Evaluating reservoir performance using a transformer based proxy model

Abstract

In reservoir simulation, proxy models have been used to explore relationships between explanatory variables (e.g., porosity, permeability, well locations and constraints) and response variables (e.g., production rates and bottom hole pressure). Compared to traditional methods used in the proxy models such as capacitance-resistance model (CRM), and interwell numerical Simulation model (INSIM), deep learning methods such as Recurrent Neural Networks (RNNs) have achieved remarkable advancement in predicting reservoir production and assessing uncertainty. However, one limitation of the RNNs is that they are hard to parallelize, which makes their training process computationally expensive. In this paper, a Transformer based proxy model, which is effective in processing sequential data, is developed to accelerate learning and simulation processes. Different types of data are embedded and concatenated as input sequences including water injection, drilling decision, well position, porosity, and permeability. Incorporating additional important information like operational and geological data is found to improve the accuracy of simulation significantly compared to sole injection data input. The model based on a sequence-to-sequence architecture can also be extrapolated to a longer horizon. Both RNNs and Transformer models are used to compare the result accuracy and computation speed. It is found that the Transformer model can be four times faster than the RNNs model under the same order of accuracy.