InterPore2022

Test of multi task XGBoost model and its application in Maokou-1 Member, east Sichuan Basin

30 May 2022, 09:40

1h 10m

Poster Presentation (MS15) Machine Learning and Big Data in Porous Media Poster

Speaker

Mr Keran Li (College of energy resources, Chengdu University of Technology)

Description

The carbonate clastic shoal reservoirs in the Middle Permian Maokou Formation has been proved to be outstanding oil/ gas-generating strata. Clastical shoal reservoirs are mainly developed in Maokou-2 and Maokou-3 Members, where Maokou-1 Member is mostly wackstone and packstone. However, with the gas producting under the instructing of in-stu gas generating and enrichment theory, unconventional gas reservoirs are new targets in Maokou-1 Member. To predict porosity, permeability, TOC and lithogy in Maokou-1 Member, east Sichuan basin, this study designs a new multi taks XGBoost model and compares it with traditional random forest models and XGBoost models in single tasks. Multi task XGBoost model has four parts. The first part is inputting all well logging data and responding porosity/permeability/TOC/lithology (labels). Then all labels re encoded. After mixing-training in one shared XGBoost model, the model splits into four independent XGBoost models. Comparision shows single XGBoost models have higher accuracy than single random forest models (the best model is selected by grid search algorithm). Multi task XGBoost model reaches higher accuracy than single XGBoost models. To tesct multi task XGBoost model, this study collects data from YF-1 and Y66-1 (untrained) and put into multi task XGboost model. Result shoew the accuracies for lithology/porosity/permeability/TOC are 91.3%, 90.4%, 94% and 92% repectively, while for single XGBoost models, the accuracies are 72%, 81%, 77.3% and 80.8%. With multi task XGBoost, central and southeasrten parts of east Sichuan Basin are the most potential zones uncobventional gas reservoirs develops based on Based on Fuzzy Evaluation Method.