Hybrid deep neural networks for reservoir production prediction

Abstract

For production prediction, existing physics-based methods often depend on some hypotheses and are confined to certain types of reservoirs. Some data-driven methods take single-source or single-type data as inputs and discard the spatial correlation of reservoir formation contained in structural measurements. Taking multi-source, mixed-type measurements into consideration, we developed an architecture of hybrid deep neural networks (HDNNs) for reservoir production prediction. Despite discrete geologic and engineering measurements, some structural data such as logging images and curves, which contains more details with reservoir formations, were taken as inputs. Hybrid architecture is the key to establish complex nonlinear relationship between target production and mixed inputs. Multiple networks were aggregated to cope with different measurements. For instance, convolutional neural networks (CNN) and multilayer perceptron (MLP) network were configured to separately extract features from structural and discrete inputs. Learned features were then concatenated and fed to subsequent networks for target learning. The implementation of the proposed network consists of four steps: data gathering, data preprocessing, model training and target inferencing. For gathered structural, numerical and categorical data, some essential preprocesses like resampling, normalizing and one-code encoding were performed to prepare dataset for model training. Applied to a development oil block and compared with a typical MLP model and CNN model, the proposed HDNN model stands out with high accuracy and good generalization. The HDNN model converges well both on training and validation dataset. The performance on test dataset demonstrates high accuracy with relative mean absolute error as 6.28%. Further application on three test wells achieved acceptable prediction results.