MACHINE LEARNING FOR THE OPTIMIZATION OF LIQUEFACTION PROCESSES IN THE PRODUCTION OF LIQUEFIED NATURAL GAS CROSS REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the priority benefit of United States Provisional Patent Application No. 63/201516, filed May 3, 2021, entitled MACHINE LEARNING FOR THE OPTIMIZATION OF LIQUEFACTION PROCESSES IN THE PRODUCTION OF LIQUEFIED NATURAL GAS, the entirety of which is incorporated by reference herein. FIELD OF THE INVENTION

[0002] The present application relates generally to the field of liquefied natural gas (LNG) production. Specifically, the disclosure relates to a methodology and framework for training a machine learning model and using a machine-learned model to optimize one or more key performance indicators (KPI) in the production of LNG, such as production quantity, volume, flow rate, or value.

BACKGROUND OF THE INVENTION

[0003] This section is intended to introduce various aspects of the art, which may be associated with exemplary embodiments of the present disclosure. This discussion is believed to assist in providing a framework to facilitate a better understanding of particular aspects of the present disclosure. Accordingly, it should be understood that this section should be read in this light, and not necessarily as admissions of prior art.

[0004] Many natural gas sources can be located a significant distance away from the end- consumers. In some instances, a cost-effective method of transporting natural gas over long distances is to liquefy the natural gas through cooling and condensation. Once liquefied, the LNG may be transported at near ambient pressure using ground-based transport or tanker ships. Once transported to the destination, the LNG can be revaporised into the gaseous state.

[0005] A mixed refrigerant (MR) loop may be used to cool and condense the natural gas during the liquification process. Components in the MR loop, such as compressors, have process variables that may be set to achieve a desired KPI. For example, the process variables of the MR loop may be set to achieve a maximum quantity, volume, flow rate, or value of the LNG produced. SUMMARY OF THE INVENTION

[0006] In one or some embodiments, a computer-implemented method for generating a machine-learned model based on a liquefied natural gas production process is disclosed. The method includes receiving a set of process variables and a corresponding set of performance indicators, where the set of performance indicators describes a state of a liquefied natural gas production process, and where the set of process variables describes a state of a refrigerant loop of the liquefied natural gas production process; applying a subset of process variables and a corresponding subset of performance indicators to a machine learning model; generating the machine-learned model based on applying the subset of process variables and the corresponding set of performance indicators to the machine learning model, where the machine-learned model is configured to output one or more estimated performance indicators based on one or more input process variables; and outputting the machine-learned model to modify operation of one or more aspects of the liquefied natural gas production process.

[0007] In one or more embodiments, a computer-implemented method for predicting a performance indicator of a liquefied natural gas production process is disclosed. The method includes receiving a set of measured process variables describing a state of a refrigerant loop of the liquefied natural gas production process; applying the set of measured process variables to a machine-learned model, where the machine-learned model is trained to output the performance indicator based on input process variables; generating the performance indicator based on applying the set of measured process variables; and outputting the performance indicator.

[0008] In one or more embodiments, a system for generating a machine-learned model based on a liquefied natural gas production process is disclosed. The system includes a processor; and a non-transitory machine readable medium comprising code. The code is configured to direct the processor to receive a set of process variables and a corresponding set of performance indicators, where the set of performance indicators describes the state of a liquefied natural gas production process, and where the set of process variables describes a state of a refrigerant loop of the liquefied natural gas production process; apply a subset of process variables and a corresponding subset of performance indicators to a machine learning model; generate the machine-learned model based on applying the subset of process variables and the corresponding set of performance indicators to the machine learning model, where the machine-learned model is configured to output one or more estimated performance indicators based on one or more input process variables; and output the machine-learned model to modify operation of one or more aspects of the liquefied natural gas production process.

BRIEF DESCRIPTION OF THE DRAWINGS [0009] The present application is further described in the detailed description which follows, in reference to the noted plurality of drawings by way of non-limiting examples of exemplary implementations, in which like reference numerals represent similar parts throughout the several views of the drawings. In this regard, the appended drawings illustrate only exemplary implementations and are therefore not to be considered limiting of scope, for the disclosure may admit to other equally effective embodiments and applications.

[0010] FIG. 1 illustrates a diagram of an example LNG production process.

[0011] FIG. 2 illustrates a diagram of another example LNG production process.

[0012] FIG. 3 illustrates an example flow chart for optimizing an LNG production process.

[0013] FIG. 4 illustrates an example flow chart for generating a machine-learned model.

[0014] FIG. 5 illustrates another example flow chart for determining process variables corresponding to an optimal performance indicator and updating a machine-learned model. [0015] FIG. 6 illustrates a further example flow chart for updating a machine-learned model.

[0016] FIG. 7 illustrates yet another example flow chart for determining a difference between an actual and an estimated performance indicator.

[0017] FIG. 8 illustrates data of process variables of a mixed refrigerant loop.

[0018] FIG. 9 illustrates a scatter plot of feed gas flow rate for different compositions of mixed refrigerant.

[0019] FIG. 10 illustrates another scatter plot of feed gas flow rate for different compositions of mixed refrigerant.

[0020] FIG. 11 illustrates a shaded contour plot of an estimated feed gas flow rate for different compositions of mixed refrigerant.

[0021] FIG. 12 is a diagram of an exemplary computer system that may be utilized to implement the methods described herein. DETAILED DESCRIPTION OF THE INVENTION

[0022] The methods, devices, systems, and other features discussed below may be embodied in a number of different forms. Not all of the depicted components may be required, however, and some implementations may include additional, different, or fewer components from those expressly described in this disclosure. Variations in the arrangement and type of the components may be made without departing from the spirit or scope of the claims as set forth herein. Further, variations in the processes described, including the addition, deletion, or rearranging and order of logical operations, may be made without departing from the spirit or scope of the claims as set forth herein.

[0023] It is to be understood that the present disclosure is not limited to particular devices or methods, which may, of course, vary. It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only, and is not intended to be limiting. As used herein, the singular forms “a,” “an,” and “the” include singular and plural referents unless the content clearly dictates otherwise. Furthermore, the words “can” and “may” are used throughout this application in a permissive sense (i.e., having the potential to, being able to), not in a mandatory sense (i.e., must). The term “include,” and derivations thereof, mean “including, but not limited to.” The term “coupled” means directly or indirectly connected. The word “exemplary” is used herein to mean “serving as an example, instance, or illustration.” Any aspect described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other aspects. The term “uniform” means substantially equal for each sub-element, within about ±10% variation. The term “nominal” means as planned or designed in the absence of variables such as wind, waves, currents, or other unplanned phenomena. “Nominal” may be implied as commonly used in the fields of seismic prospecting and/or hydrocarbon management. It is to be understood that the present disclosure is not limited to particular devices or methods, which may, of course, vary. It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only, and is not intended to be limiting.

[0024] The term “simultaneous” does not necessarily mean that two or more events occur at precisely the same time or over exactly the same time period. Rather, as used herein, “simultaneous” means that the two or more events occur near in time or during overlapping time periods. For example, the two or more events may be separated by a short time interval that is small compared to the duration of the overall operation. As another example, the two or more events may occur during time periods that overlap by about 40% to about 100% of either period.

[0025] As used herein, “obtaining” data generally refers to any method or combination of methods of acquiring, collecting, or accessing data, including, for example, directly measuring or sensing a physical property, receiving transmitted data, selecting data from a group of physical sensors, identifying data in a data record, and retrieving data from one or more data libraries.

[0026] If there is any conflict in the usages of a word or term in this specification and one or more patent or other documents that may be incorporated herein by reference, the definitions that are consistent with this specification should be adopted for the purposes of understanding this disclosure.

[0027] As discussed in the background, process variables of the mixed refrigerant (MR) loop may be set to achieve a desired KPI, such as production quantity, volume, flow rate, or value. Such process variables, which are directly controllable, may be referred to as controllable process variables. For example, a setpoint of a valve or a speed of a compressor may be controllable process variables because such setpoints are adjustable. A controllable process variable may describe a setpoint of a process element.

[0028] Other process variables of the MR loop, which may not be controlled either directly or indirectly, may be referred to as uncontrollable process variables. For example, an ambient temperature or a temperature of a heat exchanger or of refrigerant in the MR loop may be uncontrollable process variables. Despite having a significant effect on KPIs, uncontrollable process variables are measured but may not be directly changed. Instead, uncontrollable process variables may change or have a value resulting from setpoints of other controllable process variables and the associated production elements.

[0029] One goal of LNG production is to increase the quantity, volume, flow rate, or value of LNG produced by a facility. For example, capital improvements, such as debottlenecking and upgrades to the facility, may increase the quantity of LNG produced at a facility. The increased production volume may introduce new bottlenecks not addressed by the prior debottlenecking and upgrades, so that components, such as those in the MR loop, may need to be tuned to reach an optimal production state. In this case, the optimal composition of MR may be unknown because the original design MR compositions were specified based on the initial capacity of the production facility and as such are not applicable to the current production throughput. Removing the production bottlenecks (e.g., by tuning the setpoints of equipment in the facility) may further increase production or other KPIs without further capital expenditure. In one example, the mass flow of MR that can pass through the Main Cryogenic Heat Exchanger (MCHE) may be a bottleneck.

[0030] In one or some embodiments, a machine learning model may be trained with a training dataset having process variables and KPIs, resulting in a machine-learned model configured to estimate KPIs given input process variables. The machine learning model may be developed utilizing a programming language, such as Python, with machine learning modules. The training data may include process variable data measured from the LNG facility, as well as KPI data corresponding to the process variable data. For example, the KPI data may correspond to the same production run or production time as the process variable data. In another example, the KPI may be selected based on a time delay from the measurement or collection of the process variables. The delay may account for dynamics in the LNG production process (e.g., transport delays) as the process reacts to changes in the process variables. Once trained, the machine learning model may be referred to as a machine-learned model.

[0031] The machine-learned model may focus on or be adapted to predict the KPI based on process variables from one or more processes within the LNG facility. For example, the machine-learned model may be configured to estimate a KPI based on any one, any combination, or all of the MR composition, current operating conditions (e.g., ambient temperature), and other process variables. In another example, the machine-learned model may estimate the KPI based on inputs to the liquefaction process, such as an input gas feed, an acid gas removal process, and a dehydration and mercury removal process.

[0032] In some cases, the machine-learned model may be adapted to predict KPIs for other facilities beyond the LNG facility. For example, a machine learning model may be trained using process variable data from a helium liquefaction facility and configured to estimate KPIs for such a facility. In another example, a machine learning model may be trained using process variable data from an upstream process, such as a well, and configured to estimate KPIs for such a facility. In particular, a machine learning model may provide accurate estimations of KPIs for facilities where ambient temperature has a significant effect on production. Additionally, the approach may be extended to facilities with controllable process elements and accurate data collection of process variables and KPIs.

[0033] The machine-learned model may be used to predict LNG productions resulting from a variety of operating conditions. In this way, a difference between predicted and actual KPIs may be documented as either a shortfall or a production improvement. When the KPI or production level is less than estimated by the machine-learned model (e.g., a production shortfall), then the root cause may be investigated to avoid recurrence in the future. If the KPI or production level is greater than predicted by the model, then the root case may be investigated to replicate the conditions in normal operations as well as ensure the machine- learned model accounts for the potential upside in future predictions. In some cases, a certain level of drift or difference may be expected between the machine-learned model and actual production. Drift that is greater than expected or exceeds a threshold (e.g., defined as a percent difference between predicted and actual KPI over time) may be identified and act as a trigger to retrain the machine-learned model, or to indicate that a component of the LNG facility may need maintenance or replacement (e.g., ahead of proscribed maintenance intervals or schedules). In one example, the difference between the predicted and the actual KPIs may be monitored on a continual basis. A sudden increase in the difference may allow for a timely response to production deviations or shortfalls.

[0034] Further, the machine-learned model may estimate optimal controllable process variables (e.g., the optimum MR compositions) to fit operational condition changes caused by any one, any combination, or all of seasons, weather, equipment degradation/replacement, or other disturbances, yielding production improvements of +0.1% per year or more worth several million USD in revenue. As compared to other analysis methods, such as first-principles thermodynamic models or simple models considering only one process variable (such as ambient temperature), the machine-learned model is a more robust, adaptable, and accurate tool for estimating and optimizing process variables and KPIs of the LNG facility.

[0035] Machine learning models, machine-learned models, and methods of training, generating, and applying such models are described, for example, in U.S. Patent No. 10,198,535 B2, issued on February 5, 2019, U.S. Patent Application Publication No. 2020/0183041 Al, published on June 11, 2020, U.S. Patent Application Publication No. 2021/0041596 Al, published on February 11, 2021, and U.S. Patent Application Publication No. 2020/0183032 Al, published on June 11, 2020, all of which are herein incorporated by reference in their entirety.

[0036] Referring to the figures, FIG. 1 illustrates a diagram 100 of an example LNG production process. Natural gas 102 (which may be referred to as a feed gas) is input to an inlet gas system 104. In some cases, the natural gas 102 may be delivered to the inlet gas system 104 via a pipeline or other source.

[0037] The natural gas 102 is passed to an acid gas removal system 106. The acid gas removal system 106 removes impurities (e.g., hydrogen sulfide and carbon dioxide) from the natural gas 102. The acid components may be removed from the natural gas 102 to prevent impurities from freezing in a heat exchanger. Additionally, by removing the acid components, the release of sulfur may be prevented when the natural gas 102 (or the LNG 120) is burned. The impurities removed from the natural gas 102 may constitute a liquefied petroleum (LP) gas. The LP gas may be output, for example, to storage.

[0038] Once treated by the acid gas removal system 106, the natural gas 102 is output from the acid gas removal system 106 to a dehydration and mercury removal system 108. Mercury (and other impurities) may be present in the natural gas 102, for example, in small concentrations as elemental (e.g., metallic), organic, and/or inorganic compounds. Mercury is removed by the dehydration and mercury removal system 108 to avoid damage to downstream equipment. As part of a dehydration process in the dehydration and mercury removal system 108, water may be removed from the natural gas 102 to prevent freezing in a heat exchanger. Additionally, the removal of water may improve the combustion characteristics of the natural gas 102 or LNG 120. The impurities removed from the natural gas 102 may constitute a fuel gas. The fuel gas may be output, for example, to storage.

[0039] After being treated by the acid gas removal system 106 and the dehydration and mercury removal system 108, the natural gas 102 is input to a liquefaction process 110 (shown in greater detail in FIG. 2). The liquefaction process 110 converts the natural gas 102 into LNG 120. The liquefaction process 110 cools the natural gas 102 by exchanging heat with MR 112 in a MR loop 114. The natural gas 102 and MR 112 may be input to a heat exchanger (e.g., as shown in FIG. 2) to liquefy the natural gas 102.

[0040] The MR 112 may include one or more components, such as nitrogen (N2), methane (Ci), ethane (C ₂ ), propane (C ₃ ), higher order hydrocarbons (C ₄ +), and other components. After being heated by interaction with the natural gas 102, the MR 112 in the MR loop 114 may be cooled and recondensed by interaction with propane 116 in a propane refrigeration loop 118. The interaction between the MR loop 114 and the propane refrigeration loop 118 is shown in greater detail in FIG. 2.

[0041] The liquefaction process 110 transforms the natural gas 102 into LNG 120. In some cases, the LNG 120 may be output to an LNG storage and loading system 122. The LNG storage and loading system 122 may contain storage facilities for the LNG 120 as well as equipment for transferring the LNG 120 to one or more transports 124. The transports 124 may be, for example, trucks or ships adapted to transport the LNG 120 in a liquid form. In some cases, the transports 124 may move the LNG at near ambient pressure. In one case, boil off gas (BOG) may be output from the LNG storage and loading system 122 into the dehydration and mercury removal system 108 to regenerate adsorption beds. In some other cases, the boil off gas may be output as a high pressure fuel gas, for example, to feed a flare pilot or to fuel rotating machinery. [0042] Scrub column liquids 126 are adsorbed from the natural gas 102 in the liquefaction process 110. A fractionation system 128 may receive and process the scrub column liquids 126. For example, the fractionation system 128 may separate the scrub column liquids 126 into one or more fractions 134, 136. A fraction 134 (e.g., liquefied petroleum gas (LPG) may be reinjected into the LNG production process. For example, the fraction of the scrub column liquids 126 (e.g., containing ethane, propane, and/or higher order hydrocarbons) may be reinjected to the feed gas 102.

[0043] Another fraction 136 (e.g., containing ethane or propane) may be reinjected to the MR loop 112.

[0044] A remainder of the scrub column liquids, known as plant condensate 132, may be output to a condensate storage and loading system 130. The plant condensate 132 may include butane and higher order hydrocarbons. The condensate storage and loading system 130 may ensure the scrub column liquids 126 remain in a liquid state in storage to prepare for loading on a transport 124.

[0045] FIG. 2 illustrates a diagram 200 another example of the LNG production process. The LNG production process may include several sub processes, for example, any one, any combination, or all of: a feed gas 102 to LNG flow 202; a liquefaction process 110; a MR loop 114; or a propane refrigeration loop 118.

[0046] In the feed gas 102 to LNG flow 202 (including one or more elements of the liquefaction process 110, for example, the scrub column 208 and MCHE 216), the feed gas 102 may exchange heat with a section of the propane refrigeration loop 118. The feed gas 102 exchanges heat with medium pressure (MP) liquid propane refrigerant 116 in a first exchanger 204, from which medium pressure propane refrigerant vapor 116 and low pressure (LP) propane refrigerant vapor 116 may be output. The low pressure propane refrigerant vapor 116 may exchange heat with the feed gas 102 in a second heat exchanger 206, from which low pressure propane refrigerant vapor 116 may be output. The medium pressure propane refrigerant vapor 116 and low pressure propane refrigerant vapor 116 may be reintroduced into the propane refrigeration loop 118.

[0047] In the liquefaction process 110, the feed gas 102 may enter a scrub column 208. The scrub column 208 may remove scrub column liquids 126 from the natural gas 102. The fractionation system 128 (e.g., as shown in FIG. 1) may receive and separate the scrub column liquids 126. The scrub column 208 may have a reflux loop including heat exchanger 210, vessel 212, and pump 214 to recirculate condensed liquids. Feed gas 102 leaving the reflux loop may have temperatures near -35°C before being fed into the main cryogenic heat exchanger (MCHE) 216. The natural gas 102 may be cooled in the MCHE by exchanging heat with the MR 112 (e.g., liquid MR component 218 and vapor MR component 220). LNG 120 may exit the MCHE 216 at temperatures near -160°C, allowing for storage at near atmospheric pressures.

[0048] In the MR loop 114, MR 112 is cooled by the propane refrigeration loop 118, separated into a liquid MR component 218 and vapor MR component 220, passed through the MCHE 216 to extract heat from the feed gas 102, and recirculated. MR 112 leaving the MCHE 216 passes through a vessel 222 and into a compressor 224 driven by a gas turbine 226. The MR 112 leaving the compressor 224 may pass through an air cooler 228 before entering a heat exchanging section 230 with the propane refrigeration loop 118. The MR 112 leaving the heat exchanging section 230 may be cooled and enter an MR separator 232. The MR separator 232 may separate the input MR 112 into the liquid MR component 218 and the vapor MR component 220. Both the liquid MR component 218 and the MR vapor component 220 may be input to the MCHE 216 to liquefy the feed gas 102 into LNG 120. Valves 246, 248 control the flow of the liquid MR component 218 and vapor MR component 220, respectively, into the MCHE 216.

[0049] Refrigerant 116 in the propane refrigeration loop 118 cools the MR 112. In the heat exchanging section 230, heat is exchanged between the MR 112 and propane refrigerant 116 at various pressures. Refrigerant 116 may enter the heat exchanging section 230 (e.g., a first heat exchanger in a series) at a high pressure and exit the heat exchanging section 230 at a low pressure. For example, a first heat exchanger may use high pressure (HP) refrigerant 116, a second heat exchanger may use medium pressure (MP) refrigerant 116, a third heat exchanger may use low pressure (LP) refrigerant 116, and an ultimate heat exchanger may use even lower pressure (low, low pressure, or LLP) refrigerant 116.

[0050] The refrigerant 116 output from the heat exchanging section 230 may pass through vessels 234 before entering a compressor 236 driven by a gas turbine 238. The compressed refrigerant 116 output from the compressor 236 may flow into an air cooler 240 and condenser 242. Once condensed, the liquid refrigerant 116 may flow into an accumulator 244 and back into the heat exchanging section 230, completing the loop.

[0051] A control or monitoring system of the LNG production process may receive process variable data and KPI data from the various process elements. The process variables may be associated with a respective process element, sensor, and/or unit of measure. The control or monitoring system may record or store the process variable data and KPI data in a storage device, such as the storage device 1212 of FIG .12.

[0052] FIG. 3 illustrates an example flow chart 300 for optimizing an LNG production process. The acts may be performed by the computer system 1200 of FIG. 12. Additional, different, or fewer acts may be included. For example, act 320 may be omitted. The acts may be performed in a different order than shown. For example, act 330 may proceed from act 310. In another example, act 320 may proceed from act 330. Acts in the flow chart 300 may be iterative or one or more acts may be repeated. For example, after act 350 is performed, acts 310, 320, and/or 330 may be performed.

[0053] In act 310, a dataset is developed by recording operations data (process variables and KPI data) to a database. The process variable data may be associated with a particular element of the LNG production process and may indicate a source (e.g., a sensor) and a value (e.g., temperature, flowrate, valve position), in addition to other information such as measurement units, time and production shift. The operations data may be collected from the LNG production process, for example, by a control system of the LNG production process operating or monitoring one or more elements and/or sensors involved in the LNG production process. In one example, the operations data includes process variables, such as setpoints and/or levels of pumps (e.g., pump 214), compressors (e.g., compressors 224, 236), valves (e.g., valves 246, 248), heat exchangers (e.g., exchangers 204, 206, 210, exchangers in the heat exchanging section 230, condenser 242, and/or air coolers 228, 240), vessels (e.g., vessels 212, 222, 232, 234 and/or accumulator 244), flowmeters, composition analyzers (e.g., gas chromatographs or other devices configured to measure a composition of the mixed refrigerant loop) and other elements (e.g., scrub column 208 and/or MCHE 216). Additionally or alternatively, the operations data may include KPIs, such as any one, any combination, or all of production quantity, volume, flow rate, or value. For example, the KPI may be a monetary value of LNG produced by the LNG production process.

[0054] The operations data may be historical data collected over a period of time and may be stored in a database (e.g., residing on storage device 1212 of FIG. 12). In one example, the dataset may be the process variables and performance indicators received in act 410 of FIG. 4, 610 of FIG. 6 and/or act 710 of FIG. 7, described below. The dataset may be set by or received from the database residing on storage device 1212 of FIG. 12. In another example, the dataset may be analyzed to determine the ranges for controlled variables in act 510 of FIG. 5, described below. [0055] In act 320, exploratory data analysis is performed to identify process variables of interest. Based on the operation of the plant, certain process variables may have a greater importance than others. For example, some process variables (and the underlying feature or element of the LNG production process) may have a greater range of adjustment, or may have a greater influence on a particular KPI. In another example, some other process variables may have a small range of adjustment or may have a lesser influence on the particular KPI. Process variables may be identified automatically (e.g., based on the influence on the KPI) or by a user. Different subsets of the process variables in the operations data may be selected to explore or discover which process variables are more important or have a greater influence on a KPI. [0056] In act 330, filters are used to exclude unrepresentative data (e.g., process variables and KPIs) from the collected operations data. For example, operations data recorded during one or more conditions, such as startup, shutdown, maintenance, and/or testing, may be excluded. The operations data recorded during the conditions may be unrepresentative of the typical operation of the LNG production process.

[0057] In act 340, process variables are identified to build or train a machine learning model. The selected process variables may be included in a machine learning dataset along with corresponding KPIs. For example, process variables may be identified according to the feature selection of act 420 of FIG. 4, described below. The machine learning dataset may be used to train the machine learning model to estimate a KPI. Users or experts may select or exclude process variables and KPIs from the training dataset.

[0058] By applying the training dataset (e.g., set of known input vectors (the process variables) and a set of outputs (the KPIs) to a machine learning model), the model may learn to determine a KPI based on input process variables, resulting in a machine-learned model. The training dataset may be applied to the machine learning model and the machine-learned model may be generated according to one or more acts of FIG. 4, described below, such as acts 430 and/or act 440. One or more training techniques may be used to train the machine learning model. The training techniques may include supervised learning, probabilistic trees, support vector machines, radial basis functions, and other machine learning techniques. The machine learning model may be a neural network, a probability tree, a support vector machine, radial basis functions, or another suitable machine learning model. The machine learning model may include any number of hidden layers and any number of nodes per layer, as well as any other topology of neuron connections.

[0059] In act 350, performance of the machine-learned model is verified through field application and testing. For example, new process variables and KPIs may be recorded that are not part of the training dataset. The new process variables may be applied to the machine- learned model to generate an estimated KPI. The estimated KPI may be compared to the new KPI. The verification may be conducted or proceed as described with respect to one or more acts of FIG. 6, such as acts 610, 620, 630, and/or 640, described below. When there is a difference between the estimated KPI and the new KPI, the performance (e.g., accuracy) of the model may be inadequate. The newly measured process variables and KPIs may be added to a training dataset to retrain or update the machine-learned model.

[0060] In act 360, the machine-learned model may be used to optimize and forecast the performance of the LNG production process. Using the predictive power of the machine- learned model, multiple sets of process variables may be applied to the model and corresponding performance indicators estimated. The controllable process variables resulting in a most optimal performance indicator may be implemented in the LNG production process to optimize the output. Based on weather, maintenance on process elements, or other expected or unexpected events reflected in the input to the machine-learned model, a performance indicator may be generated. In this way, changes in production may be planned ahead of time. Controllable process variables resulting in an optimal performance indicator and/or production forecasts may be determined according to one or more acts of FIG. 5, described below.

[0061] FIG. 4 illustrates an example flow chart 400 for generating a machine-learned model. The acts may be performed by the computer system 1200 of FIG. 12. Additional, different, or fewer acts may be included. For example, 420 may be omitted.

[0062] In act 410, process variables (e.g., feature variables) and performance indicators (e.g., KPIs) are received. In some cases, the process variables and performance indicators may be part of a dataset of recorded operations data, such as described above with respect to act 310 of FIG. 3. The performance indicators may describe a state of an LNG production process. For example, the performance indicators may include a production quantity, volume, flow rate, and/or value of LNG produced by the LNG production process. The process variables may describe a state of the MR loop 114 of the LNG production process. For example, the process variables may include a composition of mixed refrigerant in the refrigerant loop and/or one or more setpoints of a heat exchanger, such as an input flow rate of mixed refrigerant (vapor or liquid) into a heat exchanger (e.g., the MCHE 216). In some cases, the process variables and performance indicators may be operations data recorded over time from the LNG production process. In some other cases, the process variables and performance indicators may be operations data received in real time from the LNG production process. The process variables and performance indicators may be measured by and/or received from a control or monitoring system of the LNG production process. In some cases, the process variables and performance indicators may be received from storage, such as the storage device 1212 of FIG. 12.

[0063] In act 420, features are selected or extracted from the received process variables for inclusion in a training dataset used to train the machine learning model. Different process variables may have a larger or smaller impact on performance indicators. For example, a flow rate of MR 112 through the MCHE 216 may have a greater effect on the resulting KPI than a speed of a fan for the air cooler 228. In another example, different process variables may have different sampling rates such that certain process variables may be sparser than others. To counteract the sparsity, the sparser process variables may be weighted more. Accordingly, different process variables in the training dataset (e.g., corresponding to different process elements or collected at different times) may be given a different weight or importance in the training dataset. In one case, the weighting is determined based on a correlation between the process variables and the performance indicator. Process variables having no correlation with the KPI or being unrepresentative of abnormal operation of the LNG production process may be excluded from the training dataset or may be assigned a lower weight (e.g., a weight of zero). Feature extraction, in this way, reduces the complexity of data describing the LNG production process, thereby facilitating learning and generalization of the machine learning model.

[0064] Complexity may further be reduced by excluding groups of process variables that are redundant or intercorrelated. For example, multiple sensors may measure multiple parameters, such as temperature, pressure, and mass flow rate. These process values may be highly correlated between one another. For example, mass flow rate and volumetric flow rate may be highly correlated with one another. In this way, having just one of mass flow or pressure flow obviates the need for the other quantity to be included in the training dataset. Further examples of highly intercorrelated process variables are a separation pressure in the mixed refrigerant loop, mixed refrigerant vapor (MRV) flow rate, and the methane and ethane composition of the mixed refrigerant. As a result, multiple, highly correlated process variables are redundant and may be excluded from the training dataset used to train the machine learning model. For example, one process variable from a group of highly correlated or redundant process variables may be selected for inclusion in the training dataset.

[0065] A correlation of the set of process variables on the corresponding set of performance indicators may be determined. In some cases, the correlation may be based on user input. For example, a user may specify a correlation of a process variable on the corresponding performance indicator. In some other cases, the correlation may be determined using a statistical model. For example, the process variables and performance indicators may be inputs to the statistical model used to determine the correlation. In still some other cases, a user may determine that a process variable has no correlation with the performance indicator. For example, for related variables, such as temperature, pressure, and/or flow rate, just one of the related process variables may be selected as having a correlation, while the user may determine that the other related process variables have no correlation (e.g., assign a weight of zero during weighing). By determining that one process variable of a related set of process variables has a correlation, a conflict between the related process variables (e.g., a situation where an unrealistic or impossible combination of values of the related process variables) may be avoided.

[0066] In some cases, certain process variables of the set of process variables are rejected from inclusion in the subset of process variables. The correlation of a process variable to the performance indicator may be applied to a threshold. When the correlation of the process variable is below the threshold, the process variable may be excluded. The threshold may be set to exclude insignificant process variables while retaining highly correlated process variables. Additionally or alternatively, process variables may be unrepresentative of the typical operation of the LNG production process and excluded or rejected for training the machine learning model. For example, the process variables may have been recorded while the LNG production process was operating under one or more conditions, such as at startup, shutdown, maintenance, and/or testing, may be excluded. When determining the weight of the process variables, the rejected or excluded process variables may, in some cases, be assigned a weight of zero.

[0067] The subset of process variables may be weighed from the set of process variables. The subset may include process variables having a high correlation. For example, the subset may include process variables having a correlation with the performance indicator at or above the threshold. The subset may exclude uncorrelated or rejected process variables, for example. In another example, the uncorrelated or rejected variables may be given a weight of zero. The subset of process variables may correspond with one or more of the performance indicators (referred to as the corresponding set of performance indicators). Performance indicators not associated with any of the process variables in the subset may be rejected or excluded from the subset of process indicators.

[0068] In one example, the selected features input to the model are the composition of nitrogen and propane in the mixed refrigerant, the methane feed rate to the main cryogenic heat exchanger, the ethane feed rate to the main cryogenic heat exchanger, the warm bundle temperature of the main cryogenic heat exchanger, and a proxy for ambient temperature (e.g., “Train-1 MR Smart Temperature”).

[0069] In act 430, the subset of process variables and the corresponding subset of performance indicators are applied to a machine learning model. The machine learning model may be a neural network, a probability tree, a support vector machine, radial basis functions, or another suitable machine learning model. The subset of process variables may be an input to the machine learning model. The machine learning model may leam to map the input process variables to the corresponding performance indicators. As discussed above, the machine learning model may be trained based on the applied data using one or more techniques, such as supervised learning, probabilistic trees, support vector machines, radial basis functions, and other machine learning techniques.

[0070] In act 440, the machine-learned model is generated. The machine-learned model may represent or be generated by training the machine learning model on the subset of process variables and corresponding set of performance indicators applied to the machine learning model. The machine-learned model may be configured to output one or more estimated performance indicators based on one or more input process variables. In another example, the machine-learned model may be used to determine a set of process variables that may achieve a desired KPI. A KPI may be set, and different combinations or sets of synthetic or hypothetical process variables may be applied as input to the machine-learned model. By repeatedly applying different process variable data (e.g., corresponding to different predetermined values of controllable process variables) to the machine-learned model, a set of controllable process variables resulting in an estimated KPI matching the desired KPI may be determined. The set of synthetic or hypothetical controllable process variables may then be applied to the LNG production process. Based on the synthetic or hypothetical controllable process variables as applied to an actual LNG production process, the KPI may be measured. The machine-learned model may be updated based on the process variables and measured KPI to ensure the model may accurately predict the desired KPI.

[0071] In act 450, at least one aspect of the machine-learned model may be output. Various outputs are contemplated. As one example, the machine-learned model itself may be output to storage, such as the storage device 1212 of FIG. 12. Thereafter, the machine-learned model may be accessed or retrieved from storage (e.g. by a control, monitoring, and/or alarm system of the LNG production process) for use in order to modify at least one aspect of the LNG production process. As another example, the control system may interact with the machine- learned model to change the operation of one or more aspects of the LNG production process. Or, the machine-learned model may be output to and incorporated in directly to another system. In particular, the machine-learned model may be output to or integrated with the control, monitoring, and/or alarm system of the LNG production process in order to modify at least one aspect of the LNG production process. As still another example, the one or more estimated performance indicators, which is the machine-learned model itself outputs, may in turn be output for use by the control, monitoring, and/or alarm system of the LNG production process in order to modify at least one aspect of the LNG production process. In this way, the machine- learned model may estimate a KPI based on process variables measured from the LNG production process in real time or measured over a period of time. The estimated KPI (for example, in conjunction with a measured KPI) may be used to adapt the operation of the LNG production process.

[0072] The machine-learned model may be connected to or in communication with other systems. For example, a monitoring and alarm system of the LNG production process may receive the estimated KPI from the machine-learned model and receive a KPI measured from the LNG production process. The monitoring system may compare the estimated and measured KPI to determine a difference. When the difference is significant (e.g., larger than a predetermined threshold) or increases over time, the monitoring system may create a message, alarm, or process flag. The message may indicate that a root cause analysis is to be scheduled to investigate the difference. For example, the difference between the estimated and measured KPI may be due to equipment problems requiring replacement, maintenance, or repair. The maintenance or repair indicated may be issued prior to regular or predetermined maintenance intervals, reducing the possibility of premature equipment failure or malfunction, and avoiding complete reliance on predetermined maintenance schedules.

[0073] FIG. 5 illustrates another example flow chart 500 for determining an optimal performance indicator. Given sets of possible process variables for the liquefied natural gas production process, a most optimal performance indicator is discovered. Multiple sets of hypothetic or synthetic process variables may be applied to a machine-learned model to determine a performance indicator. By repeating the process and applying different sets of possible process variables to the machine learning model, multiple performance indicators may be determined. By comparing the multiple performance indicators, a most optimal performance indicator (e.g., a highest production quantity, volume, flow rate, or value) may be identified. The set of possible controllable process variables leading to the optimal performance indicator may be recorded. In this way, the liquefied natural gas production process may be set according to the set of controllable process variables.

[0074] The acts may be performed by the computer system 1200 of FIG. 12. Additional, different, or fewer acts may be included. The acts may be performed using a machine-learned model. For example, a machine-learned model generated and/or output according to the acts of FIG. 4 may be used.

[0075] In act 510, a set of synthetic process variables are generated. The synthetic process variables represent a hypothetical state of the LNG production process. In some cases, the synthetic process variables may contain data for the same process variables that were used to train the machine-learned model (e.g., as according to FIG. 4). In one example, the set of synthetic process variables describes a hypothetical state of the refrigerant loop of the LNG production process. By constructing a hypothetical or synthetic set of process variables and applying them to the machine-learned model, a KPI for the synthetic variables (also referred to as a synthetic performance indicator) may be estimated. By changing the values of the synthetic process variables, applying them to the machine-learned model, and analyzing the estimated KPIs, a set of synthetic process variables corresponding to a desired or optimal KPI may be discovered.

[0076] The set of synthetic process variables may be generated or constructed with controllable and/or uncontrollable process variables. The process variables may be chosen from a range of possible values. For example, a controllable process variable, such as the content of a component of the mixed refrigerant loop, may have a range of values between 7% to 9%. Another controllable process variable, the content of another component of the mixed refrigerant loop, may have a range of values between 18% to 20%.

[0077] The range may be determined by observing real-world operation of the LNG production process, reviewing documentation of the LNG production process, or based on a model of the LNG production process. For example, a dataset of historical operations data, such as the dataset described with respect to act 310 of FIG. 3, described above, may be analyzed to determine minimum and maximum limits for a controllable or uncontrollable process variable. A value of the process variable in the set of synthetic process variables may be chosen within the range.

[0078] An uncontrollable process variable may, in some cases, have a specified value. For example, an uncontrollable process variable, such as an ambient temperature, may be specified at 75 degrees Fahrenheit. In one case, the temperature may be specified according to a weather forecast or historically observed temperature values so that changes in production due to weather may be anticipated.

[0079] In some cases, the optimal KPI may be a volume of LNG produced. The volume may depend on a flow rate produced. However, the volume produced may further depend on one or more uncontrollable variables, such as weather or temperature changes over a predetermined time window (e.g., the course of a day). The synthetic process variables may then include variations on controllable variables (e.g., the content of a component of the mixed refrigerant loop), as well as variations on uncontrollable variables (e.g., weather or temperature, based on a forecast or historical observations). To determine the production volume, the methodology may repeatedly apply the machine-learned model with different synthetic process variables (e.g., different synthetic MR compositions). For example, the machine-learned model calculates the flow rate at various times in the predetermined time window, namely with different values of the uncontrollable process variables, thereby accounting for temperature fluctuations over the predetermined time window. In turn, the flow rate may be aggregated to compute the production volume for the predetermined time window. The following pseudo code is listed for illustration purposes:

[0080] Flow V olume List <- EMPTY LIST

[0081] For each controllable process-variable:

[0082] Flow Volume = 0

[0083] $ Note: the loop below may account for one or more uncontrollable process variables that may be encountered over the predetermined time window. This may be derived from one or more sources, such as forecasts and/or historical readings.

[0084] For each uncontrollable process-variable:

[0085] Flow Rate <- ML_Model(all_process_vars)

[0086] Flow Volume = Aggregate(Flow Rate)

[0087] Flow V olume List. add(Flow_V olume)

[0088] Specifically, multiple sets of synthetic process variables varying in both dimensions of controllable and uncontrollable process variables may be applied to the machine-learned model (e.g., according to act 520), a flow rate may be estimated for each set (e.g., according to act 530), and the set (or sets) of synthetic process variables corresponding to an optimal flowrate may be determined (e.g., according to act 540). To determine a production volume over a time period (during which uncontrollable variables such as weather or temperature may vary), the estimated flowrates for the varying uncontrollable variables are aggregated. In this way, an overall production volume for a period of time may be determined based on optimal flow rates estimated for uncontrollable process variables varying over the time period.

[0089] The controllable and/ or uncontrollable process variables may form a group or subset of the set of synthetic process variables, such that the set of synthetic process variables may include one or more subsets. In this way, by applying the subset(s) to the machine-learned model, an estimated performance indicator is generated for the subset(s). Because the set of synthetic process variables may include multiple subsets (e.g., including different values for the controllable and/or uncontrollable process variables) the estimated performance indicators may be compared to determine the most optimal performance indicator of all the subsets (e.g., according to act 540).

[0090] Other controllable process variables not included in the set of synthetic process variables (e.g., to reduce complexity, redundancy, or intercorrelation) may be inferred from the variables included in the set. In one example, the composition of only a subset of the mixed refrigerants (e.g., nitrogen and/or propane) may be included in the synthetic process variables. However, the mixed refrigerant includes nitrogen, methane, ethane, propane, and other hydrocarbons. The compositions of the components not included in the set of synthetic process variables may be determined based on relationships to the included process variables. For example, a composition of a second component in the mixed refrigerant may be inferred because a sum of a composition of a first component and the composition of the second component may sum to a value. Because of these relationships, while less than all of the components of the mixed refrigerant are included in the set of synthetic process variables, the total composition of mixed refrigerant may be determined. When an optimal set of synthetic process variables is determined (e.g., as in act 540), the total composition of the mixed refrigerant may be determined using these relationships, despite only a subset of mixed refrigerant components being included in the set of synthetic process variables.

[0091] In act 520, the set of synthetic process variables is applied to the machine-learned model. The machine-learned model is configured to or trained to output an estimated KPI based on input process variables. By applying the synthetic process variables (e.g., describing a hypothetical state of the LNG production process) to the machine-learned model, the machine- learned model estimates a KPI corresponding to the synthetic process variables.

[0092] In act 530, a plurality of synthetic performance indicators are estimated based on the set of synthetic process variables applied to the machine-learned model.

[0093] In act 540, an optimal subset of the set of synthetic process variables (including both controllable and uncontrollable process variables) is determined or selected based on determining an optimal synthetic performance indicator estimated by the machine-learned model. The optimal synthetic performance indicator may be determined by comparing the synthetic performance indicators estimated by the machine-learned model. The most optimal synthetic production indicator may be determined based on one or more criteria. For example, the estimated synthetic process indicator with the highest production quantity, volume, flow rate, or value may be the most optimal. In another example, the synthetic performance indicator having the lowest operating cost may be the most optimal. The subset of synthetic process variables may include those synthetic process variables corresponding to the optimal synthetic performance indicator of the performance indicators estimated by the machine-learned model. Because multiple subsets of process variables may be applied to the machine-learned model (e.g., representing one or more hypothetical, possible, or actual states of the LNG production process), multiple corresponding performance indicators are generated. Accordingly, the subsets of process variables that do not correspond to the optimal performance indicator may be discarded.

[0094] In act 550, the subset of the plurality of synthetic process variables and the optimal synthetic performance indicator are output. The synthetic process variables and KPIs may be stored in, for example, the storage device 1212 of FIG.12. The synthetic process variables and KPIs may be accessed or retrieved from storage. For example, a control system of the LNG production process may use the synthetic process variables to operate the LNG production process.

[0095] To test the accuracy of the optimal KPI estimate by the machine-learned model, the controllable synthetic process variables may be implemented in the LNG production process. For example, the LNG production process may be operated according to the subset of synthetic performance variables. In this way, the actual behavior of the LNG production process may be measured and compared to the estimated KPI, thereby testing the accuracy of the machine- learned model.

[0096] FIG. 6 illustrates a further example flow chart 600 for updating a machine-learned model. The acts may be performed by the computer system 1200 of FIG. 12. Additional, different, or fewer acts may be included. The acts may be performed using a machine-learned model. For example, a machine-learned model generated and/or output according to the acts of FIG. 4 may be used.

[0097] A machine-learned model may be updated by retraining the model according to a set of process variables and corresponding KPIs. Retraining may ensure that the predictions or estimations of the machine-learned model are accurate and reflect current operating conditions of the LNG production process. Retraining may be performed dynamically, periodically, or based on user input. For example, the machine-learned model may be retrained seasonally to ensure accurate predictions with changing weather. In another example, the machine-learned model may be retrained weekly or on another period (also known as “online learning”) to continually improve the model accuracy over time.

[0098] In some cases, the occurrence of an event or fulfillment of criteria may trigger the retraining. For example, the machine-learned model may be retrained after maintenance or replacement of a process element. Retraining may be planned or triggered by predefined maintenance schedules as well as in response to or triggered by replacement of a process element outside of a maintenance interval.

[0099] In this way, the reception of process variables and performance indicators (e.g., in act 610) and/or the application of the data to the machine-learned model (e.g., as in act 620) may be performed in response to a schedule, user input, or a trigger.

[00100] In act 610, performance indicators and process variables may be received that are measured from the LNG production process. The performance indicators and/or process variables may contain data that was not present in the set of process variables and performance indicators used to train the machine-learned model. The received performance indicators and process variables may be referred to as new or unseen data. By including performance indicators and process variables that are new, the machine-learned model may be updated to more accurately estimate a performance indicator for a greater variety of operating conditions of the LNG production process. In some cases, the measured process variables may be part of a dataset of recorded operations data, such as described above with respect to act 310 of FIG. 3. The process variables and performance indicators may be received from storage, such as the storage device 1212 of FIG. 12.

[00101] In some cases, feature selection and filtering may be performed on the received performance indicators and process variables, as in act 420 and/or act 430 of Figure 4. The same kinds of process variables may be included in the data set used to retrain the machine- learned model as were present in the original training dataset (e.g., corresponding to the same sensors, measured quantities, or process elements). In some cases, different or additional process variables (e.g., corresponding to different sensors, measured quantities, or process elements) may be included in the data set used to retrain the machine-learned model as to make the model more accurate. The performance indicators and process variables, after any feature selection or filtering has been applied, may be referred to as the retraining dataset.

[00102] In act 620, the retraining dataset of performance indicators and process variables is applied to the machine-learned model. When there is a difference between the KPI estimated by the machine-learned model and the KPI measured from the LNG production process operating under the same process variables, the machine-learned model may be retrained or updated to improve accuracy. As with the original training of the machine learning model, the machine-learned model may be retrained by applying a set of input process variables and corresponding performance indicators. Because, in this case, process variables and performance indicators in the retraining dataset are not part of the original machine learning dataset used to train the machine learning model, the machine-learned model may be retrained to map the “new” input process variables to the performance indicators.

[00103] In act 630, an updated machine-learned model is generated. The updated machine- learned model may be generated based on applying the performance indicators and process variables of the retraining dataset to the machine-learned model. The updated machine-learned model may reflect the original training performed on the machine learning model as well as the retraining based on the new process variables and performance indicators.

[00104] In act 640, the updated machine-learned model is output. As described with respect to act 450, the machine-learned model may be output to storage (e.g., storage device 1212), integrated with or communicate with a control, monitoring and/or alarm system of the LNG production process. The updated machine-learned model may be accessed or retrieved from storage (e.g., by a control, monitoring, and/or alarm system of the LNG production process). For example, the control system may interact with the updated machine-learned model to change the operation of the LNG production process. In this way, the updated machine-learned model may estimate a KPI based on process variables measured from the LNG production process in real time or measured over a period of time. The estimated KPI (for example, in conjunction with a measured KPI) may be used to adapt the operation of the LNG production process. The updated machine-learned model may replace the earlier machine-learned model or may provide an estimation is addition to the earlier machine-learned model. [00105] FIG. 7 illustrates yet another example flow chart 700 for determining a difference between an actual and an estimated performance indicator. The acts may be performed by the computer system 1200 of FIG. 12. Additional, different, or fewer acts may be included. For example, act 770 may be omitted.

[00106] In act 710, a set of measured process variables is received. The process variables may be measured from and describe a state of the LNG production process (e.g., the MR loop 114). In some cases, the measured process variables may be received from a control system of the LNG production process. The measured process variables may be received in real time, or with a delay. In some other cases, the measured process variables may be part of a dataset of recorded operations data, such as described above with respect to act 310 of FIG. 3.

[00107] In act 720, the set of measured process variables is applied to the machine-learned model. The machine-learned model is trained or configured to output the performance indicator based on input process variables. By applying the process variables to the machine-learned model in real time, the production indicator or KPI may be estimated in real time.

[00108] In act 730, the performance indicator is generated based on applying the plurality of measured process variables. The performance indicator represents an estimate of the performance indicator of the LNG production process operating according to the measured process variables. The performance indicator may be a quantity, volume, flow rate, and/or value of the LNG produced by the process.

[00109] In act 740, the performance indicator is output. For example, the performance indicator may be output to storage, such as the storage device 1212 of FIG. 12. The performance indicator may be retrieved from storage and compared to measured performance indicators to retrain the machine-learned model (e.g., as discussed with respect to FIG. 5) or to discover problems with the LNG production process where measured performance indicators drift or diverge over time from performance indicators estimated by the machine-learned model.

[00110] In act 750, a measured performance indicator describing a state of the LNG production process corresponding to the plurality of measured process variables is received. The measured performance indicator describes a state of the LNG production process operating according to the plurality of measured process variables. For example, the performance indicators may include a maximum production quantity, volume, flow rate, and/or value of LNG produced by the LNG production process. In some cases, the process variables and performance indicators may comprise operations data recorded over time from the LNG production process. In some other cases, the process variables and performance indicators may be operations data may be received in real time from the LNG production process. The process variables and performance indicators may be measured by and/or received from a control or monitoring system of the LNG production process. In some cases, the process variables and performance indicators may be received from storage, such as the storage device 1212 of FIG. 12

[00111] In act 760, a difference between the performance indicator and the measured performance indicator is determined. The difference may be determined between the measured performance indicator and the estimated performance indicator for the same set of process variables. In some cases, the difference may be determined for pairs of estimated and measured performance indicators and tracked over time. An increasing difference between the measured and estimated performance indicators may indicate that an element of the LNG production process may be malfunctioning or may need maintenance. For example, one or more sensors may give an inaccurate reading, or one or more process elements (e.g., a heat exchanger or pressure vessel) may need maintenance. Where the difference is caused by lack of maintenance of process elements or malfunction, maintenance intervals may be updated or sensor outputs may be verified more often. In this way, deviations from expected KPIs (and an accompanying deviation from optimal operation of the LNG production process) may be identified and resolved.

[00112] In act 770, a message is output when the difference is above a threshold. Minor deviations between the measured and estimated performance indicators may be expected as a result of tolerances in measurements of the LNG production process or slight inaccuracies of the predictive power of the machine-learned model. Such small differences may not indicate that the model should be retrained or that the LNG production process requires maintenance. However, where the deviation meets or exceeds a pre-determined threshold, a message may be output indicating the difference. The threshold may be predetermined to filter significant differences from insignificant differences. For example, a threshold set too low may generate false positives, such as when the difference does not correspond to an error in the machine- learned model or a malfunction in the LNG production process. In another example, a threshold set too high may generate false negatives, resulting in messages not being output when a model no longer accurately predicts the production indicator or when a malfunction has happened in the LNG production process.

[00113] FIG. 8 illustrates data 800 of process variables of a MR loop. The dates 802 represent trials of the process variables indicated by the machine-learned model as corresponding to a desired or optimal KPI. In the following columns, concentrations of components 804, 806, 808, 810, and 812 in the MR loop 114 are shown. For each trial, low, target, and high concentrations of the mixed refrigerant components 804, 806, 808, 810, and 812 are shown. In column 814, a change from the prior week 804 or trial in overall concentration is shown.

[00114] After each trial, the process variables and resultant KPIs are measured and used to retrain or update the machine-learned model to more accurately predict the KPI of the LNG production process. The machine-learned model may be retrained or updated, for example, according to the acts of FIGs. 3 and 6. Once the machine-learned model is updated, another trial is performed to check the accuracy of the updated model. The successive updating and trialing of the machine-learned model follows a “design of experiments” method to verify the accuracy of the estimates of the machine-learned model.

[00115] FIG. 9 illustrates a scatter plot 900 of feed gas flow rate for different compositions of mixed refrigerant. The plot 900 of the mixed refrigerant components, component X and component Y, is based on historical measurements of the component concentrations in the mixed refrigerant. A key shows the measured LNG production corresponding to each measurement (e.g., 906, 908, 910) for a feed gas flowrate within lower 902 and upper 904 bounds, and within a temperature range. The historical measurements may be included in the training dataset used to train the machine learning model.

[00116] FIG. 10 illustrates another scatter plot 1000 of feed gas flow rate for different compositions of mixed refrigerant. The plot 1000 of the mixed refrigerant components, component X and component Y, is based on the same historical measurements of the component concentrations in the mixed refrigerant as in FIG. 9. However, the key shows an estimation of LNG production of each measurement (e.g., 1006, 1008, 1010) for a feed gas flowrate within lower 1002 and upper 1004 bounds, and within a temperature range. The LNG production for each measurement is predicted by inputting the measurements of the mixed refrigerant components to a machine-learned model. The accuracy of the estimation of the LNG production by the machine-learning model may be seen by comparing the LNG production values in FIGs. 9 and 10.

[00117] FIG. 11 illustrates a shaded contour plot 1100 of an estimated feed gas flow rate averaged over a period of one month for different synthetic compositions of mixed refrigerant. The horizontal axis represents component X in the mixed refrigerant and the vertical axis represents component Yin the mixed refrigerant. The contour regions 1102, 1104, 1106, 1108, 1110, 1112, and 1114 across the graph represent compositions components X and Y. Each contour represents the component compositions corresponding to a constant LNG production volume. The contour regions are ordered according to the different LNG production volumes, with region 1114 representing the highest LNG production volume, region 1102 representing the lowest LNG production volume, and regions 1112, 1110, 1108, 1106, and 1104 representing decreasing LNG production volumes between regions 1114 and 1102. No composition of mixed refrigerant is defined for unexplored or empty areas 1116.

[00118] Optimal setpoints of refrigerant components of the mixed refrigerant loop may be determined. Though two components, component X and component Y, are shown in the shaded contour plot 1100, concentrations of other components of the mixed refrigerant loop not shown in the shaded contour plot 1100 may be, in some cases, held constant or be based on a relationship with component X and/or component Y. In this way, concentrations of component X and component Y may be varied while the concentrations of the remaining components, for example, denoted as components Z, W, Q, of the mixed refrigerant loop may be determined dependent on the concentrations of component X and component Y. In one example, the concentrations of component X and component Z may sum to 50%, concentrations of components Y and W of the mixed refrigerant loop may sum to 49%, and a remaining component Q of the mixed refrigerant loop may be set at 1%.

[00119] As discussed above with respect to FIG. 5, synthetic process variables may be input to a machine-learned model and performance indicators generated. Predicted LNG production flowrates for varying concentrations of component X and component Y of the mixed refrigeration loop are shown in FIG. 11. In one example, the concentrations of component X and component Z may sum to a constant amount, such as 50% of the mixed refrigerant. In another example, the concentrations of components Y and W may sum to a constant amount, such as 49%. In this way, the composition of components Z and W is uniquely determined from the components X and Y. As such, components Z and W are not shown in FIG. 11. By inputting multiple subsets of process variables, including controllable variables such as concentrations of components of the mixed refrigerant loop, as well as uncontrollable variables, to the machine-learned model, multiple performance indicators are generated. While LNG production based on the current guidelines is in contour region 1108, by using a combination of component X and component Y residing within contour region 1114 and corresponding companion components Z and W, overall LNG production may be increased. [00120] In all practical applications, the present technological advancement must be used in conjunction with a computer, programmed in accordance with the disclosures herein. For example, FIG. 12 is a diagram of an exemplary computer system 1200 that may be utilized to implement methods described herein. A central processing unit (CPU) 1202 is coupled to system bus 1204. The CPU 1202 may be any general-purpose CPU, although other types of architectures of CPU 1202 (or other components of exemplary computer system 1200) may be used as long as CPU 1202 (and other components of computer system 1200) supports the operations as described herein. Those of ordinary skill in the art will appreciate that, while only a single CPU 1202 is shown in FIG. 12, additional CPUs may be present. Moreover, the computer system 1200 may comprise a networked, multi-processor computer system that may include a hybrid parallel CPU/GPU system. The CPU 1202 may execute the various logical instructions according to various teachings disclosed herein. For example, the CPU 1202 may execute machine-level instructions for performing processing according to the operational flow described.

[00121] The computer system 1200 may also include computer components such as non- transitory, computer-readable media. Examples of computer-readable media include computer- readable non-transitory storage media, such as a random access memory (RAM) 1206, which may be SRAM, DRAM, SDRAM, or the like. The computer system 1200 may also include additional non-transitory, computer-readable storage media such as a read-only memory (ROM) 1208, which may be PROM, EPROM, EEPROM, or the like. RAM 1206 and ROM 1208 hold user and system data and programs, as is known in the art. The computer system 1200 may also include an input/output (I/O) adapter 1210, a graphics processing unit (GPU) 1214, a communications adapter 1222, a user interface adapter 1224, a display driver 1216, and a display adapter 1218.

[00122] The I/O adapter 1210 may connect additional non-transitory, computer-readable media such as storage device(s) 1212, including, for example, a hard drive, a compact disc (CD) drive, a floppy disk drive, a tape drive, and the like to computer system 1200. The storage device(s) may be used when RAM 1206 is insufficient for the memory requirements associated with storing data for operations of the present techniques. The data storage of the computer system 1200 may be used for storing information and/or other data used or generated as disclosed herein. For example, storage device(s) 1212 may be used to store configuration information or additional plug-ins in accordance with the present techniques. Further, user interface adapter 1224 couples user input devices, such as a keyboard 1228, a pointing device 1226 and/or output devices to the computer system 1200. The display adapter 1218 is driven by the CPU 1202 to control the display on a display device 1220 to, for example, present information to the user such as images generated according to methods described herein. [00123] The architecture of computer system 1200 may be varied as desired. For example, any suitable processor-based device may be used, including without limitation personal computers, laptop computers, computer workstations, and multi-processor servers. Moreover, the present technological advancement may be implemented on application specific integrated circuits (ASICs) or very large scale integrated (VLSI) circuits. In fact, persons of ordinary skill in the art may use any number of suitable hardware structures capable of executing logical operations according to the present technological advancement. The term “processing circuit” encompasses a hardware processor (such as those found in the hardware devices noted above), ASICs, and VLSI circuits. Input data to the computer system 1200 may include various plug ins and library files. Input data may additionally include configuration information.

[00124] Preferably, the computer is a high performance computer (HPC), known to those skilled in the art. Such high performance computers typically involve clusters of nodes, each node having multiple CPU’s and computer memory that allow parallel computation. The models may be visualized and edited using any interactive visualization programs and associated hardware, such as monitors and projectors. The architecture of system may vary and may be composed of any number of suitable hardware structures capable of executing logical operations and displaying the output according to the present technological advancement. Those of ordinary skill in the art are aware of suitable supercomputers available from Cray or IBM or other cloud computing based vendors such as Microsoft, Amazon.

[00125] It is intended that the foregoing detailed description be understood as an illustration of selected forms that the invention can take and not as a definition of the invention. It is only the following claims, including all equivalents which are intended to define the scope of the claimed invention. Further, it should be noted that any aspect of any of the preferred embodiments described herein may be used alone or in combination with one another. Finally, persons skilled in the art will readily recognize that in preferred implementation, some or all of the steps in the disclosed method are performed using a computer so that the methodology is computer implemented. In such cases, the resulting physical properties model may be downloaded or saved to computer storage.

[00126] The following example embodiments of the invention are also disclosed: Embodiment 1. A computer-implemented method for generating a machine-learned model based on a liquefied natural gas production process, the method comprising: receiving, by a processor, a set of process variables and a corresponding set of performance indicators, wherein the set of performance indicators describes a state of a liquefied natural gas production process, and wherein the set of process variables describes a state of a refrigerant loop of the liquefied natural gas production process; applying, by the processor, a subset of process variables and a corresponding subset of performance indicators to a machine learning model; generating, by the processor, the machine-learned model based on applying the subset of process variables and the corresponding set of performance indicators to the machine learning model, wherein the machine-learned model is configured to output one or more estimated performance indicators based on one or more input process variables; and outputting, by the processor, at least one aspect of the machine-learned model in order to modify operation of at least one aspect of the liquefied natural gas production process. Embodiment 2. The method of embodiment 1, further comprising: determining, by the processor, a first correlation of the set of process variables to the corresponding set of performance indicators and a second correlation between different process variables of the set of process variables; and weighing, by the processor, the subset of process variables from the set of process variables based on the correlations of the set of process variables.

Embodiment 3. The method of embodiments 1 or 2, further comprising: rejecting, by the processor, the process variables of the set of process variables from inclusion in the subset of process variables when the first correlation is below a threshold amount, when the second correlation is above a threshold amount, or when the first correlation is below the threshold amount and the second correlation is above the threshold amount. Embodiment 4. The method of embodiments 1-3, further comprising: determining, by the processor, a sparsity of the set of process variables; or determining, by the processor, a recency of the set of process variables; and weighing, by the processor, the subset of process variables from the set of process variables based on the sparsity, the recency, or the sparsity and the recency.

Embodiment 5. The method of embodiments 1-4, further comprising: generating, by the processor, a set of synthetic process variables wherein the set of synthetic process variables describes a hypothetical state of the refrigerant loop of the liquefied natural gas production process; applying, by the processor, the set of synthetic process variables to the machine-learned model; estimating, by the processor, a plurality of synthetic performance indicators based on the set of synthetic process variables applied to the machine-learned model, wherein the plurality of synthetic performance indicators describe a state of the liquefied natural gas production process corresponding to the set of synthetic process variables; determining, by the processor, a subset of the set of synthetic process variables corresponding to an optimal synthetic performance indicator of the plurality of synthetic performance indicators based on a comparison of the plurality of synthetic performance indicators; and outputting, by the processor, the subset of the set of synthetic process variables and the optimal synthetic performance indicator.

Embodiment 6. The method of embodiments 1-5, further comprising: receiving, by the processor, a further set of process variables and a further corresponding set of performance indicators; applying, by the processor, the further set of performance indicators and the further set of process variables to the machine-learned model; generating, by the processor, an updated machine-learned model based on applying the further set of performance indicators and the further set of process variables to the machine- learned model; and outputting, by the processor, the updated machine-learned model.

Embodiment 7. The method of embodiments 1-6, wherein the set of process variables and the corresponding set of performance indicators are measured from the liquefied natural gas production process over a period of time.

Embodiment 8. The method of embodiments 1-7, wherein the set of performance indicators includes a maximum production quantity of liquefied natural gas, a maximum production volume of liquefied natural gas, a maximum production flowrate of liquefied natural gas, a maximum production value of liquefied natural gas, or a combination thereof, and wherein the operation of at least one aspect of the liquefied natural gas production process is modified based on the maximum production quantity of liquefied natural gas, the maximum production volume of liquefied natural gas, the maximum production flowrate of liquefied natural gas, the maximum production value of liquefied natural gas, or the combination thereof.

Embodiment 9. The method of embodiments 1-8, wherein the set of process variables includes a composition of mixed refrigerant in the refrigerant loop.

Embodiment 10. The method of embodiments 1-9, wherein the set of process variables includes one or more setpoints of a heat exchanger.

Embodiment 11. The method of embodiments 1-10, further comprising: operating, by a control system, the liquefied natural gas production process according to the one or more estimated performance indicators output by the machine-learned model. Embodiment 12. A computer-implemented method for predicting a performance indicator of a liquefied natural gas production process, the method comprising: receiving, by a processor, a set of measured process variables describing a state of a refrigerant loop of the liquefied natural gas production process; applying, by the processor, the set of measured process variables to a machine-learned model, wherein the machine-learned model is trained to output the performance indicator based on input process variables; generating, by the processor, the performance indicator based on applying the set of measured process variables; and outputting, by the processor, the performance indicator in order to modify operation of at least one aspect of the liquefied natural gas production process.

Embodiment 13. The method of embodiment 12, wherein the set of measured process variables is received in real time from a control system configured to operate the liquefied natural gas production process, and wherein the performance indicator is output to the control system in order for the control system to modify operation of the at least one aspect of the liquefied natural gas production process.

Embodiment 14. The method of embodiment 12 or 13, further comprising: receiving, by the processor, a measured performance indicator describing a state of a liquefied natural gas production process corresponding to the set of measured process variables; determining, by the processor, a difference between the performance indicator and the measured performance indicator; and outputting, by the processor, when the difference exceeds a threshold, a message to the control system in order for the control system to modify the operation of the at least one aspect of the liquefied natural gas production process.

Embodiment 15. A system for generating a machine-learned model based on a liquefied natural gas production process, the system comprising: a processor; and a non-transitory machine readable medium comprising code configured to direct the processor to: receive a set of process variables and a corresponding set of performance indicators, wherein the set of performance indicators describes the state of a liquefied natural gas production process, and wherein the set of process variables describes a state of a refrigerant loop of the liquefied natural gas production process; apply a subset of process variables and a corresponding subset of performance indicators to a machine learning model; generate the machine-learned model based on applying the subset of process variables and the corresponding set of performance indicators to the machine learning model, wherein the machine-learned model is configured to output one or more estimated performance indicators based on one or more input process variables; and output at least one aspect of the machine-learned model in order to modify operation of at least one aspect of the liquefied natural gas production process.

Embodiment 16. The system of embodiment 15, wherein the non-transitory machine readable medium comprises code configured to direct the processor to: generate a set of hypothetical process variables, wherein the set of hypothetical process variables describes a hypothetical state of the refrigerant loop of the liquefied natural gas production process; apply the set of hypothetical process variables to the machine-learned model; estimate a plurality of hypothetical performance indicators based on the set of hypothetical process variables applied to the machine-learned model, wherein the plurality of hypothetical performance indicators describe a state of the liquefied natural gas production process corresponding to the set of hypothetical process variables; determine a subset of the set of hypothetical process variables corresponding to an optimal hypothetical performance indicator of the plurality of hypothetical performance indicators based on a comparison of the plurality of hypothetical performance indicators; and output the subset of the set of hypothetical process variables and the optimal hypothetical performance indicator.

Embodiment 17. The system of embodiments 15 or 16, wherein the non-transitory machine readable medium comprises code configured to direct the processor to: receive a further plurality of performance indicators measured from an implementation of the subset of the set of hypothetical process variables; apply the further plurality of performance indicators and the subset of the set of hypothetical process variables to the machine-learned model; generate an updated machine-learned model based on applying the further plurality of performance indicators and the subset of the set of hypothetical process variables to the machine-learned model; and output the updated machine-learned model.

Embodiment 18. The system of embodiments 15-17, wherein the non-transitory machine readable medium comprises code configured to direct the processor to: receive a set of measured process variables describing a state of a refrigerant loop of the liquefied natural gas production process; apply the set of measured process variables to the machine-learned model, wherein the machine-learned model is trained to output the performance indicator based on input process variables; generate the performance indicator based on applying the set of measured process variables; and output the performance indicator in order to modify the operation of at least one aspect of the liquefied natural gas production process.

Embodiment 19. The system of embodiments 15-18, wherein the set of performance indicators includes a maximum production quantity of liquefied natural gas, a maximum production volume of liquefied natural gas, a maximum production flow rate of natural gas, a maximum production value of liquefied natural gas, or a combination thereof.

Embodiment 20. The system of embodiments 15-19, wherein the set of process variables includes a composition of mixed refrigerant in the refrigerant loop, one or more setpoints of a heat exchanger, or a combination thereof. Embodiment 21 : A system comprising: a processor; and a non-transitory machine-readable medium comprising instructions that, when executed by the processor, cause a computing system to perform a method according to any of embodiments 1-14.

Embodiment 22: A non-transitory machine-readable medium comprising instructions that, when executed by a processor, cause a computing system to perform a method according to any of embodiments 1-14.