Electric Power Systems Research 238 (2025) 111063

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

Simplified Finite Control Set Model Predictive Control for single-phase

grid-tied inverters with twisted parameters
Po Li, Xiaoxiao Huo ∗
School of Aerospace Engineering, Xiamen University, Xiamen, 361102, China

ARTICLE INFO ABSTRACT

Keywords: Large computational burden, time delay, and the necessity for precise modeling accuracy are the three main
Finite control set-model predictive control challenges for Finite Control Set-Model Predictive Control (FCS-MPC) in single-phase grid-tied inverters. To
Total harmonic distortion solve these issues, a twisted parameter scheme is proposed for the single-phase inverter in this article. Firstly,
Twisted parameters
the law regarding the influence of the model parameter on the current total harmonic distortion (THD) is
Single-phase inverters
outlined, emphasizing that a decrease in the inductance parameter leads to a corresponding reduction in current
Hardware-in-the-loop
THD. Second, a linear observer is constructed to identify the actual value of inductance and resistance, and
an RBF-GA (Radial Basis Function neural network-Genetic Algorithm) scheme is used to obtain the optimal
twisted parameter. Subsequently, the efficacy of the proposed methods was verified utilizing MATLAB/Simulink
simulations, with further validation conducted through hardware-in-the-loop (HIL) experiments performed on
Speedgoat performance real-time target machines. Simulation and experimental results demonstrate that within
a specific range, decreasing the inductance parameter can significantly improve the quality of the current.
Furthermore, the proposed method outperforms the traditional delay compensation method by reducing
computational complexity, minimizing prediction error, and decreasing the number of switching transitions.

1. Introduction when the state exceeds a threshold, significantly reducing computa-

tional burden and switch losses. In [16], the optimal vector is de-
Grid-tied inverters, as important energy conversion facilities in termined based on position, leading to a significant reduction in the
power grid systems, have attracted widespread attention from domestic computational complexity required for current prediction.
and international scholars in terms of the quality of grid-connected Digital control faces delay issues, worsened by high computational
current [1,2]. For grid-tied inverters, the control objective is typically
burden. The optimization process of FCS-MPC requires an amount
set to minimize computational complexity [3,4], minimize the total
of computation, that exceeds the sampling time, which results in a
harmonic distortion (THD) of the grid-connected current [5,6], and
delay between the current being sampled and the switching signal of
improve control robustness [7,8].
Model Predictive Control (MPC) is a powerful technique that can the next control period being applied, causing the predicted current
effectively solve nonlinear problems and control multiple constraints value to deviate from the reference value. A method for compensating
while satisfying performance requirements. FCS-MPC is a kind of MPC delay is proposed to minimize their impact on current prediction [17].
that allows direct control of the switching devices without the need In [18], a linear delay compensation design for closed-loop current
for an external modulator, and it has gained significant attention in control of inverters based on MPC is proposed. In [19], a two-step
recent years in the field of inverter control [9,10]. However, it has delay compensation approach based on deadbeat control is presented.
the following drawbacks: large computational burden, computational However, the above approach results in an additional computational
delay, and significant parameter sensitivity [11–13] (see Fig. 1). burden.
The times of prediction and optimization calculation of traditional Furthermore, the control performance of FCS-MPC is highly sen-
FCS-MPC in a control period depend on the number of effective switch- sitive to model parameters. If there is system uncertainty causing
ing states, the process of selecting the optimal vector is complicated
mismatches between nominal and actual parameters, the control per-
and the calculation burden is large. In [14], reinforcement techniques
formance will worsen. To address the parameter mismatch problem,
is combined with FCS-MPC to effectively reduce computational time.
researchers proposed online parameter identification and disturbance
In [15], an event-triggered MPC method, which only triggers control

∗ Corresponding author.
E-mail addresses: [email protected] (P. Li), [email protected] (X. Huo).

https://doi.org/10.1016/j.epsr.2024.111063
Received 25 May 2024; Received in revised form 13 August 2024; Accepted 7 September 2024
Available online 11 September 2024
0378-7796/© 2024 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
P. Li and X. Huo Electric Power Systems Research 238 (2025) 111063

Fig. 1. Relationship diagram between the problems of FCS-MPC and the proposed
method.
Fig. 2. Single-phase inverter topology.

observation methods such as those based on the extended state ob-

server [20], Kalman filter [21], recursive least squares [22], and model
reference adaptive system [23] to compensate for prediction errors
resulting from parameter mismatches.
Previous studies on parameter mismatch are limited to the mis-
match between the actual parameters and the nominal ones. However,
while these studies have mitigated the adverse effects of parameter mis-
match on control effect, the computational complexity on the controller
has also increased. Additionally, the impact of model parameter values
on control effect remains unexplored [24].
In view of the above research and observations, the model parame-
ter is artificially modified and named as the twisted parameter in this
paper, and the computational burden and current THD are reduced
by using the twisted parameter, which addresses the shortcomings of
prior studies on parameter mismatch. The main contributions are listed
below.

• The parameter mismatch problem of FCS-MPC is deeply ana-

lyzed and studied, filling the gaps in previous parameter mis-
match research. Through simulation, the relationship between the
Fig. 3. Control diagram of the traditional FCS-MPC.
inductance parameter and current THD is summarized.
• A method that twists the model parameter is proposed, achieving
similar prediction performance as the traditional delay compen-
sation method, while reducing computational burden compared output voltage, and 𝑒 represents grid voltage. The mathematical model
to the delay compensation method. of the single-phase grid-tied inverter can be expressed as follows
• To make full use of the proposed twisted parameter method and 𝑑𝑖
𝐿 + 𝑅𝑖 = 𝑢𝑎𝑏 − 𝑒 (1)
achieve the optimal control effect, a method combining offline 𝑑𝑡
extremum optimization based on RBF-GA (Radial Basis Function The current prediction formula can be expressed as (2) based on
neural network-Genetic Algorithm) and online parameter identi- Euler’s method.
fication based on the observer is introduced to obtain the optimal ( )
𝑅𝑇𝑠 𝑇 ( )
twisted parameter. 𝑖𝑝 (𝑘 + 1) = 1 − 𝑖(𝑘) + 𝑠 𝑢𝑎𝑏 (𝑘) − 𝑒(𝑘) (2)
𝐿 𝐿
The remainder of this article is outlined as follows. In Section 2, where 𝑇𝑠 is sampling period, 𝑖(𝑘), 𝑒(𝑘), and 𝑢𝑎𝑏 (𝑘) are the sampled
a single-phase grid-tied inverter is modeled, followed by traditional current, grid voltage, and inverter’s output voltage at 𝑡𝑘 respectively.
predictive control methods are analyzed. Section 3 analyzes the phe-
nomenon and law of the twisted parameters affecting the current THD, 2.2. Traditional FCS-MPC method
and proposes the twisted parameter method. Additionally, methods of
offline extremum optimization and online parameter identification are A single-phase inverter has four switching state combinations, and
presented for obtaining the optimal twisted parameter. Sections 4 and two of them correspond to the same 𝑢𝑎𝑏 , so the number of effective
5 verify the efficacy of the proposed methods through simulations and switching state combinations is 3, and each switching signal corre-
experiments, respectively. At last, the conclusions are summarized in sponds to 𝑖𝑝 .
Section 6. Generally, the cost function can be set as

𝑔 = |𝑖𝑝 (𝑘 + 1) − 𝑖𝑟 (𝑘 + 1)| (3)

2. Modeling
where 𝑖𝑟 (𝑘 + 1) is the reference current with amplitude of 𝐼𝑟𝑒𝑓 at time
2.1. System description 𝑡𝑘+1 , and the frequency and phase are the same as 𝑒(𝑘).
The control structure of the traditional FCS-MPC is illustrated in
The single-phase grid-tied inverter is the research object of this Fig. 3. The traditional FCS-MPC is to minimize the cost function by
paper. As depicted in Fig. 2, 𝑈𝑑𝑐 represents the DC-link voltage, 𝑖 enumeration traversal, select the 𝑢𝑎𝑏 (𝑘) corresponding to the minimum
represents the grid-connected current, 𝐿 and 𝑅 represents the filter current prediction error, and take the optimal voltage vector 𝑢𝑎𝑏 (𝑘) in
inductance and resistance, respectively, 𝑢𝑎𝑏 represents the inverter’s the next control period so that 𝑖𝑝 (𝑘 + 1) is close to 𝑖𝑟 (𝑘 + 1).

2
P. Li and X. Huo Electric Power Systems Research 238 (2025) 111063

Fig. 4. Predictive control operation diagram: (a) Without delay (ideal case); (b) With
delay (real case).

2.3. Traditional delay compensation method

In the ideal case, the controller operates instantaneously without

any computation time, minimizing the delay between sampling the
current and implementing the voltage vector. As depicted in Fig. 4(a),
current is sampled and predicted at 𝑡𝑘 , and then the optimal 𝑢𝑎𝑏 (𝑘) is Fig. 5. Improved FCS-MPC control diagram based on delay compensation.
applied immediately at 𝑡𝑘+1 .
However, as depicted in Fig. 4(b), a computational delay exists
between sampling the current and implementing the new voltage vector
in practice. During this period, the previous voltage vector continues to
be implemented. This implies that the chosen voltage vector based on
𝑖(𝑘) is implemented after 𝑡𝑘+1 , leading to a discrepancy between the
output current and the reference value.
In order to mitigate the issue of reduced prediction accuracy due to
computational delay, a two-step delay compensation is implemented in
the conventional FCS-MPC framework, as depicted in Fig. 5. The cur-
rent and the implemented voltage vector at 𝑡𝑘 are utilized to calculate
the current at 𝑡𝑘+1 as follows
( )
𝑅𝑇𝑠 𝑇 ( )
𝑖̂𝑝 (𝑘 + 1) = 1 − 𝑖(𝑘) + 𝑠 𝑢𝑎𝑏 (𝑘) − 𝑒(𝑘) (4)
𝐿 𝐿
The calculated current 𝑖̂(𝑘 + 1) and voltage vector 𝑢𝑎𝑏 (𝑘 + 1) at 𝑡𝑘+1
are utilized to predict current at 𝑡𝑘+2 . Fig. 6. Variation in THD after twisting model parameters.
( )
𝑅𝑇𝑠 𝑇 ( )
𝑖̂𝑝 (𝑘 + 2) = 1 − 𝑖̂𝑝 (𝑘 + 1) + 𝑠 𝑢𝑎𝑏 (𝑘 + 1) − 𝑒(𝑘 + 1) (5)
𝐿 𝐿
that studied the effect of system uncertainty on control effectiveness
Accordingly, the cost function is rewritten as follows
when the actual and nominal values of circuit parameters differ. In-
𝑔 = |𝑖𝑝 (𝑘 + 2) − 𝑖𝑟 (𝑘 + 2)| (6) stead, the FCS-MPC model parameters are deliberately changed, cre-
ating a discrepancy between the model and circuit parameters. This
where 𝑖𝑟 (𝑘 + 2) is the reference current at 𝑡𝑘+2 . When the switching
parameter mismatch is then utilized to create a positive impact on the
frequency significantly exceeds the grid frequency, it is reasonable to
prediction, improving current quality instead of relying on traditional
assume that 𝑖𝑟 (𝑘 + 2) = 𝑖𝑟 (𝑘 + 1). After minimizing the cost function, the
delay compensation, which would increase computational burden.
optimal 𝑢𝑎𝑏 (𝑘 + 1) is chosen and implemented in the subsequent control
Assuming there is a delay but no compensation method is adopted,
cycle.
after twisting the inductance and resistance parameters, the predicted
The steps of the delay compensation method can be summarized as
current 𝑖𝑝𝑡 (𝑘 + 1) can be written as
follows [17]
( )
𝑅𝑇 𝑇 ( )
1. sample the current at 𝑡𝑘 ; 𝑖𝑝𝑡 (𝑘 + 1) = 1 − 𝑡 𝑠 𝑖(𝑘) + 𝑠 𝑢𝑎𝑏 (𝑘) − 𝑒(𝑘) (7)
𝐿𝑡 𝐿𝑡
2. implement the voltage vector at 𝑡𝑘 ;
3. calculate the current at 𝑡𝑘+1 ; where 𝑅𝑡 and 𝐿𝑡 are the twisted resistance and inductance parameters,
4. predict the current at 𝑡𝑘+2 using 𝑖̂𝑝 (𝑘 + 1); respectively.
Fig. 6 shows the variation in THD with the change in the mismatch
5. evaluate the cost function for each 𝑖̂𝑝 (𝑘 + 2);
ratio of inductance and resistance parameters from 0.1 to 2. The
6. choose the optimal 𝑢𝑎𝑏 (𝑘 + 1) after minimizing the cost function.
mismatch ratio refers to the ratio of the twisted parameters 𝐿𝑡 and 𝑅𝑡
Compared to the traditional FCS-MPC, there is an additional estima- to the actual parameters 𝐿𝑟 and 𝑅𝑟 . The switching frequency 𝑓𝑠 was
tion step to enhance the quality of grid current. However, it also leads configured at 20 kHz with 𝐼𝑟𝑒𝑓 of 12.5 A.
to an increase in computational complexity. From Fig. 6, it is evident that the impact of inductance on THD
outweighs that of resistance. Consequently, this research will exclu-
3. Proposed twisted parameter method sively study influence of 𝐿𝑡 on current THD. When the mismatch ratio
of the inductance parameter is 0.5, the current waveforms generated
3.1. The phenomenon of twisted parameters after twisting the parameter closely overlap with those produced by
the delay compensation method, as depicted in Fig. 7.
According to (2), the current prediction model is related to the The comparison of THD and the number of switching transitions
inductance 𝐿 and resistance 𝑅. This study diverges from prior research between the proposed twisted parameter and the delay compensation

3
P. Li and X. Huo Electric Power Systems Research 238 (2025) 111063

Fig. 7. Comparison of the current waveforms. Fig. 10. Comparison of different types of neural network performance.

3.2. Offline extremum optimization

To fully utilize the law of the twisted inductance parameter and

achieve the best control effect, optimization algorithms are used in this
study to determine the optimal inductance mismatch ratio 𝜆 = 𝐿𝑡 ∕𝐿𝑟 .
Since the relationship between the current THD and mismatch ratio is
unknown, a neural network is employed for the nonlinear fitting of the
mathematical relationship.
In this study, the inductance mismatch ratio is constrained within
the specified range of 0.3 to 1, operating with a filter inductance of
Fig. 8. Simulation results of twisting inductance parameter at various reference current
values: (a) Current THD; (b) Number of switching transitions. 8 mH, a switching frequency at 20 kHz, and then the THD values of
the current are recorded after twisting. The corresponding relationship
between current THD and inductance mismatch ratio under different
reference current values is collected as a dataset for constructing a
neural network, consisting of 701 samples. Out of these, 651 samples
are chosen randomly for the training set, while the rest of the samples
are allocated for the test set. Next, preliminary tests are conducted on
the BP (Back Propagation) neural network, RBF (Radial Basis Function)
neural network, GRNN (Generalized Regression neural network), El-
man, and ELM (Extreme Learning Machine) neural network, and the
nonlinear fitting effects are measured using Mean Absolute Percentage
Error (MAPE) indicator as depicted in Fig. 10.
The MAPE calculation formula is as follows
𝑚 | (𝑖) (𝑖) |
Fig. 9. Simulation results of twisting inductance parameter at various switching
frequencies: (a) Current THD; (b) Number of switching transitions. 1 ∑ || 𝑦𝑝𝑟𝑒𝑑𝑖𝑐𝑡 − 𝑦𝑡𝑒𝑠𝑡 ||
𝑀𝐴𝑃 𝐸 = (8)
𝑚 𝑖=1 || 𝑦(𝑖)
|
|
| 𝑡𝑒𝑠𝑡 |
where 𝑦𝑝𝑟𝑒𝑑𝑖𝑐𝑡 (𝑖) is prediction result for the 𝑖th test data, and 𝑦𝑡𝑒𝑠𝑡 (𝑖) is the
method under different reference current values are depicted in Fig. 8,
accurate value of the 𝑖th test data.
where the solid lines depict the performance of twisted inductance
It can be observed from Fig. 10 that the RBF neural network has
parameter method, while the dashed lines indicate the performance the smallest fitting error among the five kinds of neural networks,
of delay compensation method. The 𝑓𝑠 was configured at 20 kHz, indicating a good fitting effect. Consequently, the study leverages the
with 𝐼𝑟𝑒𝑓 configured at 10.00 A, 11.11 A, 12.50 A, 12.90 A, and RBF neural network for nonlinear fitting purposes.
13.33 A in sequence. The number of switching transitions is a crucial The RBF neural network is distinguished by its straightforward
factor affecting the lifespan of switching devices. Limiting the increase three-layer feedforward design. The initial layer functions as the input
in the number of switching transitions minimizes switching loss and layer. The second layer is the hidden layer, where the transformation
extends the overall lifespan of the equipment. The results indicate that function of neurons, namely the radial basis function, is employed. The
within a specific range of twisted inductance parameter values, the ultimate layer functions as the output layer, tasked with processing
current THD decreases compared to the initial state (mismatch ratio input data. Within the network’s framework, the exchange of infor-
of 1). Remarkably, the THD is further reduced compared to the results mation between the hidden and output layers follows a linear pattern,
achieved through delay compensation, while simultaneously decreasing thereby diminishing computational intricacies [25,26]. In the RBF neu-
the need for additional switching transitions. ral network, the number of hidden layer nodes is set to the number of
Additional simulations were performed setting 𝐼𝑟𝑒𝑓 to 12.5 A, as training samples, and all samples are selected as the data centers of
the radial basis function, thereby achieving zero-error fitting on the
depicted in Fig. 9. The switching frequency was varied incrementally
training samples. The RBF neural network’s schematic is depicted in
at 20 kHz, 25 kHz, 32 kHz, 40 kHz, and 50 kHz. The aforementioned
Fig. 11, while Fig. 12 illustrates the fitting performance of the network.
simulations initially validate that in a specific range of inductance
The RBF neural network output 𝑦(𝑡) can be set as
mismatch ratios, the proposed twisted parameter method can serve as
an effective substitute for delay compensation. This substitution not ∑
𝐾
𝑦(𝑡) = 𝑤𝑘 (𝑡)𝜙𝑘 (𝑡) (9)
only simplifies computational tasks but also reduces current THD.
𝑘=1

4
P. Li and X. Huo Electric Power Systems Research 238 (2025) 111063

Fig. 13. Fitness curve of GA.

a very strong capability to find the global optimum. Being a parallel

iterative algorithm equipped with specific learning abilities, GA iterates
through the evaluation, selection, crossover, and mutation operators
repeatedly until it either fulfills a stopping criterion or reaches the
maximum iteration count [29–31]. The pseudocode of GA is presented
in Algorithm 1 [32,33]. Since the minimization problem is dealt with
in this study, the fitness function can be set to the current THD values
predicted by the RBF neural network fitting, and the fitness curve is
depicted in Fig. 13.

Fig. 11. Schematic of RBF neural network.

Algorithm 1: Genetic Algorithm.
Data: population size, 𝑛; maximum number of iterations, 𝐼𝑡𝑒𝑟;
crossover probability, 𝑃𝑐𝑟𝑜𝑠𝑠 ; Mutation probability, 𝑃𝑚𝑢𝑡 ;
Result: globally optimal solution, 𝑌𝑏𝑡
1 Generate initial population of 𝑛 chromosomes 𝑌𝑖 (𝑖 = 1, 2, … , 𝑛);
2 Set iteration counter 𝑡 = 0;
3 Compute the fitness value of each chromosomes;
4 while 𝑡 < 𝐼𝑡𝑒𝑟 do
5 Select a pair of chromosomes from the initial population
according to fitness;
6 Crossover operation on selected pair with 𝑃𝑐𝑟𝑜𝑠𝑠 ;
7 Mutation on the offspring with 𝑃𝑚𝑢𝑡 ;
8 The old population is replaced with the newly generated
population;
9 Increment the current iteration 𝑡 by 1;
10 end

The overall process of offline optimization for finding the extremum

Fig. 12. The fitting performance of the RBF neural network. is depicted in Fig. 14, which includes two parts: RBF neural network
fitting and GA for finding the extremum. The input values for the
optimization algorithm are the amplitude of reference current 𝐼𝑟𝑒𝑓 and
where 𝑤𝑘 (𝑡) is the 𝑘th output weight, 𝐾 is the total number of hidden the actual value of the inductance 𝐿𝑟 , and computes the inductance
neurons, and the output of the 𝑘th hidden neuron 𝜙𝑘 (𝑡) can be set as a mismatch ratio 𝜆 as the output. Finally, based on the relationship
Gaussian function: between the actual value of the inductance and the mismatch ratio
( ) for different reference current values, a lookup table (LUT) is created.
‖𝑥(𝑡) − 𝑐𝑘 (𝑡)‖2
𝜙𝑘 (𝑡) = exp − ‖ ‖ (10) Notably, this part of the optimization work operates offline and is ded-
2𝜎𝑘2 (𝑡) icated to determining the optimal parameter. Once these parameters
where 𝑥(𝑡) is the input of the RBF neural network, 𝑐𝑘 (𝑡) and 𝜎𝑘 (𝑡) are established, they do not contribute to the computational burden
represent the center vector and the width of the 𝑘th hidden neuron [27, during online operation. Hence, although the algorithms may possess a
28]. certain level of computational complexity, their impact on the system’s
After fitting the RBF neural network, a mathematical model relating real-time performance is negligible.
the inductance mismatch ratio to the current THD is established. The
3.3. Online parameter identification
next step is to use an optimization algorithm to search for the optimal
inductance mismatch ratio corresponding to the lowest current THD.
To obtain the real-time values of circuit parameters 𝐿𝑟 and 𝑅𝑟 , a
In this study, a Genetic Algorithm (GA) is utilized to search for
linear observer is constructed.
the minimum of the inductance mismatch ratio and current THD re-
Define the error between the nominal values and the actual values
lationship curve. The GA is a stochastic search method that relies on
caused by parameter mismatch as follow
a population-based approach to tackle optimization challenges, mim-
icking the mechanisms of natural evolution. More importantly, GA has 𝛥𝐿 = 𝐿0 − 𝐿𝑟 , 𝛥𝑅 = 𝑅0 − 𝑅𝑟 (11)

5
P. Li and X. Huo Electric Power Systems Research 238 (2025) 111063

Fig. 14. Overall flowchart for offline extremum optimization.

Fig. 15. Control diagram of the proposed twisted parameter method with offline
where 𝐿0 and 𝑅0 represent the nominal values of inductance and optimization and online identification.
resistance, correspondingly. Define 𝑑 as the system bias caused by
parameter mismatch, then the system equation is rewritten as
𝑑𝑖 Define error signal 𝒙̃ = 𝒙 − 𝒙,
̂ and subtract (19) from (17), the error
𝐿0 + 𝑅0 𝑖 = 𝑢𝑎𝑏 − 𝑒 + 𝑑 (12)
𝑑𝑡 can be described as
Contrasted with (1), 𝑑 can be expressed as follows
𝒙̃̇ = (𝑯 − 𝒍𝑴)𝒙̃ = 𝑲 𝒙̃ (20)
𝑑𝑖
𝑑 = 𝛥𝐿 + 𝛥𝑅𝑖 (13) 𝑅 cos 𝜔𝑡 sin 𝜔𝑡
𝑑𝑡 ⎡ − 𝐿0 − 𝑙1 𝐿0 𝐿0
⎤
Considering that the output current 𝑖 is a sinusoidal wave in ideal ⎢ 0
cos 𝜔𝑡 ⎥
where 𝑲 = 𝑯 − 𝒍𝑴 = ⎢ −𝑙2 𝐿 0 0 ⎥.
0
conditions, so that 𝑖 can be expressed as 𝑖 = 𝐼𝑟𝑒𝑓 𝑠𝑖𝑛𝜔𝑡, where 𝐼𝑟𝑒𝑓 is the ⎢ −𝑙 sin 𝜔𝑡 0 0 ⎥⎦
amplitude of 𝑖, and 𝜔 represents the fundamental angular frequency of ⎣ 3 𝐿
0
When 𝑡 → ∞, 𝑥 → 0, that is, 𝒙̂ converges to 𝒙, which means the
𝑖. Hence, (13) can be written as
system is stable. Hence, the cosine and sine components amplitudes 𝐴̂
𝑑 = 𝜔𝛥𝐿𝐼𝑟𝑒𝑓 𝑐𝑜𝑠𝜔𝑡 + 𝛥𝑅𝐼𝑟𝑒𝑓 𝑠𝑖𝑛𝜔𝑡 (14) and 𝐵̂ can be observed.
Using the Euler method, the observer model in (19) can be dis-
The bias 𝑑 is also a sinusoidal wave, so it also can be expressed as
cretized as
a combination of cosine and sine components:
⎧ 𝑇𝑠
𝑑 = 𝐴𝑐𝑜𝑠𝜔𝑡 + 𝐵𝑠𝑖𝑛𝜔𝑡 (15) ⎪ 𝑖̂(𝑘 + 1) = 𝐿0
[ − 𝑅0 𝑖̂(𝑘) + 𝐴̂ cos(𝜃(𝑘)) + 𝐵̂ sin(𝜃(𝑘))
⎪ +(𝑢𝑎𝑏 (𝑘) − 𝑒(𝑘))] − 𝑙1 (𝑖̂(𝑘) − 𝑖(𝑘)) + 𝑖̂(𝑘)
contrasted with (14), 𝐿 and 𝑅 can expressed as follows ⎨ (21)
⎪ ̂ + 1) = −𝑇𝑠 𝑙2 cos(𝜃(𝑘))(𝑖̂(𝑘) − 𝑖(𝑘)) + 𝐴(𝑘)
𝐴(𝑘 ̂
⎧ ⎪ ̂ + 1) = −𝑇𝑠 𝑙3 sin(𝜃(𝑘))(𝑖̂(𝑘) − 𝑖(𝑘)) + 𝐵(𝑘)
𝐵(𝑘 ̂
⎪ 𝐿 = 𝐿0 −
𝐴 ⎩
𝜔𝐼𝑟𝑒𝑓
⎨ 𝐵 (16)
Then the actual values of inductance and resistance can be obtained
⎪ 𝑅 = 𝑅0 − 𝐼𝑟𝑒𝑓
⎩ as follows
By observing the cosine and sine components coefficients 𝐴 and ⎧ 𝐴̂
⎪ 𝐿̂ = 𝐿0 −
𝐵, the observed values of inductance and resistance 𝐿 and 𝑅 can be 𝜔𝐼𝑟𝑒𝑓
(22)
⎨ 𝐵̂
obtained [34,35]. ⎪ 𝑅̂ = 𝑅0 − 𝐼𝑟𝑒𝑓
Define 𝒙 = [𝑥1 , 𝑥2 , 𝑥3 ]𝑇 = [𝑖, 𝐴, 𝐵]𝑇 , and 𝑦 = 𝑥1 = 𝑖. The augmented ⎩
system of the system can be expressed as
{ 4. Simulation
𝒙̇ = 𝑮𝒙 + 𝑯(𝑢𝑎𝑏 − 𝑒)
(17)
𝑦 = 𝑴𝒙 Simulations were conducted using Matlab/Simulink to validate the
where efficacy of the proposed twisted parameter method based on online
parameter identification and offline extremum optimization as depicted
⎡ − 𝑅0 𝑐𝑜𝑠𝜔𝑡 𝑠𝑖𝑛𝜔𝑡 ⎤ ⎡ 1 ⎤ 𝑇
⎢ 𝐿0 𝐿0 𝐿0 ⎥ ⎢ 𝐿0 ⎥ ⎡1⎤ in Fig. 15, with the parameters listed in Table 1.
𝑮=⎢ 0 0 0 ⎥ , 𝑯 = ⎢ 0 ⎥ , 𝑴 = ⎢ 0⎥ (18)
⎢ 0 ⎥ ⎢ 0 ⎥ ⎢ ⎥
⎣ 0 0 ⎦ ⎣ ⎦ ⎣0⎦ 4.1. Validation of extremum optimization and parameter identification
According the state space equations, a linear observer can be ex-
pressed as The first part aims to validate the effectiveness of the online param-
{ ( ) eter identification method, specifically verifying the observer’s ability
𝒙̂̇ = 𝑮𝒙̂ + 𝑯 𝑢𝑎𝑏 − 𝑒 + 𝒍(𝑦 − 𝑦)
̂ to identify the inductance and resistance parameters accurately. Under
(19)
𝑦̂ = 𝑴 𝒙̂
the condition that the amplitude of reference current value is 10 A, the
where 𝒙̂ and 𝑦̂ are the observed values of state variable 𝒙 and system FCS-MPC controller used a nominal value of 10 mH for the inductance
output 𝑦, respectively. 𝒍 = [𝑙1 , 𝑙2 𝑐𝑜𝑠𝜔𝑡
𝐿0
, 𝑙3 𝑠𝑖𝑛𝜔𝑡
𝐿0
]𝑇 is the state feedback and 0.5 Ω for the resistance as model parameters before t=1 s, while the
matrix, the convergence rate is related to the values of the coefficients actual values were 8 mH for the inductance and 1 Ω for the resistance.
𝑙1 , 𝑙2 and 𝑙3 . At t=1 s, the observer started identifying the actual values of the

6
P. Li and X. Huo Electric Power Systems Research 238 (2025) 111063

Table 1
Simulation parameters.
Symbols Parameters Values
𝑈𝑑𝑐 DC-link voltage 120 V
𝑒𝑚 Grid voltage amplitude 100 V
𝑓 Grid frequency 50 Hz
𝑓𝑠 Sampling frequency 20 kHz
𝑅𝑟 Filter resistance (actual value) 1 Ω
𝑅0 Filter resistance (nominal value) 0.5 Ω
𝐿𝑟 Filter inductance (actual value) 8 mH
𝐿0 Filter inductance (nominal value) 10 mH
𝑛 GA population size 20
𝐼𝑡𝑒𝑟 GA maximum iterations 100
𝑃𝑐𝑟𝑜𝑠𝑠 GA crossover probability 0.4
𝑃𝑚𝑢𝑡 GA mutation probability 0.1

Fig. 17. Box diagram of inductance mismatch ratio and THD reduction.

Fig. 16. Parameter identification process: (a) Inductance parameter; (b) Resistance
parameter. Fig. 18. The effect of twisted parameter method: (a) Variation chart of 𝐿 and THD;
(b) Current waveform before and after twisting.
Table 2
RBF-GA results.
𝐼𝑟𝑒𝑓 (A) 𝜆 Fitness (%) Original 𝜆 calculated by GA, and the original THD is the output current THD
THD (%)
under the traditional FCS-MPC control framework.
10.0 0.5159 4.837 5.610 Fig. 17 describes the data distribution of 𝜆 and THD reduction.
10.2 0.5682 4.784 5.539
The THD reduction represents the ratio of the difference between the
10.4 0.4798 4.578 5.460
10.6 0.5782 4.525 5.382
original THD and the fitness calculated by the GA to the original
10.8 0.5068 4.528 5.461 THD. From the figure, it can be observed that 𝜆 predominantly falls
11.0 0.5698 4.285 5.320 within the range of 0.5–0.8, indicating that twisting inductance value
11.2 0.5858 4.313 5.166 to approximately 0.5–0.8 times the initial inductance parameter can
11.4 0.5023 4.413 4.918 achieve better control performance, which results in a reduction in THD
11.6 0.5487 4.267 5.159
of approximately 15%–20%.
11.8 0.4736 4.138 4.736
12.0 0.6431 4.062 4.966
12.2 0.6080 4.064 4.752 4.2. Validation of proposed twisted parameter method
12.4 0.5279 4.030 4.646
12.6 0.5880 3.884 4.909 At t=2 s, the offline optimized result is transmitted into the control
12.8 0.5043 3.732 4.308 system, and the inductance parameter is twisted. The current THD fol-
13.0 0.6400 3.716 4.521
13.2 0.5335 3.580 4.405
lows the variation of the inductance parameter, which can be observed
13.4 0.4513 3.625 4.320 in Fig. 18(a), and the comparison of current waveform before and after
13.6 0.8047 3.437 4.208 twisting of the inductance parameter is depicted in Fig. 18(b). Before
13.8 0.7086 3.226 4.394 twisting, the THD is 5.66%, and after twisting, the THD is 5.05%,
14.0 0.7117 3.132 4.056 proving that the extremum optimization algorithm and the proposed
14.2 0.8504 2.954 4.357
twisted parameter method are effective. Due to certain limitations of
14.4 0.6075 2.968 3.673
14.6 0.5209 3.080 4.035 the optimization algorithm, there may be discrepancies between the
14.8 0.6101 2.997 3.605 actual current THD after applying the twisted parameter and the fitness
15.0 0.6480 2.610 3.846 calculated using Genetic Algorithm.

4.3. Comparison to traditional methods

inductance and resistance, as depicted in Figs. 16(a) and 16(b). The To validate the efficiency of the proposed method under various
identification results in Fig. 16 demonstrate the observer’s capability actual inductance values, the validation was conducted with inductance
of effectively observing the true values of inductance and resistance. values of 7 mH, 8 mH, 9 mH, 10 mH, 11 mH, and 12 mH, and it is
The second part involves constructing a LUT based on offline ex- compared with the conventional FCS-MPC and the delay compensation
tremum optimization. The online inductance parameter observation methods, as depicted in Fig. 19.
results 𝐿𝑟 and amplitude of reference current value 𝐼𝑟𝑒𝑓 are used as Similarly, validation was also conducted under different 𝐼𝑟𝑒𝑓 , as
inputs to the LUT, then find the optimal inductance mismatch ratio depicted in Fig. 20. It is evident that the proposed twisted parameter
𝜆. The LUT obtained through fitting with RBF neural network and method can effectively reduce current THD, even lower than the THD
using GA to find extremum is partly shown in Table 2, where fitness obtained from the delay compensation method, without increasing too
values are current THD corresponding to the optimal mismatch ratio many switching transitions.

7
P. Li and X. Huo Electric Power Systems Research 238 (2025) 111063

Fig. 19. Comparison of current THD and the number of switching transitions to Fig. 22. Comparison of current waveform when 𝐼𝑟𝑒𝑓 is 12A: (a) Before and after
traditional methods under different 𝐿𝑟 : (a) Current THD; (b) Number of switching parameter identification; (b) Before and after parameter twisting.
transitions.

Fig. 20. Comparison of current THD and the number of switching transitions to
traditional methods under different 𝐼𝑟𝑒𝑓 : (a) Current THD; (b) Number of switching
transitions.

Fig. 23. Variation of the inductance parameter and the current THD when 𝐼𝑟𝑒𝑓 is 12 A.

Fig. 21. Experimental platform: (a) Block diagram; (b) Experimental setup.

Fig. 24. Comparison of current waveform when 𝐼𝑟𝑒𝑓 is 10A: (a) Before and after
5. Experiment parameter identification; (b) Before and after parameter twisting.

To substantiate the effectiveness of the proposed methods, we con-

ducted hardware-in-the-loop (HIL) experiments on a single-phase in- current waveform before and after twisting the inductance parameter
verter system. Fig. 21(a) presents a block diagram outlining the opera- is depicted in Fig. 22(b).
tional flow of the experimental platform, while Fig. 21(b) displays the As illustrated in Fig. 22, both parameter identification and twisting
actual experimental setup. The single-phase grid-tied inverter is simu- achieved the effects of reducing the current ripple and lowering the
lated on the Speedgoat performance real-time target machine SN8767. current THD. The variation of the inductance parameter and the current
The controller is implemented utilizing the Speedgoat performance THD is depicted in Fig. 23.
real-time target machine SN8530, and the IO324 programmable FPGA Similarly, the experimental verification was conducted with a refer-
modules are utilized to input and output analog and digital signals [36]. ence current value of 10 A. After inductance identification and twisting,
The experimental parameters are the same as Table 2. the changes in the current waveform are shown in Fig. 24, and the
At t=0 s, the model parameters of FCS-MPC are initialized with variation of THD with changes in inductance parameters is depicted in
nominal values of 10 mH for inductance and 0.5 Ω for resistance. At Fig. 25. Through parameter identification, the THD decreased from the
t=1 s, an observer started functioning to online identify the actual original 7.19% to 6.30%. Furthermore, after parameter twisting, the
values of the inductance and resistance, which are determined to be THD dropped from 6.30% to 5.33%.
8 mH and 1 Ω, respectively. These observed values are then used as The experimental verification was conducted with a reference cur-
input parameters for the controller. At t=2 s, the reference current rent value of 15 A as well. After inductance identification and twisting,
amplitude and the observed inductance parameter are fed into the LUT the changes in the current waveform are shown in Fig. 26, and the
that has been created offline. The LUT then calculated the optimal variation of THD with changes in inductance parameters is depicted in
inductance mismatch ratio and used this ratio along with the actual Fig. 27. Through parameter identification, the THD decreased from the
inductance value to determine the twisted inductance parameter. This original 4.71% to 4.23%. Furthermore, after parameter twisting, the
twisted inductance parameter is then used as the model parameter for THD dropped from 4.23% to 3.24%.
FCS-MPC. The current waveform before and after the identification of Based on both simulation and experimental results, it is evident that
resistance and inductance parameters is depicted in Fig. 22(a), and the reducing the inductance within a certain range, specifically setting the

8
P. Li and X. Huo Electric Power Systems Research 238 (2025) 111063

Table 3
Evaluation of the execution time of three control
methods.
Methods Critical path delay
Traditional FCS-MPC 258.665 ns
Delay compensation 301.105 ns
Twisted parameter 276.583 ns

6. Conclusion

6.1. What is the twisted parameter method?

• Twisting the model parameters rather than the actual circuit

parameters is the focus of the investigation;
Fig. 25. Variation of the inductance parameter and the current THD when 𝐼𝑟𝑒𝑓 is 10 A. • The grid-connected current THD is more influenced by inductance
than resistance, so this study exclusively examines the twisting of
the inductance parameter.

6.2. Why adopt the twisted parameter method?

• It can effectively reduce current THD;

• It will not increase too many switching transitions;
• Compared to traditional delay compensation method, it has less
computational complexity.

6.3. How to use the twisted parameter method?

Fig. 26. Comparison of current waveform when 𝐼𝑟𝑒𝑓 is 15A: (a) Before and after
parameter identification; (b) Before and after parameter twisting.
• Offline: Fitting the relationship between the mismatch ratio 𝜆 of
𝐿𝑟 , 𝐼𝑟𝑒𝑓 , and THD using RBF neural network, then using Genetic
Algorithm to search for 𝜆 that corresponds to the minimum THD.
Finally, the optimized result is made into a LUT;
• Online: Using an observer to identify the actual value 𝐿𝑟 and 𝑅𝑟 ,
then substitute 𝐿𝑟 into the LUT to find the optimal 𝜆. Using 𝜆
along with 𝐿𝑟 to determine the twisted inductance 𝐿𝑡 , which is
then used as the model parameter.

6.4. What are the future research directions for the twisted parameter
method?

• Establish the mathematical relationship between 𝐿𝑡 and current

THD;
• Validate the compatibility of the proposed methods with other
converter types.

CRediT authorship contribution statement

Fig. 27. Variation of the inductance parameter and the current THD when 𝐼𝑟𝑒𝑓 is 15 A.
Po Li: Writing – review & editing, Conceptualization. Xiaoxiao
Huo: Validation, Software, Methodology.
inductance mismatch ratio between 0.5–0.8, can effectively decrease
Declaration of competing interest
current THD and improve the grid-connected current waveform quality.
Finally, the computational complexity of the three predictive con-
The authors declare that they have no known competing finan-
trol methods is compared. The fastest achievable rate of an algorithm is
cial interests or personal relationships that could have appeared to
determined by the longest Register-to-Register path, which is referred
influence the work reported in this paper.
to as the critical path. The critical path serves to describe the longest-
executing part of an algorithm, reflecting the overall execution speed of Data availability
the entire algorithm. Through HDL compilation, the control algorithm
is deployed onto Speedgoat’s FPGA, and the HDL Coder is utilized to No data was used for the research described in the article.
generate a critical path report. The computational complexity of the
three predictive control methods is measured based on the critical path Acknowledgments
delay, with the results presented in Table 3.
Based on the results presented above, the FCS-MPC based on delay This work was supported by the Fundamental Research Funds for
compensation exhibits an increased computational burden due to the the Central Universities, China under grants (20720220071,
additional current calculation step compared to the traditional FCS- 20720220084). In addition, the authors would like to extend their
MPC. In contrast, the proposed twisted parameter method not only heartfelt appreciation to Dr. Wujun Feng from the School of Aerospace
reduces the current THD, enhancing power quality but also lowers Engineering at Xiamen University for his expert advice and hands-on
computational time than the delay compensation approach. assistance in establishing the experimental setup.

9
P. Li and X. Huo Electric Power Systems Research 238 (2025) 111063

References [19] Luca Tarisciotti, Pericle Zanchetta, Alan Watson, Stefano Bifaretti, Jon C.
Clare, Modulated model predictive control for a seven-level cascaded H-bridge
[1] P. Narendra Babu, Adaptive grid-connected inverter control schemes for power back-to-back converter, IEEE Trans. Ind. Electron. 61 (10) (2014) 5375–5383.
quality enrichment in microgrid systems: Past, present, and future perspectives, [20] Tao Rui, Zheng Yin, Cungang Hu, Geye Lu, Pinjia Zhang, Weixiang Shen,
Electr. Power Syst. Res. 230 (2024). Wenping Cao, Xinghuo Yu, Modulated model-free predictive current control for
[2] Youjun Zhu, Jinbin Zhao, Zhiwei Zeng, Ling Mao, Keqing Qu, SISO impedance voltage source inverters with stagnation elimination and sampling disturbance
modeling and stability comparison of grid-connected inverter control system in suppression, IEEE Trans. Power Electron. 38 (6) (2023) 6996–7008.
different time domains, Electr. Power Syst. Res. 228 (2024). [21] Weichao Wang, Naoto Yorino, Yutaka Sasaki, Yoshifumi Zoka, Ahmed Bedawy,
[3] K.S. Raja Sekhar, Madhuri A. Chaudhari, Tiago Davi Curi Busarello, A PLL-less Seiji Kawauchi, A novel adaptive model predictive frequency control using
vector control technique for the single-phase grid connected inverters, Int. J. unscented Kalman filter, Electr. Power Syst. Res. 213 (2022).
Electr. Power Energy Syst. 142 (2022) 108353. [22] Anian Brosch, Oliver Wallscheid, Joachim Böcker, Long-term memory recur-
[4] Seyyed Amirhosein Saadat, Seyyed Morteza Ghamari, Hasan Mollaee, Fatemeh sive least squares online identification of highly utilized permanent magnet
Khavari, Adaptive neuro-fuzzy inference systems (ANFIS) controller design on synchronous motors for finite-control-set model predictive control, IEEE Trans.
single-phase full-bridge inverter with a cascade fractional-order PID voltage Power Electron. 38 (2) (2022) 1451–1467.
controller, IET Power Electron. 14 (11) (2021) 1960–1972. [23] Leilei Guo, Zhiye Xu, Yanyan Li, Yafei Chen, Nan Jin, Falong Lu, An inductance
[5] Inas Jawad Kadhim, Murtadha Jasim Hasan, Enhancing power stability and ef- online identification-based model predictive control method for grid-connected
ficiency with multilevel inverter technology based on renewable energy sources, inverters with an improved phase-locked loop, IEEE Trans. Transp. Electrif. 8
Electr. Power Syst. Res. 231 (2024). (2) (2022) 2695–2709.
[6] Mohammad Sharifzadeh, Hani Vahedi, Kamal Al-Haddad, New constraint in SHE- [24] Po Li, Xiaoxiao Huo, Feng Guo, Total harmonic distortion reduction method of
PWM for single-phase inverter applications, IEEE Trans. Ind. Appl. 54 (5) (2018) improved finite control set model predictive control for single-phase inverter
4554–4562. with twisted parameter, in: 2023 5th International Conference on Power and
[7] Xingang Fu, Shuhui Li, Control of single-phase grid-connected converters with Energy Technology, ICPET, 2023, pp. 284–289.
LCL filters using recurrent neural network and conventional control methods, [25] Qinghua Jiang, Lailai Zhu, Chang Shu, Vinothkumar Sekar, An efficient mul-
IEEE Trans. Power Electron. 31 (7) (2015) 5354–5364. tilayer RBF neural network and its application to regression problems, Neural
[8] Derick Mathew, Rani Chinnappa Naidu, A review on single-phase boost inverter Comput. Appl. (2022) 1–18.
technology for low power grid integrated solar PV applications, Ain Shams Eng. [26] Sanpeng Zheng, Renzhong Feng, A variable projection method for the general
J. 15 (2) (2024). radial basis function neural network, Appl. Math. Comput. 451 (2023) 128009.
[9] Ramyani Chakrabarty, Ravindranath Adda, DSTATCOM implementation using [27] Honggui Han, Wei Lu, Ying Hou, Junfei Qiao, An adaptive-PSO-based self-
reduced switch single DC source cascaded H-bridge multilevel inverter, Electr. organizing RBF neural network, IEEE Trans. Neural Netw. Learn. Syst. 29 (1)
Power Syst. Res. 199 (2021). (2016) 104–117.
[10] Mahmoud S.R. Saeed, Wensheng Song, Li Huang, Bin Yu, Double-vector-based [28] Hong-Gui Han, Jun-Fei Qiao, Qi-Li Chen, Model predictive control of dissolved
finite control set model predictive control for five-phase PMSMs with high oxygen concentration based on a self-organizing RBF neural network, Control
tracking accuracy and DC-link voltage utilization, IEEE Trans. Power Electron. Eng. Practice 20 (4) (2012) 465–476.
37 (12) (2022) 15234–15244. [29] Wu Deng, Huimin Zhao, Li Zou, Guangyu Li, Xinhua Yang, Daqing Wu, A novel
[11] Xing Liu, Lin Qiu, Youtong Fang, Kui Wang, Yongdong Li, Jose Rodriguez, collaborative optimization algorithm in solving complex optimization problems,
A fuzzy approximation for FCS-MPC in power converters, IEEE Trans. Power Soft Comput. 21 (2017) 4387–4398.
Electron. 37 (8) (2022) 9153–9163. [30] Hossam A. Abd el-Ghany, Optimal PMU allocation for high-sensitivity wide-area
[12] Teng Li, Xiaodong Sun, Gang Lei, Zebin Yang, Youguang Guo, Jianguo Zhu, backup protection scheme of transmission lines, Electr. Power Syst. Res. 187
Finite-control-set model predictive control of permanent magnet synchronous (2020).
motor drive systems—an overview, IEEE-CAA J. Automatica Sin. (2022). [31] Cheng Lu, Yunwen Feng, Rhea P. Liem, Chengwei Fei, Improved kriging
[13] Yongshu Li, Weiwei Sun, Dehai Yu, An improved three-vector coordination robust with extremum response surface method for structural dynamic reliability and
model predictive control for 3P-2L inverters, Electr. Power Syst. Res. 229 (2024). sensitivity analyses, Aerosp. Sci. Technol. 76 (2018) 164–175.
[14] Mahum Pervez, Tariq Kamal, Luis M.M. Fernandez-Ramirez, A novel switched [32] Claudio Urrea, John Kern, Exequiel Álvarez, Design of a generalized dynamic
model predictive control of wind turbines using artificial neural network-Markov model and a trajectory control and position strategy for n-link underactuated
chains prediction with load mitigation, Ain Shams Eng. J. 13 (2) (2022). revolute planar robots, Control Eng. Practice 128 (2022) 105316.
[15] Benfei Wang, Jingjing Huang, Changyun Wen, Jose Rodriguez, Cristian Garcia, [33] Sourabh Katoch, Sumit Singh Chauhan, Vijay Kumar, A review on genetic
Hoay Beng Gooi, Zheng Zeng, Event-triggered model predictive control for power algorithm: past, present, and future, Multimedia Tools Appl. 80 (2021)
converters, IEEE Trans. Ind. Electron. 68 (1) (2020) 715–720. 8091–8126.
[16] Bingtao Zhang, Weimin Wu, Yong Yang, Ning Gao, Jianming Chen, Eftichios G. [34] Po Li, Ruiyu Li, Haifeng Feng, Ying He, Jingrui Zhang, Improvement of model
Koutroulis, Henry Shu-Hung Chung, Marco Liserre, Frede Blaabjerg, A novel predictive control for single-phase inverters by using sinusoidal signal observers,
simplified finite control set repeat model predictive control for grid-connected IET Power Electron. 13 (17) (2020) 3841–3850.
inverters, IEEE Trans. Ind. Electron. 70 (11) (2023) 11324–11333. [35] Yunfeng She, Xiaoxiao Huo, Xiaoshan Tong, Chunjie Wang, Kunkun Fu, Multi-
[17] Patricio Cortes, Jose Rodriguez, Cesar Silva, Alexis Flores, Delay compensation sampling rate finite control set model predictive control and adaptive method
in model predictive current control of a three-phase inverter, IEEE Trans. Ind. of single-phase inverter, Electronics 12 (13) (2023) 2848.
Electron. 59 (2) (2011) 1323–1325. [36] Zhongqing Sang, Shaojie Li, Yuanyuan Huang, Xin Gao, Rui Qiao, Franc Mihalic,
[18] Galina Mirzaeva, Graham Goodwin, Christopher Townsend, Dealing with linear Eric Monmasson, Lahoucine Idkhajine, Miro Milanovic, Indirect matrix converter
and nonlinear time delays under model predictive control of power electronic hardware-in-the-loop semi-physical simulation based on latency-free decoupling,
inverters, in: 2016 IEEE International Conference on Automatica (ICA-ACCA), Electronics 12 (23) (2023) 4802.
IEEE, 2016.

10