Brigham Young University

BYU ScholarsArchive

Faculty Publications

2020-04-06

Benchmark Temperature Microcontroller for Process Dynamics

and Control
Junho Park
Brigham Young University

Ronald Abraham Martin

Brigham Young University

Jeffrey Kelly
Brigham Young University

John Hedengren
Brigham Young University, [email protected]

Follow this and additional works at: https://scholarsarchive.byu.edu/facpub

Part of the Chemical Engineering Commons

Original Publication Citation

https://www.sciencedirect.com/science/article/abs/pii/S0098135419310129

BYU ScholarsArchive Citation

Park, Junho; Martin, Ronald Abraham; Kelly, Jeffrey; and Hedengren, John, "Benchmark Temperature
Microcontroller for Process Dynamics and Control" (2020). Faculty Publications. 4166.
https://scholarsarchive.byu.edu/facpub/4166

This Peer-Reviewed Article is brought to you for free and open access by BYU ScholarsArchive. It has been
accepted for inclusion in Faculty Publications by an authorized administrator of BYU ScholarsArchive. For more
information, please contact [email protected].
 Benchmark Temperature Microcontroller for Process
Dynamics and Control

Junho Parka , R. Abraham Martina , Jeffrey D. Kellyb , John D. Hedengrena,∗

a Department of Chemical Engineering, Brigham Young University, Provo, Utah, USA
b Industrial Algorithms, 15 St. Andrews Road, Toronto, ON, Canada, M1P 4C3

Abstract

Standard benchmarks are important repositories to establish comparisons be-

tween competing model and control methods, especially when a new method is
proposed. This paper presents details of an Arduino micro-controller tempera-
ture control lab as a benchmark for modeling and control methods. As opposed
to simulation studies, a physical benchmark considers real process characteris-
tics such as the requirement to meet a cycle time, discrete sampling intervals,
communication overhead with the process, and model mismatch. An example
case study of the benchmark is quantifying an optimization approach for a PID
controller with 5.4% improved performance. A multivariate example shows the
quantified performance improvement by using model predictive control with a
physics-based model, an autoregressive time series model, and a Hammerstein
model with an artificial neural network to capture the static nonlinearity. These
results demonstrate the potential of a hardware benchmark for transient mod-
eling and regulatory or advanced control methods.
Keywords: benchmark, dynamics, PID tuning, model predictive control,
microcontroller

∗ correspondingauthor
Email address: [email protected] (John D. Hedengren)

Preprint submitted to Computers and Chemical Engineering October 1, 2019

1 1. Introduction

2 Benchmark problems are standard repositories in many scientific disciplines

3 such as systems biology [1, 2], reservoir modeling [3, 4, 5, 6], drilling [7, 8],
4 optimization [9, 10], dynamic optimization [11, 12], singular optimal control
5 [13, 14], combined scheduling and control [15, 16, 17, 18], and others [19, 20,
6 21]. The benchmark problems serve as a consistent measure of innovations
7 that are proposed to increase profitability or improve some aspect of control or
8 optimization performance.
9 There are many standard benchmark models for testing the performance
10 of estimation and control methods in chemical process control. Some of these
11 include a continuously stirred tank reactor (CSTR) with a single exothermic re-
12 action [22, 23, 24]. One of the most commonly cited models in chemical process
13 control is the Tennessee Eastman Process [25, 26]. The Tennessee Eastman Pro-
14 cess encapsulates valve characteristics, measurement noise, process nonlinearity,
15 and complex interactions between processing units for chemical manufacture.
16 Besides simulation, there are standard hardware benchmarks for evaluat-
17 ing control performance such as UAV control [27], process control education
18 modules [28, 29], and quadruple tank level control [30, 31, 32]. There also
19 many studies where the authors build a unique test system or implement con-
20 trol on an industrial process [33, 34] and demonstrate various control methods.
21 However, hardware benchmarks may be difficult to reproduce or the industrial
22 process may be unavailable for independent researchers to also obtain data or
23 test methods in closed-loop.
24 The purpose of this paper is to demonstrate a standard hardware benchmark
25 for control methods with a micro-controller temperature control device. This
26 Temperature Control Lab (TCLab) is used as an education module for courses
27 in process dynamics and control [35, 36]. As many have noted in assessments of
28 process control education, there is a need to give students realistic and hands-
29 on experiences with process control [37, 38, 39]. Industry desires foundational
30 and practical knowledge of control engineering concepts that are reinforced with

2
31 physical modules. Because the TCLab, as an educational module, has wide dis-
32 tribution to universities and industrial practitioners (3000 units), it has potential
33 as a standard hardware benchmark for control engineering studies. Section 2
34 gives details of the device to enable replication of the TCLab.

35 2. Temperature Control Lab Device

36 The TCLab is printed circuit board (PCB) shield that connects to an Ar-
37 duino micro-controller. The TCLab shield has two transistors as heaters and
38 two thermistor temperature sensors as shown in Figure 1. A step response of
39 the heater (0-100%) has a temperature response with an approximate dominant
o
C
40 time constant (τ ) of 2.9 min and a gain of 0.9 %heater . The process exhibits
41 second order dynamics and the two adjacent heaters create a compact multi-
42 variate control system. The Arduino micro-controller is an Arduino Uno or
43 Arduino Leonardo that includes a 10-bit Analog to Digital Converter (ADC) to
44 measure voltage of the temperature sensors in 1024 (210 ) discrete analog levels
45 and Pulse Width Modulation (PWM) with 256 (28 ) levels to change the output
46 to the heaters and LED.

Figure 1: Temperature sensors and heater transistors with connections to an Arduino

Leonardo.

3
47 The transistor heaters are TIP31C NPN Bipolar Junction Transistors (BJTs)
48 in a TO-220 package. These transistors are commonly used in audio, power, and
49 switching applications but not commonly as heaters. During the development
50 of the TCLab, the initial design was to include a MOSFET transistor (low
51 power loss switch) with a power resistor as the heating element. Instead, the
52 BJT TIP31C is able to act as both the switch and the heater, thereby simpli-
53 fying the design and reducing the cost of the hardware. The two temperature
54 sensors on the TCLab are standard TMP36GZ thermistors with an output volt-
55 age (mV ) that is linearly proportional to temperature (T o C = 0.1 mV − 50)
56 and no requirement for calibration. Typical sensor accuracy is ±1o C at room
57 temperature (25o C) and ±2o C over the −40o C to 150o C operating range.
58 As a safety and equipment protection precaution, the Arduino micro-controllers
59 come pre-programmed to shut off the heaters if the temperature rises above
60 100o C. The heaters are powered by a 5V 2A power supply for a maximum
61 power output of 10 W. A 20 AWG (American Wire Gauge) power cable reduces
62 the power dissipation compared to standard 24 AWG power cables with a barrel
63 jack connector. A USB cable connects the Arduino to a computer for serial data
64 communication. One TIP31C heater and one TMP36GZ sensor are connected
65 to each other and with a thermal heat sink attached to the TIP31C transistor.
66 The two heater units are placed in proximity to each other to transfer heat by
67 convection and thermal radiation.
68 Software interfaces to TCLab in Python, MATLAB, and Simulink are de-
69 scribed in Appendix A. The software adjusts the two heater levels between 0
70 and 100% and the LED brightness between 0 and 100% using PWM with 28
71 discrete levels. The PWM rapidly fluctuates between on and off to give nearly
72 continuous values 0, 0.392, 0.784, . . . , 99.61, 100 for actuation of the heaters and
73 LED.

4
 (a) TCLab Printed Circuit Board Layout (b) TCLab Device

Figure 2: Temperature Control Lab Design

74 3. Temperature Response Models

75 This section summarizes four simulation models that describe the dynamic
76 response of the heaters to temperature changes. The four are a lumped param-
77 eter energy balance (Section 3.1), a first-order plus dead-time (FOPDT) model
78 (Section 3.2), a higher-order autoregressive exogenous input (ARX) model (Sec-
79 tion 3.3), and an artificial neural network (ANN) steady state and linear dy-
80 namic Hammerstein model (Section 3.4). Section 3.5 compares all of the models
81 on open-loop step test data both for Single Input Single Output (SISO) and Mul-
82 tiple Input Multiple Output (MIMO) modes. Multivariate, model-based control
83 relies on an accurate simulation of the process. The models described in this
84 section are not an exhaustive list of physics-based and empirical representations.
85 Each TCLab device is slightly different so the model parameters are uniquely
86 identified. One of the principal differences is the ambient temperature where
87 the test occurs. Other potential disturbances include the power supply output,
88 air currents (e.g. nearby computer fan), and others. Figure 3 shows variability
89 due to ambient temperature differences for six tests that use the same heater
90 profile. With ±2.5o C ambient temperature difference, there is a similiar spread
91 in the heater temperature response although the trends are not parallel and
92 completely predictable, especially for heater 2 temperature.

5
 100
Heater (%) 80
60
40
20 Heater 1
Heater 2
0
0 100 200 300 400 500 600

60
T1 (oC)

40

20
0 100 200 300 400 500 600
60 Ta=22.5oC
Ta=21.0oC
50 Ta=20.8oC
Ta=19.6oC
T2 (oC)

40
Ta=19.2oC
30 Ta=17.5oC

20
0 100 200 300 400 500 600
Time (sec)

Figure 3: Variations in ambient temperature influence the temperature profiles

93 Along with measurement noise, the stochastic nature of the data is a feature
94 of the lab that portrays performance on a physical system. Reporting, plotting,
95 or controlling the starting (ambient) temperature is an important requirement
96 of the benchmark as shown in Figure 4.
97 According to the slope of the regression, an ambient temperature increase of
98 1o C equates to a 0.928 ±0.033o C rise in average temperature of the step tests.
99 One possible explanation for the slope less than unity is the radiative heat
100 transfer that has a quadratic dependence on absolute temperature and would
101 lose heat at a higher rate at elevated conditions. The main conclusion from this
102 result is that ambient temperature has a reproducible effect on the outcome of
103 benchmark tests and should be reported and controlled for repeatable results.

104 3.1. Physics-based Model

105 A lumped parameter model with convection, conduction, and thermal radi-
106 ation describes the second-order temperature response to heater changes. The
107 lumped parameter model is a simplification of a more rigorous finite element

6
 54
T = m Ta + b
95% Confidence Region
52 95% Prediction Band
Average Step Test Data
Avg Temp (oC)

50

48

46 m = 0.928 ± 0.033
b = 30.3 ± 0.7

44
16 17 18 19 20 21 22 23 24
Ambient Temp (oC)

Figure 4: Correlation of ambient temperature to average temperature during 60 step tests (10
min each)

108 analysis (FEA) that tracks the temperature distribution throughout the heat
109 sink and loss to the environment as shown in Figure 5.
110 Details of the FEA simulation are not provided here but do provide a confir-
111 mation that the temperature distribution is sufficiently uniform (< 3o C) for a
112 lumped parameter assumption. The lumped parameter model assumes that the
113 heaters (TH1 and TH2 ) and temperature sensors (TC1 and TC2 ) have a uniform
114 temperature. The temperature sensors (TC1 and TC2 ) have a small thermal
115 mass and surface area and temperature changes are driven by heat conduction
116 from the heaters (TH1 and TH2 ) where they are attached with thermal epoxy.
117 Parameters of the lumped parameter model are given in Table 1.
118 The dynamic input power to each transistor and the temperature sensed
119 by each thermistor is developed with energy balance equations (Equations 1-4)
120 that account for convection, conduction, and thermal radiation. The amount
121 of convective heat transfer from heater 1 to heater 2 is given by QC12 =

7
 Figure 5: Finite Element Analysis of the Dynamic Temperature Response.

Table 1: Lumped Parameters from Physics-based Model

Quantity Value
Initial temperature (T0 ) 296.15 K (23o C)
Ambient temperature (T∞ ) 296.15 K (23o C)
Heater output (Q1 ) 0 to 1 W
Heater factor (α1 ) 0.0131-0.0132 W/(% heater)
Heater output (Q2 ) 0 to 0.75 W
Heater factor (α2 ) 0.0063-0.0066 W/(% heater)
Heat capacity (Cp ) 500 J/kg-K
Surface Area Not Between Heaters (A) 1.0x10−3 m2 (10 cm2 )
Surface Area Between Heaters (As ) 2x10−4 m2 (2 cm2 )
Mass (m) 0.004 kg (4 gm)
Heat Transfer Coefficient (U ) 4.4-4.6 W/m2 − K
Heat Transfer Coefficient Between Heaters (Us ) 23.6-24.4 W/m2 − K
Emissivity () 0.9
−8
Stefan Boltzmann Constant (σ) 5.67x10 W/m2 − K 4
Conduction Time Constant (τc ) 21.1 − 23.3 sec

122 Us As (TH2 − TH1 ). The radiative heat transfer from heater 1 to heater 2 (or

8
 4 4


123 vice versa) is given by QR12 =  σ A TH2 − TH1 .

dTH1 4 4


m cp = U A (T∞ − TH1 ) +  σ A T∞ − TH1 + QC12 + QR12 + α1 Q1 (1)
dt

dTH2 4 4


m cp = U A (T∞ − TH2 ) +  σ A T∞ − TH2 − QC12 − QR12 + α2 Q2 (2)
dt

124 The dynamic temperature response of the two temperature sensors are pri-
125 marily by conductive heat transfer from the heaters. The temperature sensors
126 are small in mass and surface area relative to the heaters so the heat transfer by
127 other mechanisms is ignored. The time constant τc is a lumped parameter from
128 a discretized version of Fick’s Law of heat transfer with τc = ms cps ∆x/kc Acond ,

129 where ms is the mass of the sensor, cps is the heat capacity of the sensor, kc is
130 the thermal conductivity of the thermal epoxy, and ∆x is the width of the ther-
131 mal epoxy. These parameters are combined together into one parameter τc and
132 estimated from the data. The dynamic sensor temperature response expressions
133 are Equations 3 and 4.

dTC1
τc = TH1 − TC1 (3)
dt

dTC2
τc = TH2 − TC2 (4)
dt
134 The test of the physics-based model is performed in two phases that includes
135 a model fitting phase followed by validation. The model fitting adjusts the pa-
136 rameters U , Us , α1 , α2 , and τc to minimize the sum of squared error between
137 the model prediction and data as shown in Figure 6a. The model validation is a
138 simulation of the temperature profile given a different heater profile. The mea-
139 sured temperatures are not used in performing the simulation but are compared
140 afterwards to determine how well the model fitting performs on independent
141 data as shown in Figure 6b.

9
 80
T1 measured T1 measured
70 T1 energy balance 70 T1 energy balance
T1 FOPDT T1 FOPDT
60 T2 measured 60 T2 measured
T2 energy balance T2 energy balance
50 50
T ( o C)

T ( o C)
40 40

30 30
20
20
0 100 200 300 400 500 600 0 100 200 300 400 500 600

100 Q1 100 Q1
Q2 Q2
80 80

60 60
Heaters

Heaters
40 40

20 20

0 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Time (sec) Time (sec)

(a) Model Fitting (b) Model Validation

Figure 6: Dual Heater Step Response of the TCLab with Physics-based and FOPDT Model

142 3.2. First-Order Plus Dead-time Model

143 In addition to the physics-based model, a first-order plus dead-time (FOPDT)

144 model is fit to step response data. An FOPDT model includes the gain (Kp =0.92
o
145 C/%), time constant (τp =175.2 sec), and delay time (θp =15.6 sec). The FOPDT
146 model is a single differential equation as shown in Equation 5.

dTC1
τp = −TC1 + Kp Q1 (t − θp ) (5)
dt
147 The discrete solution to the FOPDT equation is Equation 6 when there is
148 a zero-order hold for the heaters between sampling intervals (∆t) between time
149 interval j and j − 1.

−∆ t −∆ t
 

TC1,j = e τp
(TC1,j−1 − TC1,0 ) + 1 − e τp Kp Q1,j−θp −1 − Q1,0 + TC1,0
(6)
150 The FOPDT model is used in this example for obtaining initial tuning pa-
151 rameters to a Proportional-Integral-Derivative (PID) controller for an optimization-
152 based tuning approach as detailed in Section 4. Heater 1 (Q1 ) is adjusted with
153 variable step sizes and heater 2 (Q2 ) remains off to generate step response data
154 for the FOPDT. The results of the temperature data and model fit is shown in
155 Figure 7a and Figure 7b for validation with a different heater profile.

10
 70 T1 measured
80 T1 measured
T1 energy balance 70 T1 energy balance
60 T1 FOPDT T1 FOPDT
T2 measured 60 T2 measured
T2 energy balance T2 energy balance
50
T ( o C)

T ( o C)
50
40 40
30
30
20
0 100 200 300 400 500 600 0 100 200 300 400 500 600

Q1 80 Q1
80 Q2 Q2
60
60
Heaters

Heaters
40
40

20 20

0 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Time (sec) Time (sec)

(a) Model Fitting (b) Model Validation

Figure 7: Single Heater Step Response of the TCLab with Physics-based and FOPDT Model

156 The physics-based model has a lower average absolute error while the FOPDT
157 model has a higher error because a first order model is fit to a higher order re-
158 sponse. The physics-based model fits the temperature response better when the
159 heater is adjusted because of the second-order model and nonlinear radiative
160 heat transfer term.

161 3.3. Linear Time Series Models

162 Auto-Regressive eXogenous input (ARX) time series models are a linear
163 representation of a dynamic system in discrete time. The ARX, Output Error
164 (OE), Finite Impulse Response (FIR), State Space (SS), and other forms are
165 common in industrial multivariate identification and control [40]. Equation 7
166 is an ARX time series model with a single heater input and single temperature
167 output with k index for the time step, i index for prediction horizon step, and
168 adjustable parameters α, β, and γ.

nα nβ
X X
TC1,k+1 = αi TC1,k−i+1 + βi Q1,k−i+1 + γ (7)
i=1 i=1

169 With nα = 3 and nβ = 2 the time series model has 5 adjustable parameters
170 and is shown in Equation 8. The ARX form uses prior temperature measure-
171 ments to predict the next temperature in the series, TC1,k+1 , while the OE

11
172 form uses prior temperature predictions to predict the next temperature in the
173 sequence. The γ1 value is adjusted to create an unbiased model prediction.

TC1,k+1 = α1 TC1,k + α2 TC1,k−1 + α3 TC1,k−2 + β1 Q1,k + β2 Q1,k−1 + γ1 (8)

174 The OE identification form is used to reduce model bias. Equations 9a and
175 9b have multiple inputs and multiple outputs for the case when nα = 2 and
176 nβ = 1.

TC1,k+1 = α1,1 TC1,k + α2,1 TC1,k−1 + β1,1 Q1,k + β1,2 Q2,k + γ1 (9a)

177

TC2,k+1 = α1,2 TC2,k + α2,2 TC2,k−1 + β2,1 Q1,k + β2,2 Q2,k + γ2 (9b)

178 An advantage of a linear time invariant (LTI) model such as SS, ARX, FIR,
179 or OE is that little or no physics-based information is required to obtain a
180 model prediction. When constraints are available, they are used to improve the
181 identification [41]. The model fit to the step test data is shown in Figure 8a and
182 the validation in Figure 8b.

Q1 80 Q1
80 Q2 Q2
60
60
MVs

MVs

40
40

20 20

0 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
70 80
TC1, meas TC1, meas
TC1, pred 70 TC1, pred
60 TC2, meas TC2, meas
TC2, pred 60 TC2, pred
50
50
CVs

CVs

40 40
30
30
20
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Time (sec) Time (sec)

(a) Model Fitting (b) Model Validation

Figure 8: Single Heater Step Response of the TCLab with Linear Time Series

183 There is insufficient data information to determine the β values associated

184 with Q2 because the value stays at zero for the duration of the test. A second

12
185 test is conducted where the second heater is also adjusted to get a multivariate
186 model from the step response data (see Figure 9).

100 Q1 100 Q1
Q2 Q2
80 80

60 60
MVs

MVs
40 40

20 20

0 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
80
TC1, meas TC1, meas
70 TC1, pred 70 TC1, pred
TC2, meas TC2, meas
60 60
TC2, pred TC2, pred
50 50
CVs

CVs
40 40

30 30

20 20
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Time (sec) Time (sec)

(a) Model Fitting (b) Model Validation

Figure 9: Dual Heater Step Response of the TCLab with Linear Time Series

187 3.4. Hammerstein Model with Artificial Neural Network

188 A final modeling approach is a Hammerstein Model with an Artificial Neural

189 Network (ANN) to predict the steady-state relationship between the heaters
190 and temperatures and a linear dynamic block that translates the steady-state
191 prediction into a dynamic prediction. The ANN is not trained directly on the
192 dynamic data because a Recurrent Neural Network or Convolutional Neural
193 Network is better suited for this type of predictive model and this is the topic
194 of future work. A diagram of the model is shown in Figure 10.
195 The parameter weights, represented by arrows connecting each of the nodes,
196 are adjusted to minimize a sum of squared error with 70 steady-state data points.
197 The steady-state data points are obtained by setting random heater values be-
198 tween 0 and 80% for 5 min, recording the temperatures, and then adjusting the
199 heater values to random levels for another data point. Although the system
200 does not fully reach steady-state ( 2 τ or 95% of change), it is judged to be
201 sufficiently close to fit the steady-state correlation. The linear dynamic part is
202 approximated as a second-order dynamic relationship between the steady-state
203 temperature outputs of the ANN and the dynamic response with τp1 =140 sec

13
 Q1
linear tanh() linear

T1,ss 2nd Order

T1
S Dynamic
Q2
linear tanh() linear

Q1
linear tanh() linear

T2,ss 2nd Order

T2
S Dynamic
Q2
linear tanh() linear

Steady State Prediction Dynamic

Figure 10: Architecture of the Hammerstein Model with a Steady-State Artificial Neural
Network and Linear Dynamics.

14
204 and τp2 =20 sec. The second order system approximates the time constant for
205 the heater and temperature sensor with heat conduction between the two.

100 Q1 100 Q1
Q2 Q2
80 80

60 60
MVs

MVs
40 40

20 20

0 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
80
TC1, meas TC1, meas
70 TC1, pred 70 TC1, pred
TC2, meas TC2, meas
60 60
TC2, pred TC2, pred
50 50
CVs

CVs
40 40

30 30
20
20
0 100 200 300 400 500 600 0 100 200 300 400 500 600
Time (sec) Time (sec)

(a) Model Fitting (b) Model Validation

Figure 11: Hammerstein Model Fitting and Validation with 2 Heaters

206 The fitting data is shown in Figure 11a and validation is shown in Figure 11b.
207 Because the steady-state data is a different data set than the dynamic fitting
208 data set, there is some offset between the predictions and data. There are many
209 ANN forms and a future case study could investigate the use of convolutional
210 or recurrent neural networks such as a network with LSTM (Long Short-Term
211 Memory) nodes to combine the dynamic and steady-state predictions into one
212 model.

213 3.5. Summary of Model Predictions with Validation

214 For model-based controllers, the choice of model depends on many factors
215 such as computation speed, ability to extrapolate outside the training region,
216 degree of nonlinearity, and others. Table 2 summarizes the model fit to data
217 with the model regression and validation tests as an average sum of absolute
218 error.

219 4. Benchmarking Closed-Loop PID Re-Tuning

220 The PID controller is a widely used basic regulatory control algorithm. PID
221 control is important in chemical engineering processes as it plays a critical role

15
 Table 2: Summary of Regression and Validation for Single Heater (SISO) and Dual Heater
(MIMO) Tests

Model Description Training Validation

o
SISO Physics-based Lumped Parame- 0.20 C 3.32 o C
ter
SISO First-order Plus Dead-time 0.41 o C 5.11 o C
SISO Second Order ARX 0.18 o C 5.16 o C
SISO Hammerstein with ANN and 3.83 o C 1.66 o C
Linear Dynamics
MIMO Physics-based Lumped Parame- 0.23 o C 0.70 o C
ter
MIMO Second Order ARX 0.26 o C 2.66 o C
MIMO Hammerstein with ANN and 1.57 o C 1.55 o C
Linear Dynamics

222 as a base regulatory layer foundation for advanced process control and opti-
223 mization systems. PID performance varies greatly on the parameters obtained
224 from tuning rules or heuristics [42, 43]. Control performance metrics such as
225 minimum variance control are common assessments of performance [44, 45].
226 Methods such as Zeigler-Nichols closed-loop tuning requires sustained oscilla-
227 tion data to obtain an ultimate gain (Ku ) and ultimate period (Pu ) [46]. To
228 avoid driving a process to the limitation of the stability region to obtain the
229 sustained oscillation data, a relay method is introduced [47]. Tuning rules are
230 a valuable starting point for further manual tuning but may not be optimized.
231 Optimization-based PID tuning is another option with prior work in extremum
232 seeking [48] algorithms, particle swarm [49, 50], and meta-heuristics such as
233 genetic algorithms [51].
234 The objective of this closed-loop PID re-tuning is to demonstrate a TCLab
235 benchmark that uses historical data to optimally re-tune a PID controller. An
236 exhaustive search method visits all feasible combinations of the PI or PID pa-

16
237 rameters to find an optimal value of the objective function without converging to
238 a local minimum for both output-error and input-move deviations. The method
239 uses simulation of the physical TCLab PID controller by: (a) re-playing back
240 the past or historical setpoint and load disturbances [52]; (b) allowing multi-
241 ple, simultaneous and probability-weighted process models to be included in
242 the simulations (i.e., multiple scenarios or situations each with specified proba-
243 bilities) for robustness; (c) including multiple and simultaneous PID controller
244 configuration formulations or even ad hoc controller designs; (d) specifying any
245 type of performance objective function criteria i.e., simultaneously minimize the
246 output-error and input-move variances, overshoot, etc. (e) adding stability rules
247 in the search to cut-off unstable sections of the closed-loop operating space and
248 (f) utilizing an indirect and constrained controller design technique [53].
249 The exhaustive search method is tested with the TCLab as a benchmark for
250 closed-loop control performance. The TCLab produces the closed-loop operat-
251 ing data with IMC PID parameters and a selected setpoint change sequence. A
252 deterministic parametric process model is then identified using an ARX struc-
253 tured model using the GEKKO dynamic optimization suite [54], estimating
254 coefficients using a least-squares prediction-error objective function. Then, the
255 exhaustive search method evaluates the range or domain of the different P , I,
256 and/or D parameters. The best search objective function found provides the P ,
257 I, and/or D. The PID controller is then run again with the temperature control
258 lab using the re-tuned PID parameters and the data recorded. There are many
259 derivations of PID formula rooted in the original continuous equation [42]. For
260 implementing PID controllers in modern digital control platforms such as a DCS
261 (Distributed Control Systems) or PLC (Programmable Logic Controllers), two
262 popular discrete forms are widely used in industry. One is the positional form
263 (Equation 10a) and the other is the velocity form (Equation 10b), which are
264 exchangeable.

t
!
∆t X P Vt−1 − P Vt
OPt = OPbias + Kc et + et + τD (10a)
τI 1 ∆t

17
265

 
∆t τD
OPt = OPt−1 + Kc (et − et−1 ) + et + (P Vt − 2 P Vt−1 + P Vt−2 )
τI ∆t
(10b)
266 where the output error et = SPt − P Vt . Whereas the positional form calculates
267 the controller output position (OP ), the velocity form calculates the change in
268 controller output (∆OP = OPt − OPt−1 ). Although the positional form is more
269 straightforward to understand as the P , I, and D terms are directly translated
270 from the original continuous form, the velocity form has several advantages from
271 the convenience perspective such as no additional logic is required for anti-reset
272 windup [55]. The positional form PI controller is used in this study while a prior
273 study [53] used a PID controller in velocity form. In both cases, an ARX model
274 is identified from closed-loop data. ARX and Box-Jenkins models have proven
275 consistency in closed-loop identification [56, 57]. The potential PID tunings are
276 re-played with the same past setpoint and load disturbance as in the process
277 data (yt ) with zt = yt − xt where, xt represents the ARX model output for
278 time-step t. The load disturbance (zt ) is super-imposed on the ARX simulated
279 process output during the search for optimal tuning parameters.
280 Two different types of objective functions are considered for PID tuning. The
281 objective functions are a variation of the PID control performance index known
282 as average IAE (Integral Absolute Error). The objective function consists of the
283 output-error (OE) term, and the input movement (IM) term. The optimization
284 solution of output error combined with input movement (or, rate of change)
285 has been analytically derived and investigated in [58] and is the simplest form
286 of move suppression. These multi-objective functions can be express in two
287 different ways. One is Archimedean and the other is the lexicographic form (or
288 goal programming) as shown in Equations 11 and 12, respectively.

Pt
i=1 (wOE kSPi − xi kn + wIM kOPi − OPi−1 kn )
min J = (11)
Kc ,τI ,τD t
289 where n is the norm w is the weighting factor for each term in the objective

18
290 function denoted as OE for output error and IM for input movement.

Pt
i=1 (kSPi − xi kn ) Subject to kOPi − OPi−1 kn ≤ U BIM
min J = (12)
Kc ,τI ,τD t

291 where U B is the upper bound of the input movement (IM ) which may be
292 initially set by the centroid PID performance. Either the Archimedean or lexi-
293 cographic form of the objective function can be used for PID controller tuning.
294 In terms of convenience, the lexicographic form is easier to use because it re-
295 quires one user input parameter, U BIM , as opposed to the Archimedean form
296 that requires two weighting factors on both OE and IM terms. One simplifi-
297 cation of the Archimedean form is to reduce the weighting factors to one by
298 dividing the objective by wOE .

299 4.1. TCLab Benchmark Validation

300 The first step of the validation is to collect the closed-loop operation data
301 and identify the ARX model parameters for identifying the ARX model. The
302 setpoint is changed from ambient temperature at the initial steady-state con-
303 dition to 50 o C, 40 o C, and then to 60 o C. The ranges of Kc and τI are
304 evaluated through the ARX model that includes the same setpoint sequences
305 and load disturbance. The performance objective functions for each Kc and τI
306 incremental combination are also calculated and stored. The `1 -norm objective
307 function in the Archimedean form is chosen for the test with weighting factors
308 wOE = 1 and wIM = 0.5. The Kc and τI combination that gives a minimum
309 value of objective function is then chosen as optimal PID tuning. The initial Kc
310 and τI are from the FOPDT model in Section 3.2 and IMC aggressive tuning
311 with Kc = 5.74 o%C and τI = 175.2 sec. Optimized values are Kc = 10.0 o%C and
312 τI = 55.0 sec as shown as the minimum value of the objective function contour
313 map (see Figure 12).
314 The objective function surface is not smooth because of the load disturbances
315 that are replayed with every PID parameter combination. Figure 13 shows
316 the measured temperature and ARX model response for both the original and

19
 6.4
6.2
6.0
Obj
5.8
5.6
5.4

10 10
20 20
30 30
40 40
τ 50 50
I
60 60 Kc

70 70
80 80

Figure 12: Average Integral Absolute Error (IAE) with Kc and τI PID Parameters.

20
317 optimized response. The validation of the optimal tuning parameter is displayed
318 as well.

100 Q1 Original
Q1 Optimal
80 Q1 Verified
Q2
60
MVs

40

20

0
0 100 200 300 400 500 600

60
TSP1
50 TC1, meas
TC1, optimal
TC1, verified
CVs

40 TC1, pred
TSP2
30 TC2, meas
TC2, optimal
TC2, pred
20
0 100 200 300 400 500 600
Time (sec)

Figure 13: ARX Simulated and TCLab Validated Performance Improvement of 5.4%.

319 The average IAE objective function is 6.09 with IMC tuning and 5.76 with
320 optimized parameters, an improvement of 5.4%. The PID improvement is simu-
321 lated with the ARX model and validated with closed-loop data from the Arduino
322 TCLab.

323 5. Multivariate Control Benchmark

324 Model predictive control (MPC) with the physics-based model, time series
325 linear model (ARX), and Hammerstein ANN model quantify multivariate con-
326 trol performance. Additional models in MPC or multivariate control strategies
327 are tested with the TCLab. This section shows benchmark performance with
328 three popular methods for multivariate control that range from linear to non-
329 linear and empirical to physics-based.

21
330 An `1 -norm objective function gives a target region for the temperature
331 range, rather than one specific target value. Equation 13 shows the `1 -norm
332 control formulation used in this work for model predictive control (MPC).

T T
min J = whi ehi + wlo elo + ∆QT c∆Q
x,CV,M V

0 = f ddTt , T, Q


s.t.
(13)
ehi ≥ T − Thi
elo ≥ Tlo − T
333 where J is the objective function, T is the temperature, Q is the heater, wlo
334 and whi are penalty matrices for solutions outside the target temperature region.
335 Slack variables elo and ehi are the error of the dead-band low and high limits,
336 respectively. Parameter c∆Q is a move suppression factor. The function f is an
337 open-equation set of model equations that include T , Q, and time derivatives
338 of T . The demand targets Tlo and Thi define lower and upper target limits for
339 temperature as shown in Figure 14.

50
Temperature (degC)

40 T1 target
T1 measured
30 T1 predicted
T1 trajectory
0 25 50 75 100 125 150 175 200
35
Temperature (degC)

30 T2 target
T2 measured
T2 predict
25 T2 range
0 25 50 75 100 125 150 175 200
100
75 Current Time
Heaters

50
Q1 history
Q1 plan
25 Q2 history
Q2 plan
0
0 25 50 75 100 125 150 175 200
Time (sec)

Figure 14: MPC with ARX Model at Cycle 81

22
340 At cycle 81, temperature 1 has just reached the target temperature setpoint
341 of 50 o C after heater 1 is ramped down from 100% to 0% at 10-15 sec prior to
342 reaching the setpoint. The model predictive controller anticipates the continued
343 rise in temperature and turns the heater off for a period of 5 seconds before
344 returning to a baseline heater value to maintain the 50 o C setpoint. The model
345 also anticipates the increase in temperature 2 due to the setpoint change to 35
o
346 C at cycle 80. The reference trajectory with time constant τ =10 sec gives a
347 guide for the fastest that the temperature should approach the new setpoint.
348 The setpoint has a ±0.2 o C range with a ±1.0 o C larger opening at the beginning
349 for less MV movement for near-term adjustments. The underlying ARX time-
350 series model coordinates the MV movements to meet both setpoints considering
351 multivariate effects.

352 6. Benchmarking Model Predictive Control

353 The multivariate models developed in Sections 3.1, 3.3, and 3.4 are compared
354 in MPC. The MPC uses an `1 -norm objective with a temperature dead-band of
355 ±0.2 o C for Thi − Tsp , Tlo − Tsp and a first-order reference trajectory of 10 sec
356 for setpoint changes. The move suppression factor c∆Q is set to 0.1, the weights
357 whi and wlo are set to 20.0, and the control and prediction horizon are 60 sec-
358 onds. The linear ARX model has a cycle time of 1 second while the nonlinear
359 physics-based and Hammerstein applications are re-computed every 2 seconds.
360 The longer cycle time is required to enable all steps of data retrieval, model
361 update, re-calculation of optimal move plan, retrieval of first step, and insertion
362 into the process. Table 3 is a numeric comparison of the methods with quan-
o
363 tified IAE rate ( C/sec) and Integral Average Move rate (%/sec) for the heater
364 adjustments. Another common performance metric is a minimum variance as
365 applied to multivariate control systems [59, 60]. Rate-based values are shown
366 in this case because of the differing cycle times between the applications.
367 The benchmark results show that all models perform equally well in terms
368 of the control performance (11.4-11.6 o C/sec) as shown in Figures 15 to 17.

23
 Table 3: Summary of Model Predictive Control Methods

Model Description IAE Avg Rate (CVs) IAE Avg Rate (∆M V s)
o
Physics-based Lumped 11.5 C/sec 2.0 %/sec
Parameter
o
Second Order ARX 11.6 C/sec 3.3 %/sec
o
Hammerstein with ANN 11.4 C/sec 2.5 %/sec
and Linear Dynamics

369 In all cases, T1 is not able to reach the setpoint of 30o C between 160-320 sec
370 because of insufficient cooling rate when Q1 is off. The ARX model has the
371 highest MV movement (3.3 %/sec) and the physics-based model has the lowest
372 MV movement even with rapid fluctuations on Q2 during the first setpoint
373 change at t = 105 sec. The values for MPC are more than the PID control
374 performance metric because there are two CVs and two MVs that accumulate
375 error approximately twice as fast and with more frequent setpoint changes.
376 The physics-based model has the potential to extrapolate to new operating
377 conditions without retuning. A physics-based MPC has the disadvantage of
378 relative difficulty in developing the model equations for complex systems. There
379 is also a potential for solver convergence problems if the physics-based model
380 is high nonlinear or does not have a suitable initial guess. This is not the case
381 for the TCLab where an approximate lumped-parameter model is an accurate
382 representation of the physical system. One drawback for the physics-based MPC
383 is that it cannot run at 1 sec cycles but does solve within a 2 second interval for
384 a 60 sec prediction horizon. The ARX control performance is shown in Figure
385 16.
386 The ARX MPC has the fastest cycle time (1 sec versus 2 sec) so that it can
387 respond more quickly to disturbances or setpoint changes. Because it is a linear
388 model, the cycle time can be faster (up to 5 Hz) due to reduced computing time.
389 The disadvantage of the ARX MPC is that it is a linear representation of the
390 slightly nonlinear TCLab. This requires re-adjustment of the move plan and

24
 60
Temperature (degC)
50

40

30
T1
T1 SP
0 100 200 300 400 500 600
50 T2
Temperature (degC)

T2 SP
40

30

0 100 200 300 400 500 600

100
Q1
75 Q2
Heaters

50
25
0
0 100 200 300 400 500 600
Time (sec)

Figure 15: MPC with Physics-based Model

60
Temperature (degC)

50

40
T1
30 T1 SP
0 100 200 300 400 500 600
50 T2
Temperature (degC)

T2 SP
40

30

0 100 200 300 400 500 600

100
Q1
75 Q2
Heaters

50
25
0
0 100 200 300 400 500 600
Time (sec)

Figure 16: MPC with ARX Time Series Model

25
391 increased cycling due to model mismatch. The ARX MPC has slight overshoot
392 due to the underestimation of process gain that leads to overly aggressive MV
393 movement as shown in Figure 17.

60
Temperature (degC)

50

40
T1
30 T1SP

0 100 200 300 400 500 600

50
T2
Temperature (degC)

T2SP
40

30

0 100 200 300 400 500 600

100
Q1
75 Q2
Heaters

50
25
0
0 100 200 300 400 500 600
Time (sec)

Figure 17: MPC with Hammerstein ANN Model

394 The Hammerstein MPC has the potential to excel in situations where the
395 process is highly nonlinear and there is not a suitable physics-based represen-
396 tation of the process. Like the physics-based MPC, it requires a slower 2 sec
397 cycle time to meet the real-time constraint. Unlike the physics-based MPC,
398 it is not expected to perform well when used outside of the training domain.
399 To facilitate the comparison, a repository of source code and Arduino firmware
400 https://github.com/APMonitor/arduino is available with all the examples
401 from this paper.

26
402 7. Conclusion and Future Work

403 The benchmark studies included in this paper are a sampling of common
404 modeling and control methods that are quantified with the TCLab shield and
405 an Arduino microcontroller. The temperature response is modeled with four
406 approaches: physics-based, FOPDT, ARX, and Hammerstein ANN with linear
407 dynamics. Separate data sets are used for training and validation. The objective
408 of the modeling is to create automatic controllers with PID and MPC. A PID
409 optimal tuning case study uses an exhaustive search as a straightforward method
410 for closed-loop retuning to improve performance by 5.4%. The optimal PID
411 parameters are selected by replaying past setpoint and load disturbances where
412 the residuals of estimation are considered as the unmeasured load disturbances.
413 A second study is the application of the three multivariate models in MPC with
414 varying degrees of nonlinearity and physics-based foundation.
415 This study presents a sample of potential modeling and control applications
416 that are quantified with the TCLab hardware benchmark. There are additional
417 potential applications for evaluating methods in estimation, data reconciliation,
418 machine learning, classification, fault detection, anomaly detection, disturbance
419 identification and rejection, integration of control and scheduling, mixed inte-
420 ger systems, stability analysis, explicit MPC, and others. Because each TCLab
421 device is slightly different, benchmark evaluations are performed on the same
422 device and with similar ambient conditions. The TCLab is an accessible hard-
423 ware platform for benchmarking models and closed-loop performance with real
424 data.

425 Acknowledgments

426 This article is prepared for a Special Issue in honor of Tom Edgar’s 75th
427 birthday and to celebrate his lifetime of accomplishments and leadership in the
428 area of Process Systems Engineering. This work is influenced by his work with
429 energy systems, optimization, control engineering education, and advancements

27
430 in model-based control among many other areas of contribution. We are grateful
431 for his contributions and continued service to the community.

432 Appendix A. Software Interface to TCLab

433 Two parts to the software interface are the firmware that runs on the Ar-
434 duino Leonardo and the serial interface to interpret and command the TCLab.
435 An important part of making the benchmark accessible is to create an inter-
436 face to software (MATLAB, Simulink, and Python) where control algorithms
437 are developed but also provide information for interfaces to other software plat-
438 forms. There is an Arduino Support Package for MATLAB and Simulink from
439 MathWorks that automatically loads firmware onto the Arduino when it is con-
440 nected for the first time. The Arduino firmware for Python is an ino file that is
441 augmented with additional sections to compile as cpp code with a gcc compiler
442 through the Arduino IDE. The TCLab is pre-loaded with the Python interface
443 firmware.
444

Listing 1: MATLAB Commands to Adjust Heaters and Display Temperatures

445
446 clear all
447 % i n c l u d e t c l a b .m
448 tclab ;
449 d i s p ( ' Turn on H e a t e r s and LED ' )
450 h1 ( 3 0 ) ; h2 ( 6 0 ) ; led ( 1 ) ;
451 pause ( 1 0 )
452 d i s p ( ' D i s p l a y Temperatures ' )
453 d i s p ( T1C ( ) )
454 d i s p ( T2C ( ) )
455 h1 ( 0 ) ; h2 ( 0 ) ; led ( 0 ) ;
456

457

28
 Figure A.18: Simulink Interface with Manual Sliders for Heater Levels.

Listing 2: Python Commands to Adjust Heaters and Display Temperatures

458
459 im po rt tclab # pip i n s t a l l t c l a b
460 im po rt time
461 # Connect t o Arduino
462 a = tclab . TCLab ( )
463 p r i n t ( ' Turn on H e a t e r s and LED ' )
464 a . Q1 ( 3 0 . 0 ) ; a . Q2 ( 6 0 . 0 ) ; a . LED ( 1 0 0 )
465 time . sleep ( 1 0 . 0 )
466 p r i n t ( ' D i s p l a y Temperatures ' )
467 p r i n t ( a . T1 )
468 p r i n t ( a . T2 )
469 a . close ( )
470

29
471 References

472 References

473 [1] A. F. Villaverde, D. Henriques, K. Smallbone, S. Bongard, J. Schmid,

474 D. Cicin-Sain, A. Crombach, J. Saez-Rodriguez, K. Mauch, E. Balsa-Canto,
475 et al., Biopredyn-bench: a suite of benchmark problems for dynamic mod-
476 elling in systems biology, BMC systems biology 9 (1) (2015) 8.

477 [2] N. R. Lewis, J. D. Hedengren, E. L. Haseltine, Hybrid dynamic optimiza-

478 tion methods for systems biology with efficient sensitivities, Processes 3 (3)
479 (2015) 701. doi:10.3390/pr3030701.
480 URL http://www.mdpi.com/2227-9717/3/3/701

481 [3] L. Peters, R. Arts, G. Brouwer, C. Geel, S. Cullick, R. J. Lorentzen,

482 Y. Chen, N. Dunlop, F. C. Vossepoel, R. Xu, et al., Results of the Brugge
483 benchmark study for flooding optimization and history matching, SPE
484 Reservoir Evaluation & Engineering 13 (03) (2010) 391–405.

485 [4] V. P. Singh, A. Cavanagh, H. Hansen, B. Nazarian, M. Iding, P. S. Ringrose,

486 et al., Reservoir modeling of CO2 plume behavior calibrated against mon-
487 itoring data from Sleipner, Norway, in: SPE annual technical conference
488 and exhibition, Society of Petroleum Engineers, 2010.

489 [5] R. W. Rwechungura, E. Suwartadi, M. Dadashpour, J. Kleppe, B. A. Foss,

490 et al., The Norne field case-a unique comparative case study, in: SPE Intel-
491 ligent Energy Conference and Exhibition, Society of Petroleum Engineers,
492 2010.

493 [6] J. Udy, B. Hansen, S. Maddux, D. Petersen, S. Heilner, K. Stevens,

494 D. Lignell, J. D. Hedengren, Review of field development optimization of
495 waterflooding, EOR, and well placement focusing on history matching and
496 optimization algorithms, Processes 5 (3) (2017) 34.

497 [7] A. N. Eaton, L. D. Beal, S. D. Thorpe, C. B. Hubbell, J. D. Hedengren,

498 R. Nybø, M. Aghito, Real time model identification using multi-fidelity

30
499 models in managed pressure drilling, Computers & Chemical Engineering
500 97 (2017) 76–84.

501 [8] R. Asgharzadeh Shishavan, C. Hubbell, H. Perez, J. Hedengren, D. S. Pix-

502 ton, et al., Combined rate of penetration and pressure regulation for drilling
503 optimization using high speed telemetry, SPE Drilling & Completion Jour-
504 nal 1 (SPE-170275-MS) (2015) 17–26.

505 [9] M. Jamil, X.-S. Yang, A literature survey of benchmark functions for global
506 optimization problems, arXiv preprint arXiv:1308.4008.

507 [10] R. V. Rao, V. J. Savsani, D. Vakharia, Teaching–learning-based optimiza-

508 tion: an optimization method for continuous non-linear large scale prob-
509 lems, Information sciences 183 (1) (2012) 1–15.

510 [11] S. M. Safdarnejad, J. D. Hedengren, N. R. Lewis, E. L. Hasel-

511 tine, Initialization strategies for optimization of dynamic sys-
512 tems, Computers & Chemical Engineering 78 (2015) 39 – 50.
513 doi:10.1016/j.compchemeng.2015.04.016.
514 URL http://www.sciencedirect.com/science/article/pii/
515 S0098135415001179

516 [12] R. Huang, Nonlinear model predictive control and dynamic real time opti-
517 mization for large-scale processes, Ph.D. thesis, Carnegie Mellon University
518 (12 2010).

519 [13] M. Hehn, R. Ritz, R. D’Andrea, Performance benchmarking of quadrotor

520 systems using time-optimal control, Autonomous Robots 33 (1-2) (2012)
521 69–88.

522 [14] W. Chen, Y. Ren, G. Zhang, L. T. Biegler, A simultaneous approach for sin-
523 gular optimal control based on partial moving grid, AIChE Journal 65 (6).

524 [15] L. D. Beal, D. Petersen, D. Grimsman, S. Warnick, J. D. Hedengren, In-

525 tegrated scheduling and control in discrete-time with dynamic parameters

31
526 and constraints, Computers & Chemical Engineering 115 (2018) 361 – 376.
527 doi:https://doi.org/10.1016/j.compchemeng.2018.04.010.
528 URL http://www.sciencedirect.com/science/article/pii/
529 S0098135418303120

530 [16] L. D. R. Beal, D. Petersen, G. Pila, B. Davis, S. Warnick, J. D. Hedengren,

531 Economic benefit from progressive integration of scheduling and control for
532 continuous chemical processes, Processes 5 (4).

533 [17] D. Petersen, L. D. R. Beal, D. Prestwich, S. Warnick, J. D. Hedengren,

534 Combined noncyclic scheduling and advanced control for continuous chem-
535 ical processes, Processes 5 (4).

536 [18] Y. Nie, L. T. Biegler, C. M. Villa, J. M. Wassick, Discrete Time Formulation

537 for the Integration of Scheduling and Dynamic Optimization, Industrial &
538 Engineering Chemistry Research 54 (16) (2015) 4303–4315. doi:10.1021/
539 ie502960p.
540 URL http://pubs.acs.org/doi/abs/10.1021/ie502960p

541 [19] P. F. Odgaard, J. Stoustrup, M. Kinnaert, Fault tolerant control of wind

542 turbines–a benchmark model, IFAC Proceedings Volumes 42 (8) (2009)
543 155–160.

544 [20] G. M. Kopanos, C. A. Méndez, L. Puigjaner, MIP-based decomposition

545 strategies for large-scale scheduling problems in multiproduct multistage
546 batch plants: A benchmark scheduling problem of the pharmaceutical
547 industry, European Journal of Operational Research 207 (2) (2010)
548 644–655. doi:10.1016/j.ejor.2010.06.002.
549 URL http://linkinghub.elsevier.com/retrieve/pii/
550 S037722171000408X

551 [21] D. Saygin, E. Worrell, M. K. Patel, D. Gielen, Benchmarking the energy use
552 of energy-intensive industries in industrialized and in developing countries,
553 Energy 36 (11) (2011) 6661–6673.

32
554 [22] L. D. Beal, J. Park, D. Petersen, S. Warnick, J. D. Hedengren, Combined
555 model predictive control and scheduling with dominant time constant com-
556 pensation, Computers & Chemical Engineering 104 (2017) 271–282.

557 [23] M. Baldea, I. Harjunkoski, Integrated production scheduling and process

558 control: A systematic review, Computers & Chemical Engineering 71
559 (2014) 377–390. doi:10.1016/j.compchemeng.2014.09.002.

560 [24] J. Kelly, J. Hedengren, A steady-state detection (SSD) algorithm to detect

561 non-stationary drifts in processes, Journal of Process Control 23 (3) (2013)
562 326–331.

563 [25] N. L. Ricker, J. Lee, Nonlinear model predictive control of the Tennessee
564 Eastman challenge process, Computers & Chemical Engineering 19 (9)
565 (1995) 961–981.

566 [26] A. Bathelt, N. L. Ricker, M. Jelali, Revision of the Tennessee Eastman

567 process model, IFAC-PapersOnLine 48 (8) (2015) 309–314.

568 [27] N. I. Vitzilaios, N. C. Tsourveloudis, Test bed for unmanned helicopters’

569 performance evaluation and benchmarking, in: IEEE/RSJ IROS 2008
570 Workshop on Performance Evaluation and Benchmarking for Intelligent
571 Robots and Systems, Citeseer, 2008.

572 [28] A. Cardoso, V. Sousa, P. Gil, Demonstration of a remote control laboratory

573 to support teaching in control engineering subjects, IFAC-PapersOnLine
574 49 (6) (2016) 226–229.

575 [29] P. K. Singh, S. Bhanot, H. K. Mohanta, V. Bansal, Self-tuned fuzzy logic

576 control of a ph neutralization process, in: 2015 21st International Confer-
577 ence on Automation and Computing (ICAC), IEEE, 2015, pp. 1–6.

578 [30] I. Alvarado, D. Limon, D. M. De La Peña, J. Maestre, M. Ridao, H. Scheu,

579 W. Marquardt, R. Negenborn, B. De Schutter, F. Valencia, A comparative
580 analysis of distributed MPC techniques applied to the HD-MPC four-tank
581 benchmark, Journal of Process Control 21 (5) (2011) 800–815.

33
582 [31] V. Kirubakaran, T. Radhakrishnan, N. Sivakumaran, Distributed multi-
583 parametric model predictive control design for a quadruple tank process,
584 Measurement 47 (2014) 841–854.

585 [32] Y. Alipouri, J. Poshtan, Optimal controller design using discrete linear
586 model for a four tank benchmark process, ISA transactions 52 (5) (2013)
587 644–651.

588 [33] B. Spivey, J. Hedengren, T. Edgar, Constrained nonlinear estimation for

589 industrial process fouling, Industrial & Engineering Chemistry Research
590 49 (17) (2010) 7824–7831.

591 [34] V. M. Zavala, L. T. Biegler, Optimization-based strategies for the operation

592 of low-density polyethylene tubular reactors: nonlinear model predictive
593 control, Computers & Chemical Engineering 33 (10) (2009) 1735–1746.

594 [35] J. Rossiter, S. Pope, B. L. Jones, J. Hedengren, Evaluation and demon-

595 stration of take home laboratory kit, IFAC-PapersOnLine 52 (9) (2019)
596 56–61.

597 [36] P. Oliveira, J. Hedengren, An APMonitor temperature lab PID control ex-
598 periment for undergraduate students, in: 24th IEEE Conference on Emerg-
599 ing Technologies and Factory Automation (ETFA), Zaragoza, Spain, IEEE,
600 2019, pp. 790–797.

601 [37] F. G. Shinskey, Process control: as taught vs as practiced, Industrial &

602 engineering chemistry research 41 (16) (2002) 3745–3750.

603 [38] T. F. Edgar, B. A. Ogunnaike, J. J. Downs, K. R. Muske, B. W. Bequette,

604 Renovating the undergraduate process control course, Computers & chem-
605 ical engineering 30 (10-12) (2006) 1749–1762.

606 [39] J. Alford, T. Edgar, Preparing chemical engineering students for industry,
607 Chemical Engineering Progress 113 (11) (2017) 25–28.

34
608 [40] S. J. Qin, T. A. Badgwell, A survey of industrial model predictive control
609 technology, Control Engineering Practice 11 (7) (2003) 733–764.

610 [41] J. Udy, L. Blackburn, J. D. Hedengren, M. Darby, Reduced order modeling

611 for reservoir injection optimization and forecasting, in: Proceedings of the
612 FOCAPO/CPC Conference, Tuscon, AZ, USA, 2017, pp. 8–12.

613 [42] K. J. Åström, T. Hägglund, PID controllers: theory, design, and tuning,
614 Vol. 2, Instrument society of America Research Triangle Park, NC, 1995.

615 [43] B.-S. Ko, T. F. Edgar, Assessment of achievable PI control performance

616 for linear processes with dead time, in: Proceedings of the 1998 American
617 Control Conference. ACC (IEEE Cat. No. 98CH36207), Vol. 3, IEEE, 1998,
618 pp. 1548–1552.

619 [44] S. J. Qin, Control performance monitoring—a review and assessment, Com-
620 puters & Chemical Engineering 23 (2) (1998) 173–186.

621 [45] B.-S. Ko, T. F. Edgar, PID control performance assessment: The single-
622 loop case, AIChE Journal 50 (6) (2004) 1211–1218.

623 [46] J. G. Ziegler, N. B. Nichols, Optimum settings for automatic controllers,

624 trans. ASME 64 (11).

625 [47] K. J. Åström, T. Hägglund, Automatic tuning of simple regulators with

626 specifications on phase and amplitude margins, Automatica 20 (5) (1984)
627 645–651.

628 [48] N. J. Killingsworth, M. Krstic, PID tuning using extremum seeking: on-
629 line, model-free performance optimization, IEEE control systems magazine
630 26 (1) (2006) 70–79.

631 [49] Z.-L. Gaing, A particle swarm optimization approach for optimum design
632 of PID controller in AVR system, IEEE transactions on energy conversion
633 19 (2) (2004) 384–391.

35
634 [50] M. I. Solihin, L. F. Tack, M. L. Kean, Tuning of PID controller using
635 particle swarm optimization (PSO), International Journal on Advanced
636 Science, Engineering and Information Technology 1 (4) (2011) 458–461.

637 [51] B. Mohanty, S. Panda, P. Hota, Controller parameters tuning of differential

638 evolution algorithm and its application to load frequency control of multi-
639 source power system, International journal of electrical power & energy
640 systems 54 (2014) 77–85.

641 [52] J. D. Kelly, Tuning digital PI controllers for minimal variance in manipu-
642 lated input moves applied to imbalanced systems with delay, The Canadian
643 Journal of Chemical Engineering 76 (5) (1998) 967–974.

644 [53] J. Park, C. Patterson, J. Kelly, J. Hedengren, Closed-loop PID re-tuning

645 in a digital twin by re-playing past setpoint and load disturbance data, in:
646 2019 (AIChE) Spring Meeting, New Orleans, LA, AIChE, 2019, pp. 1–6.

647 [54] L. Beal, D. Hill, R. Martin, J. Hedengren, Gekko optimization suite, Pro-
648 cesses 6 (8) (2018) 106.

649 [55] D. E. Seborg, D. A. Mellichamp, T. F. Edgar, F. J. Doyle III, Process

650 dynamics and control, John Wiley & Sons, 2010.

651 [56] A. Voda, I. Landau, Multi-step closed loop identification and control design
652 procedure-applications, IFAC Proceedings Volumes 27 (8) (1994) 1543–
653 1548.

654 [57] E. Jahanshahi, S. Skogestad, Closed-loop model identification and PID/PI

655 tuning for robust anti-slug control, IFAC Proceedings Volumes 46 (32)
656 (2013) 233–240.

657 [58] R. Tchamna, M. Lee, Analytical design of an industrial two-term controller

658 for optimal regulatory control of open-loop unstable processes under oper-
659 ational constraints, ISA transactions 72 (2018) 66–76.

36
660 [59] B.-S. Ko, T. F. Edgar, Performance assessment of multivariable feedback
661 control systems, Automatica 37 (6) (2001) 899–905.

662 [60] C. A. Harrison, S. J. Qin, Minimum variance performance map for con-
663 strained model predictive control, Journal of Process Control 19 (7) (2009)
664 1199–1204.

37