Machine learning-based fracturing parameter optimization for horizontal wells in Panke field shale oil

Weirong Li; Tianyang Zhang; Xinju Liu; Zhenzhen Dong; Guoqing Dong; Shihao Qian; Zhanrong Yang; Lu Zou; Keze Lin; Tao Zhang

doi:10.1038/s41598-024-56660-8

Machine learning-based fracturing parameter optimization for horizontal wells in Panke field shale oil

Sci Rep. 2024 Mar 13;14(1):6046. doi: 10.1038/s41598-024-56660-8.

Authors

Weirong Li¹, Tianyang Zhang², Xinju Liu^{3

4}, Zhenzhen Dong², Guoqing Dong², Shihao Qian², Zhanrong Yang², Lu Zou², Keze Lin³, Tao Zhang⁴

Affiliations

¹ Xi'an Shiyou University, Xi'an, 710065, China. weirong.li@xsyu.edu.cn.
² Xi'an Shiyou University, Xi'an, 710065, China.
³ China University of Petroleum (Beijing), Beijing, 102249, China.
⁴ Petrochina Changqing Oilfield Company, Xi'an, 710200, China.

Abstract

In the process of developing tight oil and gas reservoirs, multistage fractured horizontal wells (NFHWs) can greatly increase the production rate, and the optimal design of its fracturing parameters is also an important means to further increase the production rate. Accurate production prediction is essential for the formulation of effective development strategies and development plans before and during project execution. In this study, a novel workflow incorporating machine learning (ML) and particle swarm optimization algorithms (PSO) is proposed to predict the production rate of multi-stage fractured horizontal wells in tight reservoirs and optimize the fracturing parameters. The researchers conducted 10,000 numerical simulation experiments to build a complete training and validation dataset, based on which five machine learning production prediction models were developed. As input variables for yield prediction, eight key factors affecting yield were selected. The results of the study show that among the five models, the random forest (RF) model best establishes the mapping relationship between feature variables and yield. After verifying the validity of the Random Forest-based yield prediction model, the researchers combined it with the particle swarm optimization algorithm to determine the optimal combination of fracturing parameters under the condition of maximizing the net present value. A hybrid model, called ML-PSO, is proposed to overcome the limitations of current production forecasting studies, which are difficult to maximize economic returns and optimize the fracturing scheme based on operator preferences (e.g., target NPV). The designed workflow can not only accurately and efficiently predict the production of multi-stage fractured horizontal wells in real-time, but also be used as a parameter selection tool to optimize the fracture design. This study promotes data-driven decision-making for oil and gas development, and its tight reservoir production forecasts provide the basis for accurate forecasting models for the oil and gas industry.

Keywords: Fracturing parameter optimization; Integrating reservoir simulation; Machine learning; Sensitivity analysis; Shale oil.