Advanced Data
Advanced Data
net/publication/337811448
CITATION READS
1 113
2 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Mahshid Firouzi on 19 December 2019.
Mahshid Firouzi and Suren Rathnayake, School of Chemical Engineering & The University of Queensland Centre
for Natural Gas
This paper was prepared for presentation at the SPE/AAPG/SEG Asia Pacific Unconventional Resources Technology Conference held in Brisbane, Australia, 18 –
19 November 2019.
The URTeC Technical Program Committee accepted this presentation on the basis of information contained in an abstract submitted by the author(s). The contents
of this paper have not been reviewed by URTeC and URTeC does not warrant the accuracy, reliability, or timeliness of any information herein. All information is the
responsibility of, and, is subject to corrections by the author(s). Any person or entity that relies on any information obtained from this paper does so at their own
risk. The information herein does not necessarily reflect any position of URTeC. Any reproduction, distribution, or storage of any part of this paper without the written
consent of URTeC is prohibited.
Abstract
Flowing bottom-hole pressure (FBHP) is a key metric for optimising coal seam gas well performance and
enhancement of production. Downhole pressure gauges are increasingly being used to measure the FBHP.
However, they are impractical, expensive, and complex to install and maintain. Consequently, reliable
measurement and prediction of the FBHP, required to forecast well production, remains a challenge. This
paper aims to predict the flowing bottom-hole pressure in coal seam gas wells by taking advantage of the
temporal data and advanced analytics. Data-driven models have been developed to predict the FBHP by
leveraging the temporal data gathered at the surface in order to control the performance of the wells. The
data used in the study was obtained from five coal seam gas wells containing seven sensor measurements
gathered over 15 -18 months production period. For the prediction of FBHP, we applied linear regression
and neural network-based approaches. Overall, neural networks resulted in the best predictions with the
root mean squared error (RMSE) within 198 - 450 kPa for the five wells.
Keywords: FBHP, big data, advanced analytics, multiphase flow modelling
Introduction
In order to extract coal seam gas (CSG), also known as coal bed methane (CBM), from the coal seams,
groundwater needs to be removed to reduce the flowing bottom-hole pressure (FBHP—the pressure at the
base of the well), to allow gas to flow into the wellbore (Towler et al. 2016).
Depending on the gas and water production rates and dynamic interfacial structure between the gas
and liquid phases, fluid properties and flow geometry, different flow regimes are established. Although
the multiphase flows in conventional oil and gas wells have been well studied, multiphase flows in CSG
wells are yet to receive much attention. Conditions in CSG wells are different to those of conventional
wells, mainly due to the fact that the production profile of gas and liquid in CSG wells and their interfacial
interactions are significantly different to those of conventional wells. Multiple coal seams at different depths
contribute different amounts of gas and water, varying over time. This leads to different flow regimes, which
change with depth and time, affecting the FBHP in the well (Wu et al. 2019).
2 URTEC-198240-MS
The prediction of the FBHP is important for the CSG industry to forecast the well production, track
the well capacity, enhance the amount of the recoverable reserves and accelerate the well dewatering by
optimising the performance of the down-hole pump. Ideally, in CSG wells the pressure at the bottom of a
well needs to be managed carefully to maximise the gas production rate (Firouzi et al. 2016). Downhole
pressure gauges are expensive, unreliable and complex to install. Furthermore, calibration and maintenance
of these down-hole gauges is associated with an expensive and complex well intervention to replace the
gauge. Therefore, the industry relies on mathematical models to predict the FBHP.
A range of mathematical models are available in industry simulators, which are designed for conventional
wells (Duns and Ros 1963, Hagedorn and Brown 1965, Orkiszewski 1967, Beggs and Brill 1973, Gray
1978, Vo and Shoham 1989, Barnea et al. 1985, Barnea 1990, Ansari et al. 1990, Caetano et al. 1992,
Zhang et al. 2003, Hasan and Kabir 1992). Firouzi et al. (2016, 2018) developed mechanistic models using
physical principles to describe the complex flow behavior at different regimes in CSG wells. However,
these mechanistic models rely on closure relations to approximate gas and liquid interactions and their
characteristics such as liquid holdup, slug length and thickness of the liquid films (Mitchell et al. 2018).
These relations are mainly empirical, derived from data based on specific and relatively simple experimental
conditions, which can lead to a great level of uncertainty when used for different and complicated operating
conditions (Wu et al. 2017). Also, the flowrate of gas and water from each coal seam, which varies with
time as the reservoir is produced and interacts with the well conditions, must be measured and used as
the input to the mechanistic models for flow regime prediction. Moreover, due to varying flow patterns
across the well, the flow geometry needs to be segmented into many sections representing each flow pattern
and the most suitable mechanistic model for each segment must be applied, which makes the calculation
process more complicated (Chen et al. 2017). Further to all the above limitations, the fluid level in the
well also needs to be determined, which depends on produced water and gas from each coal seam and
the performance of the down-hole pump. Therefore, due to the highly complex nature of these multiphase
flows and their high dimensionality, mechanistic models become insufficient to rely on and developing a
low-cost method of predicting the FBHP remains a key challenge. Existing instruments on wells (surface
pressure, phase flow rates, pump power and temperature control systems) provide data streams of proxy
measurements, potentially revealing the nature of the flow behavior in the well, and permitting a low-cost
method of predicting the FBHP (and the downhole pump condition). Advanced data analytics allows us to
capture the most influential components by examining the relationship of individual parameters with the
FBHP as the target variable and other variables to avoid losing useful information. A wide range of advanced
data analytics techniques have been implemented for predicting the pressure in multiphase flow systems in
conventional oil and gas wells. These include support vector machines (SVM) (Chen et al. 2017), neural
networks (NN) using feed forward (Osman et al. 2005) and recurrent networks (Li et al. 2014, Antonelo et
al. 2017, Tian 2018, Spesivtsev et al. 2018), but there is no such a study to predict the FBHP in CSG wells.
Different factors including gas and water production rate, pump performance (represented by its
efficiency, torque and speed), surface pressure (which represents the back pressure of surface facilities), and
well completion can contribute to the complexity of the wellbore flow dynamics and, in turn, the prediction
of the FBHP. This study assesses the potential for predicting the FBHP in CSG wells, using only available
data.
Methodology
Our aim is to predict the FBHP using seven predictor variables; casing pressure, gas flow rate, pump torque,
pump speed, tubing pressure, water flow rate from the tubing, and water flow rate from the separator. The
data were obtained from five coal seam gas wells in two geographical areas; wells denoted by 2B, 2C, and
2E are located in one area, while the other two wells, denoted by 4A and 4B, are located in another area.
The data were obtained during 15 to 18 months operation of the wells. After removing the outliers, the raw
URTEC-198240-MS 3
data were down-sampled to a sampling rate of one minute. With all five wells combined, there were close
to 3,400,000 data points.
For the prediction of the FBHP, we have applied linear regression, and feed-forward neural networks to
the data. The accuracy of the models’ predictions was assessed using cross-validation.
Linear regression
With linear regression, the aim is to find linear relationships between a given set of independent variables, x,
in p-dimension to predict the dependent variable y, here the FBHP. The linear regression model is given by:
(1)
where βo and β are the model parameters to be estimated from data, and ϵ is the error term. The ordinary
least squared (OLS) estimates of model parameters are obtained by minimizing the residual sum squares.
Penalization techniques have been proposed to improve the OLS estimates; for example, to improve
predictions through bias-variance trade-off, perform automatic variable selection, and handle correlated
variables. LASSO (least absolute shrinkage and selection operator) proposed by Tibshirani (1996)
simultaneously performs regularization and variable selection by imposing a L1 penalty of the regression
coefficients. The ridge regression (Hoerl and Kennard 1988) imposes a L2 penalty of the regression
coefficients to improve the prediction accuracies through a bias-variance trade-off. The Elastic Net proposed
by (Zou and Hastie, 2005) combines the penalties of LASSO and ridge regression methods. The values of
α and λ, where α = [0, 1], in Equation (2) would determine the penalization method. For LASSO the value
of α is fixed to 1, and for ridge regression the value of α is fixed to 0. The optimal values for penalization
can be estimated during model fitting using the training data.
(2)
Neural networks
In addition to linear regression, we applied multilayer feedforward neural network models to predict the
FBHP. Neural networks provide a convenient approach to handle non-linear relationships in the data for
predictive analysis. Having multiple layers enables modelling of complex relationships in the data. In
feedforward neural networks the information flows from the input, x, through the successive layers to the
output, y.
In the case of multilayer feedforward neural networks, the modelling of the output y can be given by y =
f(x,θ), where the model parameters, θ are to be estimated from the data during the training stage. In the case
of a two-layer feedforward neural network, f(x) has the from f(x) = f(2)(f(1)(x)), where f (1)(x) is the output
from the first layer. The output from the jth neuron in the 1st layer can be given by ,
where wj are the weights for the inputs x for the jth node, bj is a bias term, and g is a mapping function such
as sigmoidal and hyperbolic tangent. As in the case with linear regression, we can consider the squared loss
function, and penalty terms L1 and L2 in estimating the model parameters, the weights and the bias terms.
Performance assessment
To assess the accuracy of the predictive models, we employed a five-fold cross validation strategy, where
each fold contained data from a unique well. During the ith, i = (1, …, 5), iteration of the cross-validation, we
used the ith fold as the test set for a model fitted using the remaining folds. To estimate the hyperparameters
relating the L1 and L2 penalties during each iteration of the cross-validation, we have performed internal
cross-validation only using the data from the training folds. With this scheme, no data from the test well
4 URTEC-198240-MS
were used in fitting the model. For example, to predict the FBHP in well 2B, the model was trained using
the data from other four wells.
Figure 1—Cumulative gas and water flow rates for the five wells.
In addition to the seven surface and sub-surface measurements used as predictor variables for the FBHP,
we created a dichotomous categorical variable which represents the two areas of the wells’ locations. We
first considered ordinary least square (OLS) linear regression to predict the FBHP by a) using only the
seven surface and sub-surface measurements as predictor variables, b) including the well location as a
predictor variable in the previous model; this model is denoted by OLSC, and c) in addition to the main
effects considered in a) we added all pairwise interactions, where for two given (continuous type) variables,
a new variable that corresponds to interaction between them is obtained by multiplying the two. This model
is denoted by OLSI. Further, we considered linear regression with regulation, LASSO, Ridge and the Elastic
Net, with and without considering pairwise interactions between the predictor variables. For penalized
regression, we considered values of λ in (0.1, 0.2, …, 10), while for Elastic Net we searched for optimal α
in (0.75, 0.5, 0.25, 0.1, 0.01, 0.001, 0.0001, 0.00001).
Table 1 provides a comparison of these models using the Akaike information criterion (AIC), a commonly
used criterion for model selection (Akaike 1974) and the adjusted R-squared (Adj. R2) values. In the case
of AIC, the model with the lowest AIC is selected, whereas the highest value of the adjusted R-squared
points to a better fit. With each linear regression modelling approach, inclusion of pairwise interactions
has lowered the AIC, suggesting an improved model fit. The OLS linear regression model with pairwise
interactions (OLSI) has the lowest AIC, closely followed by the LASSO with pairwise interactions. They
both have the same adjusted R-squared values.
URTEC-198240-MS 5
Table 1—Comparision of the linear regresion models. The subscript C denots that the categorical variable for location is used in the
model fitting, while the subscript I denotes that all pairwise interactions between predictor variables are included in the model fitting.
AIC Adj. R2
Neural network has the ability to model complex interactions between the predictor variables without
requiring to explicitly specify the interactions. We considered a two-layer feedforward neural network for
the prediction of FBHP. The first layer of the neural network consisted of h1 = (2, …, 8) nodes in the first
hidden-layer, and h2 = 2 nodes in the second hidden-layer. We have fixed h2 = 2 nodes as it seems sufficiently
small. For the function, g, we used a hyperbolic tangent function. For L1 and L2 penalties, we considered
values 0, 0.0001, 0.001, 0.01, 0.1, and 1. To select the optimal values for penalty terms, we used a five-
fold internal cross-validation only using the training data. When we conducted the five-fold cross-validation
(where each data from each well formed a separate fold), the neural network with h1 = 5 has the lowest
overall RMSE.
To compare the performance of all the models including the neural networks (NN), we estimated the
RMSE values for each model, summarized in Table 2. To compare the predicted FBHP obtained during
the cross-validation, we also estimated the overall RMSE by adding the squared residuals in each fold of
the cross-validation and then obtaining square root of the mean. Overall, the neural network model has the
lowest RMSE values, followed by Elastic NetI, LASSOI and RidgeI.
Table 2—Comparison of the root mean squared error (RMSE) in predicting the FBHP using linear regresion and neural
network models. The overall RMSE is obtained by collating the actual and the predicted FBHP during the cross-validation.
Elastic Elastic
Well OLS OLSC OLSI LASSO LASSOI Ridge RidgeI NN
Net NetI
2B 224 222 270 224 269 227 278 224 247 198
2C 499 494 414 499 407 453 299 490 393 248
2E 200 217 618 200 607 184 445 199 539 473
4A 635 697 515 634 511 614 562 633 501 362
4B 688 774 672 689 667 836 654 699 584 450
ALL 488 528 518 488 466 519 466 489 467 362
Tables 3 summarizes the median relative absolute error of different techniques used for the prediction of
the FBHP. The relative absolute error is defined as follows.
(3)
6 URTEC-198240-MS
Table 3—Comparison of the median absolute relative error percentages in predicting the FBHP using different
techniques. The overall estimate is obtained by collating the actual and the predicted FBHP during the cross-validation.
Elastic Elastic
Well OLS OLSC OLSI LASSO LASSOI Ridge RidgeI NN
Net NetI
2B 5 4 9 5 9 5 8 5 9 5
2C 13 11 12 13 13 12 9 12 14 8
2E 5 6 13 5 13 6 16 5 14 15
4A 29 38 17 29 17 26 23 29 18 14
4B 35 34 45 35 45 40 36 36 30 20
ALL 12 11 14 12 14 11 15 11 15 10
Tables 3 also indicates that the neural networks model’s predictions are better than the other approaches
explored in this study. Considering both RMSE and absolute relative error, the neural network approach
performed better than the linear regression models. For individual wells, the neural network has the lowest
RMSE and absolute relative error except for the well 2E. The true FBHP and the predicted FBHP using the
neural network approach is shown in the Figure 2.
Figure 2—Comparison of the predicted FBHP using the neural network approach with the actual (measured)
FBHP. When predicting the FBHP for a particular well, the data for that well were not used in the model fitting.
Conclusions
In this study, we investigated the predictability of the flowing bottom hole pressure (FBHP) using a small
set of predictors including the casing pressure, gas flow rate, pump torque, pump speed, tubing pressure,
and water flow rates from the tubing and separator in coal seam gas wells. We used a cross-validation based
approach to compare the prediction accuracies of predictive models obtained using linear regression, and
neural networks.
URTEC-198240-MS 7
The cumulative gas rate versus the cumulative water flow rate of these wells indicated that, wells 4A
and 4B demonstrate different flow characteristics compared to the wells 2B, 2C and 2E. These two groups
of wells are also located in two different areas. Including well location as a categorical variable in linear
regression did not improve the prediction accuracy. Based on the AIC, it appears that inclusion of pairwise
interactions between the predictor variables improve the model fit. In order to investigate further complex
relationships between these predictor variables and the target variable, we considered neural networks to
predict the FBHP. Overall, the prediction accuracy obtained with the neural networks appeared to be better
compared to the other approaches. The neural network approach outperformed the linear regression models
with the lowest RMSE and absolute relative errors in prediction.
Given that the linear regression models with pairwise interactions produced lower AIC values, and
the neural network approach provided the highest prediction accuracies, we conclude that the non-linear
relationships between surface and subsurface measurements are important in accurate prediction of the
FBHP.
This study provides evidence that a small set of surface and sub-surface measurements obtained at coal
seam gas wells can be used in the prediction of the FBHP. We believe further studies using data from a
larger number of wells, demonstrating different well characteristics may provide stronger evidence for the
prediction of the FBHP.
Acknowledgements
The authors gratefully acknowledge the financial support from the National Energy Resources Australia
(NERA) supported by industry (Arrow Energy, APLNG and Santos) funding through The University of
Queensland’s Centre for Natural Gas (natural-gas.centre.uq.edu.au).
References
Akaike, H. 1974. A new look at the statistical model identification. IEEE Transactions on Automatic Control, 19: 716–723.
Ansari, A. M., Sylvester N. D. et al. Comprehensive mechanistic model for upward two-phase flow in wellbores. Vol. 9,
151–165. http://dx.doi.org/10.2118/20630-PA.
Antonelo, Eric A., Eduardo Camponogara, Bjarne Foss. 2017. Echo State Networks for data-driven downhole pressure
estimation in gas-lift oil wells. Neural Networks, 85: 106–117. https://doi.org/10.1016/j.neunet.2016.09.009.
Barnea, D. 1990. Effect of bubble shape on pressure drop calculations in vertical slug flow. International Journal of
Multiphase Flow, 16 (1): 79–89. http://dx.doi.org/10.1016/0301-9322(90)90039-L.
Barnea, D., Shoham O. et al. 1985. Gas-liquid flow in inclined tubes: Flow pattern transitions for upward flow. Chemical
Engineering Science, 40 (1): 131–136. http://dx.doi.org/10.1016/0009-2509(85)85053-3.
Beggs, D. H., Brill J. P.. 1973. A Study of Two-Phase Flow in Inclined Pipes. Journal of Petroleum Technology, 25 (5):
607–617. http://dx.doi.org/10.2118/4007-PA.
Caetano, E. F., Shoham O., Brill J. P.. 1992. Upward vertical two-phase flow through an annulus—Part II:
Modeling bubble, slug, and annular flow. Journal of Energy Resources Technology, 114 (1): 14–30. http://
dx.doi.org/10.1115/1.2905916.
Chen, Wei, Qinfeng Di, Feng Ye et al. 2017. Flowing bottomhole pressure prediction for gas wells based on support
vector machine and random samples selection. INT J Hydrogen Energ, 42 (29): 18333–18342. https://doi.org/10.1016/
j.ijhydene.2017.04.134.
Duns, H., Jr., Ros N. C. J.. 1963. Vertical flow of gas and liquid mixtures in wells. Proc., 6th World Petroleum Congress,
Frankfurt am Main, Germany.
Firouzi, M., Rowlandson A. et al. 2018. Mathematical modelling of annular regimes of counter-current gas-liquid flows
in annuli. Proc., Chemeca Queenstown, New Zealand
Firouzi, M., Towler B.F., Rufford T. E . 2016. Developing new mechanistic models for predicting pressure gradient in
coal bed methane wells. Journal of Natural Gas Science and Engineering, 33: 961–972.
Gray, H.E. 1978. Vertical fow correlation in gas wells Vol. User’s Manual for API 14B Surface Controlled Subsurface
Safety Valve Sizing Computer Program, 2n Edition, (Appendix B), American Petroleum Institute, Dallas, TX
(Reprint).
8 URTEC-198240-MS
Hagedorn, Alton R., Brown Kermit E.. 1965. Experimental Study of Pressure Gradients Occurring During
Continuous Two-Phase Flow in Small-Diameter Vertical Conduits. Journal of Petroleum Technology, 17 (4). http://
dx.doi.org/10.2118/940-PA.
Hasan, A. R., Kabir C. S.. 1992. Two-phase flow in vertical and inclined annuli. International Journal of Multiphase
Flow, 18 (2): 279–293. http://dx.doi.org/10.1016/0301-9322(92)90089-Y.
Hoerl, A., Kennard R.. 1988. Ridge regression, in Encyclopedia of Statistical Sciences, Vol. 8, 129–136, Vol. 8, Wiley,
New York, pp. 129-136 (Reprint).
Xiaopeng Li, Miskimins Jennifer, Hoffman B. Todd. 2014. A Combined Bottom-hole Pressure Calculation Procedure
Using Multiphase Correlations and Artificial Neural Network Models. Proc., SPE Annual Technical Conference
Amsterdam, The Netherlands, 12. 10.2118/170683-MS.
Mitchell, T. R., Leonardi C. R., Firouzi, M. et al. 2018. Towards closure relations for the rise velocity of Taylor bubbles in
annular piping using phase-field lattice Boltzmann techniques. Proc., 21st Australasian Fluid Mechanics Conference,
Adelaide, Australia.
Orkiszewski, J. 1967. Predicting Two-Phase Pressure Drops in Vertical Pipe. Journal of Petroleum Technology, 19 (06).
http://dx.doi.org/10.2118/1546-PA.
Osman, E.A., Ayoub, M. A., Aggour M. A.. 2005. An Artificial Neural Network Model for Predicting Bottomhole Flowing
Pressure in Vertical Multiphase Flow. Proc., SPE Middle East Oil and Gas Show and Conference, Kingdom of Bahrain,
11. 10.2118/93632-MS.
Spesivtsev, Pavel, Konstantin Sinkov, Ivan Sofronov et al. 2018. Predictive model for bottomhole pressure based on
machine learning. J Petrol Sci Eng, 166: 825. https://doi.org/10.1016/j.petrol.2018.03.046.
Tian, Chuan. 2018. Machine learning approaches for permanent downhole gauge data interpretation. Ph.D, Stanford
University (June 2018).
Tibshirani, R. 1996. Regression Shrinkage and Selection via the lasso. Journal of the Royal Statistical Society, Series B
(58): 267–88.
Towler, Brian, Mahshid Firouzi, James Underschultz et al. 2016. An overview of the coal seam gas developments
in Queensland. Journal of Natural Gas Science and Engineering, 31: 249–271. https://doi.org/10.1016/
j.jngse.2016.02.040.
Vo, D. T., Shoham O.. 1989. Note on the existence of a solution for two-phase slug flow in vertical pipes. Journal of
Energy Resources Technology, Transactions of the ASME, 111 (2): 64–65. 10.1115/1.3231406.
Benjamin Wu, Firouzi Mahshid, Mitchell Travis et al. 2017. A critical review of flow maps for gas-liquid flows in vertical
pipes and annuli. Chemical Engineering Journal, 326: 350–377. https://doi.org/10.1016/j.cej.2017.05.135.
Benjamin Wu, Firouzi Mahshid, Rufford Thomas E. et al. 2019. Characteristics of counter-current gas-liquid two-
phase flow and its limitations in vertical annuli. Experimental Thermal and Fluid Science, 109: 109899. https://
doi.org/10.1016/j.expthermflusci.2019.109899.
Zhang, Hong-Quan, Qian Wang, Cem Sarica et al. 2003. Unified Model for Gas-Liquid Pipe Flow via Slug
Dynamics—Part 1: Model Development. Journal of Energy Resources Technology, 125 (4): 266–273. http://
dx.doi.org/10.1115/1.1615246.