Investigation of the Relationship between Operational Parameters and Decline Curve Characteristics in Shale Gas Wells Using Data Analytics and Machine Learning

In the current technologically advancing world, the role of data science and data analytics has been increasing throughout the last decade. Machine learning has become increasingly important in different disciplines and its application spheres include also the petroleum industry. There have been studies that revealed how the operational parameters affect the production and well performance which leads to more and more studies to be dedicated to them. One of the methods to predict the production and well performance for the future and its potential lifespan, especially in the decline phase is the decline curve analysis. This study involves an investigation of the relationship between operational parameters and decline curve characteristics based on the dataset consisting of 53 shale gas wells data provided by SPE. Via use of well data, decline curves were fit onto the production history for all 53 shale gas wells, and decline curve characteristics (which are qi, Di, b) were obtained accordingly. As a main subsequent step, the development and application of different machine learning algorithms such as Multiple Linear Regression and a tree-based method of Random Forest, has been performed for the determination of prediction models using operational parameters as an input and decline curve characteristics as an output. As the additional second part of the project, new predictive models of aforementioned types were developed for the prediction of the cumulative production after 0.5 and 1 year. The conclusion reached regarding the relationship between operational parameters and decline curve characteristics is that there is some correlation although the lack of data has complicated the decision-making procedure a lot. With a much higher amount of data, it would have been more precise to define to what extent the correlation is. When it comes to the comparison between the distinct types of models, it has been concluded that Random Forest model performed better wholistically despite in the second part of the study, the Linear Regression model outperformed the former one. Furthermore, feature importance analysis was conducted to disclose the influence level of input parameters on the output ones after the predictive models have been developed. Parameters making the most significant contribution to the results were different based on the case being analysed.