Professional Documents
Culture Documents
1
China University of Petroleum
2
University of Toronto
3
University of Oulu
Introduction
Waterflooding is a major secondary recovery method that is used after oil has been partly produced with natural energy (Heffer et al. 1997).
Normally, only 30% of the oil in a reservoir can be extracted, but waterflooding can increase the recovery and maintain the production of a reservoir
over a longer period. However, after being developed for several years with flooding, the oil fields usually suffer from high water cut of crude oil
and low oil recovery. Interwell connectivity, an important reservoir property, can represent flooding ability and thus be used to guide water injection.
For example, a large interwell connectivity can represent a high permeability zone or a dominant flow channel between two wells. Meanwhile, a
low interwell connectivity is an evidence for low permeability. Therefore, oil producers can take measures such as water plugging or fracturing
according to the interwell connectivity of the reservoir to optimize the production environment.
The most common methods used to evaluate the interwell connectivity are tracer testing, pressure test analysis, and numerical simula-
tion. Interwell tracers (Ji et al. 2002; Du and Guan 2005) can monitor the radiation capacity over time and give a profile of the waterflood-
ing path from the injection well to the production well, which helps describe the connectivity. However, it requires several months for
monitoring and is costly. Pressure test analysis (Whittle et al. 2003; Gringarten et al. 2000) interprets the pressure of stimulus wells and
monitors the pressure variations of response wells to get a pressure transmission coefficient, thereby evaluating the connectivity. This
method disturbs production and increases costs. Numerical simulation (Rwechungura et al. 2011; Emerick 2017) can evaluate the remain-
ing oil after history matching with physical models but cannot reveal the connectivity with grid directly. As such, these methods cannot
be used on a large scale for interwell connectivity.
Over the last several decades, the digital oil field has made great progress (Artun 2017). Continuous sensor data are collected from
individual wells and sent to the intelligent management system, which supports data mining. Therefore, lots of data-driven methods have
been proposed to evaluate interwell connectivity. Some studies use bottomhole pressure data (Tiab 2007; Dinh and Tiab 2008; Liu et al.
2020) to train the model for sensitivity analysis or construct the linear regression relationship (Dinh and Tiab 2013). However, there is a
lot of noise in bottomhole pressure data because of weak correlation. Therefore, it is not easy to learn the regularity from pressure data
and evaluate the connectivity with good performance.
In response to the weak correlation problem, many researchers introduce injection and production data to evaluate interwell connec-
tivity. The interwell connectivity reflects the flowing ability of a reservoir with complex geological conditions. It is difficult to fully
understand these conditions. However, these geological conditions exist stably and follow some laws. Take trap leakage for example. It
will happen in each waterflooding production and follows the complex physical laws. If we control variables of production, the reservoir
is a black-box model. The intelligence model can learn the output of this black-box model with lots of data.
Based on this idea, some studies combine machine learning methods with physical models to evaluate connectivity with noise (Orta
Aleman and Horne 2021) or time lag and attenuation (Albertoni and Lake 2003). These methods use machine learning to solve partly
tasks and coordinate with physical models. This can not only empower the model with laws of physics but also limit the learning ability
of laws not involved. For machine learning methods, the most widely used are correlation analysis (Tian and Horne 2016), multivariate
linear regression (Dinh 2009), and capacitance model (Yousef et al. 2006; Lake et al. 2007; Sayarpour 2008; Soroush et al. 2013; Soroush
2014). Kaviani et al. (2012) even propose two enhancements to increase the tolerance of capacitance model when bottomhole pressures
are unavailable. However, injection/production data and interwell connectivity are dynamically changing over time, and time series can
also be one of the important features. Therefore, some studies introduce time series-oriented deep learning methods to evaluate the con-
nectivity. Cheng et al. (2019) propose an artificial neural network-based interwell connectivity analysis method to get the mapping
Fig. 1—The waterflooding process in the well pattern. The blue nodes are injection wells, and the red nodes are production wells.
Water will flow from the injection well to the production well which makes edge directed.
Representation with Graph Structure. A well pattern can be denoted as a graph G = (V, E), where V and E denote the set of T nodes
(wells in oil field) and Ledges (distribution ratio of injection water). Let V = fv1 , : : : , vT gand E = fe1 , : : : , eL g. Each edge el 2 E can also
be presented as eij = (vi , vj ), where vi , vj 2 V are endpoints of the edge. Let Adenote the adjacency matrix of G where Aij = eij if there is
an edge between node vi and vj:
well1 wellT
2 3
well1 e11 e1T
A= .. 6 . .. 7 (1)
6 . 7
. 4 . eij . 5.
wellT eT1 eTT
Therefore, the production amount can be presented as
Methodology
In this section, we propose G-CMNN to evaluate interwell connectivity. First, we introduce transformer theory for model design. Then,
we propose a graph-based data presentation method to consider the correlation of wells in the well pattern. Finally, we design cooperation-
mission neural networks to evaluate interwell connectivity based on time series.
Transformer for Time Series. Transformer is a kind of neural network for processing sequential data based on the self-attention
mechanism (Vaswani et al. 2017). It has achieved start-of-the-art results in various neural language tasks (Dai et al. 2019; Radford et al.
2019; Devlin et al. 2018). As a powerful self-attention component, it can also be introduced to the computer version domain (Han et al.
2020; Touvron et al. 2021; Zheng et al. 2021) and achieved good performance. For its good feature extraction and memory ability in short
sequential data, we introduce it to learn the temporal information of connectivity and waterflooding.
QKT
Attention (Q, K, V) = softmax p V,
dk (3)
where Q , K , and V
are packed together from input X with fully connected neural networks as Q = XWq , K = XWk , V = XWv , where
q , Wk , and Wv are the weights of neural networks. Multihead attention is utilized to boost the results of several attention layers from
W
different mapping spaces for better performance. The function can be represented as
MultiHead (Q, K, V) = Concat head1 , : : : , headh WO , (4)
Q
headi = Attention QWi , KWK i , VWi ,
V
(5)
Qdmodel dq , WK 2 Rdmodel dk , WV 2 Rdmodel dv , WO 2 Rhdv dmodel , and i 2 Rh . In this work, we employ h = 3attention heads.
where W
i 2R i i i
Graph-Based Well Pattern Representation. To represent the well pattern with graphs, this paper presents some assumptions according to the
production characteristics.
I1 IN
2 3
P1 e11 e1n
AQ = .. 6 . .. 7 (6)
6 . 7,
. 4 . emn . 5
PM em1 eMN
where M and N are the number of production wells and injection wells in a well pattern.
Assumption 2. There is no edge if fault rock interrupts formation continuity between the injection well and the production well. The fault rock
can be identified from seismic data with the technology of formation inversion. In most oil fields, this is a basic work and can obtain the results
easily. In addition, with a deep understanding of the reservoir, the results of formation inversion will be modified. The results of fault rock should
be updated dynamically.
Cooperation-Mission Neural Networks. Based on aforementioned assumptions, there are two targets for neural networks.
Target 1. The predicted production results of production wells should match the real-world results. If we can infer the production
results of each production wells based on the injection data and interwell connectivity, the interwell connectivity results will be reliable.
Therefore, the loss of predicted production results can be used to train the model for interwell connectivity.
To train Transformer
A, we collect the features Fe(I)from the injection well (e.g., injection pressure, thickness of injection layer, and injection
mode), the features Fe(P)from the production wells (e.g., oil pressure, parameters od pump, and thickness of production layer), and well spacing
Dto build the data set. Let AQ denote the interwell connectivity matrix and start it with adjacency matrix A. The function of model training can be
summarized as follows:
Q
A(m, ) =TransformerA (Fe(Pm ), kNn=1 (Fe(In ), Dmn )),
m
(7)
AQ = kM Q
m=1 A(m, ), (8)
loss1 = RMSE(AQ EI > P), Q (9)
where Nmis the number of injection wells that may be connected to the production well Pmbased on the assumptions, EI is the injection amount
vector of all injection wells, PQ is the production amount vector of production wells in 1 month, and is to concatenate the connectivity vector of
production well group. Transformer
A can infer the interwell connectivity with production well group. After M batches, interwell connectivity
matrix AQ can be utilized to calculate the loss with root-mean-square error and train the model with backpropagation.
Target 2. The summation of water distribution ratios of one injection well should be close to 1. However, edge-bottom water and trap
leakage make the water flooding system unbalanced. P
To alleviate this issue, we set the initialized value ˛as the ratio of production volume
M
P
to injection volume which can be written as ˛ = Pm=1 N
m
. The summation of ratios of injection wells are different from each other, but they
n=1 In
are more likely to be around the initialized value ˛. Therefore, we introduce hyper-parameter ıto set the confidence interval which can
be summarized as [ ˛ ı, ˛ + ı]. The summation of ratios of each well can fluctuate within this interval with no loss. This would be the
constraints of both submodels and guide the model training with the loss of production prediction.
We collect the same features as Target 1 and organize the input as injection well group to train Transformer
Q denote the inter-
B . Let A
well connectivity matrix and start it with adjacency matrix A. The function of model training can be summarized as
AQ = kN Q
n=1 A (, n), (11)
8̂ PM PM
< m=1 AQ (m, n) (˛ + ı), if Q
m=1 A (m, n) > (˛ + ı),
P
n2N PM PM
loss2 = (˛ ı) m=1 AQ (m, n), if Q
m=1 A (m, n) < (˛ ı), (12)
n=1 :̂
0, otherwise;
After N batches, interwell connectivity matrix AQ can be utilized to calculate the loss with root-mean-square error and train the model with
backpropagation.
To fuse the results based on two targets, we introduce a fully connected neural network FNNC to predict the interwell connectivity. The
loss of FNNC is the sum of loss1and loss2. The function can be summarized as
Q
AO = Reshape(W3 ((W1 Flatten(A))k(W Q )))),
2 Flatten(A (13)
loss = loss1 + loss2 , (14)
where Flattenis to flatten the matrix to vector and Reshapeis to reshape the vector to the matrix with the original dimension. The structure of
cooperation-mission neural networks is shown in Fig. 3, and the algorithm of cooperation-mission neural networks is shown in Algorithm 1.
Experiments
Introduction to the Field Case and Data Set Generation. The Dongdong unit is the eastern extension of the CNPC-Dagang Oil Field.
The CNPC-Danggang Oil Filed is one of the most prolific oil fields with a complex fault zone shown as Fig. 4. The field was discovered
in 1965, which is the third oil field in China. Since 1970, waterflooding has been their main development mode. By 2018, the composite
water cut of the Dongdong unit is more than 91%. From 2014, the Dongdong unit starts to drive the oil with polymer. Therefore, in this
case study, we collect the production data from 2011 to 2013 for interwell connectivity evaluation.
We evaluate the interwell connectivity based on injection/production data. To reduce the influence of wells that are not in the testing
unit, we select an independent block that is surrounded by faults shown in Fig. 5. The number of this zone that we considered is 11,
including seven injection wells and four production wells. The injection, production, and oil pressure history of each months in the block
is shown in Fig. 6. The variety of injection volume empowers the data of effectiveness. The interwell connectivity is also affected by
Features Output
Injection data Number of layers, thickness, days for injection, TransformerA:
injection mode, oil pressure, casing pressure, Q ;
The adjacency matrix, A
mainline pressure, top depth, bottom depth, TransformerB:
injection volume Q
The adjacency matrix,A
Production data Number of layers, thickness, days for FNNC;
production, pump depth, pump diameter, The adjacency matrix,AO
displacement, pump stroke, pump frequency,
oil pressure, casing pressure, backpressure,
pump efficiency, top depth, bottom depth,
crude oil production
Well spacing Well spacing
Experimental settings. In our work, the production data are collected once a month which makes data scarce. Therefore, we try to infer
the interwell connectivity with lightweight neural networks. The Transformer
Aand Transformer
B have both three transformer layers and
three fully connected layers. The FNNC has three fully connected layers. The ı is the hyper-parameters, and we employ ı = 0.1 in this
work. It can be adjusted by accuracy when values of other parameters are invariant.
We test our model with real-world data. Some simulation models can also present the connectivity. For example, we can trace streamlines and
compute the average interwell connectivity. However, this method may cause some errors because of the complicated geological conditions. The
results from tracers can present the connectivity intuitively. Therefore, we evaluate the model from the following two aspects:
• Interwell connectivity can be tested with tracers. However, the tracer test cannot be applied in all wells because of the high cost. We
can use the available tracer test results to evaluate the model.
• Interwell connectivity is gradual and will not change a lot in a short time, which is called graduality. If the interwell connectivity
fluctuates sharply within several months, it shows that that model can not infer the connectivity with good physical laws.
Results and Evaluation. The accuracy calculated with tracer testing results shows the practicality of the model and the advantage of graphs. The
graduality of prediction results shows the applicability of the model. Therefore, we evaluate the model with accuracy and applicability.
Accuracy Evaluation. According to injection/production data and geological data, the model can evaluate the interwell connectivity
of a well pattern. The crosspoints in Table 2 show the interwell connectivity of injection wells and production wells. According to the
assumption of our model, the column sum of Table 2 should be a vector of 1s when the total injection of a zone is equal to the total pro-
duction. But as shown in Fig. 6, the total injection of this zone is much more than production, so lots of water is lost.
Table 3 shows the results of the tracer test. It is the flooding ability from an injection well to a production well. The interwell connectivity pre-
sented in our model is the distribution ratio that can also be considered as a flooding ability. However, they have different dimensions because of
the different testing methods. Normalization can eliminate the influence of dimension. For the flooding ability presented by tracer testing, we can
calculate the flooding volume per unit time by multiplying the layer thickness, the layer width, and the flooding speed. The target zone is in the same
sand body where reservoir heterogeneity is weak. Therefore, the swept area of each well is assumed to be approximately equal to each other, and
we can infer that the width of each layer is almost the same. By omitting the layer width, we calculate the flow ratios from testing results with the
layer thickness hand the flooding speed v. In our work, we have four tracer samples, so the function can be summarized as follows:
h v
ratioi = P4 i i .
i=1 hi vi (15)
To eliminate the influence of dimension, we also normalize the interwell connectivity in Table 4. The results show that the proposed model can
predict the interwell connectivity and match with tracer testing results. To calculate the accuracy, we design the evaluation function as
Table 4—The accuracy of the model according to the results of the trace test.
1 P
K
jOy(k) y(k) j
Accuracy = 1 K y(k)
,
k=1 (16)
where y(k)is each observed value and yO (k)is the predicted one. The accuracy shows the similarity of K observed values and K predicted values. A
low value would show a low level of correlation. Here we get 91.4% which shows the good performance of the proposed method.
Applicability Evaluation. Interwell connectivity is a characteristic of the reservoir and cannot change a lot in a short time, so we
can also verify the graduality of interwell connectivity. Table 5 presents the interwell connectivity results of four edges from June to
November 2012. We can find that interwell connectivity has a good graduality. Although the injection/production environment is chang-
ing, the model can evaluate the interwell connectivity well and match the physics laws.
Interwell Connectivity Prediction Results G282 to G6-35 G282 to G7-37 G7-36 to G6-35 G7-36 to G7-37
June 0.09727044 0.08929992 0.08032408 0.10174170
July 0.09726977 0.08929960 0.08032368 0.10174119
August 0.09726933 0.08929929 0.08032331 0.10174073
September 0.09726899 0.08929896 0.08032294 0.10174021
October 0.09726868 0.08929857 0.08032261 0.10173964
November 0.09726836 0.08929814 0.08032229 0.10173903
Table 5—The interwell connectivity prediction results from June to November 2012.
Acknowledgments
”This research was supported by the National Natural Science Foundation of China (grant 62072469), National KeyR&D Program
(2018YFE0116700), and the China Scholarship Council.”