CN116151480A - Shale oil well yield prediction method and device - Google Patents

Shale oil well yield prediction method and device Download PDF

Info

Publication number
CN116151480A
CN116151480A CN202310350371.1A CN202310350371A CN116151480A CN 116151480 A CN116151480 A CN 116151480A CN 202310350371 A CN202310350371 A CN 202310350371A CN 116151480 A CN116151480 A CN 116151480A
Authority
CN
China
Prior art keywords
static
dynamic
data
parameter data
sequence
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202310350371.1A
Other languages
Chinese (zh)
Other versions
CN116151480B (en
Inventor
李江昀
赵丹
李擎
张天翔
张毅思
苗磊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Science and Technology Beijing USTB
Original Assignee
University of Science and Technology Beijing USTB
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Science and Technology Beijing USTB filed Critical University of Science and Technology Beijing USTB
Priority to CN202310350371.1A priority Critical patent/CN116151480B/en
Publication of CN116151480A publication Critical patent/CN116151480A/en
Application granted granted Critical
Publication of CN116151480B publication Critical patent/CN116151480B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/02—Agriculture; Fishing; Mining
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Abstract

The invention provides a shale oil well yield prediction method and device, comprising the following steps: collecting historical production data of shale oil wells, and preprocessing the data, wherein the data comprises static parameter data, dynamic production arrangement parameter data and daily oil production data; inputting static parameter data into a static embedding initial bias module, extracting static embedding information, and taking the static embedding information as initial state bias from a sequence to a sequence model; inputting the static parameter data, the dynamic production scheduling parameter data and the daily oil production data into a dynamic-static complementary cross fusion module to obtain complementary fusion information of the dynamic daily oil production scheduling parameter data and the static parameter data, and inputting the complementary fusion information as a sequence to a sequence model; and (3) performing iterative prediction of any step length within a certain range of shale oil well yield by using a sequence-to-sequence model. The invention can adapt to complex geological engineering conditions, improves the yield prediction accuracy, and realizes the iterative prediction of any step length within a certain range of shale oil well yield.

Description

Shale oil well yield prediction method and device
Technical Field
The invention relates to the technical field of shale oil well yield prediction, in particular to a shale oil well yield prediction method and device.
Background
Shale oil is used as an unconventional clean energy source with huge reserves and wide application prospect, and the development and the utilization of the shale oil can effectively reduce the use amount of conventional fossil fuels such as coal, petroleum and the like. However, blindly performing the development and production of shale oil wells is not preferable; for the high benefit of shale oil development, the prediction of the production of shale oil wells is an important point in the oil and gas field.
The existing shale oil well yield prediction method comprises a prediction method based on mechanism modeling and a prediction method based on data driving. The prediction method based on the mechanism modeling mainly starts from an empirical formula and a mechanical principle, and is based on simplified models provided by different geological conditions. Such methods are generally based on the assumption of idealization, and have unsatisfactory prediction effect and weak generalization capability under complex geological and engineering conditions. The traditional data-driven yield prediction method for shale oil wells is generally established on the basis of autocorrelation of yield data in a time dimension, and a yield decreasing rule is developed and analyzed, so that the model can accurately describe the yield decreasing rule, but the model is complex in structure and difficult to solve. And the machine learning method is a novel data-driven prediction method. While some conventional machine learning methods may be used to analytically predict shale well production, these models tend to ignore the autocorrelation of the data in historical observations and do not extract timing dependent features well. The output of each time step in the long-short-period memory network can influence the output of the next time step, so that the time sequence dependent characteristics can be well extracted and the long-term dependent characteristics can be learned. In recent years, long and short term memory networks (Long Short Term Memory networks, LSTM) have been increasingly applied to production predictions for oil wells. However, the existing prediction model based on the long-short-period memory network rarely comprehensively considers the influence of factors such as geology, construction, production arrangement and the like on the yield and the complex cross-correlation relation thereof. Meanwhile, predicting a future production curve by using dynamic and static parameters such as geology, construction and the like and a known production curve is a dynamic iterative process, and the known production curve and the production curve to be predicted are not in one-to-one correspondence, namely, after an input is given, the model can meet the prediction of daily output of any step. Few models now consider this problem.
Disclosure of Invention
The invention provides a shale oil well yield prediction method and device, which are used for predicting the shale oil well yield. The technical scheme is as follows:
in one aspect, a shale oil well yield prediction method is provided, including:
collecting historical production data of shale oil wells, and preprocessing the data, wherein the data comprises static parameter data, dynamic production scheduling parameter data and daily oil production data;
inputting the static parameter data into a static embedding initial bias module, extracting static embedding information, and taking the static embedding information as initial state bias from a sequence to a sequence model;
inputting the static parameter data, the dynamic production scheduling parameter data and the daily oil production data into a dynamic-static complementary cross fusion module to obtain complementary fusion information of the dynamic daily oil production scheduling parameter data and the static parameter data, and taking the complementary fusion information as input of the sequence to a sequence model;
iterative predictions of shale well production are made using the sequence-to-sequence model.
Optionally, the main body of the static embedded initial bias module is a multi-layer perceptron MLP, and adopts a tanh activation function, and the input of the static embedded initial bias module is the static parameter data and the mining days
Figure SMS_1
The static parameter data includes: geological and fluid physical parameter data->
Figure SMS_2
Construction parameter data->
Figure SMS_3
The output is expressed as +.>
Figure SMS_4
Subsequently->
Figure SMS_5
Hidden state of the sequential feature extractor as the sequence-to-sequence model, +.>
Figure SMS_6
The cell state as the time sequence feature extractor is fed into the model as an initial state bias, and a specific formula is shown in (1):
Figure SMS_7
(1)。
optionally, the dynamic and static complementary cross fusion module includes: two major parts, namely a cross attention module and a gate control fusion module;
the cross attention module is divided into a static passage and a dynamic passage, wherein the static passage firstly determines priori influence information of the dynamic production arrangement parameter data on daily oil production through the dynamic yield interaction module
Figure SMS_8
Determining superposition influence information of the static parameter data on daily oil production by means of static channel attention>
Figure SMS_9
As an output of the static path; the dynamic path firstly determines priori influence information of the static parameter data on daily oil production through a static yield interaction module>
Figure SMS_10
Determining superposition influence information of the dynamic production schedule parameter data on daily oil production through dynamic channel attention>
Figure SMS_11
As an output of the dynamic path;
the said
Figure SMS_12
、/>
Figure SMS_13
、/>
Figure SMS_14
、/>
Figure SMS_15
And daily oil production data are input into the gating fusion module to obtain complementary fusion information +.>
Figure SMS_16
Said->
Figure SMS_17
Together with daily oil production data as input to a timing feature extractor of the sequence-to-sequence model.
Optionally, the specific operation of the static path includes:
introducing a priori influence by the dynamic yield interaction module as shown in a formula (2), wherein the formula (2) is as follows:
Figure SMS_18
(2)
daily oil production data
Figure SMS_21
By means of a linear layer->
Figure SMS_24
Mapping to look-upPolling vector->
Figure SMS_28
Dynamic production of scheduling parameter data->
Figure SMS_22
By two different linear layers +.>
Figure SMS_25
and />
Figure SMS_29
Mapped as key vectors +.>
Figure SMS_32
Sum vector->
Figure SMS_19
Then->
Figure SMS_23
and />
Figure SMS_27
Is subjected to a dot multiplication operation and divided by the channel dimension +.>
Figure SMS_31
Then generating importance weights in the time sequence dimension by using softmax function, and using the generated importance weight pairs +.>
Figure SMS_20
Re-weighting to obtain the data of the introduced dynamic production arrangement parameters->
Figure SMS_26
Daily oil production->
Figure SMS_30
Priori influence information after a priori influence +.>
Figure SMS_33
；
The said
Figure SMS_34
And the main body of the static path is input with static parameter data +.>
Figure SMS_35
Into the static channel attention, determining superposition influence information of static parameter data on daily oil production +.>
Figure SMS_36
As shown in formula (3):
Figure SMS_37
(3)
first through a linear layer
Figure SMS_39
Will->
Figure SMS_42
Mapping as +.>
Figure SMS_46
Will->
Figure SMS_40
By two linear layers
Figure SMS_43
and />
Figure SMS_47
Respectively mapped as +.>
Figure SMS_49
and />
Figure SMS_38
Then->
Figure SMS_44
Transpose and +.>
Figure SMS_48
Proceeding pointMultiplication operation and division by channel dimension +>
Figure SMS_50
Then generating importance weights in channel dimensions using a softmax function, using the generated importance weight pairs +.>
Figure SMS_41
Re-weighting to obtain the output +.>
Figure SMS_45
；
The specific operation of the dynamic path comprises the following steps:
introducing a priori influence by the static yield interaction module as shown in formula (4), wherein formula (4) is as follows:
Figure SMS_51
(4)
will be
Figure SMS_53
By means of a linear layer->
Figure SMS_57
Mapping to->
Figure SMS_61
Static parameter data->
Figure SMS_52
By two linear layers->
Figure SMS_56
and />
Figure SMS_59
Respectively mapped as +.>
Figure SMS_62
and />
Figure SMS_54
Then->
Figure SMS_60
and />
Figure SMS_64
Is subjected to a dot multiplication operation and divided by the channel dimension +.>
Figure SMS_66
Then generating importance weights in the time sequence dimension by using softmax function, and using the generated importance weight pairs +.>
Figure SMS_55
Re-weighting to obtain the introduced static parameter data +.>
Figure SMS_58
Daily oil production->
Figure SMS_63
Is->
Figure SMS_65
；
Will be
Figure SMS_67
And the main body of the dynamic path inputs dynamic production schedule parameter data +.>
Figure SMS_68
Into the dynamic channel attention, determining the superposition influence information of dynamic production schedule parameter data on daily oil production +.>
Figure SMS_69
As shown in formula (5): />
Figure SMS_70
(5)
First through a linear layer
Figure SMS_73
Will->
Figure SMS_75
Mapping as +.>
Figure SMS_78
Will->
Figure SMS_74
By two linear layers
Figure SMS_76
and />
Figure SMS_79
Respectively mapped as +.>
Figure SMS_81
and />
Figure SMS_72
Then->
Figure SMS_77
Transpose and +.>
Figure SMS_80
Performing a dot multiplication operation and dividing by the channel dimension +.>
Figure SMS_82
Then generating importance weights in channel dimensions by using softmax function, re-weighting by using the generated importance weights to obtain output of static channel->
Figure SMS_71
。
Optionally, said bringing said
Figure SMS_83
、/>
Figure SMS_84
、/>
Figure SMS_85
、/>
Figure SMS_86
And daily oil production data are input into the gating fusion module to obtain complementary fusion information +.>
Figure SMS_87
Specifically, as shown in formula (6):
Figure SMS_88
(6)
prior influence information of static path
Figure SMS_89
And a priori influence information of dynamic pathways +.>
Figure SMS_90
Daily oil production data +.>
Figure SMS_91
Stacked together, fed into a linear layer fusion for dimension reduction and using +.>
Figure SMS_92
Activating function to generate gating information->
Figure SMS_93
；
Utilizing the gating information
Figure SMS_94
Controlling the information retention degree after the superposition of the static information path, the operations comprise:
output of static path
Figure SMS_95
Through a linear layer->
Figure SMS_96
And gating information->
Figure SMS_97
By doing Hadamard product, the gating information element tends to be 1 with corresponding value +.>
Figure SMS_98
The information of (2) is preserved and the information corresponding to the value tending to 0 is discarded;
output of dynamic path
Figure SMS_99
Through a linear layer->
Figure SMS_100
and />
Figure SMS_101
By doing Hadamard product, the gating information element tends to be 1 with corresponding value +.>
Figure SMS_102
The information of (2) is preserved and the information corresponding to the value tending to 0 is discarded;
for static channel information after gate screening
Figure SMS_103
And dynamic path information->
Figure SMS_104
Performing element addition operation to obtain complementary fusion information after dynamic and static complementary cross fusion>
Figure SMS_105
。
Optionally, the sequence-to-sequence model main body architecture is composed of a time sequence feature extractor and an iteration predictor, both of which adopt a long-short-period memory network LSTM, the time sequence feature extractor uses the static embedded information extracted by the static embedded initial bias module as an initial state bias, uses the complementary fusion information output by the dynamic-static complementary cross fusion module and daily oil production data as input, and extracts the autocorrelation of the dynamic daily oil production data according to the characteristic that the output of the current time state of the LSTM is affected by the previous time state, as shown in formula (7):
Figure SMS_106
(7)
time input segment of the time sequence feature extractor
Figure SMS_107
Output +.>
Figure SMS_108
And historical daily oil production data->
Figure SMS_109
Stacked together as input to the timing feature extractor, the output of the static embedded initial bias module>
Figure SMS_110
and />
Figure SMS_111
As the initial state bias of the time sequence feature extractor, the time sequence feature extractor is utilized to extract the autocorrelation of the dynamic daily output data, and finally the output state of the time sequence feature extractor is obtained>
Figure SMS_112
and />
Figure SMS_113
；
The said
Figure SMS_114
and />
Figure SMS_115
Feeding into said iterative predictor, said iterative predictor being based on said +.>
Figure SMS_116
And
Figure SMS_117
dynamic iterative prediction of oil production is performed in combination with dynamic production schedule parameter data, as shown in equation (8):
Figure SMS_118
(8)
the iterative predictor inputs dynamic production schedule parameter data corresponding to the current time step at each time step
Figure SMS_119
And the daily oil production prediction value of the iteration predictor of the previous time step +.>
Figure SMS_120
And input the output hidden state of the previous time step +.>
Figure SMS_121
And cell status->
Figure SMS_122
When being the first starting state of the iterative predictor, the +.>
Figure SMS_123
For the output state of the timing feature extractor +.>
Figure SMS_124
Then based on this, the daily oil production of the current time step is predicted +.>
Figure SMS_125
；
The process described by equation (8) is repeated until the desired prediction step size is reached.
Optionally, the method further includes dividing data into training data and verification data, training the static embedded initial bias module, the dynamic-static complementary cross fusion module and the sequence-to-sequence model by using the training data, including:
dividing the input time segment and the predicted time segment corresponding to each sample, combining the normalized corresponding characteristics and labels, and sending the normalized corresponding characteristics and labels into each module according to the parameters required by each module, wherein the time segment step length input by each sample time sequence characteristic extractor is as follows
Figure SMS_126
The corresponding iterative prediction time slice step length of the iterative predictor is +.>
Figure SMS_127
；
And (3) carrying out loss calculation on the predicted tag value corresponding to the output of the sequence-to-sequence model and the training data, wherein the loss function selection RMSLE is shown in a formula (9):
Figure SMS_128
(9)
wherein ,
Figure SMS_129
sample data amount representing one batch, < ->
Figure SMS_130
Representing prediction->
Figure SMS_131
And updating the model parameters according to the gradient back-propagation values of the model parameters by a time step.
In another aspect, there is provided a shale oil well production prediction apparatus comprising:
the data collection preprocessing module is used for collecting historical production data of the shale oil well and preprocessing the data, wherein the data comprises static parameter data, dynamic production arrangement parameter data and daily oil production data;
the static embedding initial bias module is used for inputting the static parameter data into the static embedding initial bias module, extracting static embedding information as an initialization state of the sequence-to-sequence model, and providing initial state bias of the reaction oil reserve information for the sequence-to-sequence model;
the dynamic-static complementary cross fusion module is used for inputting the static parameter data, the dynamic production scheduling parameter data and the daily oil production data into the dynamic-static complementary cross fusion module, fully exploring the mutual influence of the dynamic-static parameter data and the superposition influence of the dynamic-static parameter data on the yield, and obtaining the complementary fusion information of the dynamic daily oil production scheduling parameter data and the static parameter data as the input of the sequence to the sequence model;
a sequence-to-sequence model for use in performing iterative predictions of shale well production using the sequence-to-sequence model.
In another aspect, an electronic device is provided that includes a processor and a memory having at least one instruction stored therein that is loaded and executed by the processor to implement the shale well production prediction method described above.
In another aspect, a computer readable storage medium having stored therein at least one instruction loaded and executed by a processor to implement the shale well production prediction method described above is provided.
Compared with the prior art, the technical scheme has at least the following beneficial effects:
1) The dynamic and static multiple influence factors are fully utilized, the mutual influence between the dynamic and static multiple influence factors and the superposition influence of the dynamic and static multiple influence factors on the yield are furthest excavated, and meanwhile, the initial state bias of the reaction oil reserve information is set, so that the model can adapt to complex geological engineering conditions, and the yield prediction accuracy is improved;
2) Extracting the autocorrelation time sequence dependence characteristic of daily oil yield by using a sequence-to-sequence model, and simultaneously decoupling input and output by considering the actual demand of oil yield prediction, thereby realizing the dynamic iteration prediction of the yield of any step length in the output length range set during model training;
3) The model is trained based on historical oil yield data, and is simple in training and good in generalization capability.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a shale oil well yield prediction method provided by an embodiment of the invention;
FIG. 2 is a diagram of an overall network architecture provided by an embodiment of the present invention;
FIG. 3 is a schematic diagram of a dynamic and static complementary cross fusion module according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a method for dividing input/output time slices according to an embodiment of the present invention;
FIG. 5 is a block diagram of a shale oil well production prediction apparatus provided by an embodiment of the invention;
fig. 6 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. It will be apparent that the described embodiments are some, but not all, embodiments of the invention. All other embodiments, which can be made by a person skilled in the art without creative efforts, based on the described embodiments of the present invention fall within the protection scope of the present invention.
As shown in fig. 1, an embodiment of the present invention provides a shale oil well yield prediction method, including:
s1, collecting historical production data of a shale oil well, and preprocessing the data, wherein the data comprises static parameter data, dynamic production arrangement parameter data and daily oil production data;
s2, inputting the static parameter data into a static embedding initial bias module, extracting static embedding information, and taking the static embedding information as initial state bias from a sequence to a sequence model;
s3, inputting the static parameter data, the dynamic production scheduling parameter data and the daily oil production data into a dynamic-static complementary cross fusion module to obtain complementary fusion information of the dynamic daily oil production scheduling parameter data and the static parameter data, and using the complementary fusion information as input from the sequence to a sequence model;
s4, carrying out iterative prediction on the shale oil well yield by using the sequence-to-sequence model.
The following describes in detail a shale oil well yield prediction method provided by the embodiment of the invention with reference to fig. 2 to 4, which comprises the following steps:
s1, collecting historical production data of a shale oil well, and preprocessing the data, wherein the data comprises static parameter data, dynamic production arrangement parameter data and daily oil production data;
collecting historical production data of shale oil wells, analyzing the data, screening the original data by considering the relativity, comprehensiveness and quantifiability of parameters and the yield, selecting proper static parameter data, dynamic production arrangement parameter data and daily oil production data as data required by yield prediction, taking the shale oil wells as an example, wherein the static parameter data comprises geological and fluid physical parameter data and construction parameter data, and the geological and fluid physical parameter data comprises: the average thickness of the oil layer, the viscosity of the crude oil on the ground, the density of the ground, the solidifying point and the like, and the construction parameter data comprise: the number of fracturing sections, the number of clusters, the fracturing pressure, the carbon dioxide addition, the number of carbon dioxide addition sections, the liquid consumption, the sand addition amount, the fiber and the like, and the dynamic production arrangement parameter data comprise: a nipple, oil pressure, etc.
Preprocessing data, including:
removing or interpolating the empty data to remove abnormal values and obtain effective and clean data;
dividing the data to form training data and verification data, respectively normalizing the training data and the verification data, eliminating the influence of dimension, and forming a data set
Figure SMS_132
and />
Figure SMS_136
，/>
Figure SMS_139
Wherein N is the number of samples, ">
Figure SMS_133
To influence the independent variable of the oil yield, the constitution is +.>
Figure SMS_140
，/>
Figure SMS_143
For geological and fluid physical parameter data, +.>
Figure SMS_146
Is->
Figure SMS_134
Individual geological and fluid physical parameter data, +.>
Figure SMS_137
For construction parameter data>
Figure SMS_142
Is->
Figure SMS_145
Construction parameter data->
Figure SMS_135
Scheduling parameter data for dynamic production,/->
Figure SMS_138
Is->
Figure SMS_141
Dynamic production schedule parameter data,/->
Figure SMS_144
Is daily oil production.
S2, inputting the static parameter data into a static embedding initial bias module, extracting static embedding information, and taking the static embedding information as initial state bias from a sequence to a sequence model;
optionally, as shown in fig. 2, the main body of the static embedded initial bias module is a multi-layer perceptron MLP, and adopts a tanh activation function, and the input of the static embedded initial bias module is the static parameter data and the mining days
Figure SMS_147
The static parameter data includes: geological and fluid physical parameter data->
Figure SMS_148
Construction parameter data->
Figure SMS_149
The output is expressed as +.>
Figure SMS_150
Subsequently->
Figure SMS_151
Hidden state of the sequential feature extractor as the sequence-to-sequence model, +.>
Figure SMS_152
The cell state as the time sequence feature extractor is fed into the model as an initial state bias, and a specific formula is shown in (1):
Figure SMS_153
(1)。
considering that the daily recoverable oil quantity, namely daily oil production, is based on oil reserves of an oil field, the oil reserves and static parameters are in a nonlinear relation, and the oil reserves are in a decreasing rule along with the increase of the exploitation time, a static embedded initial bias module is arranged, the main body of the module is a multi-layer perceptron, and a tanh activation function is adopted, so that the initialization state and the static parameters obtained by a sequential characteristic extractor of a sequence-to-sequence model are ensured to be in a nonlinear relation.
Compared with the prior art, is simpleHiding state of sequential feature extractor
Figure SMS_154
And cell status->
Figure SMS_155
Zero initialization mode, the embodiment of the invention obtains +.>
Figure SMS_156
By the method, the oil reserve information bias with the change of the production days can be provided for the main body framework of the sequence-to-sequence model, the change trend of the oil yield can be better followed on the basis, and the accuracy of yield prediction is improved.
S3, inputting the static parameter data, the dynamic production scheduling parameter data and the daily oil production data into a dynamic-static complementary cross fusion module to obtain complementary fusion information of the dynamic daily oil production scheduling parameter data and the static parameter data, and using the complementary fusion information as input from the sequence to a sequence model;
optionally, as shown in fig. 3, the dynamic and static complementary cross fusion module includes: (a) A cross attention module and (b) a gate control fusion module;
the cross attention module is divided into a static passage and a dynamic passage, wherein the static passage firstly determines priori influence information of the dynamic production arrangement parameter data on daily oil production through the dynamic yield interaction module
Figure SMS_157
Determining superposition influence information of the static parameter data on daily oil production by means of static channel attention>
Figure SMS_158
As an output of the static path; the dynamic path firstly determines priori influence information of the static parameter data on daily oil production through a static yield interaction module>
Figure SMS_159
Attention determination through dynamic channelSpecifying information about the superimposed influence of the dynamic production schedule parameter data on daily oil production>
Figure SMS_160
As an output of the dynamic path;
the said
Figure SMS_161
、/>
Figure SMS_162
、/>
Figure SMS_163
、/>
Figure SMS_164
And daily oil production data are input into the gating fusion module to obtain complementary fusion information +.>
Figure SMS_165
Said->
Figure SMS_166
Together with daily oil production data as input to a timing feature extractor of the sequence-to-sequence model.
Alternatively, the goal of the static path is to explore the data of the production schedule parameters in the dynamic state
Figure SMS_167
Daily oil production->
Figure SMS_168
Static data under a priori influence of +.>
Figure SMS_169
Daily oil production->
Figure SMS_170
Of (2), wherein->
Figure SMS_171
Representing stacking two matrix data along the attribute dimension (cocat) together, the specific operations include:
introducing a priori influence by the dynamic yield interaction module as shown in a formula (2), wherein the formula (2) is as follows:
Figure SMS_172
(2)
daily oil production data
Figure SMS_181
By means of a linear layer->
Figure SMS_175
Mapping to query vector +.>
Figure SMS_177
(Query) dynamic production scheduling parameter data +.>
Figure SMS_185
By two different linear layers +.>
Figure SMS_189
and />
Figure SMS_186
Mapped as key vectors +.>
Figure SMS_190
(Key) and value vector->
Figure SMS_184
(Value), then pair->
Figure SMS_188
and />
Figure SMS_174
Is subjected to a dot multiplication operation and divided by the channel dimension +.>
Figure SMS_180
Is prepared from (function: p->
Figure SMS_176
and />
Figure SMS_179
Scaling the dot product of (i) to avoid oversized result, entering into saturation region of softmax function), and generating importance weight (measuring dynamic production schedule parameter data by weight) in time sequence dimension by using softmax function>
Figure SMS_183
Degree of influence on daily oil production), using the generated importance weight pair +.>
Figure SMS_187
Re-weighting to obtain the data of the introduced dynamic production arrangement parameters->
Figure SMS_173
Daily oil production->
Figure SMS_178
Priori influence information after a priori influence +.>
Figure SMS_182
(the re-weighted data will enhance the value with greater impact on daily oil production and suppress the value with less impact);
the said
Figure SMS_191
And the main body of the static path is input with static parameter data +.>
Figure SMS_192
The superposition influence of the dynamic daily output arrangement parameter data on the static parameters under the prior influence of the oil output is explored in the attention of the static channel, and superposition influence information of the static parameter data on the daily output is determined>
Figure SMS_193
As shown in formula (3):
Figure SMS_194
(3)
first through a linear layer
Figure SMS_196
Will->
Figure SMS_201
Mapping as +.>
Figure SMS_204
Will->
Figure SMS_198
By two linear layers
Figure SMS_202
and />
Figure SMS_205
Respectively mapped as +.>
Figure SMS_207
and />
Figure SMS_195
Then->
Figure SMS_199
Transpose and +.>
Figure SMS_203
Performing a dot multiplication operation and dividing by the channel dimension +.>
Figure SMS_206
Then generating importance weights in channel dimensions using a softmax function, using the generated importance weight pairs +.>
Figure SMS_197
Re-weighting (attribute dimension with larger influence on daily oil production superposition and attribute dimension with smaller inhibition influence in static parameter data can be enhanced) to obtain output of a static passage>
Figure SMS_200
；
The basic operation of the dynamic path is similar to that of the static path, and the specific operation of the dynamic path comprises the following steps:
introducing a priori influence by the static yield interaction module as shown in formula (4), wherein formula (4) is as follows:
Figure SMS_208
(4)
will be
Figure SMS_210
By means of a linear layer->
Figure SMS_214
Mapping to->
Figure SMS_218
Static parameter data->
Figure SMS_212
By two linear layers->
Figure SMS_216
and />
Figure SMS_220
Respectively mapped as +.>
Figure SMS_223
and />
Figure SMS_209
Then->
Figure SMS_213
and />
Figure SMS_217
Is subjected to a dot multiplication operation and divided by the channel dimension +.>
Figure SMS_221
Then generating importance weights in the time sequence dimension by using softmax function, and using the generated importance weight pairs +.>
Figure SMS_211
Re-weighting to obtain the introduced static parameter data +.>
Figure SMS_215
Daily oil production->
Figure SMS_219
Is->
Figure SMS_222
；
Will be
Figure SMS_224
And the main body of the dynamic path inputs dynamic production schedule parameter data +.>
Figure SMS_225
Into the dynamic channel attention, determining the superposition influence information of dynamic production schedule parameter data on daily oil production +.>
Figure SMS_226
As shown in formula (5): />
Figure SMS_227
(5)
First through a linear layer
Figure SMS_230
Will->
Figure SMS_233
Mapping as +.>
Figure SMS_236
Will->
Figure SMS_231
By two linear layers
Figure SMS_234
and />
Figure SMS_237
Respectively mapped as +.>
Figure SMS_239
and />
Figure SMS_228
Then->
Figure SMS_232
Transpose and +.>
Figure SMS_235
Performing a dot multiplication operation and dividing by the channel dimension +.>
Figure SMS_238
Then generating importance weights in channel dimensions by using softmax function, re-weighting by using the generated importance weights to obtain output of static channel->
Figure SMS_229
。
Optionally, in the embodiment of the invention, superposition influence information that two paths are prior to each other is obtained
Figure SMS_240
and />
Figure SMS_241
Afterwards, the two are not directly fused but a gating fusion module is provided, said +.>
Figure SMS_242
、/>
Figure SMS_243
、/>
Figure SMS_244
、/>
Figure SMS_245
And daily oil production data are input into the gating fusion module to obtain complementary fusion information +.>
Figure SMS_246
Specifically, as shown in formula (6):
Figure SMS_247
(6)
prior influence information of static path
Figure SMS_248
And a priori influence information of dynamic pathways +.>
Figure SMS_249
Daily oil production data +.>
Figure SMS_250
Stacking (concat) together, feeding into linear layer fusion for dimension reduction, and using +.>
Figure SMS_251
Activating function to generate gating information->
Figure SMS_252
（/>
Figure SMS_253
The value of (2) is within +.>
Figure SMS_254
And most of the time tends to be 0 or 1), so embodiments of the present invention use it to generate gating information);
utilizing the gating information
Figure SMS_255
Controlling the information retention degree after the superposition of the static information path, the operations comprise:
output of static path
Figure SMS_256
Through a linear layer->
Figure SMS_257
And gating information->
Figure SMS_258
Hadamard product (Hadamard product) by which gating information elements tend to correspond to values of 1>
Figure SMS_259
The information of (2) is preserved and the information corresponding to the value tending to 0 is discarded;
output of dynamic path
Figure SMS_260
Through a linear layer->
Figure SMS_261
and />
Figure SMS_262
By doing Hadamard product, the gating information element tends to be 1 with corresponding value +.>
Figure SMS_263
The information of (2) is preserved and the information corresponding to the value tending to 0 is discarded;
for static channel information after gate screening
Figure SMS_264
And dynamic path information->
Figure SMS_265
Performing element addition operation to obtain complementary fusion information after dynamic and static complementary cross fusion>
Figure SMS_266
。
The embodiment of the invention can avoid information redundancy and transmission of error information through the gating fusion module, and ensure the effectiveness of information transmission.
S4, carrying out iterative prediction on the shale oil well yield by using the sequence-to-sequence model.
Optionally, as shown in fig. 2, the sequence-to-sequence model main body architecture is composed of a time sequence feature extractor and an iteration predictor, both of which adopt a long-short-period memory network LSTM, the time sequence feature extractor uses the static embedded information extracted by the static embedded initial bias module as an initial state bias, uses the complementary fusion information output by the dynamic-static complementary cross fusion module and daily oil production data as inputs, and extracts the autocorrelation of the dynamic daily oil production data according to the characteristic that the output of the current time state of the LSTM is affected by the previous time state, as shown in a formula (7):
Figure SMS_267
(7)
time input segment of the time sequence feature extractor
Figure SMS_268
Output +.>
Figure SMS_269
And historical daily oil production data->
Figure SMS_270
Stacked together as input to the timing feature extractor, the output of the static embedded initial bias module>
Figure SMS_271
and />
Figure SMS_272
As the initial state bias of the time sequence feature extractor, the time sequence feature extractor is utilized to extract the autocorrelation of the dynamic daily output data, and finally the output state of the time sequence feature extractor is obtained>
Figure SMS_273
and />
Figure SMS_274
；
This output state
Figure SMS_275
and />
Figure SMS_276
Based on the initial state bias, the superposition influence of the dynamic state and the static state is fused, and the autocorrelation of daily oil yield of the input time segment is extracted.
The said
Figure SMS_277
and />
Figure SMS_278
Feeding into said iterative predictor, said iterative predictor being based on said +.>
Figure SMS_279
And
Figure SMS_280
dynamic iterative prediction of oil production is performed in combination with dynamic production schedule parameter data, as shown in equation (8):
Figure SMS_281
(8)
the iterative predictor inputs dynamic production schedule parameter data corresponding to the current time step at each time step
Figure SMS_282
And the daily oil production prediction value of the iteration predictor of the previous time step +.>
Figure SMS_283
And input the output hidden state of the previous time step +.>
Figure SMS_284
And cell status->
Figure SMS_285
When being the first starting state of the iterative predictor, the +.>
Figure SMS_286
For the output state of the timing feature extractor +.>
Figure SMS_287
Then based on this, the daily oil production of the current time step is predicted +.>
Figure SMS_288
；
The process described by equation (8) is repeated until the desired prediction step size is reached.
Equation (8) describes an iterative prediction process for only one time step, and the process described by equation (8) is repeated throughout the prediction process until the desired prediction step is reached
Figure SMS_289
. Through a sequence-to-sequence architecture, the input time segment and the predicted time segment are decoupled, so that the autocorrelation of daily output data can be extracted, and the output length ++ ++set during model training can be realized>
Figure SMS_290
Dynamic yield iterative prediction of arbitrary step sizes in a range.
Optionally, the method further includes dividing data into training data and verification data, training the static embedded initial bias module, the dynamic-static complementary cross fusion module and the sequence-to-sequence model by using the training data, including:
as shown in FIG. 4, the input time segment and the predicted time segment corresponding to each sample are divided, and normalized corresponding features and labels are combined and sent into each module according to the parameters required by each module, wherein each sample is used forThe time segment step size input by the sequence feature extractor is
Figure SMS_291
The corresponding iterative prediction time slice step length of the iterative predictor is +.>
Figure SMS_292
；
And (3) carrying out loss calculation on the predicted tag value corresponding to the output of the sequence-to-sequence model and the training data, wherein the loss function selection RMSLE is shown in a formula (9):
Figure SMS_293
(9)
wherein ,
Figure SMS_294
sample data amount representing one batch, < ->
Figure SMS_295
Representing prediction->
Figure SMS_296
And updating the model parameters according to the gradient back-propagation values of the model parameters by a time step.
The visual observation shows that the input and output of the model are in non-one-to-one correspondence, and after training, the model can realize
Figure SMS_297
And outputting any step length in the range.
According to the embodiment of the invention, the trained model is verified by using verification data, the generalization and the accuracy are weighed, and the optimal model is reserved. Based on the reserved optimal model, efficient and accurate time sequence dynamic yield iterative prediction of shale oil well yield is realized.
As shown in fig. 5, the embodiment of the invention further provides a shale oil well yield prediction device, which comprises:
the data collection preprocessing module 510 is used for collecting historical production data of the shale oil well, preprocessing the data, wherein the data comprises static parameter data, dynamic production scheduling parameter data and daily oil production data;
the static embedding initial bias module 520 is configured to input the static parameter data into the static embedding initial bias module, extract static embedding information as an initialization state of a sequence-to-sequence model, and provide an initial state bias of reaction oil reserve information for the sequence-to-sequence model;
the dynamic-static complementary cross fusion module 530 is configured to input the static parameter data, the dynamic production scheduling parameter data and the daily oil production data into the dynamic-static complementary cross fusion module, fully explore the mutual influence of the dynamic-static parameter data and the superposition influence of the dynamic-static parameter data on the yield, and obtain complementary fusion information of the dynamic daily oil production scheduling parameter data and the static parameter data as input from the sequence to the sequence model;
a sequence-to-sequence model 540 for iterative prediction of shale well production using the sequence-to-sequence model.
The functional structure of the shale oil well yield prediction device provided by the embodiment of the invention corresponds to the shale oil well yield prediction method provided by the embodiment of the invention, and is not repeated here.
Fig. 6 is a schematic structural diagram of an electronic device 600 according to an embodiment of the present invention, where the electronic device 600 may have a relatively large difference due to different configurations or performances, and may include one or more processors (central processing units, CPU) 601 and one or more memories 602, where at least one instruction is stored in the memories 602, and the at least one instruction is loaded and executed by the processors 601 to implement the steps of the shale oil well yield prediction method described above.
In an exemplary embodiment, a computer readable storage medium, such as a memory including instructions executable by a processor in a terminal to perform the shale well production prediction method described above, is also provided. For example, the computer readable storage medium may be ROM, random Access Memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.
It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program for instructing relevant hardware, where the program may be stored in a computer readable storage medium, and the storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.
The foregoing description of the preferred embodiments of the invention is not intended to limit the invention to the precise form disclosed, and any such modifications, equivalents, and alternatives falling within the spirit and scope of the invention are intended to be included within the scope of the invention.

Claims (10)

1. A shale oil well production prediction method, comprising:
collecting historical production data of shale oil wells, and preprocessing the data, wherein the data comprises static parameter data, dynamic production scheduling parameter data and daily oil production data;
inputting the static parameter data into a static embedding initial bias module, extracting static embedding information, and taking the static embedding information as initial state bias from a sequence to a sequence model;
inputting the static parameter data, the dynamic production scheduling parameter data and the daily oil production data into a dynamic-static complementary cross fusion module to obtain complementary fusion information of the dynamic daily oil production scheduling parameter data and the static parameter data, and taking the complementary fusion information as input of the sequence to a sequence model;
iterative predictions of shale well production are made using the sequence-to-sequence model.
2. The method of claim 1, wherein the body of the static embedded initial bias module is a multi-layer perceptron MLP and employs a tanh activation function, and the inputs of the static embedded initial bias module are the static parameter data and days of mining
Figure QLYQS_1
The static parameter data includes: geological and fluid matterSex parameter data->
Figure QLYQS_2
Construction parameter data->
Figure QLYQS_3
The output is expressed as +.>
Figure QLYQS_4
Subsequently->
Figure QLYQS_5
Hidden state of the sequential feature extractor as the sequence-to-sequence model, +.>
Figure QLYQS_6
The cell state as the time sequence feature extractor is fed into the model as an initial state bias, and a specific formula is shown in (1):
Figure QLYQS_7
(1)。
3. the method of claim 1, wherein the dynamic and static complementary cross-fusion module comprises: two major parts, namely a cross attention module and a gate control fusion module;
the cross attention module is divided into a static passage and a dynamic passage, wherein the static passage firstly determines priori influence information of the dynamic production arrangement parameter data on daily oil production through the dynamic yield interaction module
Figure QLYQS_8
Determining superposition influence information of the static parameter data on daily oil production by means of static channel attention>
Figure QLYQS_9
As an output of the static path; the dynamic path first determines the static state through a static yield interaction modulePriori influence information of parameter data on daily oil production>
Figure QLYQS_10
Determining superposition influence information of the dynamic production schedule parameter data on daily oil production through dynamic channel attention>
Figure QLYQS_11
As an output of the dynamic path;
the said
Figure QLYQS_12
、/>
Figure QLYQS_13
、/>
Figure QLYQS_14
、/>
Figure QLYQS_15
And daily oil production data are input into the gating fusion module to obtain complementary fusion information +.>
Figure QLYQS_16
Said->
Figure QLYQS_17
Together with daily oil production data as input to a timing feature extractor of the sequence-to-sequence model.
4. A method according to claim 3, wherein the specific operation of the static path comprises:
introducing a priori influence by the dynamic yield interaction module as shown in a formula (2), wherein the formula (2) is as follows:
Figure QLYQS_18
(2)
daily oil productionQuantity data
Figure QLYQS_19
By means of a linear layer->
Figure QLYQS_23
Mapping to query vector +.>
Figure QLYQS_27
Dynamic production of scheduling parameter data->
Figure QLYQS_22
By two different linear layers +.>
Figure QLYQS_25
and />
Figure QLYQS_29
Mapped as key vectors +.>
Figure QLYQS_32
Sum vector->
Figure QLYQS_20
Then to
Figure QLYQS_26
and />
Figure QLYQS_30
Is subjected to a dot multiplication operation and divided by the channel dimension +.>
Figure QLYQS_33
Then generating importance weights in the time sequence dimension by using softmax function, and using the generated importance weight pairs +.>
Figure QLYQS_21
Re-weighting to obtain the data of the introduced dynamic production arrangement parameters->
Figure QLYQS_24
Daily oil production->
Figure QLYQS_28
Priori influence information after a priori influence +.>
Figure QLYQS_31
；
The said
Figure QLYQS_34
And the main body of the static path is input with static parameter data +.>
Figure QLYQS_35
Into the static channel attention, determining superposition influence information of static parameter data on daily oil production +.>
Figure QLYQS_36
As shown in formula (3):
Figure QLYQS_37
(3)
first through a linear layer
Figure QLYQS_39
Will->
Figure QLYQS_44
Mapping as +.>
Figure QLYQS_48
Will->
Figure QLYQS_40
By two linear layers->
Figure QLYQS_42
and />
Figure QLYQS_46
Respectively mapped as +.>
Figure QLYQS_49
and />
Figure QLYQS_38
Then->
Figure QLYQS_43
Transpose and +.>
Figure QLYQS_47
Performing a dot multiplication operation and dividing by the channel dimension +.>
Figure QLYQS_50
Then generating importance weights in channel dimensions using a softmax function, using the generated importance weight pairs +.>
Figure QLYQS_41
Re-weighting to obtain the output +.>
Figure QLYQS_45
；
The specific operation of the dynamic path comprises the following steps:
introducing a priori influence by the static yield interaction module as shown in formula (4), wherein formula (4) is as follows:
Figure QLYQS_51
(4)
will be
Figure QLYQS_54
By means of a linear layer->
Figure QLYQS_57
Mapping to->
Figure QLYQS_61
Static parameter data->
Figure QLYQS_53
By two linear layers->
Figure QLYQS_56
And
Figure QLYQS_60
respectively mapped as +.>
Figure QLYQS_64
and />
Figure QLYQS_52
Then->
Figure QLYQS_58
and />
Figure QLYQS_62
Is subjected to a dot multiplication operation and divided by the channel dimension +.>
Figure QLYQS_65
Then generating importance weights in the time sequence dimension by using softmax function, and using the generated importance weight pairs +.>
Figure QLYQS_55
Re-weighting to obtain the introduced static parameter data +.>
Figure QLYQS_59
Daily oil production->
Figure QLYQS_63
Is->
Figure QLYQS_66
；
Will be
Figure QLYQS_67
And the main body of the dynamic path inputs dynamic production schedule parameter data +.>
Figure QLYQS_68
Into the dynamic channel attention, determining the superposition influence information of dynamic production schedule parameter data on daily oil production +.>
Figure QLYQS_69
As shown in formula (5): />
Figure QLYQS_70
(5)
First through a linear layer
Figure QLYQS_72
Will->
Figure QLYQS_75
Mapping as +.>
Figure QLYQS_78
Will->
Figure QLYQS_74
By two linear layers->
Figure QLYQS_76
And
Figure QLYQS_79
respectively mapped as +.>
Figure QLYQS_81
and />
Figure QLYQS_71
Then->
Figure QLYQS_77
Transpose and +.>
Figure QLYQS_80
Performing a dot multiplication operation and dividing by the channel dimension +.>
Figure QLYQS_82
Then generating importance weights in channel dimensions by using softmax function, re-weighting by using the generated importance weights to obtain output of static channel->
Figure QLYQS_73
。
5. A method according to claim 3, wherein said bringing said into contact with said substrate is performed by
Figure QLYQS_83
、/>
Figure QLYQS_84
、/>
Figure QLYQS_85
、/>
Figure QLYQS_86
And daily oil production data are input into the gating fusion module to obtain complementary fusion information +.>
Figure QLYQS_87
Specifically, as shown in formula (6):
Figure QLYQS_88
(6)
prior influence information of static path
Figure QLYQS_89
And a priori influence information of dynamic pathways +.>
Figure QLYQS_90
Daily oil production data +.>
Figure QLYQS_91
Stacked together, fed into a linear layer fusion for dimension reduction and using +.>
Figure QLYQS_92
Activating function to generate gating information->
Figure QLYQS_93
；
Utilizing the gating information
Figure QLYQS_94
Controlling the information retention degree after the superposition of the static information path, the operations comprise:
output of static path
Figure QLYQS_95
Through a linear layer->
Figure QLYQS_96
And gating information->
Figure QLYQS_97
By doing Hadamard product, the gating information element tends to be 1 with corresponding value +.>
Figure QLYQS_98
The information of (2) is preserved and the information corresponding to the value tending to 0 is discarded;
output of dynamic path
Figure QLYQS_99
Through a linear layer->
Figure QLYQS_100
and />
Figure QLYQS_101
By doing Hadamard product, the gating information element tends to be 1 with corresponding value +.>
Figure QLYQS_102
The information of (2) is preserved and the information corresponding to the value tending to 0 is discarded;
for static channel information after gate screening
Figure QLYQS_103
And dynamic path information->
Figure QLYQS_104
Performing element addition operation to obtain complementary fusion information after dynamic and static complementary cross fusion>
Figure QLYQS_105
。
6. The method of claim 1, wherein the sequence-to-sequence model body architecture is composed of a timing feature extractor and an iteration predictor, both of which adopt a long-short-period memory network LSTM, the timing feature extractor takes the static embedded information extracted by the static embedded initial bias module as an initial state bias, takes the complementary fusion information output by the dynamic-static complementary cross fusion module and daily oil production data as inputs, and extracts autocorrelation of dynamic daily oil production data according to the characteristic that the output of the current time state of the LSTM is affected by the previous time state, as shown in formula (7):
Figure QLYQS_106
(7)/>
extracting the time sequence characteristics from the time sequence characteristicsTime input segment
Figure QLYQS_107
Output +.>
Figure QLYQS_108
And historical daily oil production data->
Figure QLYQS_109
Stacked together as input to the timing feature extractor, the output of the static embedded initial bias module>
Figure QLYQS_110
and />
Figure QLYQS_111
As the initial state bias of the time sequence feature extractor, the time sequence feature extractor is utilized to extract the autocorrelation of the dynamic daily output data, and finally the output state of the time sequence feature extractor is obtained>
Figure QLYQS_112
and />
Figure QLYQS_113
；
The said
Figure QLYQS_114
and />
Figure QLYQS_115
Feeding into said iterative predictor, said iterative predictor being based on said +.>
Figure QLYQS_116
and />
Figure QLYQS_117
Incorporating dynamic production scheduling parametersThe data were subjected to dynamic iterative prediction of oil production as shown in equation (8):
Figure QLYQS_118
(8)
the iterative predictor inputs dynamic production schedule parameter data corresponding to the current time step at each time step
Figure QLYQS_119
And the daily oil production prediction value of the iteration predictor of the previous time step +.>
Figure QLYQS_120
And input the output hidden state of the previous time step
Figure QLYQS_121
And cell status->
Figure QLYQS_122
When being the first starting state of the iterative predictor, the +.>
Figure QLYQS_123
For the output state of the timing feature extractor +.>
Figure QLYQS_124
Then based on this, the daily oil production of the current time step is predicted +.>
Figure QLYQS_125
；
The process described by equation (8) is repeated until the desired prediction step size is reached.
7. The method of claim 6, further comprising separating data into training data and validation data, training the static embedded initial bias module, the dynamic-static complementary cross-fusion module, and the sequence-to-sequence model using the training data, comprising:
dividing the input time segment and the predicted time segment corresponding to each sample, combining the normalized corresponding characteristics and labels, and sending the normalized corresponding characteristics and labels into each module according to the parameters required by each module, wherein the time segment step length input by each sample time sequence characteristic extractor is as follows
Figure QLYQS_126
The corresponding iterative prediction time slice step length of the iterative predictor is +.>
Figure QLYQS_127
The method comprises the steps of carrying out a first treatment on the surface of the And (3) carrying out loss calculation on the predicted tag value corresponding to the output of the sequence-to-sequence model and the training data, wherein the loss function selection RMSLE is shown in a formula (9):
Figure QLYQS_128
(9)
wherein ,
Figure QLYQS_129
sample data amount representing one batch, < ->
Figure QLYQS_130
Representing prediction->
Figure QLYQS_131
And updating the model parameters according to the gradient back-propagation values of the model parameters by a time step.
8. A shale oil well production prediction apparatus, comprising:
the data collection preprocessing module is used for collecting historical production data of the shale oil well and preprocessing the data, wherein the data comprises static parameter data, dynamic production arrangement parameter data and daily oil production data;
the static embedding initial bias module is used for inputting the static parameter data into the static embedding initial bias module, extracting static embedding information as an initialization state of the sequence-to-sequence model, and providing initial state bias of the reaction oil reserve information for the sequence-to-sequence model;
the dynamic-static complementary cross fusion module is used for inputting the static parameter data, the dynamic production scheduling parameter data and the daily oil production data into the dynamic-static complementary cross fusion module, fully exploring the mutual influence of the dynamic-static parameter data and the superposition influence of the dynamic-static parameter data on the yield, and obtaining the complementary fusion information of the dynamic daily oil production scheduling parameter data and the static parameter data as the input of the sequence to the sequence model;
a sequence-to-sequence model for use in performing iterative predictions of shale well production using the sequence-to-sequence model.
9. An electronic device comprising a processor and a memory having at least one instruction stored therein, wherein the at least one instruction is loaded and executed by the processor to implement the shale well production prediction method of any of claims 1-7.
10. A computer readable storage medium having stored therein at least one instruction, wherein the at least one instruction is loaded and executed by a processor to implement the shale well production prediction method of any of claims 1-7.
CN202310350371.1A 2023-04-04 2023-04-04 Shale oil well yield prediction method and device Active CN116151480B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310350371.1A CN116151480B (en) 2023-04-04 2023-04-04 Shale oil well yield prediction method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310350371.1A CN116151480B (en) 2023-04-04 2023-04-04 Shale oil well yield prediction method and device

Publications (2)

Publication Number Publication Date
CN116151480A true CN116151480A (en) 2023-05-23
CN116151480B CN116151480B (en) 2023-07-18

Family

ID=86340966

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310350371.1A Active CN116151480B (en) 2023-04-04 2023-04-04 Shale oil well yield prediction method and device

Country Status (1)

Country Link
CN (1) CN116151480B (en)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140116776A1 (en) * 2012-10-31 2014-05-01 Resource Energy Solutions Inc. Methods and systems for improved drilling operations using real-time and historical drilling data
CN113236228A (en) * 2021-06-24 2021-08-10 中海石油(中国)有限公司 Single-well yield rapid prediction method and system
CN113722997A (en) * 2021-09-01 2021-11-30 北京中地金石科技有限公司 New well dynamic yield prediction method based on static oil and gas field data
CN113962148A (en) * 2021-10-20 2022-01-21 中国石油大学(北京) Yield prediction method, device and equipment based on convolutional coding dynamic sequence network
CN113988479A (en) * 2021-12-07 2022-01-28 扬州江苏油田瑞达石油工程技术开发有限公司 Pumping well multi-well dynamic liquid level depth prediction method based on dynamic and static information feature fusion neural network
CN114021922A (en) * 2021-10-27 2022-02-08 中海石油(中国)有限公司 Oil well productivity main control factor analysis method, system, equipment and storage medium
CN114925623A (en) * 2022-07-22 2022-08-19 中国地质大学（北京） Oil and gas reservoir yield prediction method and system

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140116776A1 (en) * 2012-10-31 2014-05-01 Resource Energy Solutions Inc. Methods and systems for improved drilling operations using real-time and historical drilling data
CN113236228A (en) * 2021-06-24 2021-08-10 中海石油(中国)有限公司 Single-well yield rapid prediction method and system
CN113722997A (en) * 2021-09-01 2021-11-30 北京中地金石科技有限公司 New well dynamic yield prediction method based on static oil and gas field data
CN113962148A (en) * 2021-10-20 2022-01-21 中国石油大学(北京) Yield prediction method, device and equipment based on convolutional coding dynamic sequence network
CN114021922A (en) * 2021-10-27 2022-02-08 中海石油(中国)有限公司 Oil well productivity main control factor analysis method, system, equipment and storage medium
CN113988479A (en) * 2021-12-07 2022-01-28 扬州江苏油田瑞达石油工程技术开发有限公司 Pumping well multi-well dynamic liquid level depth prediction method based on dynamic and static information feature fusion neural network
CN114925623A (en) * 2022-07-22 2022-08-19 中国地质大学（北京） Oil and gas reservoir yield prediction method and system

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
张进双,余志清: "静态式可控偏心器旋转导向钻井工具造斜能力分析", 石油钻探技术, no. 05 *

Also Published As

Publication number Publication date
CN116151480B (en) 2023-07-18

Similar Documents

Publication Publication Date Title
Tang et al. A deep-learning-based surrogate model for data assimilation in dynamic subsurface flow problems
Anand et al. Fractional-Iterative BiLSTM Classifier: A Novel Approach to Predicting Student Attrition in Digital Academia
Bao et al. Data-driven end-to-end production prediction of oil reservoirs by enkf-enhanced recurrent neural networks
Zhang et al. Fourier neural operator for solving subsurface oil/water two-phase flow partial differential equation
He Reduced-order modeling for oil-water and compositional systems, with application to data assimilation and production optimization
Yuyang et al. Shale gas well flowback rate prediction for Weiyuan field based on a deep learning algorithm
Calvette et al. Forecasting smart well production via deep learning and data driven optimization
CN111832227A (en) Shale gas saturation determination method, device and equipment based on deep learning
Jiang et al. Use of multifidelity training data and transfer learning for efficient construction of subsurface flow surrogate models
CN114693005A (en) Three-dimensional underground oil reservoir dynamic prediction method based on convolution Fourier neural network
CN111048163A (en) Shale oil hydrocarbon retention amount (S1) evaluation method based on high-order neural network
WO2020219057A1 (en) Systems and methods for determining grid cell count for reservoir simulation
CN115204064A (en) Gas injection huff-puff parameter optimization method and system in shale oil exploitation
Omosebi et al. Development of lean, efficient, and fast physics-framed deep-learning-based proxy models for subsurface carbon storage
Xie et al. Intelligent modeling with physics-informed machine learning for petroleum engineering problems.
Liu et al. Predicting gas flow rate in fractured shale reservoirs using discrete fracture model and GA-BP neural network method
Yang et al. Using one-dimensional convolutional neural networks and data augmentation to predict thermal production in geothermal fields
Yang et al. A steam injection distribution optimization method for SAGD oil field using LSTM and dynamic programming
Maucec et al. GeoDIN-Geoscience-Based Deep Interaction Networks for Predicting Flow Dynamics in Reservoir Simulation Models
Huang et al. A deep-learning-based graph neural network-long-short-term memory model for reservoir simulation and optimization with varying well controls
CN116151480B (en) Shale oil well yield prediction method and device
Shan et al. Physics-informed machine learning for solving partial differential equations in porous media.
Lee et al. Model Regeneration Scheme Using a Deep Learning Algorithm for Reliable Uncertainty Quantification of Channel Reservoirs
Alpak et al. Augmenting Deep Residual Surrogates with Fourier Neural Operators for Rapid Two-Phase Flow and Transport Simulations
Xiao et al. Robust optimization of geoenergy production using data-driven deep recurrent auto-encoder and fully-connected neural network proxy

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant