Ai Guided Optimization

2021-07-06
Stefan Meili
smeili@noram-eng.com

Project Overview

NORAM Analytics has applied it's machine learning guided process optimization technology to increase the maximum sustained production capacity of a Huntsman Polyurethanes plant producing mononitrobenzene (MNB) by 3% and improve product quality. Originally designed by NORAM Engineering and Constructors, the Huntsman plant has been in operation for over 20 years and has been systematically debottlenecked. It continues an enviable reputation as a safe, efficient and reliable production facility.

Project Goals

Huntsman requested assistance in optimizing the operations of their MNB plant, as our machine learning technology aligns closely with the objectives of Huntsman's advanced analytics roadmap. Our shared goal was to increase capacity by applying machine learning techniques and achieve sustainable and scalable improvement. Furthermore, an improvement in product quality was highly desirable, in keeping with Huntsman's commitment to reduce environmental impact and drive towards sustainability.

NORAM Analytics applied it's machine learning guided process optimization technology to simultaneously achieve a 3% increase in capacity and a 20% reduction in byproduct concentration.

Challenges with Process Optimization

Process optimization is a serious challenge! Process control set-points rarely control only a single output parameter. Control inputs almost always interact with one another. Furthermore, optimization goals often conflict with each other. For example, increasing production capacity of an existing plant often results in a disproportionate increase in energy consumption and product loss to side reactions due to equipment limitations.

In this example, mononitrobenzene (MNB) is produced by reacting nitric acid with benzene in sulphuric acid. These react to form MNB and water and release a significant amount of heat. Product MNB is separated from weak sulphuric acid in the MNB / Acid decanter. Weak sulphuric acid is re-concentrated in a sulphuric acid flash evaporator (SAFE) under deep vacuum and then recycled to the MNB Reactors. Water removed from the SAFE is condensed and then used to wash the product.

'MNB Plant'

This first step in the process appears simple. However, tuning operations to maximize performance can be extremely complex. There are 8 major degrees of freedom in this system including feed stock rates (4x), sulphuric acid circulation rate (2x), as well as SAFE temperature and pressure.

Optimization is further complicated by the nonlinear nature of the reactions. A small change in any one of the inputs can produce a very large response in the plant. Heat losses and cooling water temperatures vary seasonally, and feed stock quality varies between shipments. A final confounding factor is that the product of the two parallel reactors is mixed before being sampled and analyzed in a lab.

Applying Our Deep Learning Technology

NORAM Analytics built a deep neural network to model Huntsman's nitration plant and trained it to suggest improved operating conditions.

Six years of site data was used to train a model that can predict not only how changes to control inputs will affect readings from sensors connected to the Distributed Control System (DCS), but how the quality of the product will be impacted. One of our challenges was connecting the two different data sets logically.

Neural networks are incredibly flexible and powerful, but care must be taken to ensure valid predictions are generated. Our deep learning model of this process was structured to match the unit operations of the process and trained using rigorous hold-out cross-validation. 80% of available data was used to train the model, while 20% was used to test it's predictions. After this, a technique known as 'transfer learning' was used to leverage a model trained the large volume of data available from the DCS and use it improve predictions about the relatively small amount of data available in the product sample analysis data set.

Process Goals

Improving a process requires not only a detailed process model, but a deep understanding of the objectives being achieved. To mathematically optimize a process, an objective function must be constructed that scores the performance of any given operating point.

The score must penalize operating points outside of safe operations like alarm and trip set points programmed into the DCS. Penalties are also applied to operating points that produce poor product quality. Conversely, rewards are given when operating in regions that improve product quality or improve production capacity. Once an objective function is constructed, an optimal operating point is suggested by using an algorithm that maximizes rewards (and minimizes penalties).

A total of over 200 constraints and goals were used to score the prediction of our MNB plant model. These included alarm and trip set points, production capacity targets, product quality specifications,

Optimizing the MNB Process

The deep learning model and objective function work together in a dashboard that suggests improved operating conditions. Clients can use this dashboard to set goals for production rate and product quality and select 'optimize' to suggest controller set points that best achieve those goals. The plot below summarizes how closely a model suggested operating point meets the targets.

'Model Predictions'

Explainability is critical in building the confidence to take an Ai generated operating suggestion and then implement it in the field. The costs and consequences of a miss-step are non trivial. At NORAM Analytics we support this critical step in two ways.

First, model predictions can be sliced to understand which inputs (or combination of inputs) generate the strongest response. Additionally, NORAM Analytics provides design level insight into the process. We help explain model predictions with engineering analysis of the process. Our model can be interrogated to understand why a suggestion is being made. Once an operating point has been suggested, the effect of changes to individual controllers in isolation can assessed. Instead of digging through time history data, the model predictions can be 'sliced' along each control axis.

'Model Predictions'

In the case of our client's operation, the model identified that a significant reduction in byproducts (nitrophenols) could be achieved by reducing the temperature of the acid flash evaporator (SAFE) in concert with several other inputs. The plot below shows the predicted response of crude byproduct concentrations when SAFE temperature is changed at 103% of the plant's maximum capacity. Reducing temperature will significantly reduce nitrophenol concentrations.

Optimized Operations

And The Result?

After careful internal review, Huntsman implemented the suggested operating conditions in a trial. First the plant production rate was increased to 103% of previous rated capacity using standard operating conditions. After several days of stable operation, the optimized operating conditions were gradually introduced in two steps. Production rate, model predictions and lab analysis of crude product are plotted below. Model predictions are presented as a median (solid blue line) and 25% - 75% quartile predictions are presented as the blue shaded area. The two trial operating points are highlighted in grey.

'Model Predictions'

The results are summarized below, comparing the predicted outcomes sent to Huntsman before the trials began with the lab sample analysis drawn from the crude product. The shaded area is a skew normal distribution ft to the statistical information predicted by our model, while the red points are the sample analysis results gathered from site.

Optimized Operations

Our model suggested an improved operating point that reduces byproduct concentration in crude product by 20% at 103% of the previously accepted maximum capacity.

Conclusion

We hope this example illustrates how we can help you improve the day-to-day operation of your industrial process. Your system may be entirely different, but our approach to solving problems and modelling complex interactions can be applied with great results. With Huntsman, we were able to demonstrate a 20% reduction in byproducts when operating at 103% of previously accepted maximum capacity. Our deep learning models make detailed predictions about how your process responds to changes in controller inputs and can be interrogated to understand why a suggested operating point is an improvement over past practice.

Whether you want to increase production capacity without downtime or new equipment, or improve efficiency when operating at alternate rates, our team at NORAM Analytics is here to help.