betaFeedback

WorksEngineering

Characterisation of the DQsym Framework Against Harmonic State-Space and Dynamic Phasor Models for Modular Multilevel Converters

Abstract

Modern power systems rely on power converters that exhibit multifrequency periodic dynamics, such as modular multilevel converters (MMCs). Such converters are commonly analysed using harmonic state-space (HSS) and dynamic phasor (DP) representations. Both approaches rely on building state matrices based on the system topology, whose size grows with the number of harmonics retained in the model. This paper shows that a symmetrical direct-quadrature (dq) dynamic phasor formulation for time-domain simulation, named DQsym, has a smaller solver state matrix than HSS and DP, and whose size is independent of the truncation order K. This reduction results from evaluating harmonic coupling outside the state matrix, in the nonlinear signal blocks. The open-loop MMC model used in this paper is derived in three harmonic-domain (HD) formulations and validated against a three-phase time-domain reference model. The models are compared in terms of state-space size and sparsity, harmonic content, steady-state accuracy and eigenvalues. For truncation at K = 4, the DQsym solver has 45 states against 135 for HSS and DP, stays within about 4% of the reference, and, once closed with its coupling blocks, yields eigenvalues identical to those of HSS and DP.

Introduction

For power systems and power electronics analysis, engineers rely on Harmonic-Domain (HD) state-space models to expose frequency coupling, such as circulating current harmonics, that dictates interactions between a power converter’s internal dynamics and the AC/DC networks it connects to.

HD modelling represents time-varying or linear time-periodic (LTP) electrical systems through the complex Fourier coefficients of their signals, referred to as harmonic vectors in harmonic state-space (HSS) modelling and as dynamic phasors (DP) in generalised averaging [1, 2]. Instead of tracking rapidly oscillating instantaneous waveforms, HD models describe the dynamics of these coefficients, which vary slowly compared with the fundamental period. In DP, each coefficient is computed over a sliding window of one fundamental period, so it is a time-varying quantity [3]. In HSS, the signal is expanded with exponentially modulated periodic components e(s+jhω0)te^{(s+jh\omega_0)t} around a periodic steady state, so the coefficients are defined in the Laplace domain [4, 5].

While both methods are frequency-lifted modelling techniques, based on generalised Fourier expansion, they differ in origin and primary focus [6]. HSS stems from the harmonic analysis framework that is structured around block Toeplitz matrices; on the other hand, DP originates from generalised averaging concepts with the primary focus on increasing the simulation time step to simulate large interconnected power system networks efficiently [3, 7].

HSS relies on a block-banded, block-Toeplitz LTI state-space representation over a contiguous harmonic range, h∈{−K,…,K}h \in\{-K,\ldots,K\}, thereby converting LTP power converters, such as a Modular Multilevel Converter (MMC), into an equivalent LTI state-space model. HSS is widely used for small-signal stability analysis and exact harmonic impedance extraction [5, 8]. On the other hand, the DP approach retains only non-negative harmonic orders h∈{0,…,K}h \in\{0,\ldots,K\} and is typically implemented in a real-valued form to reduce system computational burden [9, 10, 11, 12]. Analytically, both approaches share a structural cost: the state space matrix size scales with the truncation order KK, thereby affecting the state matrix sparsity and the resulting harmonic coupling. In addition, the sparsity and coupling structure only becomes apparent after the full harmonic-domain matrices are derived.

To avoid this dependence of the state matrix on the number of harmonic orders, the DQsym framework [13, 14, 15] represents harmonics in the signals rather than in the states. Each signal is expressed as a 3×(K+1)3 \times(K + 1) complex matrix, whose rows correspond to the positive-, negative-, and zero-sequence components and whose columns correspond to the DC term and harmonic orders 1 to KK, with each entry given in d+jqd + jq form. The system matrices of the studied model are therefore retained without augmentation, regardless of KK. Harmonic coupling is introduced only where signals are multiplied, in coupling blocks outside the solver, so the coupling structure follows directly from the block topology rather than from a derived global matrix. The computational cost is thereby shifted from the size of the solver state matrix to the width of the signals and the convolutions evaluated at each time step.

Throughout this paper, DQsym solver refers to the arm- and neutral-current dynamics integrated by the solver, and DQsym to the complete model including the capacitor and multiplication blocks. The contributions of this paper are as follows:

  • Derivation of an open-loop MMC model in the abcabc frame and in three harmonic-domain formulations: HSS, DP, and DQsym.

  • A comparison of the three formulations for the same MMC at a matched harmonic truncation order, in terms of state-vector size, state-matrix sparsity, harmonic content, time-domain accuracy against the abcabc reference, and eigenvalues.

  • A demonstration that DQsym achieves accuracy comparable to HSS and DP with a diagonal, KK-independent solver state matrix and that closing the coupling blocks around the solver recovers the same eigenvalues.

The remainder of this paper is organised as follows. Section 2 reviews the HSS and DP formulations and outlines the mathematical approach adopted in this work. Section 3 develops the MMC model in the abcabc frame and derives the corresponding HD models from the converter’s ordinary differential equations. Section 4 presents and discusses the results of the analysis. Finally, Section 5 concludes the paper.

Harmonic Domain Formulations

Building a model in HD relies on three fundamental mathematical operations:

  • Fourier Series Expansion: A time-domain signal x(t)x(t) is expanded into Fourier components using

x(t)≈∑h=−KK⟨x⟩h(t)ejhω0t(1)x(t) \approx\sum_{h=-K}^{K} \langle x\rangle_h(t)e^{jh\omega_0t} \tag*{(1)}

where ⟨x⟩h(t)\langle x\rangle_h(t) is the hh-th dynamic phasor or harmonic, and h∈Zh \in\mathbb{Z} is the harmonic order, corresponding to the frequency hω0h\omega_0, and KK is the truncation order, so that only harmonics with ∣h∣≤K\lvert h\rvert\le K are retained.

  • Derivative Property (Frequency Shift): Differentiating a signal in the time domain translates into a state derivative plus an imaginary frequency offset in the harmonic domain:

⟨dxdt⟩h=d⟨x⟩hdt+jhω0⟨x⟩h(2)\left\langle\frac{dx}{dt}\right\rangle_h = \frac{d\langle x\rangle_h}{dt} + jh\omega_0\langle x\rangle_h \tag*{(2)}

This offset creates diagonal matrices, such as the differentiation matrix N=blkdiag⁡(jhω0I)\mathcal{N}=\operatorname{blkdiag}(jh\omega_0I), that shift the system dynamics along the selected frequency.

  • Linearisation: The non-linear differential equations governing the model f(x,u)f(x,u) are expanded around a steady-state operating point using Taylor series or Jacobian linearisation to derive an LTP or LTI system.

These steps transform the system from nonlinear and time-periodic to linear and time-invariant. Fourier series expansion and the derivative property are universal core transformations shared across all three frameworks. However, Toeplitz matrices are fundamental to HSS and DP, whereas DQsym replaces global Toeplitz matrices with a specialised 4-step phase-domain multiplication algorithm that involves transforming the multiplied variables to the abcabc frame, extracting Fourier coefficients from each signal, applying convolution, and packing the result into a complex sequence-dq signal for the DQsym solver [13, 15].

Multiplication of power converter state variables by non-linear switching functions in the time domain transforms into a discrete convolution across the harmonic domain ⟨xy⟩h=∑l⟨x⟩h−l⟨y⟩l\langle xy\rangle_h=\sum_l\langle x\rangle_{h-l}\langle y\rangle_l. When expressed in matrix form, these convolutions are represented as block-Toeplitz matrices, which capture the coupling between harmonic orders.

Table 1 compares the HD formulations in terms of state-matrix structure, sparsity and retained harmonic orders.

FeatureDPDQsymHSS
State matrix formReal-valued, with frequency-shift 2×22\times2 rotation blocksComplex-valued as sequence dqpr,ndq_{p}^{r,n}Block-Toeplitz minus frequency shift, AHSS=A−N\mathcal{A}_{\mathrm{HSS}}=\mathcal{A}-\mathcal{N}
Retained harmonic orderh∈{0,…,K}h\in\{0,\ldots,K\}h∈{0,…,K}h\in\{0,\ldots,K\}contiguous h∈{−K,…,K}h\in\{-K,\ldots,K\}
State dimension(2K+1)Nbase(2K+1)N_{\mathrm{base}} (real-split)NsolverN_{\mathrm{solver}}, each state transformed into complex dqdq, independent of KK(2K+1)Nbase(2K+1)N_{\mathrm{base}}
StructureBanded, with real 2×22\times2 rotation blocksDiagonal and decoupled per sequence and harmonicBanded block-Toeplitz

Table 1. Comparison of DP, DQsym and HSS formulations

Benchmark Model of Modular Multilevel Converter (MMC)

Open-loop MMC Model in the abcabc Frame

An MMC phase leg consists of an upper and a lower arm connected between the DC bus and the AC terminal, as in Figure 1. Each arm contains a series inductor LarmL_{\mathrm{arm}}, a resistance RarmR_{\mathrm{arm}} and an equivalent capacitor Carm=CSM/NsmC_{\mathrm{arm}}=C_{\mathrm{SM}}/N_{\mathrm{sm}} representing the series connection of the submodule capacitors. The arm-averaged representation is used, with each arm described by its aggregated capacitor voltage and a controlled voltage source defined by its insertion index, which captures the main electrical behaviour without modelling individual submodules. The model is open-loop with no controllers included, and the insertion indices are treated as inputs. The AC side is represented by an ideal voltage source v⃗g\vec{v}^{g} at the converter terminals, and a star-point grounding reactor, RstarR_{\mathrm{star}}-LstarL_{\mathrm{star}} branch, is connected at the AC terminals. The converter DC side is held constant at vdcuv_{\mathrm{dc}}^{u} and vdclv_{\mathrm{dc}}^{l} with respect to ground.

Figure 1. Three-phase MMC with half-bridge submodules: upper (blue) and lower (red) arms, split DC link, AC grid with star-point grounding, and showing the used arm-averaged model.

The modulated arm voltages are, for φ∈{a,b,c}\varphi\in\{a,b,c\},

vmφu=mφuvCφu,vmφl=mφlvCφl(3)v_{m\varphi}^{u}=m_{\varphi}^{u}v_{C\varphi}^{u},\qquad v_{m\varphi}^{l}=m_{\varphi}^{l}v_{C\varphi}^{l} \tag*{(3)}

where m⃗φu\vec{m}^{u}_{\varphi} and m⃗φl\vec{m}^{l}_{\varphi} are the upper and lower insertion indices. The upper arm current iφui^{u}_{\varphi} flows from the positive pole to the AC terminal, and the lower arm current iφli^{l}_{\varphi} from the AC terminal to the negative pole. Three-phase quantities are collected in vectors, i⃗u=[iau ibu icu]T\vec{i}^{u} = [i^{u}_{a}\ i^{u}_{b}\ i^{u}_{c}]^{T}, 1⃗≜[1 1 1]T\vec{1} \triangleq[1\ 1\ 1]^{T}, and ⊙\odot denotes elementwise multiplication. Applying Kirchhoff’s voltage law in each arm, with the neutral voltage set to vn=0v_{n}=0, gives the averaged nonlinear model:

Larmdi⃗udt=vdcu1⃗−v⃗g−Rarmi⃗u−m⃗u⊙v⃗Cu(4)L_{arm}\frac{d\vec{i}^{u}}{dt}=v^{u}_{dc}\vec{1}-\vec{v}^{g}-R_{arm}\vec{i}^{u}-\vec{m}^{u}\odot\vec{v}^{u}_{C} \tag*{(4)}
Larmdi⃗ldt=−vdcl1⃗+v⃗g−Rarmi⃗l−m⃗l⊙v⃗Cl(5)L_{arm}\frac{d\vec{i}^{l}}{dt}=-v^{l}_{dc}\vec{1}+\vec{v}^{g}-R_{arm}\vec{i}^{l}-\vec{m}^{l}\odot\vec{v}^{l}_{C} \tag*{(5)}
Carmdv⃗Cudt=m⃗u⊙i⃗u(6)C_{arm}\frac{d\vec{v}^{u}_{C}}{dt}=\vec{m}^{u}\odot\vec{i}^{u} \tag*{(6)}
Carmdv⃗Cldt=m⃗l⊙i⃗l(7)C_{arm}\frac{d\vec{v}^{l}_{C}}{dt}=\vec{m}^{l}\odot\vec{i}^{l} \tag*{(7)}
Lstardi⃗ndt=v⃗g−Rstar1⃗1⃗Ti⃗n.(8)L_{star}\frac{d\vec{i}_{n}}{dt}=\vec{v}^{g}-R_{star}\vec{1}\vec{1}^{T}\vec{i}_{n}. \tag*{(8)}

With vn=0v_{n}=0 the three phases are decoupled, and i⃗n\vec{i}_{n} does not influence the arm quantities.

For interpretation, the sum and difference variables:

iφΔ=iφu−iφl,iφΣ=12(iφu+iφl),i^{\Delta}_{\varphi}=i^{u}_{\varphi}-i^{l}_{\varphi},\qquad i^{\Sigma}_{\varphi}=\frac{1}{2}\left(i^{u}_{\varphi}+i^{l}_{\varphi}\right),
vCφΔ=12(vCφu−vCφl),vCφΣ=12(vCφu+vCφl),(9)v^{\Delta}_{C\varphi}=\frac{1}{2}\left(v^{u}_{C\varphi}-v^{l}_{C\varphi}\right),\qquad v^{\Sigma}_{C\varphi}=\frac{1}{2}\left(v^{u}_{C\varphi}+v^{l}_{C\varphi}\right), \tag*{(9)}

are used, where iφΔi^{\Delta}_{\varphi} is the AC output current and iφΣi^{\Sigma}_{\varphi} the circulating current. They are obtained from the arm quantities

Parameter & DescriptionValue
VDCV_{DC}: Pole-to-pole DC voltage640 kV
RarmR_{arm}, LarmL_{arm}: Arm resistance, inductance1.07 Ω\Omega, 50 mH
CSMC_{SM}, NsmN_{sm}: Submodule capacitance, submodules per arm10 mF, 400
CarmC_{arm}: Equivalent arm capacitance25 μ\muF
RstarR_{star}, LstarL_{star}: Star-point reactor resistance, inductance10 Ω\Omega, 5 H
VgV_{g}: AC voltage (phase, peak)271.9 kV
m^\hat{m}, ϑ\vartheta: Modulation amplitude, angle0.85, −2∘-2^{\circ}
ω0\omega_{0}: Fundamental frequency2π⋅502\pi\cdot50 rad/s

by a linear transformation and are used as outputs only. The models below are formulated in terms of upper- and lower-arm variables.

The AC source is a balanced positive-sequence voltage, v⃗g(t)=Vgc⃗(ω0t)\vec{v}^{g}(t)=V_{g}\vec{c}(\omega_{0}t) where c⃗(θ)=[cos⁡θ cos⁡(θ−2π3) cos⁡(θ+2π3)]T\vec{c}(\theta)=[\cos\theta\ \cos(\theta-\frac{2\pi}{3})\ \cos(\theta+\frac{2\pi}{3})]^{T}. The insertion indices are set in open loop with a sinusoidal reference and no harmonic injection,

m⃗u=12(1⃗−m^c⃗(ω0t+ϑ)),m⃗l=12(1⃗+m^c⃗(ω0t+ϑ)),(10)\vec{m}^{u}=\frac{1}{2}\left(\vec{1}-\hat{m}\vec{c}(\omega_{0}t+\vartheta)\right),\qquad \vec{m}^{l}=\frac{1}{2}\left(\vec{1}+\hat{m}\vec{c}(\omega_{0}t+\vartheta)\right), \tag*{(10)}

and the pole potentials are vdcu=VDC/2v^{u}_{dc}=V_{DC}/2 and vdcl=−VDC/2v^{l}_{dc}=-V_{DC}/2. The parameters are listed in tbl:2.

Collecting (4), the state and input vectors are:

x⃗=[i⃗uT i⃗lT v⃗CuT v⃗ClT i⃗nT]T∈R15,\vec{x}=\left[\vec{i}^{uT}\ \vec{i}^{lT}\ \vec{v}^{uT}_{C}\ \vec{v}^{lT}_{C}\ \vec{i}^{T}_{n}\right]^{T}\in\mathbb{R}^{15},
u⃗=[v⃗gT m⃗uT m⃗lT vdcu vdcl]T∈R11.(11)\vec{u}=\left[\vec{v}^{gT}\ \vec{m}^{uT}\ \vec{m}^{lT}\ v^{u}_{dc}\ v^{l}_{dc}\right]^{T}\in\mathbb{R}^{11}. \tag*{(11)}

Every term in (4) is either linear in the states and inputs or a product of an insertion index with a state, so the model is bilinear,

x⃗˙=f⃗(x⃗,u⃗)=A(u⃗)x⃗+B0u⃗,(12)\dot{\vec{x}}=\vec{f}(\vec{x},\vec{u})=A(\vec{u})\vec{x}+B_{0}\vec{u}, \tag*{(12)}

where B0B_{0} is constant and A(u⃗)A(\vec{u}) contains the circuit parameters and the insertion indices. Since the insertion indices (6) are TT-periodic with T=2π/ω0T=2\pi/\omega_{0}, the steady state is periodic, x⃗0(t)=x⃗0(t+T)\vec{x}_{0}(t)=\vec{x}_{0}(t+T). Linearising around (x⃗0(t),u⃗0(t))(\vec{x}_{0}(t),\vec{u}_{0}(t)) gives the LTP system:

Δx⃗˙=A(t)Δx⃗+B(t)Δu⃗,(13)\Delta\dot{\vec{x}}=A(t)\Delta\vec{x}+B(t)\Delta\vec{u}, \tag*{(13)}
A(t)=∂f⃗∂x⃗∣0=A(u⃗0(t)),B(t)=∂f⃗∂u⃗∣0.A(t)=\left.\frac{\partial\vec{f}}{\partial\vec{x}}\right|_{0}=A(\vec{u}_{0}(t)),\qquad B(t)=\left.\frac{\partial\vec{f}}{\partial\vec{u}}\right|_{0}.

Because (8) is bilinear, the state Jacobian is simply A(u⃗)A(\vec{u}) evaluated along the steady-state inputs, and the same symbol AA is used for it. In open loop, A(t)A(t) depends only on the insertion indices m⃗0u(t)\vec{m}^{u}_{0}(t), m⃗0l(t)\vec{m}^{l}_{0}(t), whereas B(t)B(t) depends on the steady-state capacitor voltages and arm currents through the columns associated with Δm⃗u\Delta\vec{m}^{u} and Δm⃗l\Delta\vec{m}^{l}. Figure 2 shows the sparsity pattern of the linearised open-loop abc arm-averaged MMC model, with AabcA_{abc} (15×1515\times15) and BabcB_{abc} (15×1115\times11) each have 27 nonzeros.

Figure 2. Sparsity pattern of the linearised open-loop abc arm-averaged MMC model. Each cell is one signal of phase aa, bb, or cc. Dots are nonzero entries.

MMC Model in harmonic domain formulation

This subsection derives the three harmonic-domain models of the benchmark MMC. Linearising the open-loop state-space model in (12) around the periodic steady state and expanding the result in Fourier series gives the HSS model. Applying generalised averaging to the nonlinear model instead gives the DP model, and transforming each harmonic of the DP model to sequence dq components gives the DQsym model. All three use the same model, parameters and variables.

HSS Model

The HSS model is obtained from the linearised model (13) by representing its time-periodic matrices in the harmonic domain. The periodic matrices are expanded as

A(t)=∑hAhejhω0t,B(t)=∑hBhejhω0t,A(t)=\sum_h A_h e^{jh\omega_0t}, \qquad B(t)=\sum_h B_h e^{jh\omega_0t},

with

Ah=1T∫0TA(t)e−jhω0t dt,A_h=\frac{1}{T}\int_0^T A(t)e^{-jh\omega_0t}\,\mathrm{d}t,

the same expansion can be done to BhB_h, and the signals as

Δx⃗(t)=∑hΔX⃗he(s+jhω0)t.\Delta\vec{x}(t)=\sum_h \Delta\vec{X}_h e^{(s+jh\omega_0)t}.

Truncating to ∣h∣≤K|h|\le K and stacking X=[ΔX⃗−KT ⋯ ΔX⃗KT]T\mathcal{X}=[\Delta\vec{X}_{-K}^{T}\ \cdots\ \Delta\vec{X}_{K}^{T}]^{T}, likewise U\mathcal{U}, gives the linear time-invariant HSS model:

sX=(A−N)X+BU,(14)s\mathcal{X}=(\mathcal{A}-\mathcal{N})\mathcal{X}+\mathcal{B}\mathcal{U}, \tag*{(14)}
A=[A0A−1⋯A−2KA1A0⋯A−2K+1⋮⋮⋱⋮A2KA2K−1⋯A0],(15)\mathcal{A}= \begin{bmatrix} \mathbf{A}_0 & \mathbf{A}_{-1} & \cdots& \mathbf{A}_{-2K}\\ \mathbf{A}_1 & \mathbf{A}_0 & \cdots& \mathbf{A}_{-2K+1}\\ \vdots& \vdots& \ddots& \vdots\\ \mathbf{A}_{2K} & \mathbf{A}_{2K-1} & \cdots& \mathbf{A}_0 \end{bmatrix}, \tag*{(15)}

where N=diag⁡(−K,…,K)⊗jω0I15\mathcal{N}=\operatorname{diag}(-K,\ldots,K)\otimes j\omega_0 I_{15} and B\mathcal{B} is the block-Toeplitz matrix built from BhB_h in the same way as (15). For an MMC model with Nbase=15N_{\mathrm{base}}=15 states and Mbase=11M_{\mathrm{base}}=11 inputs, truncating at K=4K=4 yields an HSS model with 15(2K+1)=13515(2K+1)=135 complex states and 11(2K+1)=9911(2K+1)=99 complex inputs.

Because (10) contains only the DC and fundamental components, only A0\mathbf{A}_0 and A±1\mathbf{A}_{\pm1} are nonzero in (15), and each harmonic of the arm quantities couples only to its neighbours h±1h\pm1. Figure 3 shows the sparsity pattern of the HSS model for K=4K=4, AHSS−N\mathcal{A}_{\mathrm{HSS}}-\mathcal{N} has 135×135135\times135, 483 nonzeros, and BHSS\mathcal{B}_{\mathrm{HSS}} has 135×99135\times99, 1107 nonzeros. Each signal block spans the 2K+12K+1 contiguous harmonic components h∈{−K,…,K}h\in\{-K,\ldots,K\}. The banded blocks in AHSS\mathcal{A}_{\mathrm{HSS}} are the Toeplitz coupling through m⃗\vec{m}, and the full blocks in BHSS\mathcal{B}_{\mathrm{HSS}} come from the harmonic content of vC(t)v_C(t) and i(t)i(t).

Figure 3. Sparsity pattern of the HSS model for K=4K=4: AHSS−N\mathcal{A}_{\mathrm{HSS}}-\mathcal{N} and BHSS\mathcal{B}_{\mathrm{HSS}}. Dots are nonzero entries.

DP Model

Whereas HSS linearises first and then expands in Fourier series, DP first transforms the nonlinear model (12) and then linearises it.

The hh-th dynamic phasor of xx is

⟨x⟩h(t)=1T∫t−Ttx(τ)e−jhω0τ dτ,\langle x\rangle_h(t)=\frac{1}{T}\int_{t-T}^{t}x(\tau)e^{-jh\omega_0\tau}\,\mathrm{d}\tau,

with the properties

d⟨x⟩hdt=⟨dxdt⟩h−jhω0⟨x⟩h,⟨x⃗⊙y⃗⟩h=∑l⟨x⃗⟩h−l⊙⟨y⃗⟩l.(16)\frac{\mathrm{d}\langle x\rangle_h}{\mathrm{d}t}=\left\langle\frac{\mathrm{d}x}{\mathrm{d}t}\right\rangle_h-jh\omega_0\langle x\rangle_h,\qquad \langle\vec{x}\odot\vec{y}\rangle_h=\sum_l\langle\vec{x}\rangle_{h-l}\odot\langle\vec{y}\rangle_l. \tag*{(16)}

Applying (16) to (4) gives, for 0≤h≤K0\le h\le K and with Dh=ddt+jhω0D_h=\frac{\mathrm{d}}{\mathrm{d}t}+jh\omega_0,

LarmDh⟨i⃗ u⟩h=⟨vdcu⟩h1⃗−⟨v⃗ g⟩h−Rarm⟨i⃗ u⟩h−⟨v⃗m u⟩h,(17)L_{\mathrm{arm}}D_h\langle\vec{i}^{\,u}\rangle_h=\langle v_{\mathrm{dc}}^{u}\rangle_h\vec{1}-\langle\vec{v}^{\,g}\rangle_h-R_{\mathrm{arm}}\langle\vec{i}^{\,u}\rangle_h-\langle\vec{v}^{\,u}_m\rangle_h, \tag*{(17)}
LarmDh⟨i⃗ l⟩h=−⟨vdcl⟩h1⃗+⟨v⃗ g⟩h−Rarm⟨i⃗ l⟩h−⟨v⃗m l⟩h,(18)L_{\mathrm{arm}}D_h\langle\vec{i}^{\,l}\rangle_h=-\langle v_{\mathrm{dc}}^{l}\rangle_h\vec{1}+\langle\vec{v}^{\,g}\rangle_h-R_{\mathrm{arm}}\langle\vec{i}^{\,l}\rangle_h-\langle\vec{v}^{\,l}_m\rangle_h, \tag*{(18)}
CarmDh⟨v⃗C u⟩h=∑l⟨m⃗ u⟩h−l⊙⟨i⃗ u⟩l,(19)C_{\mathrm{arm}}D_h\langle\vec{v}^{\,u}_C\rangle_h=\sum_l\langle\vec{m}^{\,u}\rangle_{h-l}\odot\langle\vec{i}^{\,u}\rangle_l, \tag*{(19)}
CarmDh⟨v⃗C l⟩h=∑l⟨m⃗ l⟩h−l⊙⟨i⃗ l⟩l,(20)C_{\mathrm{arm}}D_h\langle\vec{v}^{\,l}_C\rangle_h=\sum_l\langle\vec{m}^{\,l}\rangle_{h-l}\odot\langle\vec{i}^{\,l}\rangle_l, \tag*{(20)}
LstarDh⟨in⟩h=⟨v⃗ g⟩h−Rstar1⃗T⟨i⃗n⟩h,(21)L_{\mathrm{star}}D_h\langle i_n\rangle_h=\langle\vec{v}^{\,g}\rangle_h-R_{\mathrm{star}}\vec{1}^{T}\langle\vec{i}_n\rangle_h, \tag*{(21)}

where ⟨v⃗m u⟩h\langle\vec{v}^{\,u}_m\rangle_h and ⟨v⃗m l⟩h\langle\vec{v}^{\,l}_m\rangle_h are the DPs of the modulated arm voltages (3),

⟨v⃗m u⟩h=∑l⟨m⃗ u⟩h−l⊙⟨v⃗C u⟩l,⟨v⃗m l⟩h=∑l⟨m⃗ l⟩h−l⊙⟨v⃗C l⟩l,(22)\langle\vec{v}^{\,u}_m\rangle_h=\sum_l\langle\vec{m}^{\,u}\rangle_{h-l}\odot\langle\vec{v}^{\,u}_C\rangle_l,\qquad \langle\vec{v}^{\,l}_m\rangle_h=\sum_l\langle\vec{m}^{\,l}\rangle_{h-l}\odot\langle\vec{v}^{\,l}_C\rangle_l, \tag*{(22)}

The modulation (10) has ⟨m⃗ u⟩0=⟨m⃗ l⟩0=121⃗\langle\vec{m}^{\,u}\rangle_0=\langle\vec{m}^{\,l}\rangle_0=\frac{1}{2}\vec{1} and ⟨m⃗ u⟩1=−⟨m⃗ l⟩1=14μ⃗1\langle\vec{m}^{\,u}\rangle_1=-\langle\vec{m}^{\,l}\rangle_1=\frac{1}{4}\vec{\mu}_1, where μ⃗1=4ejθ[1 a2 a]T\vec{\mu}_1=4e^{j\theta}[1\ a^2\ a]^T and a=ej2π/3a=e^{j2\pi/3}. The pole voltages enter only at h=0h=0 and the AC source only at h=1h=1.

A real-valued model is obtained by taking the real and imaginary parts of each DP as states,

X=[⟨x⃗⟩0T ℜ⟨x⃗⟩1T ℑ⟨x⃗⟩1T ⋯ ℜ⟨x⃗⟩KT ℑ⟨x⃗⟩KT]T,(23)\mathcal{X}=[\langle\vec{x}\rangle_0^T\ \Re\langle\vec{x}\rangle_1^T\ \Im\langle\vec{x}\rangle_1^T\ \cdots\ \Re\langle\vec{x}\rangle_K^T\ \Im\langle\vec{x}\rangle_K^T]^T, \tag*{(23)}

with U\mathcal{U} built from u⃗\vec{u} in the same way, which gives the nonlinear time-invariant DP model. For K=4K=4, the state-space formulation of the DP model has 135 real states and 99 real inputs.

The DP model is inherently time-invariant, and at steady state it has a constant equilibrium (X0,U0)(X_0, U_0). Linearising around this operating point gives an LTI representation.

Figure 4 shows the sparsity pattern of the DP model for K=4K = 4, ADPA_{\mathrm{DP}} has 135×135135 \times135, 699 nonzeros and BDPB_{\mathrm{DP}} has 135×99135 \times99, 1059 nonzeros. Each signal block holds ⟨x⟩0\langle x\rangle_0 and the real and imaginary parts of ⟨x⟩1,…,⟨x⟩K\langle x\rangle_1,\ldots,\langle x\rangle_K. The frequency shift hω0h\omega_0 appears as 2×22 \times2 blocks in ADPA_{\mathrm{DP}}, which makes it denser than AHSS−NA_{\mathrm{HSS}}-\mathcal{N}.

Figure 4. Sparsity pattern of the DP model for K=4K = 4. Dots are nonzero entries.

DQsym

In the DQsym implementation, the model is split into two parts. The first is the arm current dynamics of {(xref)}(4a)–(5),

Larmdi⃗udt=v⃗dcu−v⃗g−Rarmi⃗u−v⃗mu,L_{\mathrm{arm}}\frac{d\vec{i}^{u}}{dt}=\vec{v}^{u}_{\mathrm{dc}}-\vec{v}^{g}-R_{\mathrm{arm}}\vec{i}^{u}-\vec{v}^{u}_{m},
Larmdi⃗ldt=−v⃗dcl+v⃗g−Rarmi⃗l−v⃗ml.(24)L_{\mathrm{arm}}\frac{d\vec{i}^{l}}{dt}=-\vec{v}^{l}_{\mathrm{dc}}+\vec{v}^{g}-R_{\mathrm{arm}}\vec{i}^{l}-\vec{v}^{l}_{m}. \tag*{(24)}

which is linear and time-invariant. Because LarmL_{\mathrm{arm}} and RarmR_{\mathrm{arm}} are identical in all three phases, the symmetrical-component transformation

x⃗pnz=Tx⃗abc,T=13[1aa21a2a111],a=ej2π/3,(25)\vec{x}_{pnz}=\mathbf{T}\vec{x}_{abc},\qquad\mathbf{T}=\frac{1}{3}\begin{bmatrix}1&a&a^2\\1&a^2&a\\1&1&1\end{bmatrix},\qquad a=e^{j2\pi/3}, \tag*{(25)}

commutes with the arm impedance and decouples {(xref)}(16) into independent positive-, negative-, and zero-sequence channels. Each sequence component is then expressed as a sum of complex dq phasors, one per retained harmonic order h∈{0,…,K}h\in\{0,\ldots,K\},

xs(t)=∑h=0K(Xs,hd+jXs,hq)ejhω0t,s∈{p,n,z},(26)x_s(t)=\sum_{h=0}^{K}\left(X^{d}_{s,h}+jX^{q}_{s,h}\right)e^{jh\omega_0t},\qquad s\in\{p,n,z\}, \tag*{(26)}

where ω0\omega_0 is the fundamental angular frequency. Substituting {(xref)}(18) into the sequence form of {(xref)}(16) gives, for each harmonic order hh, where I\mathbf{I} is the identity matrix,

ddtX⃗h=AhX⃗h+BhU⃗h,Ah=−(RarmLarm+jhω0)I,(27)\frac{d}{dt}\vec{X}_h=A_h\vec{X}_h+B_h\vec{U}_h,\qquad A_h=-\left(\frac{R_{\mathrm{arm}}}{L_{\mathrm{arm}}}+jh\omega_0\right)\mathbf{I}, \tag*{(27)}

with states X⃗h=[i⃗pnz,hu i⃗pnz,hl i⃗pnz,hn]T\vec{X}_h=[\vec{i}^{u}_{pnz,h}\ \vec{i}^{l}_{pnz,h}\ \vec{i}^{n}_{pnz,h}]^{T} and inputs

U⃗h=[v⃗pnz,hacT v⃗u,pnz,hdcT v⃗l,pnz,hdcT v⃗u,pnz,huT v⃗pnz,hlT]T,(28)\vec{U}_h=[\vec{v}^{ac}_{pnz,h}{}^{T}\ \vec{v}^{dc}_{u,pnz,h}{}^{T}\ \vec{v}^{dc}_{l,pnz,h}{}^{T}\ \vec{v}^{u}_{u,pnz,h}{}^{T}\ \vec{v}^{l}_{pnz,h}{}^{T}]^{T}, \tag*{(28)}

giving 9 states and 15 inputs forming the solver state-space matrix. The solver matrix ADQsymA_{\mathrm{DQsym}} remains 9×99\times9 for any KK, and the harmonic order enters only through a diagonal frequency shift applied to each harmonic column of the signal. As shown in Figure 5, the solver state matrix ADQsymA_{\mathrm{DQsym}} is diagonal, and each row of the input matrix BDQsymB_{\mathrm{DQsym}} contains a single nonzero entry per input group, located in the same sequence channel as the state. Each sequence current is therefore driven only by the inputs of its own sequence, and no coupling between sequences or between harmonics appears in this part of the model.

Figure 5. Sparsity patterns of the DQsym state matrix ADQsymA_{\mathrm{DQsym}} and BDQsymB_{\mathrm{DQsym}} in pnz coordinates. Dots are nonzero entries.

The second part is the arm voltage calculation: the open-loop insertion indices {(xref)}(6) are used directly to compute the capacitor voltages {(xref)}(4c)–(7) and the modulated arm voltages {(xref)}(3),

v⃗mu=m⃗u⊙v⃗Cu,v⃗ml=m⃗l⊙v⃗Cl,(29)\vec{v}^{u}_{m}=\vec{m}^{u}\odot\vec{v}^{u}_{C},\qquad\vec{v}^{l}_{m}=\vec{m}^{l}\odot\vec{v}^{l}_{C}, \tag*{(29)}

which are transformed with {(xref)}(17) and fed to {(xref)}(16) as the inputs v⃗u,pnzu\vec{v}^{u}_{u,pnz} and v⃗pnzl\vec{v}^{l}_{pnz}. All harmonic coupling and nonlinearity are therefore confined to the arm voltage calculation, while the arm current dynamics remain diagonal and sparse.

Results

All three HD models are evaluated on the same open-loop MMC of subsection 3.1, with no circulating current control or energy control. The open-loop MMC has an AC side with a star-point reactor, as shown in Table 2, and is used to investigate three forms of a single model. DP and HSS are truncated at K=4K=4 in their state matrices. DQsym uses the same truncation, K=4K=4, but applies it to the signals. Each signal carries harmonics 0→K0\to K, and the truncation acts in the multiplication and integration blocks rather than in the solver matrix. The averaged arm model integrated directly in the abcabc frame (“abc ODE”) serves as the reference. All matrices in Figure 2 to Figure 4 are shown in the same abcabc harmonic basis, so the structural differences reflect only where the modulation-capacitor coupling is placed, not the choice of coordinates.

Matrix structure and sparsity

Table 3 summarises the sizes and sparsity. The DQsym solver holds only the arm and neutral currents (iui_u, ili_l, ini_n), giving 9 states per harmonic, 45 in total, instead of 135. The capacitor-voltage rows are evaluated outside the solver by the multiplication and integration blocks. Therefore, ADQsymA_{\mathrm{DQsym}} contains only the per-harmonic diagonal blocks of the arm RLRL branches, and it has no off-diagonal bands. In ADPA_{\mathrm{DP}} and AHSS−NA_{\mathrm{HSS}-N}, the modulation spectrum couples each arm current to its capacitor voltage through banded Toeplitz blocks (iu↔vCui_u \leftrightarrow v_{Cu}, il↔vCli_l \leftrightarrow v_{Cl}). ADPA_{\mathrm{DP}} has the same pattern with each complex entry expanded into a 2×22 \times2 real block.

DQsymDQsym†DPHSS
States / inputs9 / 1545 / 75135 / 99135 / 99
nnz⁡(A)\operatorname{nnz}(A) / nnz⁡(B)\operatorname{nnz}(B)9 / 2145 / 105699 / 1059483 / 1107
Density AA / BB (%)11.1 / 15.62.2 / 3.13.8 / 7.92.65 / 8.3
Total nnz3015017581590
State typecomplex dqcomplex dqrealcomplex

Table 3. State-space dimensions and sparsity for K=4K = 4. †Solver over all K+1K+1 harmonics; excludes capacitor states and harmonic coupling, evaluated outside the solver.

The input matrices differ in the same way. In BHSSB_{\mathrm{HSS}} and BDPB_{\mathrm{DP}}, the modulation inputs m⃗u\vec{m}_u, m⃗l\vec{m}_l enter as dense (2K+1)×(2K+1)(2K+1) \times(2K+1) blocks in both the current and the capacitor-voltage rows, and linearising the products m⃗vC\vec{m}v_C and m⃗i\vec{m}i uses the operating point of vCv_C and ii at the periodic steady state. The DQsym solver instead takes the arm voltages as inputs, which map onto the current rows harmonic by harmonic. Its input vector has 15 entries per harmonic, 75 in total, instead of 99.

Steady state

Figure 6 shows one fundamental period of the phase-aa upper and lower arm currents and arm capacitor voltages in periodic steady state. All three HD formulations overlap closely with the reference abcabc ODE model. The capacitor voltages ripple between about 608 and 673 kV around the 640 kV mean, and the upper-arm current reaches its minimum of about −660-660 A at t=10t = 10 ms in all models.

Periodic steady-state waveforms of phase-a upper and lower arm currents and arm capacitor voltages

Figure 6. Periodic steady state of the open-loop MMC over one fundamental period: phase-aa upper and lower arm currents and arm capacitor voltages. Reference abcabc ODE compared with DQsym, DP, and HSS.

Figure 7 extends the comparison to three fundamental periods and shows the error of each model with respect to the reference. The three formulations achieve comparable accuracy: for all three, the arm-current error remains below about 4% of the peak arm current, and the capacitor-voltage error remains below about 0.2% of the mean capacitor voltage. DQsym shows slightly larger peaks than DP and HSS (about 28 A versus 20 A in iu,ai_{u,a}), but its error is of the same order of magnitude and does not grow over the simulated interval. This accuracy is obtained with the diagonal DQsym solver of Figure 5, with the harmonic coupling evaluated outside it.

Figure 7. Response of the open-loop MMC over three fundamental periods: phase-aa upper and lower arm currents and arm capacitor voltages (top), and error of each model with respect to the reference abc ODE model (bottom).

Harmonic content

Figure 8 and Table 4 compare the harmonic magnitudes of the phase-aa upper arm current and capacitor voltage in the periodic steady state of Figure 6, with the abcabc ODE as reference. DP and HSS give identical spectra at every order.

QuantityModel01234
∣iu,a∣|i_{u,a}| [A]abc ODE37.089.6411.8162.140.7
DQsym38.093.4417.0164.240.4
DP / HSS37.390.2411.7161.338.9
∣vCu,a∣|v_{Cu,a}| [kV]abc ODE639.614.7016.477.521.79
DQsym640.014.9016.697.611.76
DP / HSS639.614.7216.467.491.71

Table 4. Harmonic magnitudes of ∣iu,a∣|i_{u,a}| and ∣vCu,a∣|v_{Cu,a}| for h=0,…,4h = 0,\ldots,4.

Figure 8. Harmonic magnitudes of the phase-aa upper arm current and upper capacitor voltage in periodic steady state (logarithmic scale).

All three formulations reproduce the reference closely, both for the dominant components, such as the 2nd harmonic of the arm current (about 410 A) and the dc capacitor voltage (about 640 kV), and for the smallest ones at h=4h = 4 (about 40 A and 1.8 kV). The largest deviation is 4.2% for DQsym in the 1st harmonic of iu,ai_{u,a}, and 4.5% for DP and HSS in the 4th harmonic of vCu,av_{Cu,a}. DP and HSS are within 0.8% of the reference up to h=3h = 3 and deviate only at the truncation order h=K=4h = K = 4. DQsym deviates up to 4.2% more uniformly across the orders, but is closer to the reference at h=4h = 4, −0.7%-0.7\% and −1.7%-1.7\%, against −4.5%-4.5\% for DP and HSS.

Eigenvalues

DQsym evaluates varm=m⃗vCv_{\mathrm{arm}} = \vec{m}v_C and the capacitor integration outside the solver. Writing this part as x˙C=ACxC+BCi\dot{x}_C = A_C x_C + B_C i, varm=CCxCv_{\mathrm{arm}} = C_C x_C, where BvB_v maps the arm voltages into the solver and CRLC_{RL} extracts the arm currents, and closing it around the solver (ARL,Bv,CRL)(A_{RL}, B_v, C_{RL}) gives

Acl=[ARLBvCCBCCRLAC],(30)A_{\mathrm{cl}} = \begin{bmatrix} A_{RL} & B_v C_C \\ B_C C_{RL} & A_C \end{bmatrix}, \tag*{(30)}

where ARL≡ADQsymA_{RL} \equiv A_{\mathrm{DQsym}}. AclA_{\mathrm{cl}} is similar to the linearised abc ODE model, with BvCCB_v C_C and BCCRLB_C C_{RL} as the modulation couplings. The loop has no algebraic path, and the Park and sequence transformations act as a similarity transformation, so the eigenvalues of AclA_{\mathrm{cl}} can be compared directly with DP and HSS.

The eigenvalues of the three models are identical, as seen in Figure 9(a), with only three real parts, σ=0\sigma= 0, −6-6 and −10.7 s−1-10.7\ \mathrm{s}^{-1}, and repeat at shifts of ±hfg\pm h f_g as these copies stem from the harmonic formulation, so the distinct modes are those in ∣f∣<fg/2\lvert f\rvert< f_g/2 with fg=ω0/2πf_g = \omega_0/2\pi.

Figure 9. Eigenvalues of the open-loop MMC for K=4K=4: (a) DP, HSS and DQsym (AclA_{\mathrm{cl}}); (b) DQsym solver alone and closed loop. The shaded band marks ∣f∣<fg/2|f| < f_g/2.

The solver alone, Figure 9(b), has real eigenvalues at −Rarm/Larm=−21.4-R_{arm}/L_{arm} = -21.4, −3Rstar/Lstar=−6-3R_{star}/L_{star} = -6 and 0 s−10\ \mathrm{s}^{-1}, and repeated at ±hfg\pm h f_g. The last two belong to the star branch, where the shared RstarR_{star} damps only the zero sequence, and they do not involve the capacitors and are unchanged in the closed loop. The arm pole moves to −Rarm/(2Larm)=−10.7 s−1-R_{arm}/(2L_{arm}) = -10.7\ \mathrm{s}^{-1} and becomes oscillatory, as the capacitors turn each first-order RLRL arm into an underdamped second-order one. All twelve arm modes take this value, showing that the arm resonance exists only after adding the capacitor dynamics.

Conclusion

This paper compared three harmonic-domain formulations of the same MMC model: dynamic phasors (DP), harmonic state space (HSS) and DQsym. They differ only in where the harmonic coupling between the modulation and the arm capacitor voltages is placed. HSS and DP store this coupling in the state matrix, whereas DQsym evaluates it outside the solver through multiplication and integration blocks. Closing the DQsym solver with the capacitor dynamics recovers the same linear time-periodic system, as confirmed by the identical eigenvalues of all three formulations.

The formulations also differ in size and structure. For a truncation order K=4K = 4, HSS and DP both have 135 states, and their input matrices are dominated by full (2K+1)×(2K+1)(2K+1) \times(2K+1) blocks, so the number of nonzeros grows quadratically with KK (1590 for HSS and 1758 for DP). The DQsym solver contains only the RLRL network in sequence coordinates. It is diagonal with 9 states (45 for K=4K = 4), each a complex signal of width K+1K+1, so it stays 9 for any KK. The harmonic coupling is not removed but moved into the multiplication and integration blocks. All three reproduce the steady-state harmonic content of the time-domain model up to h=4h = 4. HSS is the natural choice for analytical stability and impedance studies, DP for real-valued time-domain simulation, and DQsym when modularity and a KK-independent solver structure matter.

Because the solver alone does not contain the capacitor dynamics, stability analysis with DQsym must linearise the solver together with its multiplication and integration blocks, as done here. Future work will extend the comparison to closed-loop operation with circulating-current suppression and to asymmetric operating points, use the DQsym linearisation for small-signal stability analysis, and present a systematic procedure for assembling the DQsym system matrices.

References

References

  1. [1]Pallavi Ghimire, Samip Poudel, Mariko Shirazi, Donald M. Hummels, and Reinaldo Tonkoski. Application of dynamic phasor modeling in power systems: A review. IEEE Access, 14:28573–28594, 2026.DOI
  2. [2]Wilsun Xu, Roberto Langella, Antonio Bracale, Yuanyuan Sun, Kuo Lung Lian, Yang Wang, and Jason David. Modeling of inverter-based resources for power system harmonics studies. IEEE Transactions on Power Delivery, 40(1):166–177, 2025.DOI
  3. [3]Seth R. Sanders, J. M. Noworolski, X. Z. Liu, and George C. Verghese. Generalized averaging method for power conversion circuits. In Proc. 21st Annual IEEE Conference on Power Electronics Specialists, pages 333–340, 1990.
  4. [4]Philippe De Rua, Özgür Can Sakıncı, and Jef Beerten. Comparative study of dynamic phasor and harmonic state-space modeling for small-signal stability analysis. Electric Power Systems Research, 189:106626, 2020.DOI
  5. [5]J. B. Kwon, X. Wang, F. Blaabjerg, C. L. Bak, A. R. Wood, and N. R. Watson. Harmonic instability analysis of a single-phase grid-connected converter using a harmonic state-space modeling method. IEEE Trans. Ind. Appl., 52(5):4188–4200, 2016.DOI
  6. [6]Nicolas Blin, Pierre Riedinger, Jamal Daafouz, Louis Grimaud, and Philippe Feyel. A comparison of harmonic modeling methods with application to the interconnection and the control of switched systems. European Journal of Control, 58:245–257, March 2021.DOI
  7. [7]Jose Daniel Lara, Rodrigo Henriquez-Auba, Deepak Ramasubramanian, Sairaj Dhople, Duncan S. Callaway, and Seth Sanders. Revisiting power systems time-domain simulation methods and models. IEEE Transactions on Power Systems, 39(2):2421–2437, 2024.
  8. [8]Johanna Kristin Maria Becker, Andreas Martin Kettner, and Mario Paolone. Harmonic stability analysis of microgrids with converter-interfaced distributed energy resources, Part I: Modeling and theoretical foundations. IEEE Transactions on Smart Grid, 16(3):2316–2326, 2025.
  9. [9]Aliakbar Jhamshidifar and Dragan Jovcic. Small-signal dynamic DQ model of modular multilevel converter for system studies. IEEE Trans. Power Del., 31(1):191–199, 2016.
  10. [10]Udoka C. Nwaneto, Seyed Ali Seif Kashani, and Andrew M. Knight. An Equivalent Dynamic Phasor Model for a Single-Phase Boost Power-Factor-Correction Converter. IEEE Open Journal of Power Electronics, 6:1002–1021, 2025.DOI
  11. [11]Udoka C. Nwaneto and Andrew M. Knight. Full-Order and Simplified Dynamic Phasor Models of a Single-Phase Two-Stage Grid-Connected PV System. IEEE Access, 11:26712–26728, 2023.DOI
  12. [12]Wonbae Choi. Dynamic phasor modeling of type 3 wind turbine generators for large-scale power system transient stability studies. PhD thesis, University of British Columbia, 2017.DOI
  13. [13]Saif Alsarayreh, Robert Dimitrovski, and Aleksandra Lekić. Dynamic phasor state-space solver for hybrid AC/DC systems. Electric Power Systems Research, 262:113547, January 2027.DOI
  14. [14]Saif Alsarayreh, Robert Dimitrovski, and Aleksandra Lekić. A novel dynamic phasor-based mathematical framework for hybrid AC/DC power system simulation. IEEE Open Access Journal of Power and Energy, 13:651–663, 2026.
  15. [15]Saif Alsarayreh, Robert Dimitrovski, and Aleksandra Lekić. DQsym: A MATLAB/Simulink library for dynamic phasor simulation of AC/DC power systems. SoftwareX, 34:102719, 2026.DOI

Paper details

vlnc:W_ua3rrwyu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 5

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