Combining machine learning with domain knowledge and first principles for modeling in the process industries
Computer-based process modeling and simulation methods and systems combine first principles models and machine learning models to benefit where either model is lacking. In one example, input values (measurements) are adjusted by first principles techniques. A machine learning model of the chemical process of interest is trained on the adjusted values. In another example, a machine learning model represents the residual (delta) between a first principles model prediction and empirical data. Residual machine learning models correct physical phenomena predictions in a first principles model of the chemical process. In another example, a first principles simulation model uses the process input data and predictions of the machine learning model to generate simulated results of the chemical process. The hybrid models enable a process engineer to troubleshoot the chemical process, enable debottlenecking the chemical process, enable optimizing performance of the chemical process at the subject industrial plant, and enable automated process control.
1. A computer-implemented method of process modeling and simulation, comprising:
modeling a chemical process of a subject industrial plant, said modeling being by a processor and including generating a first principles model that predicts at least one variable of the chemical process of the subject industrial plant;
receiving a dataset comprising of measured variables of the chemical process of the subject industrial plant;
enriching the received dataset by adding at least one supplemental variable based on the first principles model;
training, using the enriched dataset, a machine learning model to generate a hybrid model that predicts behavior of the chemical process;
generating, using the trained machine learning model, a hybrid model; and
based on predictions made by the generated hybrid model, automatically modifying settings of equipment of the subject industrial plant and enabling improvements in performance of the chemical process.
2. A method as claimed in claim 1 wherein the least one supplemental variable based on the first principles model is used to enhance an input variable in the received dataset, and values of the input variable are augmented in training and developing the machine learning model.
3. A method as claimed in claim 2 further comprising:
generating a dataset of variables based on the first principles model that augment at least one of the measured variables in the received dataset, said generating resulting in augmented variables; and
using the measured variables combined with the augmented variables in training the machine learning model, the trained machine learning model producing a corresponding output variable dataset with enhanced accuracy.
4. A method as claimed in claim 1 wherein the at least one supplemental variable is a measurement of a physical property of the chemical process, and the machine learning model prediction of the measurement of the physical property is used by the generated model instead of the first principles model prediction of the physical property.
5. A method as claimed in claim 4 further comprising:
calculating a predicted value for output from a simulator, the first principles model forming the simulator; and
training and developing the machine learning model to represent differences between observed output variable values from plant data and corresponding output variable values predicted by the simulator.
6. A method as claimed in claim 1 wherein the at least one supplemental variable is a computed error of an output variable of the first principles model relative to a measured output variable from the received dataset.
7. A method as claimed in claim 6 wherein the first principles model is a simulation model, and the method further comprises:
calculating, with the machine learning model, an unmeasurable simulation model parameter or function value of the chemical process; and
using the calculated model parameter or function value as an input into the simulation model.
8. A method as claimed in claim 1 wherein the generated hybrid model predicts any one or more of: operating conditions, physical properties, and output of the chemical process; and
wherein the step of enabling improvements in performance of the chemical process include any of: enabling a process engineer to troubleshoot the chemical process, enabling debottlenecking a portion of the chemical process, and optimizing performance of the chemical process at the subject industrial plant.
9. A method as claimed in claim 1 wherein the step of enabling improvements in performance of the chemical process includes: based on predictions of progress of the chemical process made by the generated hybrid model, automatically controlling settings of equipment of the subject industrial plant.
10. A computer-based process modeling and simulation system, comprising:
a modeling subsystem configured to:
model a chemical process of a subject industrial plant, said modeling being by a processor and including generating a first principles model that predicts at least one variable of the chemical process of the subject industrial plant;
receive a dataset comprising of measured variables of the chemical process of the subject industrial plant;
enrich the received dataset by adding at least one supplemental variable based on the first principles model;
train, using the enriched dataset, a machine learning model to generate a hybrid model that predicts behavior of the chemical process; and
generate, using the trained machine learning model, a hybrid mode; and
a controller coupled to the modeling subsystem configured to automatically modify settings of equipment of the subject industrial plant based on predictions made by the generated hybrid model in a manner that enables improvements in performance of the chemical process at the subject industrial plant.
11. A system as claimed in claim 10 wherein the least one supplemental variable based on the first principles model is used to enhance an input variable in the received dataset, and values of the input variable are augmented in training and developing the machine learning model.
12. A system as claimed in claim 11 wherein the modeling sub system further:
generates a dataset of variables based on the first principles model that augment at least one of the measured variables in the received dataset, said generating resulting in augmented variables; and
uses the measured variables combined with the augmented variables in training the machine learning model, the trained machine learning model producing a corresponding output variable dataset with enhanced accuracy.
13. A system as claimed in claim 10 wherein the at least one supplemental variable is a measurement of a physical property of the chemical process, and the machine learning model prediction of the measurement of the physical property is used by the generated model instead of the first principles model prediction of the physical property.
14. A system as claimed in claim 13 wherein the modeling sub system further comprising:
calculates a predicted value for output from a simulator, the first principles model forming the simulator; and
trains and develops the machine learning model to represent differences between observed output variable values from plant data and corresponding output variable values predicted by the simulator.
15. A system as claimed in claim 10 wherein the at least one supplemental variable is a computed error of an output variable of the first principles model relative to a measured output variable from the received dataset.
16. A system as claimed in claim 15 wherein the first principles model is a simulation model, and the modeling subsystem further configured to:
calculate, with the machine learning mode, an unmeasurable simulation model parameter or function value of the chemical process; and
use the calculated model parameter or function value as an input into the simulation model.
17. A system as claimed in claim 10 wherein the generated hybrid model predicts any one or more of: operating conditions, physical properties, and output of the chemical process; and
wherein the interface enables improvements in performance of the chemical process by any of: enabling a process engineer to troubleshoot the chemical process, enabling debottlenecking a portion of the chemical process, and optimizing performance of the chemical process at the subject industrial plant.
18. A system as claimed in claim 10 wherein the interface includes a controller interface communicatively coupled to a controller of the subject industrial plant such that based on predictions of progress of the chemical process made by the generated model, the controller automatically controls settings of equipment of the subject industrial plant.
19. A computer program product, comprising:
A non-transitory computer readable medium having a memory area carrying computer code instructions embodying process modeling and simulation of a chemical process of interest in a subject industrial plant; and
the computer code instructions including instructions which when executed by one or more digital processors:
(a) model a chemical process of a subject industrial plant, said modeling being by a processor and including generating a first principles model that predicts at least one variable of the chemical process of the subject industrial plant;
(b) receive a dataset comprising of measured variables of the chemical process of the subject industrial plant,
(c) enrich the received dataset by adding at least one supplemental variable based on the first principles model,
(d) train, using the enriched dataset, a machine learning model to generate a hybrid model that predicts behavior of the chemical process,
(e) generate, using the trained machine learning model, a hybrid model, and
(f) based on predictions made by the generated hybrid model, automatically modify settings of equipment of the subject industrial plant and enable improvements in performance of the chemical process at the subject industrial plant.
20. A computer program product as claimed in claim 19 wherein the generated hybrid model predicts any one or more of: operating conditions, physical properties, and output of the chemical process; and
the computer code instructions further include instructions to a processor to implement process control by, based on predictions made by the generated model, automatically controlling settings of equipment of the subject industrial plant.
21. A computer program product as claimed in claim 19 wherein the generated hybrid model utilizes any of: (i) an input variable based on the received dataset, and values of the input variable are augmented in training and developing the machine learning model, (ii) a measurement of a physical property of the chemical process instead of a first principles model prediction of the physical property, and (iii) a quantitative representation of a physical property, derived using a machine learning model prediction, for which there is no known measurement in the received dataset.
22. A computer program product as claimed in claim 19 wherein the improvements in performance of the chemical process include any of: enabling a process engineer to troubleshoot the chemical process, debottlenecking a portion of the chemical process, and optimizing performance of the chemical process at the subject industrial plant.