# US 2023/0141060 A1 · Method and system for predicting non-linear well system dynamics performance in green energy generation utilizing physics and machine learning models

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

- Number: US 2023/0141060 A1
- Document ID: US20230141060A1
- Status: Published application
- Kind: A1
- Application number: 17/454,355
- Filed: 2021-11-10
- Published: 2023-05-11
- Earliest claimed: 2021-11-10
- Assignees: Banpu Innovation & Ventures LLC
- Applicants: Banpu Innovation & Ventures LLC.
- Inventors: Othman Elkhomri
- Web page: https://patentgrep.com/patent/US20230141060A1

## CPC

- G06N 3/084: Backpropagation, e.g. using gradient descent
- E21B 43/00: Methods or apparatus for obtaining oil, gas, water, soluble or meltable materials or a slurry of minerals from wells (applicable only to water E03B)
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/048: Activation functions
- E21B 2200/20: Computer models or simulations, e.g. for reservoirs under production, drill bits
- E21B 2200/22: Fuzzy logic, artificial intelligence, neural networks or the like
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- E21B: EARTH OR ROCK DRILLING; OBTAINING OIL, GAS, WATER, SOLUBLE OR MELTABLE MATERIALS OR A SLURRY OF MINERALS FROM WELLS

## Abstract

A method of managing a well system includes: obtaining, by a digital twin manager and based on a predetermined monitoring criterion, well dynamics behavior data of the well system; obtaining, by the digital twin manager, modeled well dynamics behavior data for the well system using a physics-based model; training, by the digital twin manager, a physics constrained machine learning model using one or more machine learning algorithms based on the well dynamics behavior data and the modeled well dynamics behavior data as input data; obtaining, by the digital twin manager, real-time well dynamics behavior data of the well system; outputting, by the digital twin manager, predicted well dynamics behavior data using the real-time well dynamics behavior data, the physics-based model, and the trained physics constrained machine learning model; and transmitting a command to the well system that adjusts a well operation based on the predicted well dynamics behavior data.

## Claims

### Claim 1

A method of managing a well system, comprising:
obtaining, by a digital twin manager and based on a predetermined monitoring criterion, well dynamics behavior data of the well system;
obtaining, by the digital twin manager, modeled well dynamics behavior data for the well system using a physics-based model;
training, by the digital twin manager, a physics constrained machine learning model using one or more machine learning algorithms based on the well dynamics behavior data and the modeled well dynamics behavior data as input data;
obtaining, by the digital twin manager, real-time well dynamics behavior data of the well system;
outputting, by the digital twin manager, predicted well dynamics behavior data using the real-time well dynamics behavior data, the physics-based model, and the trained physics constrained machine learning model; and
transmitting, by the digital twin manager, a command to the well system that adjusts a well operation based on the predicted well dynamics behavior data.

### Claim 2 (depends on 1)

The method of claim 1,
wherein the predicted well dynamics behavior data have higher time resolution than the real-time well dynamics behavior data and captures non-linear well dynamics behavior of the well system,
wherein the non-linear well dynamics behavior includes well mass flow, well pressure, and well temperature performance of a well of interest, and
wherein the well system comprises interconnected subsystems that include a compressor subsystem and a sales header subsystem.

### Claim 3 (depends on 1)

The method of claim 1,
wherein the physics-based model emulates components of non-linear well dynamics behavior that are below a predetermined frequency, and
wherein the physics constrained machine learning model is trained to predict components of the non-linear well dynamics behavior that are above, below, and include the predetermined frequency.

### Claim 4 (depends on 1)

The method of claim 1,
wherein the physics constrained machine learning model is trained based on at least six months of the well dynamics behavior data,
wherein the well dynamics behavior data include data for both normal operational conditions and shut-down conditions,
wherein an interpolation algorithm is applied to fill missing data or outliers in the input data for the physics constrained machine learning model, where the input data are normalized by subtracting signal mean value and scaling with respect to standard deviation, and
wherein a filter is applied to match frequency spectra of the input data for the physics constrained machine learning model based on the real-time well dynamics behavior data from the well system and the modeled well dynamics behavior data using the physics-based model.

### Claim 5 (depends on 4)

The method of claim 4,
wherein the interpolation algorithm is selected from a group consisting of a nearest neighbor interpolation, a linear interpolation, a piecewise cubic spline interpolation, a shape-preserving piecewise cubic spline interpolation, and a modified Akima cubic Hermite interpolation,
wherein the filter is a moving average filter applied in a sliding window along the input data to compute an average value of the input data in each sliding window,
wherein a length of the filter has a predetermined value based on the input data within a predetermined frequency spectrum.

### Claim 6 (depends on 1)

The method of claim 1,
wherein the physics constrained machine learning model is trained using a machine learning algorithm selected from a group consisting of a Levenberg-Marquardt algorithm, a Gauss-Newton algorithm, a steepest descent algorithm, and an artificial neural network.

### Claim 7

A well system, comprising:
a well site;
a physics-based modeling server that outputs modeled well dynamics behavior data for the well site based on a physics-based model; and
a digital twin manager, coupled to the physics-based modeling server and the well site, that includes a processor,
wherein the digital twin manager:
obtains, based on a predetermined monitoring criterion, well dynamics behavior data of the well system;
obtains modeled well dynamics behavior data for the well system using the physics-based model;
trains a physics constrained machine learning model using one or more machine learning algorithms based on the well dynamics behavior data and the modeled well dynamics behavior data as input data;
obtains real-time well dynamics behavior data of the well system;
outputs predicted well dynamics behavior data using the real-time well dynamics behavior data, the physics-based model, and the trained physics constrained machine learning model; and
transmits a command to the well system that adjusts a well operation based on the predicted well dynamics behavior data.

### Claim 8 (depends on 7)

The well system of claim 7,
wherein the predicted well dynamics behavior data have higher time resolution than the real-time well dynamics behavior data and captures non-linear well dynamics behavior of the well system,
wherein the non-linear well dynamics behavior includes well mass flow, well, pressure, and well temperature performance of a well of interest, and
wherein the well system comprises interconnected subsystems that include a compressor subsystem and a sales header subsystem.

### Claim 9 (depends on 7)

The well system of claim 7,
wherein the physics-based model predicts components of non-linear well dynamics behavior that are below a predetermined frequency, and
wherein the physics constrained machine learning model predicts components of the predicted well dynamics behavior that are above, below, and include the predetermined frequency.

### Claim 10 (depends on 7)

The well system of claim 7,
wherein the physics constrained machine learning model is trained based on at least six months of the well dynamics behavior data that include data for both normal operational conditions and shut-down conditions,
wherein an interpolation algorithm is applied to fill missing data or outliers in the input data for the physics constrained machine learning model, where the input data are normalized by subtracting signal mean value and scaling with respect to standard deviation, and
wherein a filter is applied to match frequency spectra of the input data for the physics constrained machine learning model based on the real-time well dynamics behavior data from the well system and the modeled well dynamics behavior data using the physics-based model.

### Claim 11 (depends on 10)

The well system of claim 10,
wherein the interpolation algorithm is selected from a group consisting of a nearest neighbor interpolation, a linear interpolation, a piecewise cubic spline interpolation, a shape-preserving piecewise cubic spline interpolation, and a modified Akima cubic Hermite interpolation,
wherein the filter is a moving average filter applied in a sliding window along the input data to compute an average value of the input data in each sliding window,
wherein a length of the filter has a predetermined value based on the input data within a predetermined frequency spectrum.

### Claim 12 (depends on 7)

The well system of claim 7:
wherein the physics constrained machine learning model is trained using a machine learning algorithm selected from a group consisting of a Levenberg-Marquardt algorithm, a Gauss-Newton algorithm, a steepest descent algorithm, and an artificial neural network.

### Claim 13 (depends on 7)

The well system of claim 7:
wherein the physics constrained machine learning model uses a misfit function which includes a well dynamics behavior prediction error, and
wherein the well dynamics behavior prediction error is selected from a group consisting of integral square error (ISE), mean error (ME), normalized ISE, and normalized ME.

### Claim 14

A non-transitory computer readable medium storing instructions executable by a computer processor, the instructions comprising functionality for:
obtaining well dynamics behavior data of a well system based on a predetermined monitoring criterion;
obtaining modeled well dynamics behavior data for the well system using a physics-based model;
training a physics constrained machine learning model using one or more machine learning algorithms based on the well dynamics behavior data and the modeled well dynamics behavior data as input data;
obtaining real-time well dynamics behavior data of the well system;
outputting predicted well dynamics behavior data using the real-time well dynamics behavior data, the physics-based model, and the trained physics constrained machine learning model; and
transmitting a command to the well system that adjusts a well operation based on the predicted well dynamics behavior data.

### Claim 15 (depends on 14)

The non-transitory computer readable medium of claim 14,
wherein the predicted well dynamics behavior data have higher time resolution than the real-time well dynamics behavior data and captures non-linear well dynamics behavior of the well system,
wherein the non-linear well dynamics behavior includes well mass flow, well pressure, and well temperature performance of a well of interest, and
wherein the well system comprises interconnected subsystems that include a compressor subsystem and a sales header subsystem.
