By: Vishnu Subash , Ranjit Chikhalkar , Ashirvad Vatsa

Real-time well rate estimation can provide clear visibility into the performance of a well and might even indicate its reliability issues. Production engineers have access to well production rates from various sources like Flowmeter Data, Allocation Data, and rate estimates from engineering workflows on physical models. Accurate real-time well production rates are not always readily available due to the following issues:

• MPFM (Multiphase flow meters) are not guaranteed in all assets, more so in Shale fields.
• Most current well rate estimation methods are based on the wells and networks physics-based model, which can go stale with changing production and reservoir conditions.
• For non-operated assets, allocated volumes are the only production numbers that come out either at the end of the day or the end of the week based on the contracts.

With the recent technological changes, AI and Machine learning-based solutions have gained momentum in the oil and gas sector. Machine Learning and advanced analytics can be used to calculate real-time well rate estimates belonging to both conventional and unconventional reservoirs. It can also be used to forecast many wells where the physics of fluid flow is uncertain, or wells’ productivity is significantly driven by completion design. The real-time rates obtained from AI/ML can validate Flowmeter data accuracy, check bad allocation data and help improve production solution accuracy.

Real-Time Well Rate Estimates based on AI/ML

Downhole gauge pressure, wellhead pressure, well test, and choke data are readily available for wells with bare minimum automation since these are regulatory requirements. Machine Learning (ML) based Well Rate Estimators can predict oil, water, and gas rates for a well by using a trained model with well test data and the above-mentioned real-time data as input.

ML models can be trained and tested continuously and automatically, using well-test data to ensure that the models are up to date with recent well test. . From the range of models generated, the best-fit ML model can be selected for actual predictions.

A data massage needs to be performed before using the data for the machine learning activity, including cleaning data, removing duplicates, identifying and isolating outliers, normalizing, and computing various running averages to smoothen abrupt changes. The massaged data is then input into the ML training algorithm for the selection, training, and testing of the model. The algorithm chosen for the process could be a polynomial regression algorithm.

In this approach, features are identified from the input that best describes the nature of the well and contributes to the rates. Since the combinations of features could be different for each well, nonrepetitive combinations of features are prepared. Using a brute-force approach, the best-fit model is determined as the one that responds best to the combination of features. Incoming well tests are split in an 80:20 ratio for training and testing the models, respectively. The model with the least average error can thus be chosen as the best-fit model for rate predictions.

Real-time data can be passed as input to the trained model, which represents the live state of the well. Using these inputs, the ML Model can predict the rates for the well, which ideally is within a +/- 5% error range giving the best-predicted rates.

The application handles the errors, logs exceptions, and saves the predictions along with the inputs. The data can then be used for easy error identification and resolution. It also enables users to view the historical data for analytical purposes.

Conclusion:

ML-based well rate estimation can provide an alternate way to evaluate production volumes. This information can then be used for surveillance, forecasting, scenario testing, and verifying the physics-based model health by comparing results with ML-based model results. Artificial intelligence is a good option to lower the investment and maintenance cost for conventional and unconventional fields in the low oil price trends and help in the optimization.