{"title":"Advances in Virtual Flow Metering Using Deep Composite Lstm-Autoencoder Network for Gas-Condensate Wells","authors":"J. Omeke","doi":"10.2523/iptc-22524-ms","DOIUrl":null,"url":null,"abstract":"\n In terms of cost and execution time, data-driven Virtual Flow Meters (VFM) are alternative solutions to traditional well testing (WT) and physical multiphase flow meters (MPFM) for production rate determination which is needed for critical decisions by operators but faced with the challenge of low accuracy due to the transient and dynamic state of multiphase flow systems. Recently, some progress has been recorded by training steady state feed-forward neural networks to learn to approximate production rate based on certain number of input features (e.g., choke opening, pressure and temperature etc.) without any recursive feedback connection between the network outputs and inputs. This disconnection has impacted their accuracy. Dynamic artificial neural network, for example, the recurrent neural networks (RNN), e.g., LSTM has shown good performance as its architecture allows for the usage of data from the past time step to predict the current time step. Forecast accuracy for RNN are limited to short period of time due to their inherent vanishing gradient issues. While majority of VFM application have been developed for oil and gas systems, little or non is applied to gas condensate system.\n In this project, a sequence-to-sequence deep composite LSTM-Autoencoders was explored and used to demonstrate the ability of leveraging on its architecture to accurately predict multiphase flow rate for some wells in a gas condensate reservoir with highly dynamic multiphase flow phenomenon. A more complicated flow system was developed using a 3D compositional simulator to simulate, as close as possible, a realistic case of compositional reservoir. A single well was used to train the model and a blind test was ran on two other wells in same reservoir whose data are not part of the training set in order to predict their flow rate with accuracy.\n Based on the actual vs predicted results demonstrated, especially the blind test case, the feature extraction and encoding process of the trained LSTM-autoencoder was actually learning the physics of fluid flow and accurately passing the encoded results to the two decoders with very good output (training and testing mean square error are 0.02 and 0.05 respectively).\n The ability to leverage on some advanced artificial intelligence framework such as a composite LSTM-autoencoder has proven that it is possible to achieve the desired accuracy needed in data driven VFM to meet the requirement of low cost, low execution time and high accuracy.\n This project has also demonstrated the ability of the data driven model to learn the complex dynamics within the temporal ordering of input sequences of production data, with an internal memory adapted to remember or use information across long input sequences, hence, yield longer and reliable forecast, unlike other networks.","PeriodicalId":10974,"journal":{"name":"Day 2 Tue, February 22, 2022","volume":null,"pages":null},"PeriodicalIF":0.0000,"publicationDate":"2022-02-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Day 2 Tue, February 22, 2022","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.2523/iptc-22524-ms","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0
Abstract
In terms of cost and execution time, data-driven Virtual Flow Meters (VFM) are alternative solutions to traditional well testing (WT) and physical multiphase flow meters (MPFM) for production rate determination which is needed for critical decisions by operators but faced with the challenge of low accuracy due to the transient and dynamic state of multiphase flow systems. Recently, some progress has been recorded by training steady state feed-forward neural networks to learn to approximate production rate based on certain number of input features (e.g., choke opening, pressure and temperature etc.) without any recursive feedback connection between the network outputs and inputs. This disconnection has impacted their accuracy. Dynamic artificial neural network, for example, the recurrent neural networks (RNN), e.g., LSTM has shown good performance as its architecture allows for the usage of data from the past time step to predict the current time step. Forecast accuracy for RNN are limited to short period of time due to their inherent vanishing gradient issues. While majority of VFM application have been developed for oil and gas systems, little or non is applied to gas condensate system.
In this project, a sequence-to-sequence deep composite LSTM-Autoencoders was explored and used to demonstrate the ability of leveraging on its architecture to accurately predict multiphase flow rate for some wells in a gas condensate reservoir with highly dynamic multiphase flow phenomenon. A more complicated flow system was developed using a 3D compositional simulator to simulate, as close as possible, a realistic case of compositional reservoir. A single well was used to train the model and a blind test was ran on two other wells in same reservoir whose data are not part of the training set in order to predict their flow rate with accuracy.
Based on the actual vs predicted results demonstrated, especially the blind test case, the feature extraction and encoding process of the trained LSTM-autoencoder was actually learning the physics of fluid flow and accurately passing the encoded results to the two decoders with very good output (training and testing mean square error are 0.02 and 0.05 respectively).
The ability to leverage on some advanced artificial intelligence framework such as a composite LSTM-autoencoder has proven that it is possible to achieve the desired accuracy needed in data driven VFM to meet the requirement of low cost, low execution time and high accuracy.
This project has also demonstrated the ability of the data driven model to learn the complex dynamics within the temporal ordering of input sequences of production data, with an internal memory adapted to remember or use information across long input sequences, hence, yield longer and reliable forecast, unlike other networks.