The Development of Artificial-Neural-Network-Based Universal Proxies to Study Steam Assisted Gravity Drainage (SAGD) and Cyclic Steam Stimulation (CSS) Processes

摘要

Steam injection is one of the most broadly deployed enhanced oil recovery techniques in heavy oil reservoirs. Numerical reservoir simulation plays a significant role in studying the mechanism and design of the field development strategies of steam injection procedures. Artificial neural network (ANN) is considered as a powerful subsidiary tool for high fidelity numerical models for its fast computational speed, especially when large volume of simulation runs are required (Monte Carlo simulation, sensitivity analysis and population-based optimization). This paper focuses on the discussion of the development of ANN-based proxy models studying steam assisted gravity drainage (SAGD) and cyclic steam stimulation (CSS) procedures. The proxy models will consider rock and fluid properties such as relative permeability and temperature dependent fluid viscosity as variables so that they will be capable of handling different types of reservoirs and formation fluids. Half of the SAGD well pair is selected as the minimum unit to study. The ANN model will predict the oil flow rate and cumulative oil production profiles of a SAGD project. To better utilize the injected heat, the CSS procedure in this paper is designed in such a way that the cycle will automatically switch when the oil flow rate in the production phase drops down to a certain threshold value. The project will be terminated when the initial flow rate of one cycle could not maintain the threshold oil flow rate. Following this design scheme, the total number of cycles within the production life will be an unknown. Given a certain set of input parameters, the proxy model will predict the number of CSS cycles and the corresponding oil flow rate and cumulative production profiles. The CSS proxy model developed in this work could be implemented in studying both conventional oil sands and naturally fractured reservoirs. Furthermore, the proxy models developed in this work could be implemented as screening tools which provide engineers with an opportunity to obtain fast recovery estimation of SAGD and CSS projects. They may also assist high fidelity model in history matching, or be employed as proxies in sensitivity analysis and population-based optimization. The ANN proxy models discussed in this paper are parts of a comprehensive ANN-based screening toolbox which is an ensemble extensive EOR processes.