Application of deep learning on well-test interpretation for identifying pressure behavior and characterizing reservoirs

Abstract

Pressure transient well test analysis is an important tool for identifying reservoir characteristics. However, the reliability of the results from well test analysis could be uncertain due to the analysts' lack of experience. This study aims to apply one-dimensional convolutional neural networks (1D CNN) and build an automatic interpretation model of well test data. The model can automatically identify not only the curve type but also the associated parameters. We integrate this automatic interpretation model with four classic well test models, with no model architecture adjustment and hyper-parameters. We validate the results that the curve classification accuracy reaches 97 %, and the median relative error of the curve parameter inversion is approximate 10 %. In addition, the performance of 1D CNN is compared to the artificial neural network (ANN) and two-dimensional convolutional neural networks (2D CNN). Results show that the 1D CNN has a faster training speed and has better accuracy in parameter inversion than ANN and 2D CNN. Finally, the automatic interpretation model is further validated with three field cases. • One-dimensional convolutional neural networks (1D CNN) are used for automatic well test data interpretation. • The non-unique solution of well-testing interpretation can be effectively alleviated by the automated interpretation. • 1D CNN outperforms the ANN and 2D CNN for parameter inversion of well test model. • The results of automatic interpretation in practical cases are close to the level of manual interpretation.