Deriving New Type Curves through Machine Learning in the Wolfcamp Formation

Abstract

Abstract Forecasting production from unconventional reservoirs is a slippery slope that could lead to unrealistic results even though a model has been history matched with production history and vetted by experienced engineers. The reasons for failing conventional decline curves lie in convoluting and heterogeneous reservoir properties, advancing drilling and completion techniques, and dynamic production and operations management. However, the initial rates, declining rate, and ultimate recovery of a well can be viewed as relatively static and predetermined properties of a declining profile. This paper will propose a machine learning-based framework to determine these properties for unconventional reservoir development. In the proposed algorithm, instead of directly data mining on the raw data from different categories and scales, we propose to convert these data into dimensionless variable groups to reduce the dimension of the problem. The dimensionless variables are developed using inspection analysis; most have physical meanings and are easy to upscale. In the case study, we used the production, completion, and petrophysical data to generate new type curves and developed a step-by-step process to explain the aspect of "engineering" code that incorporates physics into the machine learning (ML) process. Dimensionless variables are used in the machine learning process giving physical meaning and reducing the number of predictors, thus improving the speed and efficiency of the code. The results show that the quantity of cumulative oil production over time can be determined using machine learning models with R2 >= 0.90 for individual wells and R2>=0.80 for cross-validated cumulative production forecasts. We can use these determined values to assess the quality of initial rates, declining rates, and ultimate recovery to derive new type curves that incorporate physics and engineering practices. The work emphasizes the importance of accounting for completion parameters, fluid properties and rock quality, thus improving the confidence in results obtained through traditional engineering methods. The machine learning model results provide credibility and support to rates and recoveries for DCA forecasted wells. When modeling hundreds if not thousands of wells, this work shows the importance of utilizing machine learning to harness the power of the data that has been collected on them. The machine-learning-based declining profile is a promising technique and has some advantages over the classical methods based on averaging historical data. First, the determining parameters are highly scalable for newly drilled wells as the main input parameters are dimensionless variables derived from reservoir properties and well completions. Secondly, this algorithm explores not only the production data but also reservoir properties and completion data to capitalize on the advancing techniques.