Predicting long-term production dynamics in tight/shale gas reservoirs with dual-stage attention-based TEN-Seq2Seq model: A case study in Duvernay formation

Abstract

Production dynamics forecasting plays an important role in the decision-making and development scenario evaluation process throughout the entire life cycle of the unconventional tight/shale gas reservoirs. The traditional method such as decline curve analysis can't be applied prior to the wells are put into production as it heavily depends on the historical production for the estimation of parameters. In this work, a new artificial intelligence framework is proposed to predict the well behaviors by simultaneously processing the sequential and tabular data including well depth, proppant tonnage, and fracturing stages. Specifically, a time evolution network is employed first to encode the tabular features matrix into a pseudo-sequence tensor, and then an encoder-decoder architecture based on the dual-stage attention mechanism is used to extract effective information from the encoded information and capture long-term dependencies relationship. A comparison of the proposed model with the fully connected neural network (FCNN) and the long and short-term memory (LSTM) network indicates that the new framework has better generalization performance and robustness to predict well productivities, that is, the prediction errors are reduced by 65% and 50% respectively compared with LSTM and FCNN. Moreover, a bidirectional parametric rectified linear unit (BPReLU) is employed to adaptively learn the sign and magnitude of slopes. It is found that the error is further reduced by approximately 10% compared to that using PReLU. Also, four different target variables are defined, and the experimental results reveal that the average rate within the production time Vi is much easier to predict, with an average error of 19%.