Performance Prediction for Advanced Well Structures in Unconventional Oil and Gas Reservoirs Using Artificial Intelligent Expert Systems

Abstract

Abstract Advanced well structures have continued to see an increase in field applications, particularly in unconventional reservoir development. Traditional methods for well performance analyses barely capture the complex interactions between these well structures and the reservoir. In this study, a set of artificial expert systems has been developed to predict the performance of advanced well structures in tight oil reservoirs given a set of reservoir properties. The artificial expert systems are built on a system of neural networks. Data for training and validation was generated by building feature sets that contain inputs that were randomly selected within a broad range of limits to capture properties of a typical tight oil reservoir. The features set for inputs into the neural networks includes reservoir extent data, rock and fluid properties, relative permeability and capillary pressure data, as well as well design and operating condition data. Each feature set is run through a commercial numerical simulator to provide time-series data of cumulative oil, gas, water production, as well as bottom-hole pressure profile where applicable. Both the feature set and the time- series data are used for training the neural networks. Generalization tests conducted on each expert system using a validation dataset show that the outputs of the expert systems compare closely with results from the numerical simulator for the same dataset with an average Mean Absolute Error (MAE) of 8%. Other tests such as permutation tests for estimating feature importance, and sensitivity analysis were conducted. The results of the sensitivity analysis showed that the expert system is consistent within its own parametric domain. It also showed that a trained expert system has the ability to diagnose certain characteristics of flow patterns within the reservoir, particularly in the near-wellbore region. This is determined by observing the relative sensitivity of one feature to another. This study shows that an artificial expert system, once trained properly, provides a very practical, fast and robust method for predicting production performance in tight oil reservoirs using advanced well structures. This helps to eliminate the snag of developing an in-depth mathematical model to describe the complex relationships between variables involved, and also increases efficiency in well planning by providing quick, accurate estimates for production forecasts, thus reducing the iteration time spent between the reservoir engineer and other disciplines.