betaFeedback

WorksEngineering

Data-driven current control of modular multilevel converter-based HVDC transmission system

Abstract

Model-based design (MBD) is a mature framework for designing the current controllers of modular multilevel converters (MMCs) in high-voltage direct current (HVDC) transmission systems. In MBD, linear state-space models are preferred for their good trade-off between accurate dynamics prediction and simplicity. However, first-principle derivations often neglect relevant MMC dynamics, which can hinder controller performance. Data-driven (DD) methods can deliver less biased models that capture parameter uncertainties, control-loop interactions, and grid dynamics, yet they remain under-exploited for controlling MMC currents. This paper studies DD current control method based on integral Subspace Predictive Control (iSPC), synthesized directly from input–output data measured in a high fidelity real-time digital simulator (RTDS), and proposes an initial guideline outlining the main steps for its systematic application to MMC current control. The proposed iSPC is compared against state-of-the-art DD solution, and the findings are validated through simulation studies in MATLAB/Simulink.

Introduction

Nowadays, high-voltage power transmission is increasingly based on direct current (DC) technology [1]. In high-voltage direct current (HVDC) systems, modular multilevel converters (MMCs) have become the preferred solution for power conversion owing to their high efficiency, scalability, robustness, and low harmonic distortion [2]. However, the performance and safe operation of MMCs strongly depend on the design of their control systems, particularly the current controllers [3].

A mature framework for designing the current controllers of MMC is model-based design (MBD), where mathematical models and control theory are exploited to systematically synthesize controllers that offer desired closed-loop properties, such as stability, and robustness [4]. Among the examples of MBD, we can highlight optimal control [5], robust control [6], and model predictive control (MPC) [7]. Here, linear state-space (SS) models are the preferred model structure due to their good trade-off between accurate dynamics prediction and computational complexity. These features are particularly important in MPC, where prediction models are used to compute optimized control actions while explicitly considering converter constraints and future system behavior [8]. As a consequence, MPC has demonstrated the ability to enhance MMC performance, including efficiency and power quality, compared with other control methods [9].

Despite these advantages, the effectiveness of MBD strongly depends on the accuracy of the underlying model. When using first-principle SS (fSS) modeling, models rely on small-signal theory and often neglect practical effects such as digital delays, filtering, and interactions between control layers (including phase-locked loop), which may degrade performance or even induce instability [10]. These shortcomings have motivated a shift toward data-driven (DD) modeling, where frequency scanning is often used to generate input–output (I/O) data that allows to build less biased model, see e.g. [10, 11]. These models have been successfully applied to enhance stability and interaction studies; however, these advantages have not yet been leveraged in the design of MMC current controllers.

DD control is a promising alternative to MBD, synthesizing controllers from measured I/O data. Prominent examples, including data-driven LQR [12], Subspace Predictive Control (SPC) [13], and Data-Enabled Predictive Control (DeePC) [14], have been applied in power-system and power-electronics applications [15, 16, 17], yet remain unexplored for MMC current control. Their systematic implementation in this setting is still an open problem, as the closed-loop performance hinges on design choices with no counterpart in MBD [18], namely the system identification (SI) experiment design (excitation signal, amplitude, and data length), the satisfaction of persistency-of-excitation conditions [19], and the choice of horizons and regularization. In practice, they are still selected ad hoc due to the complexity of systematically applying traditional SI procedures [20], showing the need for design guidelines.

Hence, this paper proposes using DD, i.e., an integral SPC (iSPC), for current control of MMCs in HVDC systems. We first design the SI experiment to obtain informative I/O data measured in a high fidelity real-time digital simulator (RTDS®), which capture the commonly neglected dynamics. Using these data, the iSPC is designed and compared with DeePC, showing more accurate tracking and lower sensitivity to noise at a lower computation time. Based on extensive MATLAB/Simulink simulations, we further propose guidelines for tuning the iSPC, covering the selection of the prediction horizon, the data length, and excitation signal type.

MMC-based HVDC Fundamentals

Figure 1 shows the topology of a point-to-point two-terminal MMC-based HVDC link, emulating an offshore wind farm connected to an onshore AC grid. The onshore MMC1_1 controls the power flowing to the AC grid, while the offshore MMC2_2 regulates the DC voltage of the DC link. Three-level hierarchical control architectures (HCAs) are often used to control MMC [21]. At the highest level, the outer control loop regulates the active/reactive AC power or DC voltage, depending on the MMC operation mode. This control layer outputs are the current references used in the current control loop (CCL), i.e., idq∗=(idqo∗,idqc∗)i_{dq}^{*}=(i_{dq}^{o*},i_{dq}^{c*}) where the circulating is set to zero idqc∗=0i_{dq}^{c*}=0. Then, the CCL computes arm voltage references (udqo∗,udqc∗)(u_{dq}^{o*},u_{dq}^{c*}) that regulate MMC currents. At the lowest level, (udqo∗,udqc∗)(u_{dq}^{o*},u_{dq}^{c*}) are converted into modulation indices and switching signals for the submodule gates.

Figure 1. Schema of point-to-point two terminals MMC-based HVDC link including control and SI stages.

Commissioning MMC controllers is often a complex process. Besides the practical effects, MMC controllers must handle nonlinear dynamics arising from internal arm energy, modulation, and interactions between control layers. Linear fSS models neglect these dynamics by taking the arm voltages as control inputs and assuming that they equal their references (udqo∗,udqc∗)(u_{dq}^{o*},u_{dq}^{c*}) [21]. Hence, controller performance often relies on ad-hoc fine-tuning at later stages.

MMC Current Dynamics

MMCs are comprised of 3 legs with two arms, in which NN submodules (SMs) are connected in series to an equivalent resistor (RmR_m) and inductor (LmL_m), see Figure 2. The arms are named as upper and lower arm, denoted by n∈{u,l}n \in\{u,l\}, and the phases are denoted by m∈{a,b,c}m \in\{a,b,c\}. Then, Kirchhoff’s voltage law shows that MMC currents are governed by the transformer (TF) (VmrV_m^r),

Figure 2. Simplified diagram of the MMC topology per-phase m∈{a,b,c}m \in\{a,b,c\}: (a) MMC arms; (b) Half-bridge submodule.

DC-link (VdcV_{dc}), and arm voltages (umu,lu_m^{u,l}), such that:

Lmdimudt=−Rmimu+umu+Vmr+Lrdimodt+Rrimo−Vdc2,(1)L_m\frac{d i_m^u}{d t}=-R_m i_m^u+u_m^u+V_m^r+L_r\frac{d i_m^o}{d t}+R_r i_m^o-\frac{V_{dc}}{2}, \tag*{(1)}
Lmdimldt=−Rmiml+uml−Vmr−Lrdimodt−Rrimo−Vdc2,(2)L_m\frac{d i_m^l}{d t}=-R_m i_m^l+u_m^l-V_m^r-L_r\frac{d i_m^o}{d t}-R_r i_m^o-\frac{V_{dc}}{2}, \tag*{(2)}

where LmL_m and LrL_r are the arm and transformer inductance, respectively; and RmR_m and RrR_r are the arm and transformer resistance, respectively. Per phase mm, the MMC currents determine the power transfer through the output current, i.e., imo=imu−imli_m^o=i_m^u-i_m^l and circulating current, i.e., imc=12(imu+iml)i_m^c=\frac{1}{2}(i_m^u+i_m^l). Related to these currents, there are the differential voltage, i.e., umo=umu−umlu_m^o=u_m^u-u_m^l and common voltage, i.e., umc=12(umu+uml)u_m^c=\frac{1}{2}(u_m^u+u_m^l).

MMC Model: First-principle state-space derivation

Assuming an average equivalent representation of the MMC as in [8] and using the modeling procedures in [7], the dynamics of (imo,imc)(i_m^o,i_m^c) in (1) are formulated as discrete fSS models* with sampling time TsT_s where the state, control input and output vectors are defined as x:=[ido,iqo,idc,iqc]Tx := [i^{o}_{d}, i^{o}_{q}, i^{c}_{d}, i^{c}_{q}]^{\mathsf{T}}, u=[udo−Vdr/2,uqo−Vqr/2,udc,uqc]Tu = [u^{o}_{d} - V_{d}^{r}/2, u^{o}_{q} - V_{q}^{r}/2, u^{c}_{d}, u^{c}_{q}]^{\mathsf{T}}, and y=xy = x, respectively. Then, its dynamics yield:

x+=Ax+Bu,(3)x^{+} = Ax + Bu, \tag*{(3)}

with A=eAcTs\mathbf{A}=e^{\mathbf{A}_{c}T_{s}} and B=∫0TseAcτBcdτ\mathbf{B}=\int_{0}^{T_{s}}e^{\mathbf{A}_{c}\tau}\mathbf{B}_{c}\mathrm{d}\tau are state-space matrices, such that Ac=diag⁡(Aco,Acc)\mathbf{A}_{c}=\operatorname{diag}(\mathbf{A}^{o}_{c},\mathbf{A}^{c}_{c}) and Bc=diag⁡(Bco,Bcc)\mathbf{B}_{c}=\operatorname{diag}(\mathbf{B}^{o}_{c},\mathbf{B}^{c}_{c}), where

Aco=[−ReqLeq−ωω−ReqLeq],Acc=[−RmLm2ω−2ω−RmLm],(4)\mathbf{A}^{o}_{c} = \begin{bmatrix} -\frac{R_{eq}}{L_{eq}} & -\omega\\ \omega& -\frac{R_{eq}}{L_{eq}} \end{bmatrix}, \qquad \mathbf{A}^{c}_{c} = \begin{bmatrix} -\frac{R_{m}}{L_{m}} & 2\omega\\ -2\omega& -\frac{R_{m}}{L_{m}} \end{bmatrix}, \tag*{(4)}

Bco=diag⁡(1Leq,1Leq)\mathbf{B}^{o}_{c}=\operatorname{diag}\left(\frac{1}{L_{eq}},\frac{1}{L_{eq}}\right), Bcc=diag⁡(−1Lm,−1Lm)\mathbf{B}^{c}_{c}=\operatorname{diag}\left(-\frac{1}{L_{m}},-\frac{1}{L_{m}}\right), Leq=Lr+Lm2L_{eq}=L_{r}+\frac{L_{m}}{2}, Req=Rr+Rm2R_{eq}=R_{r}+\frac{R_{m}}{2} and ω=2πf\omega=2\pi f. Operator diag⁡(⋅)\operatorname{diag}(\cdot) defines a diagonal matrix where the diagonal elements are its argument.

The augmented model is derived to enforce zero-error tracking when following step references. Discrete-time derivation is denoted by the operator δ⋅\delta\cdot, such that δx(k):=x(k)−x(k−1)\delta x(k):=x(k)-x(k-1). The augmented state vector is defined as x⃗=[δxT yT]T\vec{x}=[\delta x^{\mathsf{T}}\ y^{\mathsf{T}}]^{\mathsf{T}}, and an input rate vector, i.e, δu(k)=u(k)−u(k−1)\delta u(k)=u(k)-u(k-1). Then, (2) augmented model yields

x⃗+=A⃗x⃗+B⃗δu,y=C⃗x⃗,(5)\vec{x}^{+}=\vec{A}\vec{x}+\vec{B}\delta u,\qquad y=\vec{C}\vec{x}, \tag*{(5)}

where A⃗=[A04CAI4]\vec{A}=\begin{bmatrix} A & 0_{4}\\ CA & I_{4}\end{bmatrix}, B⃗=[BCB]\vec{B}=\begin{bmatrix} B\\ C B\end{bmatrix}, C⃗=[04 I4]\vec{C}=[0_{4}\ I_{4}] and C=I4C=I_{4}. Notation I4I_{4} indicates an identity matrix of dimension 4, and 040_{4} indicates a square matrix of dimension 4 filled with zeros.

Data-driven predictive control fundamentals

MBD-MPC computes control actions, which in this case are δu\delta u, via solving at each discrete time step kk an optimal control problem (OCP), that can be represented as:

min⁡δUkJ(y(0),δUk)(6)\min_{\delta U_{k}} J(y(0),\delta U_{k}) \tag*{(6)}
s.t.x⃗h+1∣k=A⃗x⃗h∣k+B⃗δuh∣k,y=C⃗x⃗h∣k,(7)\text{s.t.}\quad\vec{x}_{h+1|k}=\vec{A}\vec{x}_{h|k}+\vec{B}\delta u_{h|k},\quad y=\vec{C}\vec{x}_{h|k}, \tag*{(7)}
ymin⁡≤y≤ymax⁡,umin⁡≤u≤umax⁡,(8)y_{\min}\le y\le y_{\max},\quad u_{\min}\le u\le u_{\max}, \tag*{(8)}

with a cost function

J(⋅)=eN∣kTQeN∣k+∑h=0N−1eh∣kTQeh∣k+δuh∣kTRδuh∣k,(9)J(\cdot)=e_{N|k}^{\mathsf{T}}Qe_{N|k}+\sum_{h=0}^{N-1}e_{h|k}^{\mathsf{T}}Qe_{h|k}+\delta u_{h|k}^{\mathsf{T}}R\delta u_{h|k}, \tag*{(9)}

where eh∣k=yh∣k−r(k+h)e_{h|k}=y_{h|k}-r(k+h) is the predicted error1, NN is the prediction horizon, (Q,R)(Q,R) are weighting matrices, (x⃗min⁡,x⃗max⁡)(\vec{x}_{\min},\vec{x}_{\max}) are state limits, (umin⁡,umax⁡)(u_{\min},u_{\max}) are input limits.

DD predictive control (DPC) follows the MBD-MPC principle, except that the prediction step, i.e., (5b), relies on an identified model or directly on measured I/O data rather than on the fSS.

On the fSS. The I/O data are collected through an SI experiment, in which an excitation signal applied to the arm voltages forms the input sequence u~\tilde{u} and the measured currents (idqo,idqc)(i^{o}_{dq},i^{c}_{dq}) form the output sequence y~\tilde{y} (see Figure 1). DPC builds on two key concepts: Hankel matrix and persistence of excitation. A Hankel matrix arranges a data sequence (y~,u~)(\tilde{y},\tilde{u}) in a fixed block structure; following [18], it is defined as

Definition 3.1. A Hankel matrix constructed using an arbitrary dataset zz, with depth NN, length TT and starting index ii is denoted as:

H[N,T]i(z)=[z(i)z(i+1)⋯z(i+T−1)z(i+1)z(i+2)⋯z(i+T)⋮⋮⋱⋮z(i+N−1)z(i+N)⋯z(i+T+N−2)]\mathcal{H}_{[N,T]}^{i}(z)= \begin{bmatrix} z(i) & z(i+1) & \cdots& z(i+T-1)\\ z(i+1) & z(i+2) & \cdots& z(i+T)\\ \vdots& \vdots& \ddots& \vdots\\ z(i+N-1) & z(i+N) & \cdots& z(i+T+N-2) \end{bmatrix}

where the samples in zz can comprise information incoming from multiple channels (e.g., multiple inputs or outputs), i.e., z(i)=[z1(i) z2(i) ⋯ zn(i)]Tz(i)=[z_{1}(i)\ z_{2}(i)\ \cdots\ z_{n}(i)]^{\mathsf{T}}.

The I/O data collected during SI, i.e., (y~,u~)(\tilde{y},\tilde{u}), are used to construct the following Hankel matrices:

Up=H[Tini,T]0(u~),Yp=H[Tini,T]0(y~),U_{p}=\mathcal{H}_{[T_{\mathrm{ini}},T]}^{0}(\tilde{u}),\qquad Y_{p}=\mathcal{H}_{[T_{\mathrm{ini}},T]}^{0}(\tilde{y}),
Uf=H[N,Tini]Tini(u~),Yf=H[N,Tini]Tini(y~),(10)U_{f}=\mathcal{H}_{[N,T_{\mathrm{ini}}]}^{T_{\mathrm{ini}}}(\tilde{u}),\qquad Y_{f}=\mathcal{H}_{[N,T_{\mathrm{ini}}]}^{T_{\mathrm{ini}}}(\tilde{y}), \tag*{(10)}

where TiniT_{\mathrm{ini}} is the "past horizon". These matrices are at the core of DPCs, namely SPC, DeePC, γ\gamma-DDPC and GDPC [18].

Remark 3.1. The past horizon TiniT_{\mathrm{ini}} is related to the system order nn and should satisfy Tini≥nT_{\mathrm{ini}}\ge n; for the current dynamics, n=8n=8, hence Tini≥8T_{\mathrm{ini}}\ge8. To ensure that the data are informative, the Hankel matrix length must satisfy T≥nu(Tini+N+n)T\ge n_{u}(T_{\mathrm{ini}}+N+n), where nu=4n_{u}=4 is the input dimension. Then, we define the data length as Tdata=Tini+T+NT_{\mathrm{data}}=T_{\mathrm{ini}}+T+N, such that (y~,u~)={((y~(0),u~(0)),…,(y~(Tdata),u~(Tdata))}(\tilde{y},\tilde{u})=\{((\tilde{y}(0),\tilde{u}(0)),\ldots,(\tilde{y}(T_{\mathrm{data}}),\tilde{u}(T_{\mathrm{data}}))\}.

Additionally, informativity can be ensured if u~\tilde{u} is persistently exciting, defined as

Definition 3.2 (Persistence of Excitation). The input data sequence u~\tilde{u} is called persistently exciting of order LL if the rank of the Hankel matrix satisfies

rank⁡(H[L,T]0(u~))=Lnu(11)\operatorname{rank}\left(\mathcal{H}_{[L,T]}^{0}(\tilde{u})\right)=Ln_{u} \tag*{(11)}

where nun_{u} is the dimension of the control inputs.

Integral Subspace Predictive Control (iSPC)

iSPC is an indirect DPC [17] that estimates, from I/O data, a prediction matrix Θ\Theta that computes NN-step-ahead, such that

Yk=[P1P2Γ]⏟=:Θ[uini(k)yini(k)δUk],(12)Y_{k}= \underbrace{\begin{bmatrix}P_{1}&P_{2}&\Gamma\end{bmatrix}}_{=:\Theta} \begin{bmatrix} u_{\mathrm{ini}}(k)\\ y_{\mathrm{ini}}(k)\\ \delta U_{k} \end{bmatrix}, \tag*{(12)}

where Yk=[y1∣kT,…,yN∣kT]TY_{k}=[y_{1|k}^{\mathsf{T}},\ldots,y_{N|k}^{\mathsf{T}}]^{\mathsf{T}}, δUk=[δu0∣kT,…,δuN−1∣kT]T\delta U_{k}=[\delta u_{0|k}^{\mathsf{T}},\ldots,\delta u_{N-1|k}^{\mathsf{T}}]^{\mathsf{T}}, uini(k)=[δuT(k−Tini),…,δuT(k−1)]Tu_{\mathrm{ini}}(k)=[\delta u^{\mathsf{T}}(k-T_{\mathrm{ini}}),\ldots,\delta u^{\mathsf{T}}(k-1)]^{\mathsf{T}} and $y_{\mathrm{ini}}(k)=[y^{\mathsf{T}}(k-T_{\mathrm{ini}}),\ldots, yT(k−1)]y^{\mathsf{T}}(k-1)]. According to (12) future outputs YkY_k are described as a linear combination of previous I/O data and δUk\delta U_k. By relying on this relation, the estimation of the unknown matrix Θ\Theta can be carried out by solving the least-squares problem:

min⁡Θ∥Yf−Θ[UpYpUf]∥Fro2(13)\min_{\Theta}\left\|Y_f-\Theta\begin{bmatrix} U_p \\ Y_p \\ U_f \end{bmatrix}\right\|_{\mathrm{Fro}}^2 \tag*{(13)}

with solution Θ∗\Theta^*, i.e.,

Θ∗=[P1P2Γ^]=Yf[UpYpUf]†(14)\Theta^*=\begin{bmatrix} P_1 & P_2 & \hat{\Gamma} \end{bmatrix}=Y_f\begin{bmatrix} U_p \\ Y_p \\ U_f \end{bmatrix}^{\dagger} \tag*{(14)}

where Γ^\hat{\Gamma} is the DD estimation of Γ\Gamma and †\dagger indicates the pseudoinverse matrix. Using (14), iSPC yields

min⁡δUkJ(y(0),δUk)(15)\min_{\delta U_k} J(y(0),\delta U_k) \tag*{(15)}
s.t.Yk=P1uini(k)+P2yini(k)+Γ^δUk,(16)\text{s.t.}\quad Y_k=P_1u_{\mathrm{ini}}(k)+P_2y_{\mathrm{ini}}(k)+\hat{\Gamma}\delta U_k, \tag*{(16)}
ymin⁡≤y≤ymax⁡,umin⁡≤u≤umax⁡,(17)y_{\min}\le y\le y_{\max},\quad u_{\min}\le u\le u_{\max}, \tag*{(17)}

where the cost function J(y(0),δUk)J(y(0),\delta U_k) stays as in (6). The iSPC implementation is summarized by the following algorithm.

Algorithm 1 SPC algorithm

Require: Up,Yp,Uf,Yf,(N,Tini,Q,R,r)U_p,Y_p,U_f,Y_f,(N,T_{\mathrm{ini}},Q,R,r)

Before closing the control loop:

  1. Obtain Θ∗\Theta^* from (14) and extract P1P_1, P2P_2 and Γ^\hat{\Gamma}.

  2. Construct controller as in (12).

During the control loop:

  1. Measure y(k)y(k) and construct uini(k)u_{\mathrm{ini}}(k) and yini(k)y_{\mathrm{ini}}(k).

  2. Solve control law (12) and obtain δUk∗\delta U_k^*.

  3. Apply u(k)=u(k−1)+δu0∣k∗u(k)=u(k-1)+\delta u^*_{0|k} to the system.

Case Study

The iSPC performance is studied in a point-to-point, two-terminal MMC-based HVDC link, i.e., CIGRE DCS1 RTDS® benchmark [22], rated at 0.8 GVA and 50 Hz, with a DC-link voltage of 400 kV. On the AC side, both MMCs operate at 220 kV and are interfaced through TFs to the 380 kV onshore and 145 kV offshore grids; the TFs have a resistance of 0.006 pu and an inductance of 0.18 pu. Each MMCs, modeled with a type 5 average value model [22], have 200 SMs per arm with a 10 mF capacitance, Rm=0.15 ΩR_m=0.15\ \Omega and Lm=29L_m=29 mH. CIGRE DCS1 benchmark [22] is a high fidelity model that includes SM-level dynamics, practical effects, and a three-level HCA, detailed in [22 Ch. 3]. iSPC is proposed to replace only the CCL PI controllers of the control architecture.

System identification and validation

Figure 3 presents the excitation input data u~\tilde{u}, i.e, a random binary sequence (RBS) superimposed on white noise:

Figure 3. Input data sequence, i.e., u~(t)={u~do,u~qo,u~dc,u~qc}\tilde{u}(t)=\{\tilde{u}^{o}_{d},\tilde{u}^{o}_{q},\tilde{u}^{c}_{d},\tilde{u}^{c}_{q}\}.

u~j(k)={−α+εj(k),if ηj(k)≤0.5,α+εj(k),if ηj(k)>0.5,j∈{1,…,4},(18)\tilde{u}_j(k)= \begin{cases} -\alpha+\varepsilon_j(k), & \text{if }\eta_j(k)\le0.5,\\ \alpha+\varepsilon_j(k), & \text{if }\eta_j(k)>0.5, \end{cases} \qquad j\in\{1,\ldots,4\}, \tag*{(18)}

where α\alpha is the RBS amplitude, ηj(k)∼U(0,1)\eta_j(k)\sim U(0,1) is a uniformly distributed random number whose seed depends on channel jj, and εj(k)∼N(0,σ2)\varepsilon_j(k)\sim\mathcal{N}(0,\sigma^2) is zero-mean white noise. Whereas multisine signals are common in the literature, e.g., for frequency scans, the RBS concentrates its energy at low frequencies and captures the steady-state behavior, while the added white noise reduces the correlation between samples and improves the data PE. Choosing an α\alpha large enough to excite the system without violating operating constraints or driving it toward instability is challenging. We recommend α\alpha between 2% and 6% of the steady-state value; larger submodule capacitance and arm resistance allow amplitudes toward the upper end of this range. For εj\varepsilon_j, a small variance on the order of var⁡(εj)≈10−5\operatorname{var}(\varepsilon_j)\approx10^{-5} is suggested to preserve identification accuracy. Figure 4 compares the current measured in RTDS

MMC currents generated by $\tilde{u}$

Figure 4. MMC currents generated by u~\tilde{u} in: (–) RTDS, (–) fSS, and (–) dSS.

data-driven SS (dSS) model obtained by subspace identification (Appendix A). The measurements and dSS matrices reveal dynamics neglected by the fSS model: cross-coupling between the output and circulating current loops, switching and modulation ripple, and pre-processing effects. The dSS model closely tracks the measured current, particularly its low-frequency content, as confirmed by its lower mean squared prediction error (MSEp_p) in Table 1. The current spectra in Figure 5 show this more clearly: the fSS model underestimates the 0 Hz component, which sets the gain of the current dynamics, and overestimates the 50 Hz and 100 Hz components relative to RTDS and dSS. Higher-order linear models could reproduce better the RTDS time response, but would likely overfit rather than yield an unbiased predictor. These mismatches do not necessarily compromise stability or feasibility, as fSS models have been used effectively in MBD [7]; nevertheless, they reveal a modelling gap that dSS models can close, notably for controller design with theoretical guarantees.

MSEp_p [pu2^2]idoi_{d}^{o}iqoi_{q}^{o}idci_{d}^{c}iqci_{q}^{c}
fSS0.07230.08280.07840.1410
dSS0.00440.00100.00490.0024

Table 1. MSEp comparison of fSS and dSS models

Figure 5. FFT analysis of idoi_{d}^{o} and iqci_{q}^{c} currents in Fig. 4.

iSPC Design and Verification

Before its implementation in the RTDS, the iSPC is designed and verified in MATLAB. The weighting matrices are chosen empirically as Q=I4Q = I_{4} and R=10I4R = 10 I_{4}, and the outputs y=(idqo,idqc)y = (i_{d q}^{o}, i_{d q}^{c}) and inputs u=(udqo,udqc)u = (u_{d q}^{o}, u_{d q}^{c}) are constrained to −1.5≤y≤1.5-1.5 \leq y \leq1.5 pu and −1.5≤u≤1.5-1.5 \leq u \leq1.5 pu, respectively. During the design, we examine how TT and NN affect the trade-off between robustness, control performance, and computational burden, with Tini=20T_{\mathrm{ini}} = 20 fixed. Robustness is assessed by adding output measurement noise at different signal-to-noise ratios (SNRs), i.e., y~(k)=y(k)+v(k)\tilde{y}(k) = y(k) + v(k), where v(k)v(k) is zero-mean white noise. Figure 6 shows the resulting mean absolute tracking error (MAEt_t) to step response of the CCL for different SNR, TT, and NN.

Figure 6. Comparison of the iSPC tracking performance, i.e., MAEt_t, with Q=I4Q=I_4, R=10I4R=10I_4, and Tini=20T_{\mathrm{ini}}=20 under different conditions of SNR, TT and NN.

For SNR=12 dB, the added noise v(k)v(k) has a variance of var⁡(v)≈0.1022\operatorname{var}(v) \approx0.1022 and a peak amplitude of max⁡∣v∣=1.363\max|v| = 1.363 pu. Under this highly noisy condition, the iSPC performance degrades considerably, the MAEt_t increases by up to a factor of 100 and even yielding infeasible OCP for N=2N=2. Increasing TT beyond 20Tini20T_{\mathrm{ini}} yields no substantial improvement under high noise. Instead, increasing NN hold by N<TiniN < T_{\mathrm{ini}} proves to be the most effective way to improve iSPC performance at a lower computational burden. The same holds under cleaner conditions (SNR = 52 dB and 102 dB), where increasing NN remains the best strategy, whereas for short prediction horizons a larger TT does not guarantee better performance. Finally, increasing TiniT_{\mathrm{ini}}, and consequently TT, yields the best (stable/feasible) performance even under very noisy conditions (SNR ≤\leq 10), but at a considerably higher computational cost. Figure 7 and Figure 8 show the CCL response to successive reference steps under iSPC in a low-noise scenario (SNR ≈52\approx52 dB), the references are tracked without constraint violation.

MMC currents under iSPC in four stacked time-series plots

Figure 7. MMC currents under iSPC

Figure 8. CCL Control input under iSPC

Comparison DeePC vs iSPC

We compare the iSPC against DeePC [14], a DPC method widely adopted in the DD control community. Unlike iSPC, DeePC is a direct method that solves the OCP using the data directly, without an explicit identification step; implementation details can be found in [14, 18]. For a fair comparison, both DPCs use the same parameters as the previous section. For DeePC, the slack variable on the yini(k)y_{\mathrm{ini}}(k) is penalized with λy=104\lambda_y=10^4, and no regularization on gg is applied. Regularized variants were tested but yielded no significant improvement under the low to moderate noise conditions considered here, while considerably increasing the computation time; hence, they are not reported. Figure 9 and Figure 10 show CCL response under iSPC and DeePC with SNR =102=102 dB, Tini=20T_{\mathrm{ini}}=20, and T=20TiniT=20T_{\mathrm{ini}}, where the same step references as in § 4.2 is considered. Both controllers track the reference within the constraints; with identical tuning, the iSPC responds faster but with a higher overshoot, whereas DeePC is slower and more sensitive to output noise. In terms of MAEt_t, the iSPC achieves 1.82×10−51.82 \times10^{-5} pu, while iDeePC has a 2.29×10−42.29 \times10^{-4} pu. The integral of the absolute control input (IAU) is 7.7846 pu for the iSPC and 6.3353 pu for DeePC. In MATLAB environment, the average computation time2 per control step is 1.15 ms for the iSPC and 268 ms for DeePC.

Figure 9. MMC currents under (—) iSPC and (—) iDeePC.

Figure 10. CCL Control input under (—) iSPC and (—) iDeePC.

The iSPC computation time in MATLAB exceeds the 50 μ\mus required for real-time implementation, mainly because the

OCP was solved via YALMIP [24], a general-purpose modeling toolbox that is not optimised for speed. However, the iSPC structure allows the OCP to be solved with Hildreth’s algorithm [25] coded in C++, which has proven effective for MMC control [7]. Future work will therefore implement the iSPC directly on RTDS.

Conclusions

This paper investigated the use of iSPC for current control of MMCs in HVDC systems. The SI experiment revealed notable cross-coupling between the output and circulating currents dynamics, often neglected in FSS models. Consequently, dSS model identified from I/O data were proposed to predict the current dynamics more accurately. Using these I/O data, we designed the iSPC and proposed a guideline for its systematic application, covering the design of the excitation signal, the selection of TiniT_{\mathrm{ini}} and TT, and the tuning of the prediction horizon. Apart from data quality, the prediction horizon NN proved to be the most influential design parameter: increasing NN yielded the best trade-off between control performance and computational burden, whereas increasing TT beyond 20Tini20T_{\mathrm{ini}} brought no substantial improvement in low to moderate noise conditions. Compared with DeePC under identical tuning, the iSPC achieved a roughly 13 times lower tracking MSE at the cost of a moderately higher control effort, while reducing the computation time by more than two orders of magnitude. Future work will therefore implement the iSPC directly on RTDS.

Acknowledgements

The work is supported by Horizon Europe project PROS-ECCO, under grant agreement 101160687, and by NWO Veni project SAFE-GRID, with project number 20248.

Appendix

fSSf_{\mathrm{SS}} matrices derived from CIGRE B4.57 DCS1 benchmark:

A=[0.9995−0.0157000.01570.999500000.99950.031400−0.03140.9995],(19)A = \begin{bmatrix} 0.9995 & -0.0157 & 0 & 0 \\ 0.0157 & 0.9995 & 0 & 0 \\ 0 & 0 & 0.9995 & 0.0314 \\ 0 & 0 & -0.0314 & 0.9995 \end{bmatrix}, \tag*{(19)}
B=[0.0615−0.0005000.00050.06150000−0.1042−0.0016000.0016−0.1042].(20)B = \begin{bmatrix} 0.0615 & -0.0005 & 0 & 0 \\ 0.0005 & 0.0615 & 0 & 0 \\ 0 & 0 & -0.1042 & -0.0016 \\ 0 & 0 & 0.0016 & -0.1042 \end{bmatrix}. \tag*{(20)}

dd from CIGRE B4.57 DCS1 benchmark:

A=[0.67450.06150.3082−0.03100.03960.9825−0.03680.0108−0.02550.01041.0237−0.0065−0.00470.00360.00440.9975],(21)A = \begin{bmatrix} 0.6745 & 0.0615 & 0.3082 & -0.0310 \\ 0.0396 & 0.9825 & -0.0368 & 0.0108 \\ -0.0255 & 0.0104 & 1.0237 & -0.0065 \\ -0.0047 & 0.0036 & 0.0044 & 0.9975 \end{bmatrix}, \tag*{(21)}
B=[0.19610.94371.44603.4511−0.0565−0.1211−0.2036−0.35030.01700.07740.17020.25830.0061−0.00450.02300.0496].(22)B = \begin{bmatrix} 0.1961 & 0.9437 & 1.4460 & 3.4511 \\ -0.0565 & -0.1211 & -0.2036 & -0.3503 \\ 0.0170 & 0.0774 & 0.1702 & 0.2583 \\ 0.0061 & -0.0045 & 0.0230 & 0.0496 \end{bmatrix}. \tag*{(22)}

References

References

  1. [1]N. Ahmed, A. Haider, D. Van Hertem, L. Zhang, and H.-P. Nee, “Prospects and challenges of future hvdc supergrids with modular multilevel converters,” in Proc. 14th Eur. Conf. Power Electron. Appl. (EPE). IEEE, 2011, pp. 1–10.
  2. [2]R. Marquardt, “Modular multilevel converters: State of the art and future progress,” IEEE Power Electron. Mag., vol. 5, no. 4, pp. 24–31, 2018.DOI
  3. [3]J. Beerten, O. Gomis-Bellmunt, X. Guillaud, J. Rimez, A. Van der Meer, and D. Van Hertem, “Modeling and control of hvdc grids: A key challenge for the future power system,” in Proc. Power Syst. Comput. Conf. (PSCC). IEEE, 2014, pp. 1–21.
  4. [4]V. D. R. Dreke, Y. P. Marca, M. G. Roes, and M. Lazar, “Lmi synthesis of multivariable current controllers for ac/ac modular multilevel converters,” IEEE Trans. Control Syst. Technol., 2026.DOI
  5. [5]M. Ayari, M. M. Belhaouane, X. Guillaud, and N. B. Braiek, “Multivariable optimal robust control strategy for mmc converter,” in Proc. Int. Conf. Adv. Syst. Electr. Technol. (IC_ASET). IEEE, 2017, pp. 210–215.DOI
  6. [6]S. D. Tavakoli, E. Prieto-Araujo, S. Fekriasl, J. Beerten, H. Mehrjerdi, and O. Gomis-Bellmunt, “Robust multivariable control design for hvdc systems considering ac grid impedance uncertainties,” Int. J. Elect. Power Energy Syst., vol. 141, p. 108152, 2022.DOI
  7. [7]A. Shetgaonkar, L. Liu, A. Lekić, M. Popov, and P. Palensky, “Model predictive control and protection of mmc-based mtdc power systems,” Int. J. Elect. Power Energy Syst., vol. 146, p. 108710, 2023.DOI
  8. [8]V. D. Reyes Dreke and M. Lazar, “Long-horizon nonlinear model predictive control of modular multilevel converters,” Energies, vol. 15, no. 4, p. 1376, 2022.DOI
  9. [9]J. Böcker, B. Freudenberg, A. The, and S. Dieckerhoff, “Experimental comparison of model predictive control and cascaded control of the modular multilevel converter,” IEEE Trans. Power Electron., vol. 30, no. 1, pp. 422–430, 2014.DOI
  10. [10]R. Rane, A. Kermansaravi, P. Vergara, and A. Lekić, “Transfer learning framework for impedance characterization of modular multilevel converters,” IEEE Trans. Ind. Appl., vol. 61, no. 2, pp. 2421–2433, 2025.DOI
  11. [11]H. Gong, D. Yang, and X. Wang, “Impact analysis and mitigation of synchronization dynamics for dq impedance measurement,” IEEE Trans. Power Electron., vol. 34, no. 9, pp. 8797–8807, 2018.DOI
  12. [12]G. R. Gonçalves da Silva, A. S. Bazanella, C. Lorenzini, and L. Campestrini, “Data-driven lqr control design,” IEEE Control Syst. Lett., vol. 3, no. 1, pp. 180–185, 2019.DOI
  13. [13]W. Favoreel, B. De Moor, and M. Gevers, “SPC: Subspace predictive control,” IFAC Proc. Vol., vol. 32, no. 2, pp. 4004–4009, 1999.DOI
  14. [14]J. Coulson, J. Lygeros, and F. Dörfler, “Data-enabled predictive control: In the shallows of the DeePC,” in Proc. Eur. Control Conf. (ECC), 2019, pp. 307–312.arxiv.org/abs/1811.05890
  15. [15]D. Bilgic, A. Koch, G. Pan, and T. Faulwasser, “Toward data-driven predictive control of multi-energy distribution systems,” Electr. Power Syst. Res., vol. 212, p. 108311, 2022.DOI
  16. [16]R. Leng, L. Huang, H. Xin, P. Ju, X. Wang, E. Prieto-Araujo, and F. Dörfler, “A data-driven optimal control architecture for grid-connected power converters,” IEEE Trans. Power Syst., 2026.arxiv.org/abs/2508.08578
  17. [17]I. Kraayeveld, T. de Jong, and M. Lazar, “Offset-free data-driven predictive control for grid-connected power converters in weak grid faults,” Electr. Power Syst. Res., vol. 263, p. 113551, 2027.DOI
  18. [18]P. Verheijen, V. Breschi, and M. Lazar, “Handbook of linear data-driven predictive control: Theory, implementation and design,” Annu. Rev. Control, vol. 56, p. 100914, 2023.DOI
  19. [19]J. C. Willems, P. Rapisarda, I. Markovsky, and B. L. M. De Moor, “A note on persistency of excitation,” Syst. Control Lett., vol. 54, no. 4, pp. 325–329, 2005.
  20. [20]L. Ljung, System Identification: Theory for the User, 2nd ed. Prentice Hall, 1999.
  21. [21]S. Du, A. Dekka, B. Wu, and N. R. Zargari, Modular Multilevel Converter: Analysis, Control and Applications, T. Samad, Ed. John Wiley and Sons, 2018.
  22. [22]CIGRE Working Group B4.57, “Guide for the development of models for HVDC converters in a HVDC grid,” CIGRE, Paris, France, Technical Brochure 604, 2014, working Group B4.57.
  23. [23]M. ApS, The MOSEK optimization toolbox for MATLAB manual. Version 10.1., 2024. [Online]. Available: http://docs.mosek.com/latest/toolbox/index.html
  24. [24]J. Löfberg, “YALMIP: A toolbox for modeling and optimization in MATLAB,” in Proc. IEEE Int. Conf. Comput. Aided Control Syst. Des. (CACSD), Taipei, Taiwan, 2004, pp. 284–289.DOI
  25. [25]L. Wang, Model predictive control system design and implementation using MATLAB®. Springer, 2009.DOI

Paper details

vlnc:W_h552yzbk
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

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