Estimation of Gas Initial In-Place Utilizing Well Surface Data Via Supervised Machine Learning Approach

Abstract

Abstract Machine learning techniques are being implemented across many industries, including oil and gas, disrupting traditional workflows. The objective of this paper is to provide an effective, robust and an alternate data-driven methodology to estimate gas initial in-place utilizing well surface data. A machine leaning assisted workflow has been established to transform variable rate & wellhead pressure history into constant rate & pressure response, which is subsequently translated into flowing P/Z for in-place calculation. The study concept is based on developing a relationship of flowing wellhead pressure (FWHP) against independent variables via supervised machine learning approach. Multiple linear regression algorithm (MLR) was adopted which models the linear relationship between a single dependent continuous variable and multiple predictor variables. In this study flowing wellhead pressure (FWHP) was modelled against gas rate and cumulative production. FWHP vs gas rate training was performed to establish FWHP response with changing gas rate reflecting well productivity, while FWHP vs cumulative production training was performed to determine FWHP reduction with depletion. The methodology involved typical machine learning pipeline steps including data collection and cleaning, feature engineering, model training & tuning and model deployment. Feature engineering was the most critical step where representative variables were identified and manipulated to improve model prediction accuracy. The data set of each well was split into test/train set (~ 40/60%) and model accuracy was determined via R-squared technique. The best-fit model was then used to generate FWHP profile against constant gas rate, which was then transformed into flowing P/Z to calculate gas initial in-place GIIP. The above procedure was performed on several gas producers. It was identified that FWHP was predicted with reasonable degree of accuracy when trained against feature set consistent for all wells, derived from gas rate and cumulative production. The initial gas in-place subsequently estimated was in line with conventional techniques in all cases validating the reliability of this approach. It was also identified that ~ 20-25% production data was adequate to develop robust ML model providing reliable GIIP estimates. Conventional hydrocarbon initial-in-place estimation techniques require acquisition of downhole data resulting in frequent well shut-in and/or utilization of commercially available applications. The above explained machine learning approach provides equally reasonable in-place estimation utilizing merely surface data, reducing the requirement of extensive downhole acquisition.