0% found this document useful (0 votes)
51 views22 pages

Identification and Process Control With Artificial Neural Networks and PID Controller

This document proposes using an artificial neural network (ANN) integrated with a PID controller to control a complex industrial process. The ANN would identify the process dynamics while the PID controller acts on the actuator. Key process variables are selected as inputs to the ANN. The ANN architecture and training method are described. Simulation results show the ANN approach improves control performance over manual control in terms of error and response time. The ANN solution could be implemented on existing control systems with PLCs to enhance process control.

Uploaded by

Antonio Ceron
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
51 views22 pages

Identification and Process Control With Artificial Neural Networks and PID Controller

This document proposes using an artificial neural network (ANN) integrated with a PID controller to control a complex industrial process. The ANN would identify the process dynamics while the PID controller acts on the actuator. Key process variables are selected as inputs to the ANN. The ANN architecture and training method are described. Simulation results show the ANN approach improves control performance over manual control in terms of error and response time. The ANN solution could be implemented on existing control systems with PLCs to enhance process control.

Uploaded by

Antonio Ceron
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 22

Identification and Process Control

with Artificial Neural Networks and


PID Controller
Eduardo Viera.
AGENDA
• Introduction .
• Process Description.
• Solution Presentation.
• Variable Selection.
• ANN Conformation.
• Actuator Model Estimation.
• Simulations and Results.
• Conclusion.
• Questions.
Introduction
MISO (Multiple Inputs, Single Output Systems)
𝑥𝑛 (𝑡)

𝑥𝑛+1 (𝑡) 𝑦(𝑡)
Inputs System Output
(Physical Variables) (Process) (Final Control Element)

Fig 1 – MISO Block Diagram.


Problematics:
• Single control loop are difficult to implement,
• Changes in the output generate changes in all system
variables, because of multivariable dependency.
• Complex dynamic model of the process.
• Manual operation is key to keep the process controlled.
Introduction
MISO (Multiple Inputs, Single Output Systems)
𝑥𝑛 (𝑡)

𝑥𝑛+1 (𝑡) 𝑦(𝑡)
Inputs System Output
(Physical Variables) (Process) (Final Control Element)

Fig 1 – MISO Block Diagram.


Considerations:
• Apply the knowledge and best practices of the operation.
• Apply system with “learning” capability.
• Focus the solution in control action without secondary and
cumulative effects as mechanical damage in main equipment.
• Keep controlled all the variables with a unique control action.
Introduction
MISO (Multiple Inputs, Single Output Systems)
𝑥𝑛 (𝑡)

𝑥𝑛+1 (𝑡) 𝑦(𝑡)
Inputs System Output
(Physical Variables) (Process) (Final Control Element)

Fig 1 – MISO Block Diagram.


Solutions:
• Apply Artificial Intelligence as Artificial Neural Networks
(ANN).
• Integrate ANN with classic control strategies as PID
controller.
Process Description
Case of study: BIDIRECTIONAL DUCT
CONTROL AI % SO2

PI VALVE PI TI

TI PI PI SI

DISTRIBUTION
RADIATIVE COOLING CHAMBER ACID PLANT
PI PI
LIRES
FI

TENIENTE CONVERTER BLOWING AIR INDUCED DRAFT FAN


ELCTROSTATIC PRECIPITATOR VTI
No. 1 Y No.2

Fig 2 – Gas Transportation Block Diagram.


• Metallurgical gases generation.
• Gases temperature reduction.
• Electrostatic Filtration.
• Gas Transport.
• Gas processing.
Solution Presentation

Process Signals ANN + PID(t) Plant Model


-

Process Identification
Gain

VTI Control Loop

Fig 3 – Proposed Solution Block Diagram.

• System separated into Plant Identification and Actuator


Control Loop.
• Plant Identification has internal process.
• ANN response is the dynamic consign to keep the process
balanced.
Variable Selection
• Key variables selection was based by the criticality of the
consequences because of his changes.
• Requirement of certain variables by other processes.
• Patterns in secondary variables by the change in some
variables.
Table 1 Entry Variables of ANN
Variable Type Signal Description Range Average Signal
Flow Blowing air (oxigen rich air) to CT [xxxx, xxx] zzz Input
Pressure Hood gases exhaust CT [xxxx, xxx] Zzz Input
Pressure Outlet electrostatic precipatator 1 [xxxx, xxx] Zzz Input
Pressure Outlet electrostatic precipatator 2 [xxxx, xxx] Zzz Input
Pressure Distribution chamber, VTI outlet [xxxx, xxx] Zzz Input
SO2 Sulfur dioxide concentration [xxxx, xxx] Zzz Input
Speed Speed rotor VTI [xxxx, xxx] zzz Output
Variable Selection

Fig 4 – Graph SO2 Concentration and Blowing Air Flow to CT.


Variable Selection

Fig 5 – Graph VTI Speed and Gas Train Pressure.


Variable Selection

Fig 6 – Graphic VTI Speed and Trian Gases Pressure (Differential Pressure).
Conformation of Artificial Neural Network
x1 Synapsis
wi1 In1
Cell Body Axon In2
x2 wi2 yi In3

xj wij ∑ f( ) In4 Speed


Output Set Point
In5
win In6
Inputs -1 Ɵi
In7
xn
Dentres Threshold

Fig 7 – Artificial Neuron Structure. Fig 8 – ANN Architecture.

• Inputs are represented by the input variables selected.


• Network type Feedfordward.
• Weight Connection calculated by “Backpropagation”
Algorithm.
• ANN has 7 input layer, 11 hidden layer, 1 output layer
Actuator Estimation Model
• Developed through input and output data.
• Data obtained from the historical trends of the process
data.
• N4SID (numerical algorithms for subspace state space
system identification) Algorithm used.
• Model obtained in space-state form.
Actuator Estimation Model

Fig 9 – VTI Step Response .

𝑥(𝑡) = 𝐴𝑥(𝑡) + 𝐵𝑢
𝑦(𝑡) = 𝐶𝑥(𝑡) + 𝐷𝑢

Space-State Model
Actuator Estimation Model

Fig 9 – VTI Step Response .

Accuracy Stabilization Overshoot Sample


Input Output
Model Time Time
99.1% X,X min 40.9% 13.3 min xxx RPM xxx RPM
Simulation and Result
• Test with 3 scenarios:

Test with Random


Variables

Historical Records as
Scenarios
Inputs in the System

Scenario 2 and Real


VTI Response
Comparison and
Performance Index
Simulation and Result
• Test with Random Variables.
• Random varibles generated based in a Gaussian
probability distribution.
• Simulation with a integrated solution.
Simulation and Result
Scenario 1: Test with Random Variables.

Fig 10 – Result Simulation 1.


Simulation and Result
Scenario 2: Test with Historical Records as Inputs in the
System.

Fig 11 – Result Simulation 2.


Simulation and Result
Scenario 3: and Real VTI Response Comparison and Performance Index
.

Fig 11 – Result Simulation 2.


Simulation and Result
Scenario 3: and Real VTI Response Comparison and Performance Index
.
𝑛 2

𝑖=0 𝑂𝑖 − 𝐶𝑖
𝐼𝑆𝐸 = 𝑒 𝑡 2 𝑑𝑡 𝑅𝑀𝑆 =
0
𝑛

ISE - Integral Square Error RMS - Residual Mean


Square
76400
0,0123

59500

0,0109

ISE – SYS. WITH ANN ISE - MANUAL SYS. RMS –– SYS. WITH ANN RMS - MANUAL SYS.

Fig 11 – Result Simulation 2.


Conclusion
• According the simulations, is possible to implement IA in process
without major interventions in the existent control systems.
• Settling times are shortest in comparison with manual control
actions.
• Large error are suppressed trying to achieve a little error in the
system output.
• Once ANN has been configured, easily this could by configured in
any PLC (keeping in mind the capability to handle float variables,
memory and process capacity).
• This kind of control strategies, could apply just in “steady state”and
cannot be used to start or shutdown a plant.

You might also like