3.1. Pairing Analysis for the SIMO Converter
The first step in the control design process is to perform a pairwise analysis to determine which output is most affected by a given plant input. There are several ways to perform pairwise analysis on this topic. Nevertheless, the effective relative gain array (ERGA) is easy to implement and understand, and is more efficient than other arrays found in the literature, e.g., relative gain array (RGA) or dynamic relative gain array (DRGA) [
20].
The ERGA method is commonly used to determine the sensitivity of each output to each input. This procedure is provided by [
37,
39]. The main steps of the ERGA-based pairwise analysis are as follows:
- (1)
Determine the DC gain array using the DC gain for each element of the transfer function matrix.
- (2)
Obtain the cutoff frequency for each element of the transfer function matrix.
- (3)
Compute the energy transmission ratio array based on the DC gain and cutoff frequencies.
- (4)
Compute the effective gain matrix that indicates the sensitivity between each I/O pair.
- (5)
Compute the ERGA number and check for diagonal head dominance.
The effective gain matrix shown in (9) as a result of the ERGA pair analysis shows the optimal pairing, i.e., the first (V
1) and the second output (V
2) are, respectively, more sensitive to the first (d
1) and the second input (d
2).
Therefore, a simulation evaluation is performed to check the coupling between the system outputs. So, first a pulse is applied to the duty cycle of the switch Q1 and the system is run in open loop (without any controller) mode, and then the same process is repeated for the switch Q2.
Figure 2 shows the resulting coupling for the SIMO converter test system. The results show that V
2 is very sensitive to changes in duty cycle d
2, while V
1 is very sensitive to changes in duty cycle d
1. However, V
1 and V
2 are greatly affected by changes in d
1, although changes in d
2 do not have much relevance in the outputs of the system. This finding reflects the coupling between the loops of the system, which shows the efficient pairing of the ERGA analysis.
3.2. Linear Quadratic Regulator Design
The LQR control is a well-known optimal-based approach to state feedback controller that can solve some problems in systems like instability or low stability margins, manipulating the system inputs and avoiding some typical problems on the stabilization problem like saturation control signal and fast actuator degradation [
40]. This controller design is based on the choice of the weight matrices of the states and control signal, Q and R, respectively, which is used in the minimizing cost function.
The minimization of the cost function results in a feedback control law.
Positive-definite matrix “S” is a solution of the discrete-time algebraic Ricatti equation.
In this structure, the LQR control on the closed loop has not guaranteed perturbation rejection in low frequency; however, this problem can be solved by an integrator addition on system input, creating a different model called “velocity model” [
40]. The discrete integrator
is inserted on the system plant, by this way creating a new state variable ∆x(k).
Calling the new state vector
and applying it in the incremental model, the velocity model is presented.
Inserting the augmented space state in the Ricatti transforms the feedback control law structure on a rejection perturbation problem.
Then, the control law of the digital LQR for stabilization and perturbation rejection problem is presented.
The correct selection of Q and R weight matrices is important because it is responsible for the penalization of the dynamic states and input signal, respectively, in a way that Q and R inform the priority minimization order in the cost function, giving a trade-off balance between conflicting control objectives, such as minimizing control effort and regulation [
41,
42]. In this work, this selection is based on the loop-shaping of process singular values for improving robustness in high and low frequencies where stability Phase Margin, PM, and Gain Margin, GM, are used in the robustness design, and bandwidth frequency is used in the converter velocity. This method is an interesting solution to the controller’s synthesis for multiple input multiple output process because of the difficulty in the frequency analysis in each loop [
43]. For the analysis of minimal stability margins (Phase Margin and Gain Margin) in the frequency singular values, the following constraints formulation, proposed in [
44], is used:
where
and
, with
meaning the maximum singular value of a given transfer function matrix.
For LQR design purposes, assuming that the SIMO converter operates in a steady state around the operation point:
0.485 A and
= 7 V. Then, the corresponding matrices A, B, C, and D of the linearized state-space model are as follows:
By using the A, B, C, and D matrix values given by (20), the corresponding open-loop transfer function matrix
has been computed. For the discrete system model with sample time (ts = 0.5 ms), the resulting maximum singular values for the sensitivity and complementary matrices transfer functions, S(z
−1) and T(z
−1), for the uncompensated closed-loop system (i.e., with an initial controller matrix K
0 = I
2×2) are presented in
Figure 3.
As can be seen, the maximum singular values for the sensitivity and complementary sensitivity transfer function matrices show very large peaks in high frequency, meaning that the uncompensated converter plant is very sensitive to high-frequency noise and could become unstable due low attenuation property in high-frequency band. Another problem in this uncompensated plant is its low capacity in tracking reference setpoint due to the reduced gain presented by T(s) at the low-frequency band. For this uncompensated system, the computed margin index shows low stability margins and , which confirms the singular value analysis.
For this case, low stability margins could be a problem, because uncertainty in circuit parameters (e.g., capacitors, inductors, and resistors) and in the input voltage are very common. Other problems are associated with noisy measurement (high frequency) and perturbations in low frequency. Aiming at improving the stability margins, the initial specified design requirements were as follows:
,
and
, where
is the angular frequency such that the modulus of T(s = jω) is −3 dB. By using the augmented state-space matrices given by (15), along with choosing weighting states and control matrices,
and
, in LQR cost function (10), the computed value for the control vector gain was as follows:
The resulting maximum singular values for the sensitivity and complementary matrix transfer functions S(z
−1) and T(z
−1), for the compensated closed-loop system, are presented in
Figure 4.
As can be seen, in
Figure 4, the computed controller is able to decrease the sensibility peaks which improves the process robustness. Furthermore, the perturbation rejection is also improved due to decreasing sensibility gain values in the low-frequency band. For this compensated system, the computed minimal stability margins and bandwidth frequency of the system are
,
, and
, which satisfy the design specifications.
3.3. Decoupled Multiloop PI Controller
The presentation is limited to two-input, two-output systems. Our approach was to explore the standard PI multivariable (multi-loop) tuning [
45] and observe what can be achieved by adding simple interactions between the feedback loops. The proposed scheme is based on a simple decoupling, which means that it can be easily implemented at the loop level [
46]. The advantage is that it improves performance in frequency ranges that modeling predictive control typically cannot handle.
The designable controller is a static decoupler combined with a decentralized PI controller with setpoint weighting. The control law can be written as follows:
where U is the control signal, Y the process output, and Y
r the reference. The decoupler:
is a constant matrix. The PI controller
is different from
to allow for setpoint weighting [
43]. The controllers are of the form:
The static decoupler is given by:
where G(0) is non-singular. The transfer function of the decoupled system is given by:
A Taylor series expansion of the transfer function Q(s), for a small s, gives the following:
for some constants k
12 e k
21. Then, it is possible to introduce the interaction indices:
where M
s1 and M
s2 are the maximum sensitivities of each loop. The indices k
1 and k
2 describe the interaction between the loops. The indices are the result of two terms: one is system-dependent and the other is directly the integral gain of the PI controller. Therefore, the interaction can be reduced by decreasing the controller gains [
41]. To find the decentralized PI controllers, we need to consider the diagonal terms of Q(s). Hence, standard methods can be used to design the PI controllers for each transfer function q
kk(s).
By considering the same operating point used for the design of the LQR controller (
0.485 A, and
= 7 V) the following values were computed for the transfer functions
:
Thus, it is possible to determine the matrix static decoupler D by using (27), the computed value obtained of D was as follows:
The transfer function of the decoupled system can be obtained considering Equation (28) to Equation (31); hence, to design the decentralized PI controllers, we have to consider the diagonal terms of Q(s):
Therefore, for each transfer function
and
, the root-locus design method for the PI controllers was used with the chosen requirements: settling time less than 0.5 s and damping factor greater than 0.2 (these nominal specifications give reasonable performance for DC-DC power converters [
47]), and the corresponding parameters were obtained as follows:
Thus, the corresponding transfer functions of the controllers PI
1 and PI
2 have the structures of Equations (25) and (26). Thus, the corresponding state-space representation is as follows:
where
Thereby, the control law equation can be represented like
where