1R. Pitchumani, R. Paleja, S. Nair, K.-C. Goh, J. Williamson 2L. I. Chukwu, E. Osemwinyen, O. Ajayi, E. Krohn

1Shell Global Solutions 2Nigeria LNG Limited

Raghuram Pitchumani, Shell

With capital intensive LNG plants, fluctuations in feed gas and ambient conditions are some of the challenges that necessitate operators to explore ways to increase their production margins. A C3/MR (propane precooled mixed refrigerant) plant is one of the widely used refrigeration processes for onshore LNG plants. From an energy intensity and value perspective, the mixed refrigerant (MR) loop has the largest opportunity for optimization and drives the core of the LNG plant, the main cryogenic heat exchanger (MCHE) for liquefying natural gas. The simplified maximisation of throughput based on Advanced Process Control is well established in Nigeria LNG, using technology and support from the Shell Group. However recently, advances in applied data analytics in technology driven industries have attracted attention to the Oil & Gas industry with applications in process and production optimisation. The current work applies data analytics to operational data to identify/reaffirm key optimisation handles in a MR loop and develop empirical models for LNG production and MR compressor power as a function of key independent variables. These models were incorporated in the design of the MCHE advanced process controller (APC), to further operate the plant more efficiently resulting in ~1-2 % additional LNG production.

REAL TIME PRODUCTION OPTIMISATION OF MIXED REFRIGERANT (MR) LOOP BY DATA ANALYTICS

INTRODUCTION

Variation in feed gas and ambient conditions pose a challenge for liquefied natural gas (LNG) plants and drive

different operating strategies to produce the maximum amount of LNG using the available power. The propane

mixed-refrigerant (C3/MR) process is a widely used liquefaction scheme in LNG plants with the liquefaction circuit

consisting of the mixed refrigerant (MR) loop and the Main Cryogenic Heat Exchanger (MCHE) being the heart of

the process. The MR loop and the MCHE are the most energy intensive part of the process due to high energy

demand for compressing the refrigerants required for liquefaction.

In the LNG business and specifically for the MCHE, it is very common to derive optimal operating conditions using

mathematical methods, thermodynamic principles and simulation techniques which are then applied to real time

process optimization using a suitable controller. Kyuangtae et al (Kyungtae Park, 2016) used a simulation strategy

to find the optimal design and operation solution to increase LNG production rate under scenarios of fixed MR

composition, available turbine power, changes in ambient temperature and heat transfer coefficient. Khan et al

(Mohd Shariq Khan, 2013) used Aspen Hysys dynamic simulator to derive mixed refrigerant compositional flow rate

that maximises exergy.

The research reported in the literature is purposeful to meet the design criteria but does not deal with sublime

optimization opportunities that may be required to de-constrain the MCHE in real time and gain extra production

with the available power. We discuss how using historical, closed loop data led to new insights - using data

analytics and Machine Learning - about factors that impact production. The insights were subsequently captured in

a new process controller that applies real time optimization to increase production by 1-2%.

ADVANCED PROCESS CONTROL AND REAL TIME OPTIMIZATION

In Shell, Advanced Process Control (APC) and Real Time Optimization (RTO) are standard tools applied to

stabilise and optimise complex processes in real time (Seborg, Edgar, & Mellichamp, 2004). Note that although the

term Real Time Optimization is commonly used in the industry, we will use the term Real Time Production

Optimization (RTPO) here, given the focus on production, as opposed to an economic objective function that is

often applied in RTO. The APC controller and the Real Time Production Optimiser (RTPO) lie at the apex of the

Automation pyramid shown in Figure 1.

Most commonly, APC uses dynamic linear multi-input, multi-output process models for closed loop predictive

control that stabilises complex processes by simultaneously manipulating and controlling several interactive

variables. APC models predict changes in the dependent variables due to changes in the independent variables.

The independent variables are commonly referred to as Manipulated variables (MVs) and Disturbance Variables

(DVs). The dependent variables are commonly referred to as Control Variables (CVs).

For this project, Platform for Advanced Control and Estimation (PACE) software was used for the APC application.

PACE uses special type of dependent variables other than CVs. These are referred to as Intermediate Variables

(IVs) and these variables are dependent variables that are used as inputs in the model to other downstream

dependent variables. The use of IVs facilitates building of grey-box modes within PACE which allows more

engineering knowledge to be embedded in the controller (Amrit, et al., 2015).

Raghuram Pitchumani, Shell

The APC dynamic models are derived from plant data via step testing. The dynamic models are then used to

compute moves of the APC controller based on the current dynamic state of the plant (the prediction of where all

the dependent variables are going) and the targets and limits of the CVs. Due to the feedback process, APC

controllers in practice are robust to inaccuracies inherent in the use of simplified linear models. APC controllers

can also be used to drive simple optimization objectives, such as to maximise throughput. In Shell, Advanced

Process Control has been applied to the LNG MCHE / Refrigerant Loop progress since the 2000s. (Hupkes, Lin,

Silve, & Vink, 2004), (Alers & van Dijk, 2010), (Den Bakker, Dejsupa, Beeby, Azodi, & Silve., 2006).

For more complex processes, including those where there are significant nonlinearities, including energy costs, and

products with varying economics; optimization models are required to capture key interactions, including

nonlinearities, between the inputs (manipulated and disturbance variables) and outputs (controlled variables) of the

process. A significant number of Real Time Production Optimisers have been utilized for both upstream (wells)

production and downstream manufacturing plants (Cramer, Stoever, Mehrotra, Berendschot, & Goh, 2013).

Presently, data driven models have been proven more economical and sustainable for upstream (Briers, et al.,

2016). However, for more complex processes in downstream manufacturing, first principles “White Box” or “Grey

Box” models have predominated (Onstott & Linn, 2000). For the MCHE / Mixed Refrigerant Loop in LNG, progress

to set up a model suitable for RTPO has had limited success partly due to the large numbers of interacting inputs

and outputs required for any RTPO model and the complexity of modelling the MCHE in the mixed refrigerant

liquefaction process, and developing a high-fidelity plant data rated model.

There was uncertainty as to whether there was an opportunity to increase production above that already derived

from the existing APC Systems. However, with the realization by the oil and gas industry of the potential for

modelling using Data Analytics, it was deemed worthwhile to examine this area with a view towards obtaining a fit

for purpose model for RTPO. A very exhaustive list of references for the optimization for the Gas Liquefaction

process, and the LNG plants in general, is found in (Austbøa, Weidemann Løvsethb, & Gundersena, 2014).

Figure 1 : Typical Process Control pyramid (hierarchy).

INTRODUCTION TO DATA ANALYTICS

Data analytics or Machine Learning is the science of discovering hidden patterns in the raw data and making

predictive models. This stream of science has been widely used in the healthcare industry for a long time to find

novel drugs using both large volumes of structured and unstructured data. More recently, the retail, travel, banking,

finance and cyber-security sectors have used data analytics to efficiently serve their customers’ needs as the cost

of storing and processing large volumes of data has significantly declined with time.

A variety of techniques are available as shown in Figure 2 (ML cheat sheet, 2017) and the choice of technique

depends on the type of data and the problem that needs to be addressed. Supervised learning techniques map

input variables to one or more response variables while unsupervised techniques are used when there is no

response variable. When the response is numeric (as for all the responses in this paper), some of the most

appropriate techniques that can be applied are decision tree, linear regression, random forest, neural network and

gradient boosting tree

Essential steps in modelling are data collection, data summary, visualization, model building, model simplification

and testing the model by predicting the response in dataset unseen by the model. A simple model that is easy to

interpret is preferred to a Black Box model approach which does not allow us to learn about the directional impact

of independent variables. Techniques to simplify models are well documented in the Machine Learning literature

(Crawley, 2011) but it is advisable to combine these approaches with input from experts who have process and

process controls background to enable correct choice of variables in the model. This is the approach followed in

the work reported, which is the result of close collaboration between Shell Group Data Scientists, Process

technology, Process Control and Optimisation engineers and NLNG (Nigera LNG) site Process/Process Control

technologists.

Figure 2: Machine Learning algorithms cheat sheet (ML cheat sheet, 2017).

THE MCHE MR PROCESS

The MCHE MR loop is the heart of the LNG liquefaction plant. Figure 3 shows a typical MCHE. The flow scheme in

blue shows the MR loop. Natural gas, here shown in green, enters through the tube side at the bottom of the

MCHE at ambient temperature and exits the top as Liquefied Natural Gas at a rundown temperature of -150 to -145

degC, depending on the control system settings. Liquefaction of natural gas is achieved by flowing mixed

Refrigerant, a mixture of liquid nitrogen (N2), methane (C1), ethane (C2) and propane (C3) on the Shell side of the

heat exchanger. The mixed refrigerant is divided into two streams; Light Mixed Refrigerant (LMR) and Heavy Mixed

Refrigerant (HMR). LMR provides the cooling duty in the top section of the heat exchanger which is also called the

Cold Bundle. HMR provides the cooling duty in the middle and bottom sections of the heat exchanger which are

respectively called the Intermediate and Warm Bundles.

As the mixed refrigerant exits the MCHE in gaseous form, a standard compressor marked as “KT” compresses the

refrigerants, which are then cooled to -33 degC by air coolers and propane kettles marked “C3” in the figure. A

typical compressor is designed to deliver up to 100 MW of power.

Figure 3: Schematic representation of MCHE-MR loop.

USING DATA SCIENCE TO MODEL LNG THROUGHPUT AS A FUNCTION OF INDEPENDENT VARIABLES

In order to maximise LNG throughput, a model relating independent variables to throughput is needed. It is known

from a process and operational perspective that the LNG throughput is a function of the flow rate of LMR(F-LMR),

HMR (F-HMR), the composition of MR (Nitrogen:C-N2, Methane: C-C1, Ethane: C-C2 and Propane:C-C3), run-

down temperature (T-R) and inlet guide vane position of the MR compressor “KT” (V-K). These are manipulated

variables (MVs) that are generally already used in any APC application. To discover new hidden patterns that will

help to increase the LNG throughput, first a quantitative model for LNG throughput was derived using the historical

closed loop data of the above MVs. In other words, no new experiments were conducted with the process controller

taken offline but use of historical data that had been accumulated with the process controller online was made to

derive the model. Given that we have a response that is numeric, a variety of supervised regression techniques

were available. The choice boiled down to partial least square regression, PCA regression (de Jong, 1993) or

simple least squares regression (Goldberger, 1964). With the first two techniques, it became difficult to prevent

overfitting and determine genuine effect of some MR compositional factors. This is because both these techniques

transform the original MV data into linear combination of all MVs. Given that there was little excitation/variability in

some of the MVs, using them to explain the variability in LNG throughput would lead to incorrect conclusions.

The least squares method involves developing a model as shown in equation 1

𝑦 = 𝑿𝛽 + 𝜀 (equation 1)

Where 𝑿 is the matrix of MVs, 𝛽 is a vector of unknown parameters that we need to estimate and 𝜀 is the error in

the model. Estimation of 𝛽 is done using equation 2

�̂� = (𝑿𝑇𝑿)−1𝑿𝑇𝑦 (equation 2)

The standard least squares methodology was adapted to remove the MVs from 𝑿 through cross validation. In this

modified technique, the historical data was divided into k-blocks and 80% of the blocks were randomly chosen to

build a parameterized model in equation 1 and 2. The process of randomly selecting 80% of the blocks was

repeated many times to establish which MVs had statistically significant impact on LNG throughput. This was done

by selecting only those MVs that gave coefficients with a p-value <0.05 and consistent sign in all the random tests.

This cross-validation scheme is slightly different to that described in the literature (Gavin C. Cawley, 2010) in that

only the influential MVs were identified using 80% of the blocks. The predictive performance of the model was

assessed by applying the model to a separate data set which was not part of the model building exercise.

The MVs that were statistically significant in all the random tests were used to derive a single representative model

for LNG throughput. The residuals, or the difference between the actual LNG throughput and the fitted throughput,

were treated as the variability in LNG throughput that cannot be explained by the MVs alone. The residuals were

then modelled against all possible intermediate process tags alone or their interaction with MVs already isolated.

Those process tags that had statistically significant impact on the residuals were isolated for discussion with

process engineers to understand if they could be controlled with the existing MVs or whether new MVs outside the

scope of the MCHE loop would be required. Two new intermediate variables were thus discovered. These were (a)

NG Process temperature in the warm bundle (T-WB) and (b) Process temperature of the MR composition (T-MR).

Figure 4: Left panel: Residuals (y-axis) vs T-WB(x-axis). Right panel (Residuals (y-axis) vs T-MR.

Figure 4 above shows the influence of T-WB and T-MR on residuals of LNG throughput model with MVs only. We

can see that the effect of T-WB on LNG throughput is non-linear and only temperatures that are neither warm nor

cold are best suited for increasing the LNG throughput. MVs from the MCHE/MR loop or other control loops which

help to maintain T-WB in its mid-range were identified and incorporated in the new controller. T-MR, on the other

hand has a linear effect on LNG throughput and colder T-MR is desirable to maximise production. MVs that help to

achieve T-MR as cold as possible were identified and incorporated into the new controller to maintain LNG

throughput as high as possible. The model for LNG throughput is shown in Table 1 below. The impact of all the

variables is standardised with the strongest contributor given the highest rank.

Table 1: Model for LNG

throughput using MVs and IVs.

The overall model after the inclusion of the intermediate variables (IVs) was robust with 92% of variability in LNG

throughput explained. The IVs were mapped to MVs from MCHE-MR loop or to the other loops such as MCHE-

Propane and Acid Gas Removal Unit (AGRU) by building models for IVs as a function of MVs.

Maximising the MR Compressor Power usage To maximise LNG throughput, it is logical that all the power of compressor marked “KT” in Figure 3 is appropriately

utilised by maximising the flow of HMR and LMR (F-HMR and F-LMR) at the right composition. However, if a

process constraint is violated, the APC will back off the “pushing” of the HMR and LMR, and therefore the

throughput. Some of the constraints which restrict the utilisation of MR Power include the pressure drop across the

MCHE (P-MCHE) and the suction pressure of the MR compressor (Ps-Comp). Both P-MCHE and Ps-Comp

have an upper limit and when exceeded, trigger the controller to cut down F-HMR, F-LMR and LNG throughput rate

in order to preserve the operation of the process within its preset bounds.

Figure 5 below shows a typical example of the behavior of the controller over a day when P-MCHE was

constraining. The horizontal red line in the middle panel shows the upper limit. We see that when the constraint

limit is hit, the controller would leave up to 10-20% of compressor power under-utilised and LNG throughput would

fall.

Covariate Type Units Rank pValue

F-HMR MV T/D 3 <1e-10

F-LMR MV T/D 6 <1e-10

V-K MV % 8 <1e-10

C-C1 MV mol% 9 <1e-10

T-MR IV degC 2 <1e-10

T-R MV degC 1 <1e-10

T-WB IV degC 4 <1e-10

F-HMR*F-LMR MV T/D^2 7 <1e-10

T-WB^2 IV decC^2 5 <1e-10

Root Mean Squared Error: 103 R-squared: 0.921, Adjusted R-Squared 0.921 F-statistic vs. constant model: 1.1e+04, p-value = 0

Figure 5: Top panel: P-MCHE, Middle Panel: un-utilised MR Compressor Power; Bottom panel: LNG throughput.

An empirical model using historical closed loop data for the constraint P-MCHE, like for LNG throughput, was

developed to understand which handles should be used to keep P-MCHE below the control limit. The model is

shown in Table 2 below.

Table 2: Model for P-MCHE using MVs and IVs.

Covariate Type Units Rank pValue

F-HMR MV T/D 4 <1e-10

F-LMR MV T/D 2 <1e-10

C-N2 MV mol% 5 <1e-10

C-C1 MV mol% 7 <1e-10

C-C2 MV mol% 8 <1e-10

T-MR IV degC 6 <1e-10

T-R MV degC 1 <1e-10

V-K MV % 9 <1e-10

F-HMR*F_LMR MV T/D^2 3 <1e-10

Root Mean Squared Error: 7.4 R-squared: 0.85, Adjusted R-Squared 0.85 F-statistic vs. constant model: 4.45e+04 p-value = 0

The overall model is reasonable with 85% of variability in P-MCHE explained. The impact of all the variables is

standardised with the strongest contributor given the highest rank following application of the cross-validation

technique as described in the preceding section. We notice from the model that MR composition has material

impact on P-MCHE. Joint influence of C-N2 and T-R on P-MCHE is shown in Figure 6. The red region

corresponds to P-MCHE upper limit. We see that C-N2 must be continuously adjusted as a function of T-R to

remain below the P-MCHE upper limit. A model for Ps-comp was similarly constructed.

Figure 6: T-R (x axis), C-N2 (Y axis). Colours show P-MCHE. Red area indicates P-MCHE has been exceeded. Area below diagonal line is safe operating region.

INTEGRATING THE DATA SCIENCE MODEL WITH THE ADVANCED PROCESS CONTROL LAYER Given that the Data Analytics models derived from the closed loop data were unsuitable for dynamic closed loop

control, the challenge was how to utilise the resulting recommendations for MVs such as F-HMR, F-LMR, V-K etc.

as plant conditions change, and when LNG production was not feed gas supply constrained. As an initial

implementation, it was decided to bypass a true real time RTO or RTPO layer and interpret the results from Data

Science for direct implementation in the APC system. To fully capture the benefits of the work, it was decided to

utilise the PACE APC system recently introduced in Shell, ref (Amrit, et al., 2015).

The models and insights developed using historical data provide directional relationships of MVs to IVs and CVs

such as LNG throughput, P-MCHE and P-s. The influence of MVs was modeled to the CVs often through IVs for

the MCHE-MR loop and other loops as described in the preceding section. The effect size of all MVs( 𝛽s) on LNG

throughput, P-MCHE and P-s is static and fixed. The new controller was built to include the dynamic influence of

all the MVs from MCHE-MR loop and other loops derived by conducting targeted open loop experiments.

Guidance from the Data Science modelling was incorporated in the APC configuration, for example, by setting

limits on manipulated variables.

Figure 7 below offers a simplified view of how the data analytics results were incorporated into the APC controller.

Figure 7: RTPO implementation process.

THE IMPACT OF DATA ANALYTICS AND NEW CONTROLLER

LNG throughput rate and the response of changes in MVs on CVs (constraints) such as P-MCHE and P-s were

monitored over time with the new controller. A comparison of the performance of the old and new MCHE-MR loop

strategies was carried out to evaluate the gain in LNG throughput.

Top panel of figure 8 shows the distribution of LNG throughput rate under the old (in green) and new (in black)

control strategies. As highlighted, a shift in peak production from left to right (relatively low under the old control

strategy to high under the new control strategy) is observed. With the old controller, the production profile is far

spread as there were many instances of reduced throughput due to violation of P-MCHE/P-s constraint, T- WB

either too cold or too warm or T-MR too warm. The new controller was designed to ensure these factors were

properly addressed, and as a result, the production profile is slim and there are less instances of reduced

production.

The vertical dotted lines show the mean production under the old and new controller. The mean production with the

new controller (in black) has shifted to the right compared to the old controller (in green). The MCHE-MR loop is

now able to produce 1-2% extra LNG as a result of the new controller and optimisation using empirical models

fitted to the historical data.

Figure 8: Top panel: Frequency density of LNG throughput with the old controller (green) and new controller

(black). Bottom panel: The variability in production with old (green) and new (black) controller.

SUMMARY AND CONCLUSIONS

The paper discusses the challenge of optimising in real time, the production throughput of the MCHE / Mixed

Refrigerant Loop central to the production of Liquefied Natural Gas from Natural Gas. There is an established

practice to push throughput of the MCHE / MR loop using Advanced Process Control (APC) technology. Efforts to

fully model the process from first principles for Real Time Production Optimisation (RTPO) have proven difficult due

to modelling complexities associated with phase changes of natural gas and mixed refrigerant in the MCHE. It is

reported that there are different data analytics techniques that can be applied to determine new insights and build

predictive RTPO models for the MCHE/MR process. In order to gain insights to increase LNG throughput, least

squares regression was applied to historical closed loop data. The data revealed that besides the usual MVs, some

intermediate variables such as T-WB and T-MR are important. The pressure-drop across MCHE, P-MCHE and

the suction pressure P-s of the compressor were also modelled and ways to effectively control them below their

respective threshold limits determined. The new insights into the MCHE/MR Loop were used to update the APC

solution after conducting step tests and targeted open loop test (data analytics validation) to understand the

dynamic effects of the identified MVs. The updated APC controller provided 1-2% extra LNG throughput rate. The

results reported here are the result of close collaboration between Shell Group Data Scientists, Process

technology, Process Control and Optimisation engineers and NLNG Process/Process Control technologists.

ACKNOWLEDGEMENT

The authors would like to gratefully acknowledge Jasper Stolte, Emma Ross, Yongsong Cai for their contributions.

References Alers, P., & van Dijk, F. (2010). Method of cooling a hydrocarbon stream and an apparatus therefor.

Patent WO2010031737 A2 2010-03-25 . Amrit, R., Canney, W., Carrette, P., Linn, R., Martinez, A., Singh, A., Cott, B. (2015). Platform for

Advanced Control and Estimation (PACE): Shell’s and Yokogawa’s Next Generation Advanced Process Control Technology. 9th International Symposium on Advanced Control of Chemical Processes. Whistler, BC, Canada: International Federation of Automatic Control (IFAC).

Austbøa, B., Weidemann Løvsethb, S., & Gundersena, T. (2014). Annotated bibliography—Use of optimization in LNG process design and operation. Computers and Chemical Engineering, 71, 391–414.

Bolker, B. M. ( 2009). Generalized linear mixed models: a practical guide for ecology and evolution. Trends in Ecology & Evolution,, 24(3), 127-135.

Briers, J., Goh, K., Sniekers, A., Schotanus, D., Hofland, J., & Adun, D. (2016). Looking Back 2006 - 2016 - Ten Years of Data Driven Well Rate Estimates for Real-Time Surveillance and Optimization. SPE Intelligent Energy International Conference and Exhibition. Aberdeen, UK: Society of Petroleum Engineers.

Cramer, R., Stoever, M., Mehrotra, S., Berendschot, L., & Goh, K. (2013). Upstream/Downstream real time surveillance and optimization systems, two sides of the same coin, or never the twain shall meet. SPE Digital Energy Conference (pp. SPE-163710-MS). The Woodlands, Texas: SPE.

Crawley, M. J. (2011). The R Book. John Wiley & Sons Ltd. de Jong, S. (1993). SIMPLS: An Alternative Approach to Partial Least Squares Regression. Chemometrics

and Intelligent Laboratory Systems. , Vol. 18, pp. 251–263. Den Bakker, K., Dejsupa, B., Beeby, C., Azodi, F., & Silve. (2006). A Step Change in LNG Operations

through Advanced Process Control. 23rd World Gas Conference. Amsterdam. Gavin C. Cawley, N. L. (2010). On Over-fitting in Model Selection and Subsequent Selection Bias in

Performance Evaluation. Journal of Machine Learning Research, vol 11, 2079-2107. Goldberger, A. S. (1964). Classical Linear Regression. New York: John Willey & Sons. Hupkes, W., Lin, P., Silve, R., & Vink, K. (2004). Process of liquefying a gaseous, methhane-rich feed to

obtain liquefied natural gas. Patent WO2004068049 A1 2004-08-12 [WO200468049]. Kyungtae Park, W. W. (2016). Effects of varying the ambient temperature on the performance of a single

mixed refrigerant liquefaction process. Journal of Natural Gas Science and Engineering, 958-968. ML cheat sheet. (2017, 4 12). Opgehaald van SAS Machine learning home page:

https://blogs.sas.com/content/subconsciousmusings/2017/04/12/machine-learning-algorithm-use/

Mohd Shariq Khan, S. L. (2013). Knowledge based decision making method for the selection of mixed refrigerant systems for energy efficient LNG Process. Applied Energy, 1018-1031.

Onstott, B., & Linn, R. (2000). ROMeo Provides State-of-the-Art Optimization. World Refining, Nov 2000. Seborg, D. E., Edgar, T. F., & Mellichamp, D. A. (2004). Process Dynamics and Control, 2nd Edition. John

Wiley and Sons.