Application of the long short-term memory networks for well-testing data interpretation in tight reservoirs

Abstract

Conventional interpretation of the well-testing data cannot effectively detect the real formation response from the noise to characterize the reservoir. Machine learning techniques provide a new data interpretation approach to discover the relationships between production rates and pressure responses of the well. In this study, the long short-term memory networks (LSTMNs) are explored to analyze the field permanent downhole gauge (PDG) data for better reservoir characterization and modelling. Unlike the conventional recurrent neural networks (RNNs), LSTMNs are designed to learn the long-term dependencies among the sequential data sets. More specifically, an approximate analytical model is firstly proposed to describe the flow rate and pressure behavior in a tight reservoir with natural fractures. The synthetic flow rate and pressure data with noise generated by the analytical model are then used to train the LSTMNs. Prediction accuracy of the LSTMNs is first validated by using the field data sets collected from Montney Formation, and their applicability for reservoirs with different types of boundaries are tested by the synthetic flow rate and pressure data sets generated from the analytical models. The field cases have proved that this data mining technique is able to capture the well shut-in operations in Montney liquid-rich tight reservoirs, where the well bottomhole pressure can be accurately predicted when feeding the model with both gas rate and condensate rate. In addition, the LSTMNs are able to learn the pressure behavior from the noisy data sets and do not require a denoising procedure when predicting the pressure response for a given flow rate. In summary, this work first uses the LSTMNs to interpret well-testing data in naturally fractured tight reservoirs. The LSTMNs not only shows a great tolerance to the noise of datasets, but also can capture the pressure responses characterized by well shut-in, boundary effects, and stress-sensitive fractures. It is proved that this data-driven model can discover the patterns and relationships between the flow rate and pressure through the data mining process, while not requiring the prior knowledge of physical model or mathematical assumptions.