APC AND

ARTIFICIAL
INTELLIGENCE
How artificial intelligence can improve advanced
process control.
INTRODUCTION
The development of manufacturing industries
has seen a constant drive to improve
operational efficiency.

Growing the business means outpacing your competitors to bring product to

market quicker and cheaper. Central to driving those efficiency gains has been the
development and adoption of new technology. However, having new technology
doesn’t mean that existing technology is superseded. Often the challenge we have
is to work out how to use both new and existing technology in a way that we can get
the best out of both.

A technology which has shown great benefit in manufacturing industries has been
the application of algorithms which use data from the process to identify new
improvement opportunities – a “data-driven” approach. A good example of a data-
driven, real-time algorithm is Advanced Process Control (APC), built on multivariable,
model-based predictive control (MVPC) techniques. While APC remains the dominant
technique for optimising continuous processes, new technologies are emerging.
New data-driven empirical methods, popularly described as Artificial Intelligence
(AI) or Machine Learning (ML), expand the toolset we have. An open question though
is how best to use these techniques. Do they replace APC? Do they assist APC? Do
they solve a whole new class of problems? Let’s look at some ways that AI/ML can
be used in the process industries, and how that impacts the traditional APC space.

APC and AI Whitepaper | www.honeywell.com | 2

WHAT IS
APC
Advanced Process Control (APC) is a real-time production optimization technology
which has been proven to provide tangible benefits on manufacturing plants for
over 30 years. Using multi-variable, model-based predictive control algorithms, it
can simultaneously control and optimize complex processes in industry. By taking a
closed-loop control approach to adjusting plant operation, it can increase throughput,
maximise yield and reduce operating costs.

The multi-variable predictive control algorithm is one of the most practical examples
of the use of empirical, data-driven analytic techniques to bring production benefits.
It has particular strengths in being able to handle continuous processes, with highly
interactive variables and a mix of dynamics. This accounts for many of the processes
commonly encountered in industries such as refining, petrochemicals and minerals
processing.

THE APPLICATION
OF AI/ML
The benefits from AI/ML are not as straightforward as simply replacing what used to
be done with APC with an AI/ML approach. A mix of technologies is needed to properly
address the broad range of production optimization opportunities in industry. For
linear, interactive problems, traditional APC can do a great job. It will identify and hold
an optimum operating point consistently, the methodology is well established with
low implementation risk, and the costs are generally low enough to generate a good
return on investment. Choosing the simplest and most cost-effective solution to a
problem makes good business sense.

Creating applications which utilise a traditional APC approach combined with AI/ML
techniques allows us to broaden the type of problems APC can handle and improve
the results from our APC application. One particular advantage which APC offers AI/
ML is APC’s well-established connectivity to field data, and the ability to download
optimal setpoints while managing process dynamics. It also provides the real-time
orchestration platform to execute AI/ML routines, feeding data to the model and
taking results from it. In Honeywell Forge APC, this is achieved through embedded
toolkits in the runtime environment, which can provide a range of supporting
functions to the APC application. This includes the ability to run Python scripts as part
of the APC application execution cycle. The strength of the AI/ML algorithm in this
case is its ability to process data that APC cannot deal with directly, such as discrete
data, or image data. In this hybrid approach, the AI/ML routine processes that non-
analog data, and can provide guidance to the APC application, in terms of improved
inferential measurements, dynamically modifying tuning or limits, or changing
optimization strategies.

APC and AI Whitepaper | www.honeywell.com | 3

PRACTICAL EXAMPLES
OF AI/ML WITH APC
The area where AI/ML techniques are finding most promise is in enabling and
enhancing the capabilities of APC. AI/ML can help address some of the most
common challenges with APC. Some of those opportunities are in the design and
development phase, and some of those are in the online execution phase.

Here are three examples of using AI/ML to enhance APC benefits:

1. in the design phase to improve model identification from historical data

2. in the implementation phase, to identify models from operating data in real-time

3. in the online phase, to dynamically identify and manage changes in the process
state (for example, due to a different feed or operating mode).

EXAMPLE 1
Improving model identification from historical data

Having an accurate process model is fundamental to achieving tight control. Better

models mean better predictions, and better predictions mean better ability to control.
Traditionally, developing and maintaining the APC models are one of the most time-
consuming activities in an APC project. It requires time and effort over days or weeks
to manually step-test the process, extract data, and conduct model identification.

New analytic techniques offer the opportunity to greatly improve this model
development process. Firstly, we can try to get more information from our existing
historical data, before we start active testing. Surely all those setpoint changes which
the operator makes during normal operation is just “free” data? Traditionally this
approach has proven problematic, because the historical data contains some useful,
representative data, but also contains low quality data which is not representative
of normal process behaviour. This can be due to upsets in the process, inaccurate
process readings or simply long periods of data without any operator actions. Sifting
through that data set manually to identify some “good” periods of operation will
probably take longer than going ahead and doing some step testing on the process.

Honeywell has developed new capabilities in Honeywell Forge APC Engineering

Studio to help with this process. The model identifier can select periods of high-
quality data out of the historical data set, using similar rules to how a human
might approach such a data mining task. Additionally, techniques like wavelet
decomposition provide an accurate way to locate those useful operator set point
changes. The model identification routine can then run using just this high-quality
data and generate a reliable and representative process model.

APC and AI Whitepaper | www.honeywell.com | 4

EXAMPLE 2
Learning new models in real-time

An extension of model identification from historical data is to learn the model from
operating data, in real-time. In this case, the APC application continues to perform its
primary function of control and optimization of the process, but in the background,
the application is also identifying new and potentially improved process models from
the live operating data.

The tool which allows this real-time model identification is known as Honeywell
Forge APC Online Modeller. There are several important functions that make up this
learning process. Firstly, small perturbations in the setpoints are necessary to create
some level of process response that the identifier can use. These small steps are
implemented automatically by the modeller and do not require any manual actions by
the user. The APC application remains in control of the process, and if any potential
constraint violation is predicted, it will take normal action with manipulated variables
to keep the process within constraints.

At a regular interval, the modelling algorithm will run, and identify new models from
the real-time operating data. This algorithm, known as Global Multi-Stage (GMS), is
fully capable of handling the closed-loop data collected in normal controller operation,
including highly correlated data. The algorithm will present the identification results
to the user, with detailed statistics indicating the confidence in the model results, and
how the model response has been changing over time. Models which are considered
to be accurate can then be downloaded to the running controller.

This ability to identify and refine the model in real-time offers great benefit to the user.
Firstly, it opens new application implementation methodologies, allowing an initial
model to be developed from historical data (a “seed” model), and then refined during
the commissioning phase to come up with the final, high-quality model. This greatly
reduces the time to get a new APC application installed and commissioned.

Secondly, once the controller is in service, it allows the user to quickly and easily
update any suspect models. No manual actions are required to step test or model,
with the Online Modeller simply presenting its results to the user on a regular basis.
The user’s day-to-day controller maintenance activities are greatly streamlined,
resulting in a more accurate process model, with less engineering effort required.

APC and AI Whitepaper | www.honeywell.com | 5

EXAMPLE 3
Use to identify change in process state

APC is effective when dealing with a process that displays consistent model
responses. However, many things can cause a process to move away from its
predicted process responses. A common example is a change in feed quality. Using
a denser feed, a coarser feed, a feed with higher moisture content or a feed with a
different composition may all require the operating point of the APC to be adjusted.
In practice, this often requires the operator to recognise the change themselves, and
manually change the limits on controlled or manipulated variables. With this guidance,
the APC can return to optimising the process.

Analytics can help in this situation by identifying a change in feed type, and by
selecting new limits to use for the APC application. In one specific example for a
minerals processing operation, the application used a clusterization model to identify
a change in feed quality. Input variables for the clusterization model include feed
characteristics and operational conditions.

When a change in feed quality is detected, an optimization routine using a simulation

of possible operating decisions identifies an optimum set of APC range and setpoints
to maximise production. In this case, the clusterization models and optimization both
run online, in real-time, and provide their recommendations to the operator as an open-
loop advisory. The operator can then download these new operating conditions to the
APC, which manages the dynamics of moving the process to its new operating point.

Using this approach, we can exploit the relative strengths of both AI/ML and APC.
The AI/ML algorithms handle the detection of a change in feed type (a classification
problem), and the selection of optimum limit sets (which has non-linear and discrete
characteristics). The APC provides the framework to implement those objectives,
ensuring the process is optimised within bounds. Experience on a SAG mill for copper
processing has shown this combination can deliver a 3.5% increase in production and
0.9% improvement in recovery over APC alone.

APC and AI Whitepaper | www.honeywell.com | 6

SUMMARY
These examples show there are a broad range of techniques which can be used for
production optimization. Likewise, there are a broad range of production challenges
to be addressed in industry, each with their individual characteristics. There are no
silver bullets but plenty of options for the skilled engineer to consider when choosing
a production optimization approach.

The right choice will come through gaining a detailed understanding of the process,
and then selecting a technique to apply. Both APC and AI/ML approaches can be
complementary when correctly applied. Keep an open mind and know the technology,
and the expected benefits for your next application will certainly be achieved.

This document is a non-binding, confidential document that contains valuable proprietary and confidential information of Honeywell
and must not be disclosed to any third party without our written agreement. It does not create any binding obligations on us to
develop or sell any product, service or offering. Content provided herein cannot be altered or modified and must remain in the format
as originally presented by Honeywell. Any descriptions of future product direction, intended updates or new or improved features or
functions are intended for informational purposes only and are not binding commitments on us and the sale, development, release
or timing of any such products, updates, features or functions is at our sole discretion. All pictures shown in this document are for
illustration purposes only; actual product may vary.

Honeywell® is a trademark of Honeywell International Inc. Other brand or product names are trademarks of their respective owners

Honeywell Connected Enterprise

715 Peachtree Street
Atlanta, GA 30308
1 (877) 841-2840 APC and AI Whitepaper | 03/21
www.honeywell.com © 2021 Honeywell International Inc.