An Efficient Spatial-Temporal Convolution Recurrent Neural Network Surrogate Model for History Matching

Surrogate modeling has shown to be effective in improving the solving efficiency for history matching in the development of oil and gas, but the traditional surrogate models are difficult to directly deal with the high-dimensional spatial uncertain parameters, such as the permeability field. In this paper, we introduce a new deep-learning-based surrogate modeling framework, image-to-sequence regression, which can directly predict the production data from the high-dimensional spatial parameters. Specifically, a spatial-temporal convolution recurrent neural network surrogate model is proposed based on a densely connected convolutional neural network (CNN) model and a stacked multilayer long short-term memory (LSTM) model. And a surrogate-based history-matching workflow is then developed by combining the proposed surrogate model with an improved ensemble smoother data assimilation algorithm. Two case studies on a 2D and a 3D reservoir model demonstrate that the proposed surrogate model can effectively predict production data and improve the efficiency of history matching.

Automated summary

This paper introduces a deep learning-based surrogate modeling framework that can directly predict production data from high-dimensional spatial parameters. By combining this surrogate model with an improved data assimilation algorithm, a surrogate-based history-matching workflow is developed.