Interaction Analysis

for
Multi-Loop Control
Multivariable Process Control
• Industrial Processes
• multivariable (multiple inputs influence same output)
and exhibit strong interaction among the variables
• Conventional Control scheme
• Multiple Single Input Single Output PID controllers used
for controlling plant (Multi-Loop Control)
• Consequences
• Loop Interactions
• Lack of coordination between different PID loops
• Neighboring PID loops can co-operate with each other
or end up opposing / disturbing each other
Shell Control Problem
Shell Control Problem
• Control Scheme 1
• PID-1: (y1) Top End Point - (u1) Top Draw
• PID-2: (y2) Side Endpoint - (u2) Side Draw
• PID-3: (y3) Bottom Reflux Temperature – (u3) Bottom
Reflux Duty
• Control Scheme 2
• PID-1: (y2) Side End Point - (u1) Top Draw
• PID-2: (y1) Top Endpoint - (u2) Side Draw
• PID-3: (y3) Bottom Reflux Temperature –(u3) Bottom
Reflux Duty
How to examine above options systematically and reach
a decision ?
Tennesse Eastman Problem

Primary Controlled Variable: G conc in Product and Product flow rate

Input - Output Variables
Loop Interactions

• Large loop interactions :

• can lead to poor quality of control due to lack of
coordination among PID controllers

• Solution strategy:
• Choose controller pairing with minimal interactions
• De-tune the controllers to minimize loop interactions
• Design multi-variable controllers, which simultaneously
Change all inputs by considering errors in all the outputs
Quadruple Tank Process

Schematic of Quadruple Tank Process

Block Diagram : Quad Tank
Loop Interactions : Quad Tank
Interaction Analysis
Assume a loop pairing say y1-u1 and perform the
following experiments
• With all loops open, make a step change in u1 to u1+ Δu
and measure the change in output Δy1.
• We will term this as a direct effect.
• With all loops except the u1-y1 loop closed, repeat the
same change in u1.
• There will be change in y1 because of the direct effect
but also there will be a retaliatory effect because u2
changes to keep y2 constant.
• We will term this change as Δy1+Δy1r
Interaction Analysis
• Ratio of these two terms can be defined as λ11 (for the
y1-u1 pairing) as
λ11 = Δy1/( Δy1+Δy1r)
This is called relative gain
• Compute relative gain for each assumed input/output
pairing
• Depending on the values of this index for various
assumed loop pairings (step 1), decision can be taken
on the final loop pairing.
• This decision making is based on steady state analysis
only.
Relative Gain Array (RGA)
Calculation of RGA
Calculation of RGA
Calculation of RGA
Wood-Berry Column
Wood-Berry Column
Wood-Berry Column
Multi-Loop PID Control
Change in closed loop behavior due to loop interactions
RGA calculation for MIMO Process
MIMO RGA Calculation
Analysis of RGA
• If λ11 =1 ⇒ retaliatory action is not present. So
assumed loop pairing is correct because there is no
interaction from the other loop.

• If 0 < λ11 < 1 ⇒ retaliatory action is comparable to

the direct action but is in the same direction. The
assumed loop pairing may be chosen only if the
index is closer to 1 (say 0.8).

• If λ11 =0 ⇒ Retaliatory action is much greater than

the direct action. The assumed loop pairing is
incorrect. The loop pairing u1-y2 is preferable
 𝐾𝐾𝑑𝑑
• 𝜆𝜆 =
𝐾𝐾𝑑𝑑 +𝐾𝐾𝑟𝑟
• 𝜆𝜆 = 1 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 𝐾𝐾𝑟𝑟 = 0
• 𝜆𝜆 < 1 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 𝐾𝐾𝑟𝑟 𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓
• 𝜆𝜆 > 1 𝑤𝑤𝑤𝑤𝑤𝑤𝑤 |𝐾𝐾𝑑𝑑 | > |𝐾𝐾𝑟𝑟 | but direction of Kr is opposite
to Kd
• Control objective : m,x by manipulating m1 and m2
• Desired value for m and x is m* and x*. Find RGA
m1,x1
m,x

m2,x2
Analysis of RGA
• If λ11 > 1 ⇒ Retaliatory action is in opposite
direction to the direct action but is smaller in
magnitude than the direct. The assumed loop
pairing may be chosen only if the index is close to 1.

• If λ11 < 0, Retaliatory effect is larger and opposite in

direction to the main effect. Do not choose this
loop pairing.
Suggested Loop Pairing
Refinery Distillation Column
RGA analysis : Non-square System
Singular Value Analysis
Powerful analytical tool for
• Selection of controlled, measured and manipulated
variables
• Determination of best multi-loop configuration
• Evaluation of robustness (insensitivity to changes in
plant behavior) of a control scheme
Singular Value Analysis
• Non-square systems (number of inputs not equal to
number of outputs)
• SVA can be used to find a square subset with least
difficulties in control

• Larger condition number implies difficulties in

controlling a system
• Among multiple possibilities, choose subset with
minimum condition number

• Limitation: Singular values are dependent on scaling of

input and output variables
Singular Value Decomposition
K = W Σ VT
Σ is diagonal matrix of singular values
(σ1, σ2, …, σr)
The singular values are the positive square roots of the
eigenvalues of
KTK (r = rank of KTK)
W,V are input and output singular vectors Columns of W and
V are orthonormal. Also
WWT = I
VVT = I
Calculate Σ, W, V using MATLAB (svd = singular value
decomposition)
Condition number (CN) is the ratio of the largest to the
smallest singular value and indicates if K is ill-conditioned.
Singular Value Analysis
SVD of K : 𝐾𝐾 = 𝑊𝑊 Σ 𝑉𝑉 𝑇𝑇
Pairing rule:
Output associated with the largest magnitude element
(without regard to sign) of the W1 vector with the
manipulated variable associated with the largest magnitude
element (without regard to sign) of the V1 vector. Same logic
will apply for W2 – V2 … Wn – Vn vectors.
Alternate Method
• Arrange the singular values in order of largest to smallest
and look for any σi/σi-1 > 10; then one or more inputs (or
outputs) can be deleted.
• Delete one row and one column of K at a time and
evaluate the properties of the reduced gain matrix.
Example
 0.48 0.90 −0.006   −2.4376 3.0241 0.4135 
K =  0.52 0.95 0.008  
RGA=  1.2211 −0.7617 0.5407 


 0.90 −0.95 0.020   2.2165 −1.2623 0.0458 

 0.5714 0.3766 0.7292  1.618 0 0 

W  0.6035 0.4093 −0.6843 ∑ = 0 1.143 0 
 
 −0.5561 0.8311 0.0066   0 0 0.0097 

 0.0541 0.9984 0.0151

V=  0.9985 −0.0540 −0.0068
 −0.0060 0.0154 −0.9999 
Furnace Control
4-Component Distillation Column
Niederlinsky Index
• Consider MIMO system whose inputs and outputs are
paired as 𝑦𝑦1 − 𝑢𝑢1 , 𝑦𝑦2 − 𝑢𝑢2 , … … … , 𝑦𝑦𝑛𝑛 − 𝑢𝑢𝑛𝑛 i.e, Transfer
Function matrix G(s) is arranged such that transfer functions
relating paired inputs and outputs are arranged along the
diagonal.
• Further, let each element of G(s) be (a) rational and (b) open
loop stable.
• Also, let n SISO feedback controllers with integral action be
designed such that each SISO loop is stable when all the rest
( n-1) loops are open.
• Under these assumptions, the multi-loop control system will
be unstable for all possible values of controller parameters if
det[𝐺𝐺 0 ]
the Niederlinsky Index (NI) defined as 𝑁𝑁𝑖𝑖 = 𝑛𝑛 <0
�𝑖𝑖=1 𝐺𝐺𝑖𝑖𝑖𝑖 (0)
Furnace Control
Multi-Loop PID Control
• After selection of loop pairings with minimum
interactions, one can design controllers for
individual loops.

• Presence of interaction and retaliatory effects from

other loops may require that the controller be
detuned for acceptable performance.
• BLT Detuning method
• Sequential Loop Tuning
• Independent Loop Tuning
Biggest Log Modulus Tuning (BLT)
SISO Loop Design Review:
• Characteristic Eqn. 1 + 𝑔𝑔𝑐𝑐 𝑠𝑠 𝑔𝑔𝑝𝑝 𝑠𝑠 = 1 + 𝐺𝐺𝑂𝑂𝑂𝑂 𝑠𝑠 = 0
• Nyquist Plot: depicts real part of GOL(s) on X-axis and
imaginary part of GOL(s) on Y-axis as 𝜔𝜔 → ∞
Nyquist Stability Criteria:
• A feedback control system will be unstable if the
Nyquist plot of G(j𝜔𝜔) encircles point (-1,0) as 𝜔𝜔 → ∞
• The number of encirclements correspond to the
number of roots of the characteristic equation that lie
in R.H.P. of s –plane assuming that the process is open
loop stable.
BLT Method
A measure of distance of GOL(j𝜔𝜔) contour from (-1,0) is
given as
𝐺𝐺𝑂𝑂𝑂𝑂 𝑠𝑠
𝐿𝐿𝑐𝑐 𝑠𝑠 = 20 log
1 + 𝐺𝐺𝑂𝑂𝑂𝑂 𝑠𝑠
Suggested design specification for Log Modulus: 𝐿𝐿𝑚𝑚𝑚𝑚𝑚𝑚 𝑐𝑐 ≤ 2 𝑑𝑑𝑑𝑑
Log Modulus Design:
Iteratively choose PI controller parameters 𝐾𝐾𝑐𝑐 , 𝜏𝜏𝐼𝐼 such that the
design specification is met.

Multivariable Process :
−1
𝑌𝑌 𝑠𝑠 = 𝐼𝐼 + 𝐺𝐺𝑐𝑐 𝑠𝑠 𝐺𝐺𝑝𝑝 𝑠𝑠 𝐺𝐺𝑝𝑝 𝑠𝑠 𝐺𝐺𝑐𝑐 𝑠𝑠 𝑅𝑅(𝑠𝑠)
Characteristic Eqn. : det 𝐼𝐼 + 𝐺𝐺𝑐𝑐 𝑠𝑠 𝐺𝐺𝑝𝑝 𝑠𝑠 =0
BLT Method
Define 𝑓𝑓 𝑠𝑠 = −1 + det 𝐼𝐼 + 𝐺𝐺𝑐𝑐 𝑠𝑠 𝐺𝐺𝑝𝑝 𝑠𝑠
Encirclement of (-1,0) by f(jω) would indicate instability.
Define a multivariable closed loop log modulus as
𝑚𝑚
𝑓𝑓(𝑠𝑠)
𝐿𝐿𝑐𝑐 = 20 log
1 + 𝑓𝑓(𝑠𝑠)
Suggested design specification for Log Modulus: 𝐿𝐿𝑚𝑚 𝑐𝑐 ≤ 2𝑛𝑛 𝑑𝑑𝑑𝑑 for n-
dimension system
Tuning Procedure
• Calculate Ziegler -Nichol's tuning for n individual PI controllers
• Assume a factor F such that 2 ≤ F ≤ 5
• De - tune PI controllers as follows:
• 𝐾𝐾𝑐𝑐,𝑗𝑗 = 𝐾𝐾𝑗𝑗𝑍𝑍𝑍𝑍 /𝐹𝐹 𝜏𝜏𝐼𝐼, 𝑗𝑗 = 𝜏𝜏𝐼𝐼𝑍𝑍𝑍𝑍 𝐹𝐹 for j=1,2,.. N
• Iteratively choose F such that criteria Lcm = 2n is satisfied.
BLT: Example
BLT: Example
BLT: Example
Principles of Decoupling
Simplified Decoupling
Simplified Decoupler Design
Example: Wood Berry Distillation Column
Difficulties in simplied Decoupler design
Generalized Decoupler Design
Example: Wood Berry Distillation Column
Steady State Decoupler Design
Example: Wood Berry Distillation Column
Simplified Vs Generalized Decoupler
Challenges in Decoupler Design
Challenges in Decoupler Design
Example Problem
Example Problem
Alternate Solution
Partial Decoupling
Partial Decoupling Example