Novel Machine Learning and Data Analytics Approach for History Matching Giant Mature Multilayered Oil Field

Abstract

Abstract This paper discusses an integrated reservoir study utilizing structured and novel machine learning and data analytics approach for history matching a giant mature multilayered oil field in Mahakam Delta of Indonesia. The unique reservoir modeling challenges and novelty of the data science methods will be presented, along with preliminary results and lessons learnt. One of the most important elements in reservoir characterization and history matching process is integration between static and dynamic modeling. With numerous layers as study perimeter, a large number of uncertain parameters is unavoidable, from dynamic uncertainties such as hydrocarbon contacts and communication between regions to static properties like porosity, water saturation, etc. These in turn will create hundreds of possible scenarios during History Matching. Using python scripting embedded in the reservoir simulator and the agile reservoir modeling (ARM) approach, these uncertainties can be handled quickly, and each ensemble can be analyzed easily with Data Analytics and Machine Learning based proxy approaches. The case study presented here is a giant mature oilfield with more than 50 zones and 100 contact regions. With more than 45 years of production and injection history, the conventional reservoir modeling where each zone is modeled individually and independently, assuming no communication between the regions, has been deemed as taking too much time and effort. The integrated approach bypasses this challenge by allowing simultaneous reservoir modeling as well as quick sensitivity analysis and history matching quality checking. Uncertainties were managed early on; for instance, the porosity model was generated through available algorithms and hydrocarbon contacts with Latin-hypercube sampling method. Preliminary results showed that overall time required to perform the modeling has been reduced significantly, while also establishing communication between the regions. Analytical aquifer modeling and communication between the regions were observed as the most sensitive parameters especially when matching the pressure behavior. Moreover, embedded python script and Data Analytics Dashboard have made it possible to perform fast and systematic analysis, thus more effort and time can be allocated to plan the way forward. The Machine Learning results will be further finalized at the next gate review, considering the project was initially proposed into several gates. Static modeling using Machine Learning, coupled with dynamic modeling workflow and data analytics, has created a complete loop of reservoir study and characterization. All of these are conducted in a structured cloud-based platform, ensuring time-efficient process and repeatability while at the same time enabling hybrid approach by combining conventional method and advanced data driven approach