Steam injection distribution optimization in SAGD oil field using reinforcement learning and web-based GIS

Abstract

Steam injection distribution optimization refers to the process of distributing certain amount of high temperature steam in steam-assisted gravity drainage (SAGD) oil field to maximize the total oil production. In this thesis, two novel optimization methods are presented to solve the steam injection distribution optimization problem. The first optimization method is dynamic programming (DP) method, and the second optimization method is Q-learning method. In the two proposed methods, long short-term memory (LSTM) neural network is used to construct the prediction model to predict oil production of the wells. A web-based geographical information system (GIS) called Petroleum Explorer is developed to support the two proposed methods. Comparison experiments have been conducted using the real-world SAGD production data to test the performance of the proposed methods and the influence of parameter settings on the optimization result. The experiments demonstrate that LSTM model gives better prediction result than other five existing models and both optimization methods can improve the oil production of the oil field. The result also shows the performance of Q-learning method is better than the DP method.