Development of shale gas production prediction models based on machine learning using early data

The estimate ultimate recovery (EUR) of shale gas in individual well is affected by many factors so that it is difficult to predict accurately. Data-driven methods based on geological and engineering parameters are currently one of the mainstream methods for predicting EUR. However, the importance of early data from gas wells is often overlooked. Therefore, this research set out to use early data, including production and flowback rate data, to develop machine learning models. With the ability to analyze the data by machine learning, the controlling factors on EUR have been analyzed quantitatively. Four schemes have been designed to develop the model, and various machine learning techniques (K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Decision Tree (GBDT)) were applied to process the complex patterns in the data. The results show that, except 30-day flowback rate, the most important factor for EUR is the early production data. The relationship between the early flowback rate and EUR is poor. It is not enough to predict EUR accurately provided that only using the flowback rate data. Good prediction results have been obtained by choosing the most important factors. Among the four algorithms, SVM is considered to be the most reliable model because it is suitable for small data sets and performs well in dealing with nonlinear relationships between variables. The mean absolute percentage error is 13.41% in the test set 49 wells, which can be used as the optimal algorithm for EUR prediction only based on early data. (c) 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Automated summary

This research aims to predict the ultimate recovery of shale gas by developing machine learning models using early data, including production and flowback rate data. The importance of early production data for EUR has been analyzed quantitatively, and four schemes with various machine learning techniques have been designed to process complex patterns in the data. The results show that early production data, except for the 30-day flowback rate, is the most important factor for EUR. Choosing the most important factors leads to good prediction results.