# US 2022/0035346 A1 · Predictions For A Process In An Industrial Plant

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2022/0035346 A1
- Document ID: US20220035346A1
- Status: Published application
- Kind: A1
- Application number: 17/413,020
- Filed: 2019-11-25
- Published: 2022-02-03
- Earliest claimed: 2018-12-13
- Assignees: ABB Schweiz AG
- Applicants: ABB Schweiz AG
- Inventors: Mehmet Mercangoez; Andrea Cortinovis
- Web page: https://patentgrep.com/patent/US20220035346A1

## CPC

- G05B 13/048: {using a predictor}
- G05B 19/4183: {characterised by data acquisition, e.g. workpiece identification}
- G05B 13/027: {using neural networks only}
- G05B 17/02: electric
- G05B 19/41885: {characterised by modeling, simulation of the manufacturing system}
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0499: Feedforward networks
- G06N 3/088: Non-supervised learning, e.g. competitive learning
- G06N 3/09: Supervised learning
- Y02P 90/80: Management or planning
- G05B: CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS (systems for controlling or regulating non-electric variables G05D; systems for regulating electric or magnetic variables G05F; control devices or systems insofar as characterised by mechanical features only G05G)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- Y02P: CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS

## Abstract

To generate real-time or at least near real-time predictions for a process in an industrial plant, a set of neural networks are trained to create a set of trained models. The set of trained models is then used to output the predictions, by inputting online measurement results in an original space to two trained models whose outputs are fed, as reduced space inputs and reduced space initial states, to a third trained model. The third trained model processes the reduced space inputs to reduced space predictions. They are fed to a fourth trained model, which outputs the predictions in the original space.

## Claims

### Claim 1

A computer implemented method comprising:
receiving online measurement results comprising, for a plurality of variables, values that are measured from a process in an industrial plant;
inputting the measurement results in an original space to a first trained model of the process in the industrial plant and to a second trained model of the process in the industrial plant;
processing, by the first trained model, the measurement results to reduced space inputs;
processing, by the second trained model, the measurement results to reduced space initial states;
feeding the reduced space inputs and the reduced space initial states to a third trained model of the process in the industrial plant;
processing, by the third trained model, the reduced space inputs and the reduced space initial states to reduced space predictions;
feeding the reduced space predictions to a fourth trained model of the process in the industrial plant;
processing, by the fourth trained model, the reduced space predictions to predictions in the original space; and
outputting the predictions.

### Claim 2 (depends on 1)

The computer implemented method of claim 1, further comprising:
receiving an update at least to one of the trained models; and
updating the at least one of the trained models correspondingly.

### Claim 3 (depends on 1)

The computer implemented method of claim 1, wherein the first trained model and the fourth trained model are based on one or more variational auto-encoders, the third trained model is based on one or more recurrent neural networks, and the second trained model is based on one or more feedforward artificial neural networks.

### Claim 4

A computer implemented method for creating a set of trained models for a process in an industrial plant, the set of trained models including a first trained model, a second trained model, a third trained model and a fourth trained model, the method comprising:
using past measurement results from the process in the industrial plant as training data;
creating the first trained model and the fourth trained model by training simultaneously an encoding artificial neural network for dimensionality reduction to learn the first trained model and a decoding artificial neural network for dimensionality restoring to learn the fourth trained model, wherein the dimensionality reduction is from an original space to a reduced space and the dimensionality restoring is from the reduced space to the original space; and
creating the second trained model and the third trained model by training neural networks either sequentially, first a recurrent neural network between the first and fourth trained models to create the third trained model for reduced space predictions and then a feedforward artificial neural network to create the second trained model for reduced space initial states, or training simultaneously the recurrent neural network between the first and fourth trained models and the feedforward artificial neural network.

### Claim 5 (depends on 4)

The computer implemented method of claim 4, wherein the encoding artificial neural network and the decoding artificial neural network are variational auto-encoders; and the first trained model and the fourth trained model are learned in a supervised learning process.

### Claim 6 (depends on 4)

The computer implemented method of claim 4, further comprising, when creating the second trained model and the third trained model by training neural networks sequentially:
using, when training the recurrent neural network, reduced space outputs produced by the first trained model from the past measurements results and zeroes as initial states as inputs to the recurrent neural network;
training the artificial neural network with objective to generate better projected results by using past measurement results from the process in the industrial plant and by adjusting outputted reduced order initial states.

### Claim 7 (depends on 4)

The computer implemented method of claim 4, further comprising when creating the second trained model and the third trained model by training neural networks simultaneously:
creating the second trained model and the third trained model by training in parallel an artificial neural network for initial states to create the second trained model and a recurrent neural network to create the third trained model; and
using, for training the recurrent neural network, reduced space outputs produced by the first trained model from the past measurement results, and using, for training the artificial neural network for initial states, the past measurement results.

### Claim 8 (depends on 4)

The computer implemented method of claim 4, wherein the recurrent neural network comprises long short-term memory units or gated recurrent units.

### Claim 9 (depends on 4)

The computer implemented method of claim 4, further comprising:
optimizing further the third trained model by inputting reduced space outputs from the recurrent neural network to an optimization algorithm and inputting results of the optimization algorithm as reduced space inputs to the recurrent neural network.

### Claim 10 (depends on 4)

The computer implemented method of claim 4, further comprising:
repeating the training using updated past measurement results from the process in the industrial plant as training data to retrain at least one of the trained models; and
causing updating the set of trained models after retraining.

### Claim 11 (depends on 4)

A computer implemented method for controlling a process in an industrial plant, the method comprising:
receiving, in a controller implementing a model predictive control, online measurement results including, for a plurality of variables, values that are measured from the process in the industrial plant, and predictions in an original space, which predictions are based on the online measurement results and outputted by a set of trained models, which are created by the method of claim 4 for the process in the industrial plant;
using, in the controller implementing the model predictive control, as an internal model in the model predictive control, the set of trained models, to determine, from the received online measurement results and predictions, constraints for an optimization problem of the model predictive control;
determining, by the optimization problem, based on the constraints determined using the trained models, one or more optimized actions to manipulate the process; and causing, by the controller, manipulation of the process according to the one or more optimized actions.

### Claim 12

A non-transitory computer readable medium comprising one of a first set of program instructions, a second set of program instructions, and a third set of program instructions;
wherein the first set of program instructions causes a computing equipment to perform:
receiving online measurement results comprising, for a plurality of variables,
values that are measured from a process in an industrial plant;
inputting the measurement results in an original space to a first trained model of the process in the industrial plant and to a second trained model of the process in the industrial plant;
processing, by the first trained model, the measurement results to reduced space inputs;
processing, by the second trained model, the measurement results to reduced space initial states;
feeding the reduced space inputs and the reduced space initial states to a third trained model of the process in the industrial plant;
processing, by the third trained model, the reduced space inputs and the reduced space initial states to reduced space predictions;
feeding the reduced space predictions to a fourth trained model of the process in the industrial plant;
processing, by the fourth trained model, the reduced space predictions to predictions in the original space; and
outputting the predictions;
wherein the second set of program instructions causes a computing equipment to perform:
using past measurement results from the process in the industrial plant as training data;
creating the first trained model and the fourth trained model by training simultaneously an encoding artificial neural network for dimensionality reduction to learn the first trained model and a decoding artificial neural network for dimensionality restoring to learn the fourth trained model, wherein the dimensionality reduction is from an original space to a reduced space and the dimensionality restoring is from the reduced space to the original space; and
creating the second trained model and the third trained model by training neural networks either sequentially, first a recurrent neural network between the first and fourth trained models to create the third trained model for reduced space predictions and then a feedforward artificial neural network to create the second trained model for reduced space initial states, or training simultaneously the recurrent neural network between the first and fourth trained models and the feedforward artificial neural network;
wherein the third set of program instructions causes a computing equipment to perform:
receiving, in a controller implementing a model predictive control, online measurement results comprising, for a plurality of variables, values that are measured from the process in the industrial plant, and predictions in an original space, which predictions are based on the online measurement results and outputted by a set of trained models, which are created according to the second set of program instructions for the process in the industrial plant;
using, in the controller implementing the model predictive control, as an internal model in the model predictive control, the set of trained models, to determine, from the received online measurement results and predictions, constraints for an optimization problem of the model predictive control;
determining, by the optimization problem, based on the constraints determined using the trained models, one or more optimized actions to manipulate the process; and causing, by the controller, manipulation of the process according to the one or more optimized actions.

### Claim 13

Equipment comprising at least one processor and at least one memory including computer program code, wherein the at least one memory and the computer program code are configured, with the at least one processor, to cause the equipment to perform:
receiving online measurement results comprising, for a plurality of variables, values that are measured from a process in an industrial plant;
inputting the measurement results in an original space to a first trained model of the process in the industrial plant and to a second trained model of the process in the industrial plant;
processing, by the first trained model, the measurement results to reduced space inputs;
processing, by the second trained model, the measurement results to reduced space initial states;
feeding the reduced space inputs and the reduced space initial states to a third trained model of the process in the industrial plant;
processing, by the third trained model, the reduced space inputs and the reduced space initial states to reduced space predictions;
feeding the reduced space predictions to a fourth trained model of the process in the industrial plant;
processing, by the fourth trained model, the reduced space predictions to predictions in the original space; and
outputting the predictions.

### Claim at position 14

14-15. (canceled)

### Claim 16

The equipment comprising at least one processor and at least one memory including computer program code, wherein the at least one memory and the computer program code are configured, with the at least one processor, to cause the equipment to perform:
using past measurement results from a process in an industrial plant as training data;
creating a first trained model and a fourth trained model by training simultaneously an encoding artificial neural network for dimensionality reduction to learn the first trained model and a decoding artificial neural network for dimensionality restoring to learn the fourth trained model, wherein the dimensionality reduction is from an original space to a reduced space and the dimensionality restoring is from the reduced space to the original space; and
creating a second trained model and a third trained model by training neural networks either sequentially, first a recurrent neural network between the first and fourth trained models to create the third trained model for reduced space predictions and then a feedforward artificial neural network to create the second trained model for reduced space initial states, or training simultaneously the recurrent neural network between the first and fourth trained models and the feedforward artificial neural network.

### Claim 17

The equipment comprising at least one processor and at least one memory including computer program code, wherein the at least one memory and the computer program code are configured, with the at least one processor, to cause the equipment to perform:
receiving, in a controller implementing a model predictive control, online measurement results comprising, for a plurality of variables, values that are measured from a process in an industrial plant, and predictions in an original space, which predictions are based on the online measurement results and outputted by a set of trained models, which are created by the equipment in the industrial plant, the equipment including at least one processor and at least one memory including computer program code, wherein the at least one memory and the computer program code are configured, with the at least one processor, to cause the equipment to perform:
using past measurement results from a process in an industrial plant as training data:
creating a first trained model and a fourth trained model by training simultaneously an encoding artificial neural network for dimensionality reduction to learn the first trained model and a decoding artificial neural network for dimensionality restoring to learn the fourth trained model, wherein the dimensionality reduction is from an original space to a reduced space and the dimensionality restoring is from the reduced space to the original space; and
creating a second trained model and a third trained model by training neural networks either sequentially, first a recurrent neural network between the first and fourth trained models to create the third trained model for reduced space predictions and then a feedforward artificial neural network to create the second trained model for reduced space initial states, or training simultaneously the recurrent neural network between the first and fourth trained models and the feedforward artificial neural network;
using, in the controller implementing the model predictive control, as an internal model in the model predictive control, the set of trained models, to determine, from the received online measurement results and predictions, constraints for an optimization problem of the model predictive control;
determining, by the optimization problem, based on the constraints determined using the trained models, one or more optimized actions to manipulate the process; and causing, by the controller, manipulation of the process according to the one or more optimized actions.

### Claim 18

A system comprising at least:
one or more sensors providing measurement results comprising values for a plurality of variables on at least one process in an industrial plant;
memory for storing the measurement results; and
equipment configured to:
receive the measurement results;
input the measurement results in an original space to a first trained model of the at least one process in the industrial plant and to a second trained model of the at least one process in the industrial plant;
process, by the first trained model, the measurement results to reduced space inputs;
process, by the second trained model, the measurement results to reduced space initial states;
feed the reduced space inputs and the reduced space initial states to a third trained model of the at least one process in the industrial plant;
process, by the third trained model, the reduced space inputs and the reduced space initial states to reduced space predictions;
feed the reduced space predictions to a fourth trained model of the at least one process in the industrial plant;
process, by the fourth trained model, the reduced space predictions to predictions in the original space; and
output the predictions.

### Claim 19 (depends on 18)

The system as claimed in claim 18, wherein the at least one process in the industrial plant is in one of a power plant, manufacturing plant, chemical processing plant, power transmission system, mining and mineral processing plant, pulp and paper, upstream oil and gas system, data center, ship and transportation fleet system.
