An Attention-Based Temporal and Spatial Convolution Recursive Neural Network for Surrogate Modeling of the Production Curve Prediction

Reservoir numerical simulation is a time-consuming and expensive procedure, especially for production optimization and history matching, which requires multiple calls to the reservoir numerical simulator. The surrogate model based on deep learning can provide a proxy solution to the process of calculating the production curve by the numerical reservoir simulator with approximate accuracy and higher computational efficiency, while an attention mechanism can better capture the local time characteristics of production time series. The attention mechanism is introduced based on deep CNN and LSTM to build an attention-based temporal and spatial convolution recursive neural network for surrogate modeling which is capable of extracting spatial features of reservoir static parameters and handling temporal data, and establishing an image-sequence mapping relationship from reservoir static parameters to reservoir production curve which is used to predict the reservoir production curve. The constructed surrogate model can fast and accurately predict production curves and improves the computational efficiency of production optimization and history matching.