Rate of penetration forecasting while drilling using a transformer-based deep learning model
A method for forecasting a rate of penetration while drilling includes training a transformer-based machine learning model with historical drilling data obtained from a plurality of drilled wells to establish relationships between measured drilling parameters and ROP; acquiring short context drilling data while drilling the subterranean wellbore, the short context drilling data including a plurality of measured drilling parameters and a corresponding ROP; evaluating the short context drilling data using the trained transformer-based machine learning model to update the relationships between the measured drilling parameters and the ROP; and forecasting a future ROP using the short context drilling data and the updated relationships.
1 . A method for forecasting a rate of penetration (ROP) while drilling a subterranean wellbore, the method comprising:
training a transformer-based machine learning model with historical drilling data obtained from a plurality of drilled wells to establish relationships between measured drilling parameters and ROP, the training comprising:
evaluating the historical drilling data to identify seasonal patterns in the historical drilling data; and
dividing the historical drilling data into a plurality of sub-sequences based on the identified seasonal patterns;
acquiring short context drilling data while drilling the subterranean wellbore, the short context drilling data is used to update the relationships between the measured drilling parameters and the ROP and includes a plurality of measured drilling parameters and a corresponding ROP;
forecasting a future ROP using the short context drilling data and the updated relationships; and
deploying the trained transformer-based machine learning model in a downhole tool configured for deployment in a drill string.
2 . The method of claim 1 , further comprising adjusting one or more drilling parameters while drilling the subterranean wellbore to change the ROP, wherein the adjusting is based on the forecast future ROP.
3 . The method of claim 1 , further comprising comparing a previously forecast ROP with actual ROP measurements to backtest the transformer-based machine learning model.
4 . The method of claim 1 , further comprising evaluating the short context drilling data is performed using a transformer-based encoder and a feed-forward dense decoder that extracts current relationships between the measured drilling parameters and the ROP to update the relationships between the measured drilling parameters and the ROP.
5 . The method of claim 4 , wherein the transformer-based encoder is configured to extract mechanical specific energy related information from the short context drilling data.
6 . The method of claim 4 , wherein the short context drilling data is configured and provided to the transformer-based encoder as a tensor having a preselected time window.
7 . The method of claim 6 , wherein the forecasting the future ROP forecasts ROP for a future time period equal in length to the preselected time window.
8 . The method of claim 4 , wherein the forecasting the future ROP is performed using a transformer-based encoder and a transformer-based decoder that is configured to capture sequential patterns and establish temporal relationships between the measured drilling parameters and the ROP.
9 . The method of claim 1 , wherein the measured drilling parameters comprise at least one of standpipe pressure, a drilling fluid flow rate, rotary torque, wellbore depth, weight on bit, and a drill string or top drive rotation rate.
10 . The method of claim 1 , wherein the measured drilling parameters comprise at least one parameter combination selected from the group consisting of a sum of wellbore depth and weight on bit, a product of wellbore depth and torque, a product of differential pressure and a drilling fluid flow rate, a product of a drill string rotation rate and torque, and a product of a drill string rotation rate and differential pressure.
11 . A system for forecasting a rate of penetration (ROP) while drilling a subterranean wellbore, the system comprising:
a downhole tool configured for deployment in a drill string, the downhole tool including a trained transformer-based machine learning model, wherein the trained transformer-based machine learning model is trained with historical drilling data obtained from a plurality of drilled wells to establish relationships between measured drilling parameters and ROP, the training comprising:
evaluating the historical drilling data to identify seasonal patterns in the historical drilling data; and
dividing the historical drilling data into a plurality of sub-sequences based on the identified seasonal patterns;
a processor configured to:
receive short context drilling data from a plurality of downhole sensors while drilling the subterranean wellbore, the short context drilling data including a plurality of measured drilling parameters and a corresponding ROP;
evaluate the short context drilling data using the trained transformer-based machine learning model to update the relationships between the measured drilling parameters and the ROP;
forecast a future ROP using the short context drilling data and the updated relationships; and
deploy the trained transformer-based machine learning model in the downhole tool configured for deployment in the drill string.
12 . The system of claim 11 , wherein the trained transformer-based machine learning model is deployed in a printed wiring assembly that is deployed in the downhole tool.
13 . The system of claim 11 , wherein the short context drilling data comprises time-series data having a measurement interval of less than one second.
14 . The system of claim 11 , wherein the evaluate the short context drilling data is performed using a transformer-based encoder and a feed-forward dense decoder that extracts current relationships between the measured drilling parameters and the ROP to update the relationships between the measured drilling parameters and the ROP.
15 . The system of claim 14 , wherein the forecast the future ROP is performed using a transformer-based encoder and a transformer-based decoder that is configured to capture sequential patterns and establish temporal relationships between the measured drilling parameters and the ROP.
16 . A method for forecasting a rate of penetration (ROP) while drilling a subterranean wellbore, the method comprising:
training a transformer-based machine learning model with historical drilling data obtained from a plurality of drilled wells to establish relationships between measured drilling parameters and ROP, the training comprising:
evaluating the historical drilling data to identify seasonal patterns in the historical drilling data;
dividing the historical drilling data into a plurality of sub-sequences based on the identified seasonal patterns; and
encoding the sub-sequences as tokens using a transformer-based architecture to capture the relationships between the measured drilling parameters and the ROP and thereby obtain a trained transformer-based machine learning model;
acquiring short context drilling data while drilling the subterranean wellbore, the short context drilling data including a plurality of measured drilling parameters and a corresponding ROP;
evaluating the short context drilling data using the trained transformer-based machine learning model to update the relationships between the measured drilling parameters and the ROP;
forecasting a future ROP using the short context drilling data and the updated relationships; and
deploying the trained transformer-based machine learning model in a downhole tool configured for deployment in a drill string.
17 . The method of claim 16 , wherein the short context drilling data is acquired from a plurality of downhole sensors while drilling the subterranean wellbore.
18 . The method of claim 16 , wherein the evaluating the short context drilling data is performed using a transformer-based encoder and a feed-forward dense decoder that extracts current relationships between the measured drilling parameters and the ROP to update the relationships between the measured drilling parameters and the ROP.
19 . The method of claim 18 , wherein the forecasting the future ROP is performed using a transformer-based encoder and a transformer-based decoder that is configured make use of the identified seasonal patterns and establish temporal relationships between the measured drilling parameters and the ROP.