You are on page 1of 24

APC Foundation Tools, Techniques and Applications

during the pre-commissioning checkout of an advanced control system it is often helpful to


create a dynamic process simulation that interacts with the control system.

The objectives of process control system changes generally fall into one of three categories:

1. Response to an economic incentive (reduced cost or increased revenue)

2. Safety and environmental compliance

3. Equipment protection

I need to collect: 气源站照片、气源站原理动画

two things are critical in achieving economic objectives: defining the optimal operating point,
and reducing process variation to allow operation closer to the optimal conditions.

During the design and construction of a new plant there are strong economic incentives to
minimize the number of transmitters, automatic regulating valves, and control loops that must
be installed and commissioned. Thus, in most cases the majority of the automatic control
installed in a new plant is based on single loop PID feedback control.

To help minimize the process variation that is experienced with single loop PID control,
manufacturers have introduced advanced tools that allow control performance to be
automatically evaluated (as will be discussed in Chapter 3, Evaluating Control System
Performance). These tools can be used to identify measurement and actuator problems that can
impact performance that may be achieved using single loop feedback control. Also, some tools
automatically identify where process variation could be reduced by tuning (adjusting)
parameters associated with the PID (that is, the gain, reset, and rate parameters) to achieve the
best controller response to setpoint and disturbance input changes.

The procedure used to set the PID tuning parameters can be challenging when the process
response is slow or includes significant process delay. Thus, most modern process control
systems provide some capability to automatically establish loop tuning (as addressed in Chapter
4, On-Demand Tuning). When the process gain or dynamics change with changes in operating
conditions, advanced control tools for adaptive control can be applied as will be addressed in
Chapter 5, Adaptive Tuning.

Multi-Loop Techniques

When single loop control is not sufficient to minimize process variation or to

maintain the process at an operating limit, multi-loop techniques such as

feedforward, cascade, and override control should be applied.


One of the main reasons for implementing cascade control is that the secondary PID at each
point in the cascade can react quickly to disturbance inputs to its associated process.

If the PID responds quickly enough, changes introduced by disturbance inputs will have little or
no impact on the downstream processes. The process associated with the secondary PID loop
responds to changes in its manipulated input faster than the process associated with the
primary loop, so the above condition is satisfied naturally. A feature of cascade control is that
the secondary loop response time should be several times faster (most often 3–5 times) than
the response time of the primary loop.

In some cases, cascade control can be implemented to compensate for the nonlinear installed
characteristic of a regulating valve.

Single-loop PID feedback and multi-loop traditional control techniques such as feedforward,
cascade, override, split range, and valve position control can be used to address a wide variety
of control requirements. However, if a process is characterized by non-linear response, then
adaptive control may be required to improve control performance as described in Chapter 5.
When a faster response to setpoint and load disturbance cannot be achieved using PID control,
then fuzzy logic control may be required as detailed in Chapter 6. An intelligent PID capability
may be needed to improve recovery from process saturation or to use wireless transmitters in
closed loop control as addressed in Chapter 8. When a quality parameter needed in control
cannot be measured on-line, then on-line property estimation may be utilized as described in
chapters 7, 9, and 10. In power house installations, best operating efficiency may be achieved
through the use of online process optimization as described in Chapter 13. More complex
processes are often characterized by extremely long process delays, a high degree of process
interaction, and multiple operating constraints and measured process disturbances. The control
of these more complex processes is effectively addressed by advanced control techniques such
as model predictive control (MPC). (MPC for both small and large process is more fully
addressed in Chapters 11 and 12). Pulp bleaching and reformer temperature control are
examples of two applications that benefit from the use of model predictive control.

In general, advanced control techniques and tools should be considered when:


 The control objectives cannot be achieved through the improvement of traditional
control techniques.
 Traditional control strategies are more difficult to maintain at optimal performance
because of their complexity.
 Advanced control solutions provide greater economic benefits and improved product
quality, safety, compliance, and equipment protection.

By documenting the economic and operations benefits that can be achieved through the
application of advanced control, it may be possible to justify the cost of installing and
maintaining advanced control tools.

EVALUATING CONTROL SYSTEM PERFORMANCE

Thus, in many cases process and control engineers and instrument technicians at the plant level
spend a great deal of time addressing the most critical control and measurement problems that
plant operations personnel have identified as being potentially disruptive to production or
quality. There is usually no time or resources to examine the control performance to proactively
address problems that could later impact operation efficiency or limit plant production.

Supports continuous monitoring and detection of abnormal control and field device operation.
For older distributed control systems, a similar capability can be provided by layering
performance monitoring applications on top of the existing control system. In some cases,
however, the communication speed of a layered application can limit the system’s ability to
monitor faster control loops. These limitations are inherent in any layered applications designed
for older distributed control systems and programmable logic controller (PLC) systems.
Figure 3-3. Tools for Examining Control Loop Utilization
In summary, through the use of performance monitoring tools, it is possible to effectively
analyze and diagnose control performance with minimum manpower. Applying such tools can
have a significant positive impact on plant throughput and product quality.

Adaptive control can be used to improve loop tuning by:

 Establishing initial loop tuning without testing process, this is particularly useful when
the process output responds slowly to controller output changes in the process input.
 Identifying changes in process gain and dynamics as the process operating conditions
change.
 Automatically changing loop tuning to compensate for changes in plant operating
conditions.
A change in process gain and dynamics with a change in throughput is one of the reasons that
control utilization decreases when there is a significant reduction in production rate.

A perfectly-tuned PID controller can degrade over time and perform poorly or become
oscillatory. There are two main reasons for these changes:

1. The controlled process is non-linear, and the process operation has entered a region with
significantly different process parameters than those used during the tuning.

2. The process parameters have changed since the auto-tuning was performed.

Fuzzy Logic Control

For some specific process applications, fuzzy logic control enables faster setpoint recovery with
less overshoot than PID control for both setpoint and load changes (Figure 6-1). Fuzzy logic is
best suited for controlling processes that are characterized by large time constants and little or
no deadtime and where the control objective is to reach setpoint in the minimum time with no
overshoot.
Fuzzy logic control (FLC) can perform the same functions as classical linear control: simple
control feedback loops, interactive loops, or multivariable processes. In many cases fuzzy logic
control can provide better performance, less process variability, or better robustness. The
difference derives from calculations based on fuzzy logic. By applying fuzzy logic calculations, it
is possible to design a non-linear controller without a detailed knowledge of operating point
nonlinearity, as is required for a classical linear control design.
Table 6-2. Two-Valued Logic Operations vs. Multi-Valued Logic Operations
Neural Networks for Property Estimation

To address the non-linear response of product quality parameters to changes in process inputs,
the estimator can be based on a neural network model. This chapter addresses the
development of property estimators that can be used with non-linear processes.

The PIDPlus described in Chapter 8, Intelligent PID, may be used with sampled, non-periodic
measurement values obtained through lab analysis. Where the capability to automatically
correct process operation based on lab results is supported, an interface in the control system is
normally provided that allows the operator or lab technician to manually enter lab results into
the control system.

Alternatively, when a Lab Information Management System (LIMS) is in use, lab testing
equipment can be interfaced directly with the control system to allow the results of the lab
analysis to be automatically communicated to the control system as soon as the grab sample is
processed.

Function blocks that utilize partial least square or multiple linear regression models to provide a
linear estimator may be provided in the control system as detailed in Chapters 9 and 10 – Batch
and Continuous Data Analytics. To address the nonlinearity associated with some properties, a
non-linear estimator block based on neural network technology can be used to estimate the
property as illustrated in Figure 7-1 – see more details on function blocks in (Blevins, Wojsznis,
Tzovla, and Thiele, 2002).
When a property estimator capability is available in a control system, a collection of lab analysis
results is used in conjunction with a history collection of upstream measurement conditions to
create the estimator. Once the estimator has been created, lab analysis can continue to be used
to verify the accuracy of the estimated value.

Significant improvements in the tools used to develop neural networks for softsensor
applications have been made over the last years. Soft sensors are inferential estimators,
drawing conclusions from process observations when hardware sensors are unavailable or
unsuitable. These tools are specifically designed to take into account the dynamic nature of the
processes for which a soft sensor is to be developed. The difference in the traditional approach
compared to what is available today is illustrated in Figure 7-2.

Through the use of on-line data analytics, on-line decision support for operations personnel can
provide:

Product quality predictions: Identify quality problems while there is time to make on-line
corrections to prevent them.
Early process fault detection: Detect abnormal process operation and/or equipment problems
before they affect production. Provide root cause analytics to direct operations or maintenance
personnel to quickly correct the cause of problems. (Mason and Young, 2002)

Various techniques may be used to provide a prediction of quality parameters associated with a
continuous and batch process:

Neural Network (NN) Model: For use with non-linear processes, as addressed in Chapter 7,
Neural Networks. Application is restricted to the data range used in the development of the
model.

Multiple Linear Regression (MLR) Model: May be used with linear processes over a wide
operating range. Can be used to extrapolate outside the data range used in model development.

Partial Least Squares (PLS) Model: A linear technique that is preferred for use with collinear
data. (Collinear means that multiple process input changes impact similar process output
measurements.) Can be used to extrapolate outside the data range used in model development.

Process interactions that potentially impact quality parameter prediction and fault detection
may be accounted for through the use of principal component analysis (PCA). This multivariate
technique allows the fundamental underlying changes in the process to be identified. The
benefit that this multivariate technique provides over the univariate statistical analysis for fault
detection can be illustrated by considering two parameters of a collinear process at one point in
time, as shown in Figure 9-2.

MPC as a Replacement for PID


Model Predictive Control may be used for the control of single input-single output (SISO) as well
as multiple input-multiple output (MIMO) processes. In its simplest form, MPC may be used to
control a SISO process. For example, MPC can be used to provide an effective means of
controlling a SISO process that is described as “deadtimedominant” (i.e., when the process
deadtime is equal to or greater than the process time constant). The control performance
achieved may not be satisfactory when PID feedback control is applied to a deadtime-dominant
process because as the ratio of deadtime to time constant increases, the proportional and reset
gains must be decreased for stability. As a result, the control response to changes in setpoint
and load disturbances is often slower than that required for best process operation. In such
cases, control performance can be improved by replacing PID feedback control with model
predictive control as illustrated in Figure 11-1.
Process model identification

In addition, the MPC identification tool provides a comparison of the measured process output
values to those calculated based on the process inputs. If a significant error is identified in the
model during its verification, the easiest solution may be to collect more process data and
perform model identification and verification again. However, as in the case of disturbance
inputs, further collection of data may not lead to a better result. Therefore, correcting the
associated response models based on process knowledge would be a practical solution. The
MPC software provides options for editing the step response graphically or numerically by
entering the step response parameters (e.g., gain, lags, and deadtime).
PM and PE

Penalty on Move: Control sensitivity to changes in a process is

determined by the controller robustness. The Penalty on Move (PM)

parameter used in controller generation has the most impact on

robustness. The PM defines how much the MPC controller is

penalized for a change in the manipulated output (MV). The PM

parameter is typically defined independently for every MV. High PM

values result in a slow controller with a wide stability margin. With

such settings, control is relatively insensitive to changes in the

process parameters over time or to model errors. Low PM values

result in a fast controller with a narrow stability margin.

The PM value most affects the controller performance when the

model does not match the real process.

Penalty on Error: The Penalty on Error (PE) factor allows more

importance to be placed on a specific controlled variable. Typically,

the default value for the PE gives equal weight to all controlled

parameters and provides good control for most applications. If the

control strategy clearly indicates that one of the controlled variables

should be of lower priority, the associated PE can be decreased.

However, in general the PE should not be used to change overall

control performance. In most cases, it is recommended to change the

PE only after adjusting the MPC controller using the PM and then
testing MPC control in simulation. The primary criterion for

adjusting the PE is acceptable variability on a specific controlled

parameter.

In addition to the parameters used during controller generation, the user can

influence control behavior and robustness during on-line operation. One way is

to apply a reference trajectory (i.e., the filtered predicted setpoint value), which

can also be used for funnel or range control (Figure 11-41). In funnel or range

control, the process predicted CV values that are within an acceptable setpoint

range or within the funnel are not penalized (B, C area), thus PE is applied to the

values outside the funnel.

11.7.5 MPC Simulation and Tuning

Control Evaluation and Tuning adjustment

The MPC operator interface is configured automatically for the MPC module.

The trending display serves as the main means for evaluating MPC operation and performance.
After the MPC controller is turned on, the user should observe its operations for various
conditions: setpoint changes, constrained control, disturbances compensation and optimization.
If controller performance is not satisfactory after applying a setpoint filter, the user should
regenerate the controller with new controller settings. In such cases, thorough model revision is
advised as well.

For an engineer, a good way to be introduced to MPC is by showing an analogy with PID or with
a feedback controller in general. Model Predictive Control uses modeling and a specific type of
feedback that makes it distinctive. As opposed to a traditional control loop where the controller
uses a difference (error) between the setpoint and the recent values of the measurement as its
input, the MPC controller applies the difference between the future trajectory of the set point
(whether a constant or a known function of time) and the predicted trajectory of the controlled
variable as its input. The difference is expressed not by a single value as in a traditional feedback
loop, but as a vector of error values from setpoint from the present time to a set time in the
future, usually defined to cover the settling time of the process.

Figure 11-32 shows an MPC controller for a process with two inputs and one output, in a form
that allows one to see the analogy with a typical feedback control loop. The process has a
manipulated variable (MV) and a disturbance variable (DV) on the input and a controlled
variable (CV) on the output. A simple MPC controller used in this configuration has three basic
components:
 A process model that predicts the process output for 120 or more scans ahead. (For
consistency, we will use a prediction horizon of 120 scans throughout this discussion).
 A future trajectory of the setpoint for the same number of scans as the trajectory of the
predicted process output.
 A control algorithm for computing a control action based on the error vector as the
difference between the future trajectories of the setpoint and the predicted process
output.

The controller output is a manipulated variable (MV) applied to the inputs of the
process and the process model. A measured load upset to the process input is
also applied as a disturbance variable (DV) to the model input. The process
model computes a predicted trajectory of the controlled variable (CV) that is the
process output. After this trajectory is corrected for any mismatch between the
predicted value and an actual measured value of the controlled variable, the
predicted trajectory is subtracted from the future trajectory of the setpoint to
form an error vector as shown in Figure 11-33.
11.9.1 The Basics of Process Modeling

You might also like