Generating production-type curves for new horizontal wells in unconventional reservoirs is an evolving process that requires continuous calibration to maintain the most accurate forecast over time. History matching production alone is no longer sufficient to maintain such models. Obstacles to creating production type curves are attributed to the complexities in heterogeneous reservoir properties, improved drilling and completion techniques, and evolving production and operation procedures. This paper will highlight improvements to a proposed machine-learning algorithm to generate production type curves for new wells in oil and gas unconventional reservoirs.
The algorithm utilizes dimensionless groups created from the raw data in different categories and scales, thus reducing the dimension of the problem, decreasing the processing time, and improving the efficiency of the machine-learning model. The dimensionless groups are developed using inspectional and dimensional analysis depending on the data available for feature inputs. Many of the dimensionless groups have physical meanings and can be upscaled. We advanced the ability of the previously developed algorithm utilizing production, completion, and petrophysical data from both oil and gas reservoirs to generate new type curves by using the "engineering" code that was laid out in our previous case study. The algorithm incorporates physics into the machine learning (ML) process supporting the outputs with math and science.
When using multiple reservoirs from different formations in the data, the algorithm utilizes logic in the code to determine between oil and gas wells. The quality of the results is impacted when using data from reservoirs with phase envelopes that are not similar, for example, a heavy oil and a dry gas reservoir. The algorithm is updated to include logic that can determine the major phase to predict oil and gas production more accurately. The quantity of oil and gas production is more accurately predicted using cumulative production rates rather than over time. The machine learning model maintains an R2 >= 0.8 when cross-validating both cumulative oil and gas production. The algorithm consistently predicts cumulative production over time on test data with R2 >=0.8. The predicted rates for new type curves are compared to conventional production type curves, thus validating the quality and goodness of fit for production rates, decline profile, and ultimate recovery. The results demonstrate how late-time production can be either extrapolated using the machine learning algorithm or combining traditional methods by utilizing hyperbolic and exponential declines where training data is unavailable for the machine learning model to perform late-time forecasting.
The algorithm of the ML model is proving to be a supplementary tool when generating new production type curves. The speed and efficiency provide support to the DCA generated type curves. It is versatile in its ability to combine data from multiple formations and discern between the major phase, thus providing production type curves we have confidence. The scalability of the dimensionless input parameters can account for changes in completions and reservoir properties within minutes of updating the database hence providing insight in near real-time for engineers.
