Analysis Of Injection-Production Relationship In Fire Flooding Based On Machine Learning

Fire flooding is an efficient method for oil reservoir development,which can create a high-temperature and high-pressure environment in a short time,and accelerate crude oil to flow,and improve oil recovery.However,during the fire flooding process,the uneven distribution of substances and temperature in the oil reservoir often leads to unstable oil recovery rates and reduced oil recovery efficiency.Due to the advantages of stable production well response and sensitivity,previous research on water injection development has mostly ignored the complexity of fire flooding production injection development.To address this issue,it is necessary to study the injection-production relationship between fire flooding wells.Traditional methods of analyzing static reservoir data are costly and have a certain impact on the reservoir environment.The use of dynamic production data and machine learning can efficiently and economically determine the correlation between injection and production wells.In this paper,a simple analysis of fire flooding and gas injection production and water injection production is conducted.The study area is the pilot test area of Hongqian 1 well,and the injection-production relationship between fire flooding wells is analyzed using both static and dynamic production data.Firstly,static data and production overviews are used to preliminarily understand the injection-production situation among the fire flooding wells in the test area.Secondly,actual fire flooding production data is selected,and three machine learning algorithms(Decision Tree,Random Forest,and Ada Boost)are used to establish corresponding models for each algorithm and to screen the main influencing indicators(primary features)of each well group that influence fire flooding production.Then,a primary feature-multivariate injection-production correlation model suitable for the mine fire flooding production is established,and error caused by the time-delay effect of injection-production data is analyzed through a first-order linear system,quantifying the well-to-well connectivity of fire flooding injection-production wells.The results are compared with actual mine results.Finally,using the primary feature variables combined with the multivariate injection-production correlation model and three machine learning models,the oil production of production wells is predicted.The results showed that the injection-production relationship between fire flooding wells in the pilot test area of Hongqian 1 well is complex,and it is difficult to accurately describe it through traditional dynamic production data and injection-production correlation analysis.Due to different reservoir properties,the main influencing indicators in the injection-production relationship are also different for each well group.Choosing appropriate primary features for the purpose of oil production helps to clarify the dynamic relationship between fire flooding injection-production wells and helps to establish a correlation model.The established primary feature-multivariate injection-production correlation model effectively obtained the well-towell connectivity coefficient between fire flooding injection-production wells.After optimization through a first-order linear system,the model had a higher fitting accuracy and could more accurately describe the well-to-well connectivity of fire flooding injectionproduction wells.The results were consistent with the actual production characteristics and tracer results of the mine.Neural networks showed advantages in predicting oil production,and an appropriate number of learning samples and machine learning models could effectively improve the prediction results and clarify the injection-production relationship between fire flooding wells.