Scribd Logo
Upload

Welcome to Scribd!

  • 24 views

    HW4 Pete550

    Uploaded by

    Muhammad Habiburrahman
    This document summarizes a study that explored using artificial intelligence algorithms to predict multiphase flow rates from downhole sensor measurements. The study tested different neural network models, including feedforward, LSTM, and TCN models. Based on a simulated production system, the TCN model achieved the best prediction accuracy but required more training time than other models. Incorporating past flow rate values using a recurrent scheme improved predictions but also increased training time. Overall, the study demonstrated that neural networks are a promising approach for virtual flow metering to monitor production remotely.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    HW4 Pete550

    Uploaded by

    Muhammad Habiburrahman
    24 views5 pages
    This document summarizes a study that explored using artificial intelligence algorithms to predict multiphase flow rates from downhole sensor measurements. The study tested different neural network models, including feedforward, LSTM, and TCN models. Based on a simulated production system, the TCN model achieved the best prediction accuracy but required more training time than other models. Incorporating past flow rate values using a recurrent scheme improved predictions but also increased training time. Overall, the study demonstrated that neural networks are a promising approach for virtual flow metering to monitor production remotely.

    Original Title

    HW4_PETE550

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    This document summarizes a study that explored using artificial intelligence algorithms to predict multiphase flow rates from downhole sensor measurements. The study tested different neural network models, including feedforward, LSTM, and TCN models. Based on a simulated production system, the TCN model achieved the best prediction accuracy but required more training time than other models. Incorporating past flow rate values using a recurrent scheme improved predictions but also increased training time. Overall, the study demonstrated that neural networks are a promising approach for virtual flow metering to monitor production remotely.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    24 views5 pages

    HW4 Pete550

    Uploaded by

    Muhammad Habiburrahman
    This document summarizes a study that explored using artificial intelligence algorithms to predict multiphase flow rates from downhole sensor measurements. The study tested different neural network models, including feedforward, LSTM, and TCN models. Based on a simulated production system, the TCN model achieved the best prediction accuracy but required more training time than other models. Incorporating past flow rate values using a recurrent scheme improved predictions but also increased training time. Overall, the study demonstrated that neural networks are a promising approach for virtual flow metering to monitor production remotely.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 5

    King Fahd University of Petroleum and Minerals

    College Of Petroleum Engineering and Geosciences


    Petroleum Engineering Department

    222–PETE–550–01 – Smart Wells and Reservoir Monitoring

    Homework #4

    Name : Muhammad Aufa Habiburrahman


    ID : G202214880
    Instructor : Dr. Sami Alneaim
    PETE550 – Homework 4 – G202214880 – Muhammad Aufa Habiburrahman
    Reference journal: Gryzlov, A., Mironova, L., Safonov, S., and Arsalan, M. " Artificial Intelligence and Data
    Analytics for Virtual Flow Metering." Paper presented at the SPE Middle East Intelligent Oil and Gas
    Conference and Exhibition, Manama, Bahrain (2021). doi: https://doi.org/10.2118/204662-MS.
    Gryzlov et al. (2021) conducted a study that explored the potential for using artificial intelligence in virtual
    metering to monitor multiphase production in a highly dynamic wellbore. The study found that it is
    feasible to predict multiphase flowing data using data-driven algorithms trained on a set of historical
    data. The algorithms were able to predict flow rates by training a set of parameters where the flow rate
    is a function of these parameters. The output data for the algorithm is the flow rate, with the input
    parameters consisting of different variables based on the type of instrumentation used. For example, the
    study describes a combination of sensors that include a Venturi tube, gamma densitometer, and two sets
    of electrical impedance sensors integrated into a single meter body. Additionally, the study shows that it
    is possible to measure multiphase rates in a wellbore equipped with different types of instrumentation
    located at a significant distance from each other. The following relationship can be expressed as an
    indirect measurement of the flow meter, which can then be modeled using an artificial intelligence
    predictive model.
    𝑄 = 𝑓(𝑑𝑃, 𝜖𝑚𝑖𝑥 , 𝑈𝑥𝑐𝑜𝑟 , 𝑃1 , 𝑇1 , 𝑃2 , 𝑇2 , 𝑃3 , 𝑇3 , 𝐶ℎ𝑜𝑘𝑒%, … )
    Time-Series Modelling
    The workflow for developing the time-series model involves pre-processing the available data, which
    includes removing outliers, filtering noise, and processing missing or non-available readings. The
    measurements are further analyzed to extract implicitly embedded information through mathematical
    transformation, which enables the discovery of additional input features that improve the predictive
    capabilities of the algorithm. The input parameters and multiphase flow rates vectors are introduced, and
    the process of flow rate predictions is formulated. The available input data is split into a finite number of
    overlapping sequences, and a VFM model is established by fitting a mathematical function that relates
    the output value of multiphase rates to the given time series of input features. It is feasible to choose the
    value of the indentation step as one, to include all consequent values of flow rates. Recurrent feature
    engineering can also be used to include past values of multiphase rates to potentially improve future
    predictions.
    Data Driven Algorithm
    This study explores the use of different types of neural networks to model variables. By adjusting input
    and output patterns, this study examines various scenarios, such as predicting output time series shifted
    by a predetermined number of time steps or over a predetermined time interval. One of the most popular
    machine learning techniques is the feedforward artificial neural network (also called a multi-layered
    perceptron, or MLP), but it has some limitations when it comes to learning long-term dependencies and
    tracking dynamical trends in time-dependent datasets. To overcome these limitations, the Long Short-
    Term Memory (LSTM) network was developed, which regulates information flow using gates and can
    account for both long-term memory and working memory. In addition, the study employs the Temporal
    Convolutional Network (TCN) to improve the late Convolutional Neural Network (CNN) model, which was
    initially developed for image recognition, to tackle the difficulty of time series forecasting by training filters
    corresponding to certain repeating sequences in the data. The TCN uses dilated convolutions in a 1D fully
    convolutional network and causal convolutions. The study evaluates the performance of these machine
    learning techniques for predicting multiphase flow rates in various production scenarios.
    Case Example: Simulated Production Data
    A simple system for simulating production data was performed. The system configuration consists of a
    long horizontal well connected to a vertical section by a segment that inclines slightly downward. (Figure
    1). At the entry point of the vertical well section, pressure and temperature gauges are installed, and the
    wellbore flow can be controlled with a choke. Using only the two downhole measurements, it is necessary
    to estimate the oil, gas, and water flow rates from the data provided by the topside multiphase meter
    using a data-driven virtual flow metering (VFM) paradigm.

    Figure 1. Schematic description of the production system for VFM


    application

    To evaluate the quality of predictions, the available data is split into two parts: the first 50% is used to
    train the VFM model, and the remaining data is used to test the accuracy of predicted flow rates. The
    virtual flow metering setup can be formulated as follows: with the transient measurements of pressure
    and temperature downhole, it is necessary to predict the topside flow rates of oil, gas, and water.
    Feedforward MLP, LSTM, and TCN are integrated into a single predictive model to perform the predictions.
    The results for predicting oil, gas, and water rates are compared in terms of prediction accuracy and
    training time. The error represents the average of the three multiphase rates used in testing data (Table
    1). The results show that TCN performs best with a medium amount of training time. However, the regular
    neural network achieves acceptable accuracy at a significantly lower computational cost when training
    the model. The LSTM fails to meet expectations in both multiphase rate prediction and overall efficiency.
    Accuracy can be improved using a recurrent scheme of prediction, i.e., the use of the known flow rates
    from the previous time steps, though training time has increased.
    Table 1. Performance comparison of data-driven methods

    The example of prediction with TCN is given in Figure 2, and it is important to note that the quality of
    predictions is worse for the test part of the data than for training, as anticipated. The effect of the
    recurrent scheme is demonstrated in Figure 3, where the uncertainty is plotted for both methods used to
    predict oil rate. For the regular approach, the relative error can be as high as 40% at certain time instants,
    however, these peak values are significantly reduced once the past values of multiphase rates are used to
    improve the predictions. Another observation is that although the relative error can be quite high locally,
    still the uncertainty averaged over the whole testing dataset is reasonable.

    Figure 2. Comparison of estimated gas flow rate (a), water flow rate (b), and oil flow rate (c) using TCN model and regular
    prediction scheme.
    Figure 3. Regular versus Recurrent estimation method for oil
    rate.
    Conclusion
    The authors state that neural networks can be used to predict multiphase rates from measurement data
    and can be a valuable tool for production monitoring. They discuss various data-driven algorithms and
    their practical implementation for time-series prediction. Different types of neural networks have
    different benefits depending on the problem formulation, with more advanced methods performing
    better but requiring more computational time and effort to prepare a model. Recurrent use of features
    can improve prediction quality, but it is sensitive to data quality and should be used with caution. For
    future research, the authors suggest the application of data-driven methods for field use and further
    investigation into the effects of noise on real-life scenarios.

    You might also like