Abstract
Unconventional reservoir development is a multidisciplinary challenge due to complicated physical system, including but not limited to complicated flow mechanism, multiple porosity system, heterogeneous subsurface rock and minerals, well interference, and fluid-rock interaction. With enough well data, physics-based models can be supplemented with data driven methods to describe a reservoir system and accurately predict well performance. This study uses a data driven approach to tackle the field development problem in the Eagle Ford Shale.
A large amount of data spanning major oil and gas disciplines was collected and interrogated from around 300 wells in the area of interest. The data driven workflow consists of:
Descriptive model to regress on existing wells with the selected well features and provide insight on feature importance, Predictive model to forecast well performance, and Subject matter expert driven prescriptive model to optimize future well design for well economics improvement.
To evaluate initial well economics, 365 consecutive days of production oil per CAPEX dollar spent (bbl/$) was setup as the objective function. After a careful model selection, Random Forest (RF) shows the best accuracy with the given dataset, and Differential Evolution (DE) was used for optimization.
Using recursive feature elimination (RFE), the final master dataset was reduced to 50 parameters to feed into the machine learning model. After hyperparameter tuning, reasonable regression accuracy was achieved by the Random Forest algorithm, where correlation coefficient (R2) for the training and test dataset was 0.83, and mean absolute error percentage (MAEP) was less than 20%. The model also reveals that the well performance is highly dependent on a good combination of variables spanning geology, drilling, completions, production and reservoir. Completion year has one of the highest feature importance, indicating the improvement of operation and design efficiency and the fluctuation of service cost. Moreover, lateral rate of penetration (ROP) was always amongst the top two important parameters most likely because it impacts the drilling cost significantly. With subject matter experts’ (SME) input, optimization using the regression model was performed in an iterative manner with the chosen parameters and using reasonable upper and lower bounds. Compared to the best existing wells in the vicinity, the optimized well design shows a potential improvement on bbl/$ by approximately 38%.
This paper introduces an integrated data driven solution to optimize unconventional development strategy. Comparing to conventional analytical and numerical methods, machine learning model is able to handle large multidimensional dataset and provide actionable recommendations with a much faster turnaround. In the course of field development, the model accuracy can be dynamically improved by including more data collected from new wells.
Prediction and Chemical Interpretation of Singlet-Oxygen-Scavenging Activity of Small Molecule Compounds by Using Machine Learning
