5)}

approximates the eigenfunction, as the projection of (2) onto the span of the observables dictionary [10].

Test system and its basin

Our test system (Figure 1) is a classical generator model, parameterized with H=5H = 5 s, D=10D = 10 pu, connected to an infinite bus through a series impedance. We do not neglect the resistance RR, which causes classical energy functions1 to break down. With R>0R > 0, path-dependent terms have no closed form [17], and Chiang [18] showed that no general energy function exists for lossy multimachine systems. Conversely, RR does not affect our proposed method. With the state x=[δ ω]Tx = [\delta\ \omega]^\mathsf{T} shifted so that the stable equilibrium is at the origin (that is, x⋆=0x^\star= 0),

Figure 1. One-line diagram of the single-machine infinite-bus system from [11]. Vˉ1,Vˉ2\bar{V}_1,\bar{V}_2 are complex-valued voltage phasors in nodes 1 and 2, respectively. PP (QQ) is the real (reactive) power injected into node 1. RR, XX, and Xd′X'_d are parameters. We choose R=0.05R = 0.05, X=0.30X = 0.30, Vˉ1=1.05\bar{V}_1 = 1.05, Vˉ2=1.00\bar{V}_2 = 1.00, P=0.80P = 0.80, and Xd′=0.20X'_d = 0.20, all values in pu.

\begin{bmatrix} \dot{\delta}\\ \dot{\omega} \end{bmatrix} = \begin{bmatrix} \omega\\ c_{00}+c_0\omega+c_1\cos\delta+c_2\sin\delta \end{bmatrix}, \tag*{(6)}\end{bmatrix}

c00=21.3649c_{00} = 21.3649, c0=−1c_0 = -1, c1=−21.3649c_1 = -21.3649, and c2=−78.5772c_2 = -78.5772. See [11] for details on (6). The Jacobian at the origin, J(x⋆)=[01c2c0]J(x^\star) = \begin{bmatrix}0&1\\c_2&c_0\end{bmatrix}, has λ1,2=−0.5000±j8.8503\lambda_{1,2}=-0.5000\pm j8.8503. The accelerating torque of (6) at ω=0\omega= 0,

P(δ)=c00+c1cos⁡δ+c2sin⁡δ=c00+mcos⁡(δ−a),(7)P(\delta) = c_{00}+c_1\cos\delta+c_2\sin\delta= c_{00}+m\cos(\delta-a), \tag*{(7)}

m=c12+c22=81.4299m = \sqrt{c_1^2+c_2^2} = 81.4299 and a=atan2⁡(c2,c1)=−1.8363a = \operatorname{atan2}(c_2,c_1) = -1.8363. Recall that (7) is the familiar power-angle equation [19 §16.4], but with a lossy line (R>0R > 0) offset. P(δ)=0P(\delta)=0 if and only if

cos⁡(δ−a)=−c00/m.(8)\cos(\delta-a) = -c_{00}/m. \tag*{(8)}

Since ∣−c00/m∣=0.2624<1|-c_{00}/m| = 0.2624 < 1, the cosine satisfies (8) at two points of each period, symmetric about the maximum angle δmax⁡=a\delta_{\max}=a. From (8), δ=a±cos⁡−1(−c00/m)=a±b\delta=a\pm\cos^{-1}(-c_{00}/m)=a\pm b, where b=cos⁡−1(−c00/m)∈(0,π)b=\cos^{-1}(-c_{00}/m)\in(0,\pi). The equilibrium at δ⋆=a+b\delta^\star=a+b has P′(δ⋆)=−msin⁡(δ⋆−a)=−msin⁡(b)<0P'(\delta^\star)=-m\sin(\delta^\star-a)=-m\sin(b)<0, a restoring torque; that is—the stable fixed point. At δu=a−b=−3.6725\delta_u=a-b=-3.6725, P′(δu)>0P'(\delta_u)>0, and the fixed point is a saddle; δu+2π=2.6107\delta_u+2\pi=2.6107 is also a saddle. The saddles at (δ,ω)=(−3.6725,0)(\delta,\omega)=(-3.6725,0) and (2.6107,0)(2.6107,0) have unstable eigenvalue λu=8.3785\lambda_u=8.3785, and their stable manifolds form the boundary of the basin of the origin [11], [17, §4.2].

Eigenfunctions as functions of the state

A fixed point satisfies Stx⋆=x⋆S_tx^\star=x^\star for all tt. If ϕ\phi is an eigenfunction, ϕ(Stx⋆)=eλtϕ(x⋆)\phi(S_tx^\star)=e^{\lambda t}\phi(x^\star), so (eλt−1)ϕ(x⋆)=0(e^{\lambda t}-1)\phi(x^\star)=0 for all tt, and ϕ(x⋆)=0\phi(x^\star)=0 whenever λ≠0\lambda\ne0. Hence, a Koopman eigenfunction with λ≠0\lambda\ne0 vanishes at the equilibrium.2 The approximation (5) inherits this property only if the lifted fixed point is invariant under KK. That is, if

Kg(x⋆)=g(x⋆).(9)Kg(x^\star)=g(x^\star). \tag*{(9)}

Multiplying (9) by ξkT\xi_k^\mathsf{T} yields ξkTKg(x⋆)=ξkTg(x⋆)\xi_k^\mathsf{T}Kg(x^\star)=\xi_k^\mathsf{T}g(x^\star). From the EDMD matrix eigendecomposition, ξkTKg(x⋆)=μkξkTg(x⋆)\xi_k^\mathsf{T}Kg(x^\star)=\mu_k\xi_k^\mathsf{T}g(x^\star). From the last two expressions, we have (μk−1)ξkTg(x⋆)=0(\mu_k-1)\xi_k^\mathsf{T}g(x^\star)=0, and from (5), (μk−1)ϕ^k(x⋆)=0(\mu_k-1)\hat{\phi}_k(x^\star)=0, with ϕ^k(x⋆)=0\hat{\phi}_k(x^\star)=0 for every μk≠1\mu_k\ne1. The least-squares solution (4) satisfies (9) trivially when g(x⋆)=0g(x^\star)=0, which holds for monomials in (δ,ω)(\delta,\omega) and for (10) below, but only approximately otherwise. For example, if gex=[δ ω sin⁡δ cos⁡δ ωsin⁡δ ωcos⁡δ]Tg_{\mathrm{ex}}=[\delta\ \omega\ \sin\delta\ \cos\delta\ \omega\sin\delta\ \omega\cos\delta]^\mathsf{T} [11 Eq. (13)], then gex(0)=[0 0 0 1 0 0]Tg_{\mathrm{ex}}(0)=[0\ 0\ 0\ 1\ 0\ 0]^\mathsf{T}, and the spectrum contains a spurious eigenvalue 0.05290.0529. Centering resolves the issue at no cost, and

gLie=[δ ω sin⁡δ (cos⁡δ−1) ωsin⁡δ ω(cos⁡δ−1)]T(10)g_{\mathrm{Lie}}=[\delta\ \omega\ \sin\delta\ (\cos\delta-1)\ \omega\sin\delta\ \omega(\cos\delta-1)]^\mathsf{T} \tag*{(10)}

spans, up to the constant function, the same space as gexg_{\mathrm{ex}}, and the spurious eigenvalue disappears.

Dictionaries of observable functions

We consider five different dictionaries (Table I). All of them are sets of monomials in the lifted variables z1=δz_1=\delta, z2=ωz_2=\omega, z3=sin⁡δz_3=\sin\delta and z4=cos⁡δ−1z_4=\cos\delta-1, that is,

gi(x)=δα1ωα2(sin⁡δ)α3(cos⁡δ−1)α4,(11)g_i(x)=\delta^{\alpha_1}\omega^{\alpha_2}(\sin\delta)^{\alpha_3}(\cos\delta-1)^{\alpha_4}, \tag*{(11)}

{α1,…,α4}∈Z≥0\{\alpha_1,\ldots,\alpha_4\}\in\mathbb{Z}_{\geq0}, so that a dictionary is a q×4q\times4 integer matrix of exponents. The gradient of (11) is explicit, so the q×2q\times2 Jacobian Jg(x)\mathbf{J}_g(x) of every dictionary is available in closed form. Using (5),

∇ϕ^k(x)=∑iξk,i∇gi(x)=ξkTJg(x).(12)\nabla\hat{\phi}_k(x)=\sum_i \xi_{k,i}\nabla g_i(x)=\boldsymbol{\xi}_k^{\mathsf{T}}\mathbf{J}_g(x). \tag*{(12)}

By (2), inserting the gradient (12), the Lie derivative of an EDMD eigenfunction along the model is

Lfϕ^k(x)=ξkTJg(x)f(x),(13)\mathcal{L}_f\hat{\phi}_k(x)=\boldsymbol{\xi}_k^{\mathsf{T}}\mathbf{J}_g(x)f(x), \tag*{(13)}

with ff the right-hand side of (6). Finally, define the residual

rk:=Lfϕ^k−λ^kϕ^k(14)r_k:=\mathcal{L}_f\hat{\phi}_k-\hat{\lambda}_k\hat{\phi}_k \tag*{(14)}

of the eigenfunction equation, so that the relative residual,

εk(x)=∣rk(x)∣∣λ^kϕ^k(x)∣,λ^k≠0, ϕ^k(x)≠0,(15)\varepsilon_k(x)=\frac{|r_k(x)|}{|\hat{\lambda}_k\hat{\phi}_k(x)|},\qquad\hat{\lambda}_k\ne0,\ \hat{\phi}_k(x)\ne0, \tag*{(15)}

measures, at every point, how far ϕ^k\hat{\phi}_k is from satisfying (2).

Basin of attraction from the eigenfunctions

Lyapunov candidates from the EDMD spectrum

The first candidate is V1=∣ϕ^1∣V_1=|\hat{\phi}_1|, motivated by (3). Define β(t):=∣ϕ^1(x(t))∣2\beta(t):=|\hat{\phi}_1(x(t))|^2 such that β=∣ϕ^1∣\sqrt{\beta}=|\hat{\phi}_1|. Then, for ϕ^1≠0\hat{\phi}_1\ne0,

V˙1=ddt∣ϕ^1∣=dβdβdβdt=12βdβdt=12∣ϕ^1∣ddt∣ϕ^1∣2.(16)\dot{V}_1=\frac{d}{dt}|\hat{\phi}_1|=\frac{d\sqrt{\beta}}{d\beta}\frac{d\beta}{dt}=\frac{1}{2\sqrt{\beta}}\frac{d\beta}{dt}=\frac{1}{2|\hat{\phi}_1|}\frac{d}{dt}|\hat{\phi}_1|^2. \tag*{(16)}

The derivative along the true vector field ff of (6) yields

ddtϕ^1=∇ϕ^1⋅f=Lfϕ^1,(17)\frac{d}{dt}\hat{\phi}_1=\nabla\hat{\phi}_1\cdot f=\mathcal{L}_f\hat{\phi}_1, \tag*{(17)}
ddt∣ϕ^1∣2=(Lfϕ^1)ϕ^1∗+ϕ^1(Lfϕ^1∗)=2Re⁡(ϕ^1∗Lfϕ^1).(18)\frac{d}{dt}|\hat{\phi}_1|^2=(\mathcal{L}_f\hat{\phi}_1)\hat{\phi}_1^*+\hat{\phi}_1(\mathcal{L}_f\hat{\phi}_1^*)=2\operatorname{Re}(\hat{\phi}_1^*\mathcal{L}_f\hat{\phi}_1). \tag*{(18)}

Substituting (18) into (16), V˙1=Re⁡(ϕ^1∗Lfϕ^1)/∣ϕ^1∣\dot{V}_1=\operatorname{Re}(\hat{\phi}_1^*\mathcal{L}_f\hat{\phi}_1)/|\hat{\phi}_1|. From (14),

V˙1=Re⁡(ϕ^1∗(λ^1ϕ^1+r1))∣ϕ^1∣=Re⁡(λ^1)∣ϕ^1∣+Re⁡(ϕ^1∗r1)∣ϕ^1∣.(19)\dot{V}_1=\frac{\operatorname{Re}\left(\hat{\phi}_1^*(\hat{\lambda}_1\hat{\phi}_1+r_1)\right)}{|\hat{\phi}_1|}=\operatorname{Re}(\hat{\lambda}_1)|\hat{\phi}_1|+\frac{\operatorname{Re}(\hat{\phi}_1^*r_1)}{|\hat{\phi}_1|}. \tag*{(19)}

Write ϕ^1=∣ϕ^1∣ejθ\hat{\phi}_1=|\hat{\phi}_1|e^{j\theta}, with ejθe^{j\theta} the unit complex number pointing in the direction of ϕ^1\hat{\phi}_1. Then, ϕ^1∗/∣ϕ^1∣=e−jθ\hat{\phi}_1^*/|\hat{\phi}_1|=e^{-j\theta}, the residual term in (19) becomes Re⁡(ϕ^1∗r1)/∣ϕ^1∣=Re⁡(e−jθr1)=:r1∥\operatorname{Re}(\hat{\phi}_1^*r_1)/|\hat{\phi}_1|=\operatorname{Re}(e^{-j\theta}r_1)=:r_{1\parallel}, and

V˙1=Re⁡(λ^1)∣ϕ^1∣+r1∥,(20)\dot{V}_1=\operatorname{Re}(\hat{\lambda}_1)|\hat{\phi}_1|+r_{1\parallel}, \tag*{(20)}

where r1∥r_{1\parallel} is the component of the residual along ϕ^1\hat{\phi}_1.

The second candidate comes from a different eigenpair of KK. Let μE\mu_E be the real eigenvalue of KK whose continuous-time counterpart λ^E=ln⁡(μE)/Δt\hat{\lambda}_E=\ln(\mu_E)/\Delta t is closest to 2Re⁡(λ1)=−12\operatorname{Re}(\lambda_1)=-1, with left eigenvector ξET\boldsymbol{\xi}_E^{\mathsf{T}}, and let ϕ^E=ξETg(x)\hat{\phi}_E=\boldsymbol{\xi}_E^{\mathsf{T}}g(x), the EDMD counterpart of ∣ϕ1∣2|\phi_1|^2. Because ξE\boldsymbol{\xi}_E is defined only up to sign, we set VE=sϕ^EV_E=s\hat{\phi}_E, with s=sgn⁡ϕ^E(x0)s=\operatorname{sgn}\hat{\phi}_E(x_0) for a fixed reference point x0x_0 of the training lattice; ss is a single constant that makes VE(x0)>0V_E(x_0)>0. From (14) and (17), V˙E=sddtϕ^E=sLfϕ^E=s(λ^Eϕ^E+rE)=λ^E(sϕ^E)+srE=λ^EVE+srE\dot{V}_E=s\frac{d}{dt}\hat{\phi}_E=s\mathcal{L}_f\hat{\phi}_E=s(\hat{\lambda}_E\hat{\phi}_E+r_E)=\hat{\lambda}_E(s\hat{\phi}_E)+sr_E=\hat{\lambda}_EV_E+sr_E.

Lemma 4.1. Let Re⁡(λ^1)<0\operatorname{Re}(\hat{\lambda}_1)<0 and λ^E<0\hat{\lambda}_E<0. At every xx with ϕ^1(x)≠0\hat{\phi}_1(x)\ne0, V1(x)<0V_1(x)<0 if ε1(x)<−Re⁡(λ^1)/∣λ^1∣\varepsilon_1(x)<-\operatorname{Re}(\hat{\lambda}_1)/|\hat{\lambda}_1|. At every xx with VE(x)≠0V_E(x)\ne0, VE(x)<0V_E(x)<0 if εE(x)<1\varepsilon_E(x)<1.

Proof. ∣r1∥∣=∣Re⁡(e−jθr1)∣≤∣e−jθr1∣=∣r1∣|r_{1\parallel}|=|\operatorname{Re}(e^{-j\theta}r_1)|\leq|e^{-j\theta}r_1|=|r_1|. Thus, V˙1=Re⁡(λ^1)∣ϕ^1∣+r1∥≤−∣Re⁡(λ^1)∣∣ϕ^1∣+∣r1∣=−∣Re⁡(λ^1)∣V1(1−ε1∣λ^1∣/∣Re⁡(λ^1)∣)\dot{V}_1=\operatorname{Re}(\hat{\lambda}_1)|\hat{\phi}_1|+r_{1\parallel}\leq-|\operatorname{Re}(\hat{\lambda}_1)||\hat{\phi}_1|+|r_1|=-|\operatorname{Re}(\hat{\lambda}_1)|V_1(1-\varepsilon_1|\hat{\lambda}_1|/|\operatorname{Re}(\hat{\lambda}_1)|). The prefactor −∣Re⁡(λ^1)∣V1-|\operatorname{Re}(\hat{\lambda}_1)|V_1 is strictly negative, so V˙1<0\dot{V}_1<0 whenever (1−ε1∣λ^1∣/∣Re⁡(λ^1)∣)>0(1-\varepsilon_1|\hat{\lambda}_1|/|\operatorname{Re}(\hat{\lambda}_1)|)>0, that is, whenever ε1<−Re⁡(λ^1)/∣λ^1∣\varepsilon_1<-\operatorname{Re}(\hat{\lambda}_1)/|\hat{\lambda}_1|. For the second candidate, V˙E≤−∣λ^E∣VE(1−εE)\dot{V}_E\leq-|\hat{\lambda}_E|V_E(1-\varepsilon_E).

For our system (6), −Re⁡(λ^1)/∣λ^1∣≈0.056-\operatorname{Re}(\hat{\lambda}_1)/|\hat{\lambda}_1|\approx0.056 (e.g., 0.059 for EDMD-Lie), and the modulus of the principal eigenfunction decreases wherever the relative residual is below 6%. For the second candidate, the real eigenfunction decreases wherever the residual is below 100%. The difference is the eigenvalue. For a complex pair, decay is a small fraction of the rotation rate, and only the decay competes with the residual. For a real negative eigenvalue, the whole of ∣λ^E∣|\hat{\lambda}_E| does. Thus, VEV_E certifies larger sets than V1V_1 in Section V, and it suggests a design rule: real eigenfunctions with eigenvalues 2pRe⁡(λ1)2p\operatorname{Re}(\lambda_1), p≥1p\geq1, are the natural Lyapunov candidates in the EDMD spectrum.

Remark 4.2. Let V=∣ϕ1∣2V=|\phi_1|^2. For any xx and any c≥0c\geq0, ∣ϕ1∣≤c|\phi_1|\leq c if and only if ∣ϕ1∣2≤c2|\phi_1|^2\leq c^2, since squaring is strictly increasing on [0,∞)[0,\infty) and ∣ϕ1∣|\phi_1| takes values there. Thus, the set {x:V1(x)≤c}\{x:V_1(x)\leq c\} is the same set of points {x:V(x)≤c2}\{x:V(x)\leq c^2\}. VV and V1V_1 yield the same family of sublevel sets; only the number attached to each set changes, from cc to c2c^2. Both sets have the same boundaries. Since squaring does not change the sublevel sets, ∣ϕ^1∣2|\hat{\phi}_1|^2 certifies exactly the sets of V1V_1 in Table II. Therefore, the gap between the V1V_1 and VEV_E columns measures the EDMD error in the complex mode, not the candidate shape.

Certified sublevel sets

Proposition 4.3 (Sublevel-set certificate [20 §8.2]). Let V∈C1(W)V\in C^1(W) on an open set W≠∅W\ne\varnothing with V(0)=0V(0)=0, and let Ωc\Omega_c be the connected component of {x∈W:V(x)≤c}\{x\in W:V(x)\leq c\} that contains the origin. If Ωc\Omega_c is compact, V>0V>0 and V˙<0\dot{V}<0 on Ωc∖{0}\Omega_c\setminus\{0\}, then Ωc\Omega_c is positively invariant and Ωc⊂B(x∗=0)\Omega_c\subset\mathcal{B}(x^*=0). V1V_1 is not differentiable at the origin, so we apply Proposition 4.3 to VV, whose sublevel set at level c2c^2 is Ωc\Omega_c (cf. Remark 4.2). Both conditions of Proposition 4.3 are monotone in cc because sublevel sets nest, so the largest certified level c∗c^* is found by bisection. We evaluate VV and V˙\dot{V} from (5) and (13) on the lattice δ∈[−4,4]\delta\in[-4,4], ω∈[−20,18]\omega\in[-20,18] with steps 0.025 rad and 0.1 rad/s. We compute Ωc\Omega_c from a flood fill3 from the origin, and reject a level if the component touches the border of the window or if any of its grid points other than the origin violates a condition. Let ρ\rho be the half-diagonal of a lattice cell, so that every point of WW lies within ρ\rho of a grid point (Figure 2). For a level cc, let GcG_c be the set of grid points reachable from the origin through 8-adjacent grid points xix_i with V(xi)≤c+2LVρV(x_i)\le c+2L_V\rho, LV∈RL_V\in\mathbb{R}.

A grid point and its eight neighbors; a connected component of a sublevel set running along the diagonal of the lattice

Figure 2. Left. A grid point xix_i and its eight neighbors, the four that share an edge of a cell with it (solid) and the four that share only a corner (dashed). Right. A connected component Ωc\Omega_c of a sublevel set (green) running along the diagonal of the lattice. Filled markers are grid points inside Ωc\Omega_c. Shaded cells meet Ωc\Omega_c; consecutive ones share a corner, and every corner of every shaded cell (filled and hollow markers) satisfies V≤c+2LVρV \le c+2L_V\rho and belongs to the collar component GcG_c of Lemma 4.4.

Lemma 4.4 (From a discrete lattice to the continuum [4]). Let V∈C1(W)V\in C^1(W) with V(0)=0V(0)=0, and let VV and V˙\dot V be Lipschitz on WW with constants LVL_V and LV˙L_{\dot V}. Suppose that

(a) V>0V>0 and V˙<0\dot V<0 on Br0(0)∖{0}B_{r_0}(0)\setminus\{0\}, an open disk of radius r0>ρr_0>\rho;

(b) every xi∈Gcx_i\in G_c with ∣xi∣≥r0−ρ|x_i|\ge r_0-\rho satisfies V(xi)>LVρV(x_i)>L_V\rho and V˙(xi)≤−η\dot V(x_i)\le-\eta, with η>LV˙ρ\eta>L_{\dot V}\rho;

(c) no point of GcG_c lies on the border of the lattice.

Then, Ωc\Omega_c satisfies Proposition 4.3, and Ωc⊂B(0)\Omega_c\subset B(0).

Proof. Let QQ be the set of closed lattice cells that meet Ωc\Omega_c. If QQ could be split into two families such that no cell of one touches a cell of the other, the unions of the two families would be disjoint closed sets covering Ωc\Omega_c, contradicting its connectedness; hence any two cells of QQ are linked by a chain of cells in QQ in which consecutive cells share a corner. Every cell of QQ contains a point with V≤cV\le c, and each of its corners lies within 2ρ2\rho of that point, so every corner of every cell of QQ satisfies V≤c+2LVρV\le c+2L_V\rho. The origin is such a corner, so all these corners belong to GcG_c. By (c), no cell of QQ is an outermost cell. Thus, Ωc\Omega_c lies in the closed inner rectangle and, being closed in WW, it is compact. Now, let x∈Ωc∖{0}x\in\Omega_c\setminus\{0\}. If ∣x∣<r0|x|<r_0, (a) applies. Otherwise, let xix_i be the corner nearest to xx of a cell of QQ containing xx. Then, ∣x−xi∣≤ρ|x-x_i|\le\rho, xi∈Gcx_i\in G_c, and ∣xi∣≥∣x∣−ρ≥r0−ρ|x_i|\ge|x|-\rho\ge r_0-\rho, so (b) and Lipschitz continuity yield V(x)≥V(xi)−LVρ>0V(x)\ge V(x_i)-L_V\rho>0 and V˙(x)≤V˙(xi)+LV˙ρ<0\dot V(x)\le\dot V(x_i)+L_{\dot V}\rho<0. ■

Remark 4.5 (Families): Positively invariant subsets of the basin are closed under union, so the union of certified sets over both candidates and all dictionaries is itself certified, as in the Lyapunov-function-family approach of [22].

Remark 4.6: Lemma 4.4 gives the conditions under which the grid test of this subsection certifies the continuum. We evaluate the extension along the flow in SIV-C at the grid points by numerical integration, and the basin estimate it produces holds at the grid resolution, up to integration error.

Extension along the flow

The certified set is small when the training data are confined to the linear region because ϕ^1\hat\phi_1 is accurate only near the origin. The semigroup property repairs this without new data, at the cost of numerical integration. This matters because power systems have multiple attractors, operate near a stable fixed point most of the time, and one seeks to estimate the basin from measured ring-downs after small disturbances. Let Ω0⊆Ωc∗\Omega_0\subseteq\Omega_{c^*} be the connected component of {x:∣ϕ^1(x)∣≤c0}\{x:|\hat\phi_1(x)|\le c_0\} that contains the origin, 0<c0≤c∗0<c_0\le c^*, chosen so that ϵ1≤ϵtol\epsilon_1\le\epsilon_{\mathrm{tol}} on Ω0\Omega_0. Define

τ(x):=inf⁡{t≥0:Stx∈Ω0}∈[0,∞],(21)\tau(x):=\inf\{t\ge0:S_t x\in\Omega_0\}\in[0,\infty], \tag*{(21)}

the hitting time of the seed. Since StxS_t x is the state at time tt of the trajectory starting at xx, the set {t≥0:Stx∈Ω0}\{t\ge0:S_t x\in\Omega_0\} collects all times when that trajectory is in the seed set Ω0\Omega_0. The infimum of the set in (21) therefore defines, for each state xx, the first time the trajectory starting at xx enters Ω0\Omega_0. Define

ϕ~1(x):=e−λ^1τ(x)ϕ^1(Sτ(x)x)on {τ<∞}.(22)\widetilde{\phi}_1(x):=e^{-\hat\lambda_1\tau(x)}\hat\phi_1(S_{\tau(x)}x)\qquad\text{on }\{\tau<\infty\}. \tag*{(22)}

Proposition 4.7 (Properties of the extension): Let Ω0\Omega_0 and ϕ~1\widetilde{\phi}_1 be as before. Then:

(i) Ω0\Omega_0 is compact and positively invariant, contains a neighborhood of the origin, and {τ<∞}=B(0)\{\tau<\infty\}=\mathcal{B}(0).

(ii) For every T≥0T\ge0, {τ≤T}=S−T(Ω0)\{\tau\le T\}=S_{-T}(\Omega_0). These sets are compact, positively invariant, nested in TT, and their union is B(0)\mathcal{B}(0). The hitting time is continuous on B(0)\mathcal{B}(0), continuously differentiable on B(0)∖Ω0\mathcal{B}(0)\setminus\Omega_0 with τ˙=−1\dot\tau=-1 there, and τ(x)→∞\tau(x)\to\infty as x→∂B(0)x\to\partial\mathcal{B}(0).

Proof. (i) Proposition 4.3 applied to V=V12V=V_1^2 at level c02c_0^2 yields compactness and positive invariance, since V1>0V_1>0 and V˙1<0\dot V_1<0 on Ωc∗∖{0}⊇Ω0∖{0}\Omega_{c^*}\setminus\{0\}\supseteq\Omega_0\setminus\{0\}. The open set {V1<c0}\{V_1<c_0\} contains the origin, and its connected component containing the origin lies in Ω0\Omega_0. If τ(x)<∞\tau(x)<\infty, the infimum in (21) is attained because Ω0\Omega_0 is closed, so Sτ(x)x∈Ω0⊂B(0)S_{\tau(x)}x\in\Omega_0\subset\mathcal{B}(0) and x∈B(0)x\in\mathcal{B}(0); conversely, Stx→0S_t x\to0 for x∈B(0)x\in\mathcal{B}(0), so Stx∈Ω0S_t x\in\Omega_0 for large tt. (ii) τ(x)≤T\tau(x)\le T if and only if STx∈Ω0S_Tx\in\Omega_0: the “if” is the definition of τ\tau, the “only if” is positive invariance of Ω0\Omega_0. The vector field of (6) is globally Lipschitz (its Jacobian is bounded), so its flow is complete [20 Thm. 3.2], S−TS_{-T} is a homeomorphism of R2\mathbb{R}^2, and {τ≤T}=S−T(Ω0)\{\tau\le T\}=S_{-T}(\Omega_0) is compact. Positive invariance of Ω0\Omega_0 means Ω0⊆S−s(Ω0)\Omega_0\subseteq S_{-s}(\Omega_0) for s≥0s\ge0; applying S−TS_{-T} yields S−T(Ω0)⊆S−(T+s)(Ω0)S_{-T}(\Omega_0)\subseteq S_{-(T+s)}(\Omega_0), so the sets are nested in TT. For 0≤t≤T0\le t\le T, St(S−T(Ω0))=S−(T−t)(Ω0)S_t(S_{-T}(\Omega_0))=S_{-(T-t)}(\Omega_0), and for t>Tt>T it equals St−T(Ω0)⊆Ω0S_{t-T}(\Omega_0)\subseteq\Omega_0; both lie in S−T(Ω0)S_{-T}(\Omega_0) by nestedness, so each set is positively invariant. Their union is {τ<∞}=B(0)\{\tau<\infty\}=\mathcal{B}(0) by (i). Since V˙1<0\dot V_1<0 on ∂Ω0\partial\Omega_0, every trajectory from B(0)∖Ω0\mathcal{B}(0)\setminus\Omega_0 enters Ω0\Omega_0 transversally through ∂Ω0\partial\Omega_0, and by continuous dependence on initial conditions nearby trajectories enter at nearby times; the implicit function theorem applied to V1(Stx)=c0V_1(S_tx)=c_0 at the entry time then gives that τ\tau is C1C^1 on B(0)∖Ω0\mathcal{B}(0)\setminus\Omega_0, and the same argument provides continuity on B(0)\mathcal{B}(0). Differentiating τ(Stx)=τ(x)−t\tau(S_tx)=\tau(x)-t at t=0t=0 yields τ˙=−1\dot\tau=-1. Finally, for every TT, the set {τ≤T}\{\tau\le T\} is a compact subset of the open set B(0)\mathcal{B}(0), so a sequence approaching ∂B(0)\partial\mathcal{B}(0) eventually leaves each one of them, and τ→∞\tau\to\infty. ■

Since ϕ~1(Stx)=ϕ~1(x)−t\widetilde{\phi}_1(S_tx)=\widetilde{\phi}_1(x)-t for t≥τ(x)t\ge\tau(x), (22) satisfies ϕ~1(Stx)=eλ^1tϕ~1(x)\widetilde{\phi}_1(S_tx)=e^{\hat\lambda_1t}\widetilde{\phi}_1(x) outside Ω0\Omega_0. If ϕ1\phi_1 is the exact eigenfunction, the semigroup property yields ϕ1(x)/ϕ~1(x)=ϕ1(y)/ϕ~1(y)\phi_1(x)/\widetilde{\phi}_1(x)=\phi_1(y)/\widetilde{\phi}_1(y) at the entry point y=Sτ(x)xy=S_{\tau(x)}x, with the seed error carried along the flow unchanged, while the error of

‡^{\ddagger}13 after removing z32=−z42−2z4z_3^2=-z_4^2-2z_4.

eigenvalue enters through e(λ^1−λ1)τ(x)e^{(\hat{\lambda}_1-\lambda_1)\tau(x)}. This is the mechanism behind Laplace averages [7] and the construction of [23], with a finite integration time because EDMD supplies an accurate seed. The identity {τ≤T}=S−T(Ω0)\{\tau\le T\}=S_{-T}(\Omega_0) is the trajectory-reversing construction of [3]. EDMD yields a larger seed, hence a shorter horizon, and isostables and isochrons on Ω0\Omega_0 without integration. A trajectory that starts at a small distance d>0d>0 from the basin boundary, that is, from the stable manifold of a saddle, is first drawn toward the saddle and only then expelled along its unstable direction, so it lingers near the saddle for a time that grows like λu−1ln⁡(1/d)\lambda_u^{-1}\ln(1/d) before heading to the seed. The hitting time is therefore

τ=λu−1ln⁡(1/d)+O(1)\tau=\lambda_u^{-1}\ln(1/d)+O(1)

as d→0d\to0. Since the trajectory enters the seed through ∂Ω0\partial\Omega_0, where ∣ϕ^1∣=c0|\hat{\phi}_1|=c_0, (22) yields

∣ϕ~1(x)∣=c0e−Re⁡(λ^1)τ(x)|\tilde{\phi}_1(x)|=c_0e^{-\operatorname{Re}(\hat{\lambda}_1)\tau(x)}

outside Ω0\Omega_0, and substituting the hitting time,

∣ϕ~1(x)∣=κe−Re⁡(λ^1)λu−1ln⁡(1/d)=κdRe⁡(λ^1)/λu,|\tilde{\phi}_1(x)|=\kappa e^{-\operatorname{Re}(\hat{\lambda}_1)\lambda_u^{-1}\ln(1/d)} =\kappa d^{\operatorname{Re}(\hat{\lambda}_1)/\lambda_u},

with κ\kappa bounded away from zero and infinity. The exponent is negative because Re⁡(λ^1)<0\operatorname{Re}(\hat{\lambda}_1)<0, so the modulus grows without bound as we approach the boundary. This is the mechanism by which the eigenfunction fails to extend beyond the basin [7]. The level sets of ∣ϕ~1∣|\tilde{\phi}_1|, the sets S−T(∂Ω0)S_{-T}(\partial\Omega_0), are the isostables, and those of arg⁡ϕ~1\arg\tilde{\phi}_1 the isochrons [8].

Numerical Results

The training data comprise 45 trajectories with initial states on the lattice δ∈{−0.50:0.25:0.50}\delta\in\{-0.50:0.25:0.50\}, ω∈{−1.00:0.25:1.00}\omega\in\{-1.00:0.25:1.00\}, sampled every Δt=5\Delta t=5 ms for 0.8 s (7200 snapshot pairs). Note that the training data come from a linear region of the state space near the stable equilibrium point.

We compare the results obtained with the five dictionaries in Table 1, along with the linearization (ϕ^1=ℓ1x0)(\hat{\phi}_1=\ell_1x_0). Table 2 lists the principal eigenvalues recovered by all dictionaries within 1.3% in frequency and 3.6% in decay rate, and the real eigenvalue λ^E\hat{\lambda}_E closest to 2Re⁡(λ1)2\operatorname{Re}(\lambda_1), which appears as soon as ω2\omega^2 is in the span. Note that the six-function Lie dictionary does not contain it, its closest real eigenvalue being −2.835-2.835.

DictionaryDescription
EDMD-pχ\chiAll monomials in (δ,ω)(\delta,\omega) of degree 1 to χ∈{2,4,6}\chi\in\{2,4,6\}
EDMD-LieThe dictionary in (10), which is adapted from [11]
EDMD-Lie-p2All monomials of degree one and two in (z1,…,z4)(z_1,\ldots,z_4)

Table 1. Dictionaries considered in this letter

Dictionaryqqλ^1\hat{\lambda}_1λ^E\hat{\lambda}_EV1V_1VEV_E
linearization2−0.5000±j8.8503-0.5000 \pm j8.8503—8.3%—
EDMD-p25−0.5011±j8.7504-0.5011 \pm j8.7504−1.0048-1.004823.9%0.0%
EDMD-p414−0.4949±j8.8487-0.4949 \pm j8.8487−1.0001-1.000134.3%60.0%
EDMD-p627−0.5000±j8.8504-0.5000 \pm j8.8504−1.0000-1.000055.9%66.1%
EDMD-Lie6−0.5182±j8.7347-0.5182 \pm j8.7347absent0.2%—
EDMD-Lie-p213‡13^{\ddagger}−0.5153±j8.7655-0.5153 \pm j8.7655−0.9930-0.99306.9%16.5%

Table 2. EDMD EIGENVALUES AND FRACTION OF THE BASIN CERTIFIED BY THE SUBLEVEL SETS OF V1V_1 AND VEV_E

Figure 3(a) shows the modulus of the EDMD-Lie principal eigenfunction and Figure 3(b) shows the local decay rate of ∣ϕ^1∣|\hat{\phi}_1| along the true flow. Note that the exact eigenfunction would give the constant 0.5 on the whole basin, whereas the EDMD eigenfunction gives it in a cross-shaped region around the axes and fails, with the wrong sign, in the first and third quadrants beyond ∣δ∣≈0.3|\delta|\approx0.3 and ∣ω∣≈4|\omega|\approx4. This is because the true isostables follow, to first approximation, the level sets of the transient energy 12ω2+U(δ)\frac{1}{2}\omega^2+U(\delta), whereas (10) is linear in ω\omega.

EDMD eigenfunctions evaluated on the state space

Figure 3. EDMD eigenfunctions evaluated on the state space. (a) log⁡10∣ϕ^1∣\log_{10}|\hat{\phi}_1| for EDMD-Lie, with the certified sublevel set of V1V_1 in white. (b) Local decay rate −V˙1/V1-\dot{V}_1/V_1 of the EDMD-Lie eigenfunction along the true vector field (exact value 0.5), with V˙1=0\dot{V}_1=0 in magenta and the certified set in black. (c) The scaled residual (15) of EDMD-Lie, with the seed set Ω0\Omega_0 in cyan. (d) log⁡10∣ϕ^1∣\log_{10}|\hat{\phi}_1| for EDMD-p6 with its certified set. In all panels, red dots are fixed points; the dotted rectangle is the training lattice; the black dashed line is the stable manifold of the saddles (exact basin boundary), computed by integrating (6) backward in time from the saddles along their stable eigenvectors [1].

Note that U(δ)U(\delta) is the potential energy of (6) [17]. The Lie span of cos⁡δ−1\cos\delta-1 contains −ω2sin⁡δ-\omega^2\sin\delta, which has no counterpart in the span, and the corresponding residual term grows as ω2δ\omega^2\delta. Monomials of degree six, Fig. 3(d), reproduce the eigenfunction over a much larger region because the lattice data identify the Taylor coefficients of ϕ1\phi_1 to that order.

Table 2 and Figure 4 report the certified sublevel sets; every one is a strict subset of the true basin. With V1V_1, the estimates grow with the polynomial degree, from 8% for the quadratic Lyapunov function of the linearization to 56% for EDMD-p6, and the Lie dictionary is the most conservative at 0.2%, for the reason discussed in the previous section. With VEV_E, the real eigenfunction approximates ∣ϕ1∣2|\phi_1|^2, an energy-like function quadratic in the lifted variables, and it certifies 60% and 66% of the basin for EDMD-p4 and EDMD-p6, with a set that reaches the saddle at (2.61,0)(2.61,0). For EDMD-p2, the mode is present, but its eigenfunction carries small spurious linear terms, so it is not positive definite and certifies nothing. EDMD-Lie-p2 improves on EDMD-Lie (17% with VEV_E) yet remains below the monomials. Indeed, on data from the linear region, sin⁡δ\sin\delta, cos⁡δ−1\cos\delta- 1, and their products are nearly collinear with the monomials they approximate, and the condition number of G0G_0 grows from 7×1047 \times10^4 (EDMD-Lie) to 7×1097 \times10^9 (EDMD-Lie-p2) and 6×10176 \times10^{17} for the degree-three lifted set, which is why the family stops at degree two.

Certified sublevel sets of $V_1$ and $V_E$

Figure 4. Certified sublevel sets of (a) V1V_1 and (b) VEV_E, drawn as nested fills over the true basin (gray) and labeled by the dictionary in Table II. The dashed white line in (a) is the linearization.

Sampling the training trajectories over the basin (200 trajectories from δ∈[−2,2]\delta\in[-2, 2], ω∈[−12,12]\omega\in[-12, 12], 5 s each) does not help. The least-squares fit is then dominated by large-amplitude data outside the span of every dictionary. No dictionary certifies more than 0.1% of the basin. This dictionary non-invariance [10] motivates the extension in Section IV-C.

The seed set ∣ϕ^1∣≤c0=0.3139|\hat{\phi}_1| \le c_0 = 0.3139 of the EDMD-Lie eigenfunction (certified level c⋆=0.4981c^\star= 0.4981, ε≤5%\varepsilon\le5\%) contains 54 grid points, 0.09% of the basin. All grid points are integrated with a vectorized fourth-order Runge–Kutta scheme (step 5 ms, 40 s), recording the first entry into the seed set. Every point of the true basin, 58,654 of the 122,301 grid points, reaches it within 8.09 s and no point outside does, so the basin estimate, the colored region of Figure 5(a), coincides with the basin at the grid resolution. The short hitting times reflect the strong unstable eigenvalue of the saddles: trajectories that start near the stable manifolds are expelled within a fraction of a second, and ∣ϕ~1∣|\tilde{\phi}_1| grows by a factor of only 66 between the seed set and the basin boundary, Figure 5(a), consistent with the exponent −0.062-0.062 of Section IV-C. The isochrons of Figure 5(b) wind around the origin with the spiral of a lightly damped focus and accumulate along the saddle manifolds.

Extension of the EDMD-Lie eigenfunction, showing isostables and isochrons

Figure 5. Extension (22) of the EDMD-Lie eigenfunction. (a) ∣ϕ~1∣|\tilde{\phi}_1| on a logarithmic scale. The white lines are the isostables labeled by their hitting time τ=2,4,6\tau= 2, 4, 6 s (τ=8.09\tau= 8.09 s lies within the basin-boundary line width). (b) arg⁡ϕ~1\arg\tilde{\phi}_1 (isochrons). In both, the colored region is the basin estimate {x:τ(x)<∞}\{x : \tau(x) < \infty\}, which coincides with the true basin (dashed) at the grid resolution.

Conclusion

EDMD eigenfunctions can be used beyond the snapshots on which they are learned once they are written as functions of the state. For the single-machine infinite-bus system, two Lyapunov candidates built from the spectrum give certified inner estimates of the basin of attraction, the real eigenfunction with eigenvalue 2Re⁡(λ^1)2\operatorname{Re}(\hat{\lambda}_1) reaching two-thirds of the basin with monomials of degree six learned in the linear region. The Lie-lifted observables that excel in Koopman Kalman filtering produce the smallest certificate, because the dictionary is linear in the rotor speed, and a second lifting round is limited by conditioning. The semigroup property removes the dependence of the basin estimate on the dictionary. Indeed, extended along the flow from a small certified seed, the EDMD-Lie eigenfunction is defined exactly on the basin, and the basin of attraction is recovered at the grid resolution. We show that centering the lifted variables so that the fixed point is invariant under the EDMD matrix, and verifying the decrease of the candidate with the model rather than assuming it from the spectrum, are decisive. Future work will address multimachine power systems and sum-of-squares replacements for the grid-based check.

References

  1. [1]H.-D. Chiang, F. Wu, and P. Varaiya, “Foundations of direct methods for power system transient stability analysis,” IEEE Trans. Circuits Syst., vol. 34, no. 2, pp. 160–173, 1987.DOI
  2. [2]M. Anghel, F. Milano, and A. Pahristodoulou, “Algorithmic construction of Lyapunov functions for power system stability analysis,” IEEE Trans. Circuits Syst. I: Regul. Pap., vol. 60, no. 9, pp. 2533–2546, 2013.DOI
  3. [3]R. Genesio, M. Tartaglia, and A. Vicino, “On the estimation of asymptotic stability regions: State of the art and new proposals,” IEEE Trans. Autom. Control, vol. 30, no. 8, pp. 747–755, 1985.
  4. [4]G. Cavraro, Y. Chen, and A. Bernstein, “Lyapunov-based iterative learning of regions of attraction for autonomous systems,” IEEE Open J. Control Syst., vol. 5, pp. 334–348, 2026.
  5. [5]A. Mauroy, I. Mezić, and Y. Susuki, Eds., The Koopman Operator in Systems and Control: Concepts, Methodologies, and Applications. Springer, 2020.
  6. [6]M. D. Kvalheim and S. Revzen, “Existence and uniqueness of global Koopman eigenfunctions for stable fixed points and periodic orbits,” Phys. D: Nonlinear Phenom., vol. 425, p. 132959, 2021.
  7. [7]A. Mauroy and I. Mezić, “Global stability analysis using the eigenfunctions of the Koopman operator,” IEEE Trans. Autom. Control, vol. 61, no. 11, pp. 3356–3369, 2016.
  8. [8]A. Mauroy, I. Mezić, and J. Moehlis, “Isostables, isochrons, and Koopman spectrum for the action–angle representation of stable fixed point dynamics,” Phys. D: Nonlinear Phenom., vol. 261, pp. 19–30, 2013.
  9. [9]M. O. Williams, I. G. Kevrekidis, and C. W. Rowley, “A data-driven approximation of the Koopman operator: Extending dynamic mode decomposition,” J. Nonlinear Sci., vol. 25, no. 6, pp. 1307–1346, 2015.
  10. [10]M. Korda and I. Mezić, “On convergence of extended dynamic mode decomposition to the Koopman operator,” J. Nonlinear Sci., vol. 28, no. 2, pp. 687–710, 2018.
  11. [11]M. Netto, Y. Susuki, V. Krishnan, and Y. Zhang, “On analytical construction of observable functions in extended dynamic mode decomposition for nonlinear estimation and prediction,” IEEE Control Syst. Lett., vol. 5, no. 6, pp. 1868–1873, 2021.
  12. [12]F.-G. Bierwart and A. Mauroy, “A numerical Koopman-based framework to estimate regions of attraction for general vector fields,” Nonlinear Dyn., vol. 113, no. 13, pp. 15761–15774, 2025.
  13. [13]B. Umäthe and U. Vaidya, “Spectral Koopman method for identifying stability boundary,” IEEE Control Syst. Lett., vol. 8, pp. 247–252, 2024.
  14. [14]U. J. Römer and M. Breitbühler, “A rational ansatz for the approximation of Koopman eigenfunctions,” Proc. Appl. Math. Mech., vol. 24, no. 3, p. e202400187, 2024.
  15. [15]S. A. Deka and D. V. Dimarogonas, “Supervised learning of Lyapunov functions using Laplace averages of approximate Koopman eigenfunctions,” IEEE Control Syst. Lett., vol. 7, pp. 3072–3077, 2023.
  16. [16]A. R. Matavalam, B. Hou, H. Choi, S. Bose, and U. Vaidya, “Data-driven transient stability analysis using the Koopman operator,” Int. J. Electr. Power Energy Syst., vol. 162, p. 110307, 2024.
  17. [17]A. Llamas, “Assessment of direct methods in power system transient stability analysis for on-line applications,” Ph.D. dissertation, Virginia Polytechnic Institute and State University, 1992.
  18. [18]H.-D. Chiang, “Study of the existence of energy functions for power systems with losses,” IEEE Trans. Circuits Syst., vol. 36, no. 11, pp. 1423–1429, 1989.
  19. [19]J. J. Grainger and W. D. Stevenson, Power System Analysis. McGraw-Hill, 1994.
  20. [20]H. K. Khalil, Nonlinear Systems, 3rd ed. Prentice Hall, 2002.
  21. [21]R. C. Gonzalez and R. E. Woods, Digital Image Processing, 3rd ed. Pearson, 2008, ch. 9, pp. 645–647.
  22. [22]T. L. Vu and K. Turitsyn, “Lyapunov functions family approach to transient stability assessment,” IEEE Trans. Power Syst., vol. 31, no. 2, pp. 1269–1277, 2016.
  23. [23]M. Korda and I. Mezić, “Optimal construction of Koopman eigenfunctions for prediction and control,” IEEE Trans. Autom. Control, vol. 65, no. 12, pp. 5114–5129, 2020.

Paper details

vlnc:W_hp6q32eh
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