BLOG Series: Frequently asked questions on NAPCON Advisor and machine learning - Part 1
Samuli Bergman, PRODUCT MANAGER / 25 Oct 2021
Machine learning and first principles models
Digital operator assistant is the best new coworker for a panel operator in chemical process industries, such as petroleum refineries, petrochemical plants and biorefineries. The role of a digital operator assistant is to give justified and reasonable suggestions and advice to panel operators. The goal is to help operators run the plant in a more optimal and safe way and meet the production targets.Our NAPCON Advisor is such a digital operator assistant.
The core of NAPCON Advisor is a set of machine learning models that have been trained with process data. The input to the models is streamed process data that they get continuously from DCS and other systems. These models use the current and historical data to predict the behavior of the production process several hours in the future. What I find particularly significant is the fact that the machine learning models don’t just cover the chemical and physical phenomena of the production process, but they also span the process automation layer.
NAPCON Advisor is a fairly new product that interests many of NAPCON’s customers. I have noticed that when we start to discuss with production managers, plant engineers, automation engineers and other customer’s experts, we typically get very similar questions. I will sum up a few of the most typical questions – and answers – in this three part blog series.
How can data-based ML models beat
100 years of chemical engineering?
We get a question very often, can data-based machine learning models be better than first principles models? How could a basically black box model give better results than the one that is based on known chemical and physical models? As a chemical engineer I love these questions, because they force us to contemplate what models are about and what they are used for?
I don’t think there is or there should be a competition between first principles models and data-based, dynamic, machine learning models. It all depends on the use case. I think we should always apply such a model structure that best suits the purpose. Machine learning models that are used by NAPCON Advisor to predict the process behavior in the future are closely related to differential equations. I think they have four qualities that make them especially suitable for the digital operator assistant use.
– NAPCON Advisor’s machine learning models don’t just cover the chemical processes. Models also capture process automation; sensors, valves and actuators as well as process control applications.
The most important difference between a typical first principles model and a data-based machine learning model is that process control and automation are inherently incorporated into the machine learning model. Hence, the machine learning model is not just a model on the chemical and physical phenomena in the plant, but it also captures things like measurement noise, the dynamics of valves and actuators and controller tuning. This all makes the model more realistic for everyday use.
The purpose of the Advisor’s model is to predict the future values of related process measurements, when an intended change in one controller set point is introduced. The prediction takes into account the current state and the recent history of process measurements. Building and running a first principles model for this purpose would be unfeasible, in my opinion.
– NAPCON Advisor’s machine learning models describe the true plant behavior based on data. Slow deterioration, such as fouling or catalyst deactivation, is part of the model.
NAPCON Advisor’s machine learning models describe the true plant behavior based on data. They are not idealistic, simplified nor theoretical. They also take into account the nonideal and slow phenomena occurring at the plant. Again, an important factor for this type of real time use as a digital operator assistant.
– State-of-the-art machine learning models are able to generalize. Machine learning models can assess their own uncertainty.
Recent developments in machine learning have improved the ability of ML models to generalize. Modern machine learning models can interpolate and extrapolate outside their training data. What I have found especially important is that modern machine learning models can assess their own uncertainty. In practice, this means that the output of the model is not just a vector of future process values but an area or an envelope that describes where the future process values will be in high probability. This is something that is familiar from our everyday life; from weather forecasts showing the confidence limits for future temperatures.
– Machine learning models are faster to construct than many other model types.
It is well-known that machine learning models require a lot of data. Fortunately, training a machine learning model is a systematic process. That means that training and validation of machine learning models can be automatized. My experience is that, at the end of the day, this automatization makes the work process to construct machine learning models faster than the work processes for many other model types – despite the vast amount of data.
Next in this blog series
- How do machine learning models handle erroneous or invalid data?
- How can you guarantee the benefits of a digital operator assistant?
About the author:
My first experiences on artificial intelligence applied on process data were in the year 2000 when I used self-organizing maps to build an application to detect and identify abnormal situations in a refinery unit. However, I have worked most of my career with advanced process control. Currently I am product manager, NAPCON Improve, at Neste Engineering Solutions. My product portfolio contains software for advanced process control (APC), dynamic real time optimization (DRTO) and other real time software for panel operator use. I have a MSc(Eng) degree in chemical engineering and process control from Helsinki University of Technology.