WO2019190003A1 - Method for optimizing locations of multiple wellbores in oil and gas reservoir by using artificial neural network - Google Patents

Abstract

The present invention relates to a method for determining optimal locations of multiple wellbores in an oil and gas reservoir. The present invention proposes a selection method using an artificial neural network which replaces an existing computational model. Further, there are a practical limit and increasing uncertainty in directly applying an artificial neural network to a case in which the amount of data is very large. In this regard, the present invention proposes a method in which an artificial neural network can be applied sequentially and by stages. Therefore, it is expected that optimal locations of multiple wellbores can be effectively and reliably selected using the method even when the amount of data is very large.

Description

Optimization of Multiple Position Estimation Using Artificial Neural Networks in Oil Gas Reservoir

TECHNICAL FIELD The present invention relates to techniques in the field of energy and resources, and in particular, to a method of optimizing a plurality of drilling locations by modeling in an oil and gas reservoir.

In order to produce oil and gas, so-called 'production modeling' is performed. In other words, the simulation is conducted using a computational model for oil fields or gas fields. For example, the gas field is partitioned into a three-dimensional grid to form a three-dimensional grid having a plurality of cells. The gas content is set in each cell, and the amount of oil and gas discharged as a function is inputted as the pressure in the cell is reduced. More specifically, in the case of the CBM gas field, when the production well is constructed, the pressure of the cell is reduced while methane gas is desorbed from the coal. The amount of desorption can be designated as a function of pressure. The desorbed methane gas is then transported along the crack network in the grid to the production wells. The state of the cell changes over time, resulting in dynamic modeling. The whole process of development of oil gas field is modeled using computer model.

On the other hand, research on artificial neural networks (ANN) is very active recently. Artificial neural network is composed of input layer, hidden layer, and output layer by imitating human brain, and constructs model by learning nonlinear relationship between input value and output value. Once a nonlinear relationship model between input and output values is built, the model is used to predict output values when new input values are given. Here, the input-output data used for learning are secured through the existing computer model. The neural network has the advantage of being very fast in processing the input and deriving the output. In addition, the continuous use of neural network models has the advantage of more elaborate development of nonlinear relationship models between input and output values as data is accumulated.

In order to model oil gas reservoirs, research is being conducted to use artificial neural networks. However, in the existing neural network, when the output value is found to be the maximum or minimum input value, the actual value (the value obtained by computational modeling) for the input value found due to the prediction error of the artificial neural network model is the maximum or minimum value in the domain. There was a limit not to be.

In other words, using the conventional method, it is possible to reduce the area (search space) for searching for a global solution through an artificial neural network. By applying this process repeatedly (sequential), the search space can be reduced sequentially, and finally a true global solution can be derived.

However, even with the above method, if the search space is too large, there is a limit that the possibility of dropping the global sea area increases due to uncertainty in the process of reducing the search space using an artificial neural network.

The present invention provides a modeling method that can stably derive a wide area even when the exploration space increases exponentially, and in particular, by applying the above method to the modeling of the oil gas reservoir, it is possible to optimize a plurality of drilling estimates installed in the reservoir. The purpose is to provide a method for determining location.

It is also an object of the present invention to provide a computer-readable recording medium having recorded thereon a program for executing the above method.

On the other hand, other objects not specified in the present invention will be further considered within the scope that can be easily inferred from the following detailed description and effects.

A plurality of drilling position optimization method using an artificial neural network in the oil gas reservoir according to the present invention for achieving the above object,

(a) setting an entire area of the basic grid system formed by dividing an oil or gas storage layer into a plurality of basic grids as a search space, and subdividing the search space into a grid having a size larger than the basic grid;

(b) applying an artificial neural network evaluation model to determine which one of the plurality of gratings in the search space is suitable for a plurality of grating estimates, and indicating the suitability of each grating as a rank;

(c) If the current grid size used in the evaluation step is larger than the base lattice size by a predetermined range or more, proceed to the next step to rebuild the search space, if within a certain range, the result of the evaluation step to the optimal position Determining;

(d) reconstructing the lattice recorded in a higher rank up to a predetermined rank in the goodness-of-fit of the lattice in the evaluation step into a new search space;

(e) re-segmenting the reconstructed search space into grids of size greater than or equal to the basic grid size but smaller than the grid size used in the preceding evaluation step;

The grid size in the search space is gradually reduced to become the same as the size of the basic grid within a predetermined range, so that the optimal positions of the plurality of drilling estimates are determined in step (c). It is characterized by repeatedly performing steps up to).

In the present invention, the artificial neural network model is used in selecting the optimal positions of the plurality of drilling estimates in the oil gas reservoir, and the optimal position can be selected very effectively even when the number of data is very large by applying the artificial neural network sequentially and stepwise. .

In addition, when the optimal position is selected by applying the present invention, there is an advantage that the reliability of the optimal position selection is increased as the probability of matching the result of the computational model is higher than when the artificial neural network model is directly applied at one time.

On the other hand, even if the effects are not explicitly mentioned here, it is added that the effects described in the following specification and the provisional effects expected by the technical features of the present invention are treated as described in the specification of the present invention.

1 is a schematic flowchart of a plurality of drilling position optimization method using an artificial neural network in the oil gas reservoir according to an embodiment of the present invention.

2 is an image showing that the storage layer to be the object of the present invention lattice through a conventional computer model.

3 to 7 illustrate a method according to the present invention, and FIG. 3 shows a basic grid system and a first stage grating system.

FIG. 4 shows the upper ranks evaluated through the neural network based on the first stage grid system and the area to be expanded in the second stage. FIG. 5 shows the search space and the second stage grid system determined in the second stage. It is shown.

FIG. 6 shows the upper ranks found through the artificial neural network and the area to be expanded in the second stage, and FIG. 7 shows the search space and the third stage grid system in the third stage.

8 is a flowchart of "a method for calculating a global solution using an artificial neural network".

FIG. 9 is an image of a storage layer serving as an application example of a method of calculating a global solution using the artificial neural network.

10 to 13 are graphs comparing output prediction data output through an artificial neural network model and actual output data obtained by inputting a search group into a computational model according to a method of calculating a global solution using an artificial neural network.

FIG. 14 illustrates a first stage grating system of a storage layer for explaining an embodiment of the present invention.

15 (a) to 15 (d) show a process of constructing and rebuilding an artificial neural network through the process shown in FIG. 8.

16 shows the upper ranks (hatched gratings) after selecting the optimal position in the first stage grating system.

FIG. 17 illustrates a second stage grating system, and FIG. 18 illustrates a result of obtaining a global solution for the second stage grating system.

19 shows the result of constructing the third stage grating system.

20 (a) to 20 (d) illustrate a process of selecting an optimal position in a third-stage grating system using the artificial neural network of FIG.

FIG. 21 shows the result of the optimum position found by applying the process of FIGS. 20 (a) to 20 (d) for the third stage grating system.

FIG. 22 compares the existing production profiles of six production profiles and production profiles including two additional production locations optimal positions through a computational model.

The accompanying drawings show that they are illustrated as a reference for understanding the technical idea of the present invention, by which the scope of the present invention is not limited.

A plurality of drilling position optimization method using an artificial neural network in the oil gas reservoir according to the present invention for achieving the above object,

(a) setting an entire area of the basic grid system formed by dividing an oil or gas storage layer into a plurality of basic grids as a search space, and subdividing the search space into a grid having a size larger than the basic grid;

(b) applying an artificial neural network evaluation model to determine which one of the plurality of gratings in the search space is suitable for a plurality of grating estimates, and indicating the suitability of each grating as a rank;

(c) If the current grid size used in the evaluation step is larger than the base lattice size by a predetermined range or more, proceed to the next step to rebuild the search space, if within a certain range, the result of the evaluation step to the optimal position Determining;

(d) reconstructing the lattice recorded in a higher rank up to a predetermined rank in the goodness-of-fit of the lattice in the evaluation step into a new search space;

(e) re-segmenting the reconstructed search space into grids of size greater than or equal to the basic grid size but smaller than the grid size used in the preceding evaluation step;

The grid size in the search space is gradually reduced to become the same as the size of the basic grid within a predetermined range, so that the optimal positions of the plurality of drilling estimates are determined in step (c). It is characterized by repeatedly performing steps up to).

According to the present invention, when the search space is partitioned into a grid having a size larger than that of the basic grid in the step (a) or (e), the plurality of basic grids are bundled to form a single grid.

When reconstructing the search space in step (d), the basic area occupied by the upper rank grids and the additional area extended within a predetermined range from the basic area to the surrounding area are reconstructed into the search space.

In this case, the additional region preferably extends only to a range not including the central portion of the lattice immediately adjacent to the upper rank lattice. More specifically, when determining the additional area, it is preferable to expand by the unit of the basic lattice in the grid.

Meanwhile, the evaluation method using the artificial neural network model,

An initial learning material generation step of classifying input data selected from a plurality of analysis object data and actual output data outputted by inputting the input data to a computational model constructed to analyze the analysis object data as learning data;

An initial search group generation step of classifying the analysis target data into a search group;

Construct an artificial neural network model for predicting the data output from the computational model using the training data, and determine whether to reconstruct the artificial neural network model based on the data output when the search group is input to the artificial neural network model, A model building step of reconstructing the artificial neural network model by resetting training data when it is determined that reconstruction of the artificial neural network model is necessary; And

A wide area calculation step of calculating a wide area solution when it is determined that reconstruction of the artificial neural network model is not necessary through the model building step. Here, the analysis target data is data on the position of the grid in which the plurality of drilling estimates are installed.

In an example of the present invention, the model building step may include: a first modeling step of constructing an artificial neural network model for predicting data output from the computational model using the training material; A second modeling step of obtaining output prediction data by inputting respective data included in the search group to the artificial neural network model; A third modeling step of determining whether to reconstruct the artificial neural network model based on the output prediction data; And a fourth modeling step of determining a new learning material including the learning material and a new search group when it is determined that the artificial neural network model needs to be reconstructed through the third modeling step. The first modeling step is repeated from the first modeling step using the new search group.

The third modeling step may include: a first selecting step of selecting output prediction data corresponding to a predetermined selection criterion among the output prediction data calculated through the second modeling step; A selection data extraction step of extracting input data corresponding to the output prediction data selected through the first selection step, respectively; A non-data extraction step of extracting input data not included in the learning data used for constructing the artificial neural network among the input data extracted through the selection data extraction step; And reconstructing the artificial neural network model if the number of input data extracted in the non-data extraction step is less than a predetermined number of determination criteria, and determining that the input data number selected in the non-data extraction step is preset. If the reference number or more, re-establishment determination step of determining that the neural network model needs to be rebuilt; includes.

In the third modeling step, the number of decision criteria may be equal to the number of data selected as the input data among the analysis target data in the initial learning material generation step.

In an example of the present invention, the fourth modeling step may further include: a second selecting step of selecting some of the input data extracted in the extracting data; A material addition step of generating new learning material by adding the input data selected in the second selection step to the learning material; And

And a group adding step of setting the input data selected in the extracting data not included step and the input data included in the learning material as a new search group.

In the second selection step, data from the input data extracted in the non-containing data extraction step is smaller than the number of input data selected as the learning material or the number of input data selected as the learning material in the initial learning material generation step. Select.

The data addition step may further include: a data addition acquisition step of obtaining actual output data output by inputting the input data selected in the second selection step into the computational model; And a material generation step of generating new learning material by adding the input data selected in the second selection step and the actual output data acquired in the data addition obtaining step to the learning material.

In the following description of the present invention, when it is determined that the subject matter of the present invention may be unnecessarily obscured by the person skilled in the art with respect to the related well-known functions, the detailed description will be omitted.

The present invention relates to a modeling method for simulating which position is best suited for installing a plurality of drilling wells in gas fields, oil fields and the like. Therefore, the method according to the present invention is implemented in the form of software to be executed on a computer.

In the present invention, a plurality of drilling wells may be a plurality of production wells, for example, for producing gas or oil. In addition, a plurality of injection wells may be used in the case of injecting fluid into the reservoir in order to promote the production of gas or oil. Similarly, Carbon Capture and Storage (CCS) technology for the storage of carbon dioxide can be a carbon dioxide injection well.

The present invention is mainly an oil field or a gas field, which is being developed or under development, but may be applied to a storage layer for simply storing carbon dioxide, such as in the CCS field.

Furthermore, the present invention is not necessarily limited to modeling for determining a plurality of drilling estimates, and may be used to solve various problems through modeling in a grid system composed of a plurality of grids. More specifically, it can be applied to all examples in which the grid itself serves as input data in the grid system. For example, in an example of the present invention, two grids (identification numbers for each grid) form one data set, so that each grid may be located in which grid.

Hereinafter, with reference to the accompanying drawings, it will be described in more detail with respect to a plurality of drilling position optimization methods (hereinafter, 'position optimization method') using the artificial neural network in the oil gas reservoir according to an embodiment of the present invention.

1 is a schematic flowchart of a position optimization method according to the present invention. Referring to FIG. 1, the location optimization method assumes a basic grid system.

The grid system refers to a grid system in which a certain space is divided into a plurality of grids. That is, in this example, the storage layer is composed of a plurality of lattices arranged successively up, down, left and right. Each grid may be assigned an identification number or a separate identification number according to the coordinates.

An example in which the reservoir is formed as a basic grid system using a computer model program is shown in FIG. 2. Referring to FIG. 2, it can be seen that the storage layer is formed in a curved surface and is formed of a plurality of layers to have a three-dimensional structure. Six arrows indicate the production boats currently installed.

Figure 3 is a simplified simplified view of the plane of the basic grid system in order to easily explain the present invention. That is, the plane of the actual basic lattice system is generally composed of curved surfaces as shown in FIG. 2, which is represented as a simple plane in FIG. 3, and the number of grids is greatly reduced.

The smallest grid of squares represented by thin solid lines in Fig. 3 is the basic lattice a of the basic lattice system. In the field of energy and resources, the computational model and the lattice system described above are widely used, such as the amount of residual modeling and the amount of production modeling.

In the present invention, first, the entire area (plane area) of the basic grid system is set as the search space. The search space is a target area in which a plurality of drilling estimates can be installed. That is, in the first step, the target region in which the plurality of drilling estimates can be formed is set to the widest range.

And we build a new grid system for the above search space. Since it is the first grid system newly constructed from the basic grid system, it is called a 1-stage grid system. The basic grid system is made up of a plurality of basic grids (a), but the first stage grid system differs in that the size of the grid is formed larger than the basic grid in the search space.

In FIG. 3, a thick solid line is used to form the first stage grating b. As can be seen from the figure, the first-stage grating (b) is a grid in which the basic grid (a) is grouped by 5 * 5. The grid size is 25 times larger than the basic grid. Of course, the size of the first stage grating (b) is larger than the basic grid (a) and the size is determined by the user. Preferably, the first stage grating (b) is specified in the form of a bundle of the basic grid (a). For example, 5 * 5, 6 * 6, 7 * 7 and 5 * 7 are formed by grouping the basic grid a as a unit.

When the first stage grating system is formed in this way, the number of gratings is significantly reduced compared to the basic lattice system. That is, in FIG. 3, in the basic grid system, there are 2500 basic grids (a) of 50 * 50, but the grid (b) of the first stage grid system is reduced by 25 times to 100 of 10 * 10.

The present invention seeks to find an optimal location for a plurality of drilling estimates. The position here means the grid. For example, in the first lattice system, there is a first stage grating (b) numbered from 1 to 100, which evaluates which of the two gratings is most advantageous for the production of the drilling estimate. Two grids, each of which two drilling estimates are arranged, become one data. The total number of data is ₁₀₀ C _{2, which} is 4,851. If the basic grid system is used instead of the first phase grid system, the number of data increases exponentially. In other words, if you use the basic grid (a) is ₂₅₀₀ C ₂ , which results in 3,123,750 data.

In this example, two drilling estimates are taken as an example, but the calculation of the optimal positions of three or more drilling estimates will increase the number of data exponentially.

The larger the number of data, the higher the uncertainty in the evaluation using an artificial neural network, which will be described later, which increases the probability of not finding a global maximum. In order to compensate for this, the present invention provides a step-by-step solution. First, a plurality of basic grids (a) are bundled to form a first-stage grid system to reduce the number of cases to derive a global maximum through an artificial neural network, which will be described later. After setting the maximum value and the upper ranks to the search space, the process of repeatedly searching for the solution is performed repeatedly by decreasing the size of the grid again. It demonstrates in detail below.

As shown in FIG. 3, a first-stage grid system is constructed, and a plurality of drilling estimates are evaluated through an artificial neural network model. The optimal position can be represented by a rank. For example, if you evaluate the location of the drilling estimate where the highest yield is expected, you can quantify the yield and rank the above 4,851 data points. The neural network model will be described in detail later.

Figure 4 shows the top ranks from the first to the fifth position in the result obtained through the artificial neural network model. Since the two first-stage gratings b are one data, up to the fifth position, a total of ten first-stage gratings b are designated. In FIG. 4, nine first-level gratings b are designated because one first-level grating b is overlapped in two pieces of data. For example, (numbers 14-94) and (numbers 24-94) are all ranked in the top five and contain all of the number 94 grids.

Here, the number of upper ranks may be arbitrarily determined by the user. As in the present embodiment, the present invention may be limited to the fifth position, and further extended to the lower ranks.

For reference, in FIG. 3 and FIG. 4, black indicates an area where the production well is located as a representative value of the grating.

As described above, after the upper ranks of the positions where the plurality of drilling estimates are to be selected are determined using the first-stage grid system, the search space is set again.

The search space may be designated as the first-stage gratings (b) recorded in the upper ranks in the evaluation through the previous neural network model. However, it is preferable to slightly expand the search space to the surroundings as in the present embodiment. This is because the probability of finding the maximum global solution can be further increased by extending the region around the upper rank grid.

Referring to FIG. 4, it can be seen that the search space is extended to the dotted line outside the hatched first stage grating b. However, even if the search space is extended, it is not desirable to extend to the center of the adjacent grid. For example, since the first stage grating (b) is composed of a 5 * 5 basic grid (a), even if the search space is extended to the surrounding grid, it extends only up to the basic grid (a) of two columns or two rows. As shown in FIG. 4, it can be seen that the region extended to the dotted line is extended only to the basic grid a of two columns and two rows. Of course, you can expand only one column or one row. In addition, when the search space is extended, it is extended in units of the basic grid (a).

As described above, when the search space is redefined, the hatched area and the area extended to the dotted line in FIG. 4 become a new search space. This is a much smaller area than the original search space.

For the reconstructed search space, as shown in FIG. 5, a second stage grating system is again constructed. Referring to FIG. 5, in the second stage grating system, the second grating c is formed by grouping the basic grids a by 3 * 3 units. The important point here is that as you go through the steps, you will have to gradually reduce the size of the grid. That is, in the first step, the basic grid (a) is grouped by 5 * 5, but in the second step, the grid size is reduced by grouping it by 3 * 3. In the third stage, further reduction is required, and in the final stage, the unit should be aligned with the basic grid unit (a). Narrow the search space, but seek sophistication by increasing the resolution of the narrowed area.

However, referring to FIG. 5, when the basic grid (a) is grouped by 3 * 3, there may be an insufficiency of the search space. That is, referring to FIG. 5, although the second stage grid c may be partially out of the search space or may not be partially filled, the second stage grid c may be understood as a concept of partially expanding or partially reducing the search space. In this case, however, it is preferable to include all the grid areas recorded in the upper rank.

When the search space and the grid system are constructed in the second stage, the optimal position is evaluated using the artificial neural network model as in the first stage, and the result is shown in FIG. 6 according to the number of higher ranks set in advance. It is. The seven second stage gratings c hatched in FIG. 6 are shown. Now repeat the same process again. That is, as shown in FIG. 7, the area is extended to the periphery of the grids in which the upper ranks are recorded, again defining a search space, and constructing a third-level grid system. The third stage grating used in the third stage grating system uses the basic grid (a) as it is. And we evaluate the optimal location through the neural network model. Here, the optimum positions of the plurality of drilling estimates are evaluated to determine the optimum positions.

In the above, the overall process of the present invention and the meaning of each process have been described. Since the present invention is implemented through software in a computer, the description will be made back to FIG. 1 by computer algorithm.

Referring to the algorithm of FIG. 1, the basic grid system is already constructed in the “START” state. “I = 1” means the first step. Accordingly, the first stage grid system (first grid system) is constructed. Here, the number of first stage gratings (ex: 5 * 5) is specified. The global neural network model is used to calculate the global solutions, which are represented by ranking.

After deriving the results, we compare the size of the grid with the current grid. As mentioned above, in the present invention, a plurality of basic grids are initially formed to form a lattice, and the process is repeated while gradually decreasing the size of the lattice while repeating orders (steps).

In the first step, since the grid size is larger than the basic lattice, the algorithm proceeds to the next step. In the first stage evaluation, a new search space is constructed based on the upper ranks. Then, by adding 1 to “i”, the order is increased, and after the grid system is constructed, the evaluation is performed through the artificial neural network. This process repeats cyclically.

Repeating the above process with increasing steps, as described above, comes a step where the grid of the current step is the same size as the basic grid. If the grid sizes are the same, the final solution is determined as the global solution obtained from the neural network at that stage. That is, the wide area solution obtained at this stage is determined as an optimal position where a plurality of drilling estimates are installed.

For reference, in another embodiment, the grid of the present stage is slightly larger than the basic grid, but if the size of the grid is large within a predetermined range, the cycle may be terminated and the algorithm may be terminated. That is, the present invention does not necessarily have to be forcibly repeated until the size of the base lattice and the base lattice at the present stage are the same, and the range may be further extended.

As described above, in the present invention, when the number of data is too large to derive a wide area solution in one process through an artificial neural network, it is to lower the uncertainty that a true optimal solution will not be derived. In the first step, a large grid is used to reduce the number of data, i.e. reduce the number of cases, so that the region corresponding to the optimal position is not missed. In the narrower search space, the optimal position is detected precisely while increasing the resolution, and finally, the global solution is found. This not only significantly increases the probability of finding the actual optimal solution, but also simplifies the process of finding the optimal solution, which has a big advantage at the process level.

Hereinafter, the artificial neural network model used in the present invention will be described. In the present invention, an artificial neural network model is used to evaluate and determine the optimal position of a plurality of drilling estimates. Genetic algorithms, statistical algorithms, etc. may be used as the artificial neural network model, but in this embodiment, the artificial neural network model developed by the researchers of the present invention is used. The artificial neural network model has been filed in the Republic of Korea Patent Application No. 10-2017-0017703 "method of calculating a wide area solution using the artificial neural network", the state is not disclosed at the time the present invention is filed. Hereinafter, an artificial neural network model used in the present invention will be described with reference to the accompanying drawings.

However, the present embodiment is to determine the optimum position of the plurality of drilling estimates, but for convenience of description, it will be described below by selecting an optimal position of one drilling estimate as an example.

8 is a flowchart illustrating a method for calculating a global solution using an artificial neural network according to the present invention.

Referring to the drawings, a method for calculating a global solution using an artificial neural network includes an initial learning material generation step (S100), an initial search group generation step (S200), a model building step (S300), and a wide area calculation step (S400). Include.

Initial learning material generation step (S100) is a step of generating learning material for building an artificial neural network model. First, some of the plurality of analysis target data are selected as input data. In this case, the number of input data to be selected may be determined according to the number of data to be analyzed, and it is preferable to select 10 or 20 input data among the data to be analyzed. In accordance with the present invention, the data to be analyzed means a grid. For example, when only one drilling estimate is selected by applying the above-described first-stage grid system, the number of cases, that is, the number of analysis data, becomes 100 pieces of analysis data. If there are multiple drilling estimates, the number of data is much larger than before.

Next, input data is input to a computational model constructed to analyze the data to be analyzed to obtain actual output data. Computational model is to analyze the analysis target data through computer simulation. When inputting information about drilling position, Computational model such as oil gas production simulation model according to drilling position that calculates oil gas production is calculated. Apply. Using the computational model, accurate output data can be obtained. However, when the number of data to be analyzed is large, it takes much time to calculate the computational model, and the present invention reduces the computation time of the global solution by using the learned artificial neural network model. I would like to.

As described above, the selected input data and the actual output data are classified into training materials for constructing an artificial neural network model and stored in a database (not shown).

The initial search group generation step (S200) is a step of classifying the analysis target data into search groups. In this step, all analysis data including the selected input data are classified into search groups.

In the model building step (S300), the artificial neural network model is constructed to predict the data output from the computational model using the training data, and the artificial neural network model is reconstructed based on the output data when the search group is input to the artificial neural network model. If the neural network model is determined to be reconstructed, the training data is reset to reconstruct the artificial neural network model. The model building step includes first to fourth modeling steps (S340).

The first modeling step (S310) is a step of constructing an artificial neural network model for predicting data output from a computational model using training materials. An artificial neural network model is constructed based on learning materials using an algorithm for constructing an artificial neural network model. Since the algorithm is generally used to build an artificial neural network model, a detailed description thereof will be omitted.

The second modeling step (S320) is a step of obtaining output prediction data by inputting respective data included in the search group to the artificial neural network model. The operator inputs data included in the search group into the artificial neural network model and outputs output prediction data.

The third modeling step (S330) is a step of determining whether to rebuild the artificial neural network model based on the output prediction data, the first selection step (S331), the screening data extraction step (S332), the data extraction step (S333) And determining whether to rebuild (S334).

The first selection step S331 is a step of selecting output prediction data corresponding to a predetermined selection criterion among the output prediction data calculated through the second modeling step S320. If the output prediction data is a number such as oil gas production, in the first selection step (S331), the output prediction data are ranked in descending or ascending order, and the selected criteria are ranked from the first rank to the predetermined reference rank. Set it.

In this case, the reference rank is set to a value corresponding to the product of the number of output prediction data output in the second modeling step S320 and a preset calculation ratio. The calculation ratio may be set to 20% or 30%. However, the calculation ratio may be arbitrarily set according to the number of data to be analyzed, without being limited thereto. For example, when 100 output prediction data are calculated in the second modeling step S320 and the calculation ratio is set to 20%, the reference rank is 20th. In the first selection step S331, the determined ranking Selects output prediction data corresponding to ranks 1 to 20 in the.

On the other hand, when the output data having the maximum value is calculated as the global solution in the first selection step S331, the output prediction data are ranked in descending order. In another embodiment of the present invention, the output data having the lowest value can be calculated as a global solution, and it is preferable to rank the output prediction data in ascending order.

The screening data extraction step S332 is a step of extracting input data corresponding to the output prediction data selected through the first selection step S331.

The non-data extraction step (S333) is a step of extracting input data (T) which is not included in the learning data used for constructing the artificial neural network among the input data extracted through the screening data extraction step (S332). The operator stores the extracted input data in the database.

The reconstructing determination step (S334) is a step of determining whether to rebuild the artificial neural network model based on the number of input data extracted in the non-contained data extraction step (S333). That is, if the number of input data extracted in the non-data extraction step (S333) is less than the predetermined number of determination criteria, it is determined that reconstruction of the artificial neural network model is not necessary, and the number of input data selected in the non-data extraction step (S333) If it is more than the predetermined number of criteria, it is determined that the neural network model needs to be rebuilt.

Here, the number of determination criteria may be set equal to the number of data selected as input data among the analysis target data in the initial learning material generation step (S100), but is not limited thereto, and an appropriate number may be arbitrarily set according to the number of analysis data. have. For example, if 20 input data are selected in the initial learning material generation step (S100), 20 is applied as the number of criteria.

In the fourth modeling step S340, when it is determined that the artificial neural network model needs to be reconstructed through the third modeling step S330, the new modeling step and the new search group including the learning material are determined. A step S341, a material adding step S342, and a group adding step S343 are included.

The second selection step S341 is a step of selecting some of the input data extracted in the non-data extraction step S333. In this step, it is preferable to select data from the input data extracted in the data extraction step (S333) not including as many as the data selected as the learning data of the analysis target data in the initial learning material generation step (S100), but is not limited thereto. Instead, the data may be selected from the input data extracted in the non-included data extraction step (S333) by the number smaller than the number of data selected as the learning material in the initial learning material generation step (S100).

For example, if 20 input data are selected in the initial learning material generation step (S100), in the second selection step (S341), 20 pieces of input data extracted in the non-included data extraction step (S333) are randomly selected.

The data adding step (S342) is a step of generating a new learning material by adding the input data selected in the second selection step (S341) to the learning material, and includes a data addition obtaining step (S344) and a material generating step (S345). .

The additional data acquisition step S344 is a step of acquiring the actual output data output by inputting the input data selected in the second selection step S341 to the computational model. The computational model is applied to the computational model used in the initial learning material generation step (S100).

The data generation step S345 is a step of generating new learning material by adding the input data selected in the second selection step S341 and the actual output data acquired in the data addition obtaining step S344 to the learning material.

The group addition step S343 is a step of setting the input data selected in the non-included data extraction step S333 and the input data included in the learning material as a new search group.

Meanwhile, if it is determined that the neural network model needs to be reconstructed through the third modeling step (S330), the model building step (S300) uses the new training material and the new search group, and then the fourth modeling step (S310) to the fourth. It is preferable to repeat the modeling step (S340). In addition, in the model building step S300, the first modeling step S310 to the fourth modeling step S340 are repeated a plurality of times. The input data extracted in the data extraction step S333 without the third modeling step S330 is performed. If the number of them is less than the predetermined number of criteria, it is determined that reconstruction of the artificial neural network model at this time is not necessary.

The wide area calculation step (S400) is a step of calculating a wide area solution when it is determined that reconstruction is not necessary through the model building step. The input data calculation step (S410), the output data calculation step (S420), and the completion step (S430). ).

The input data calculation step S410 is a step of calculating input data corresponding to the output prediction data selected in the first selection step S331. That is, input data input to the artificial neural network model is calculated so that the output prediction data selected in the first selection step S331 is output.

The output data calculating step S420 is a step of calculating actual output data for the input data selected in the input data calculating step S410 by using a computational model. That is, the input data calculated in the input data calculating step S410 is input to the computational model to obtain the actual output data. At this time, the computational model is applied to the computational model used in the initial learning material generation step (S100).

Completion step (S430) is input data corresponding to the actual output data having the largest or smallest value among the actual output data and the actual output data constituting all the learning data calculated in the output data calculation step (S420) Calculation step. In this step, when the output data having the maximum value is calculated as the global solution, the input data corresponding to the actual output data having the largest value is calculated, and when the output data having the lowest value is calculated as the global solution, the smallest value is obtained. It is preferable to calculate the input data corresponding to the actual output data having the value.

Hereinafter, an application example of a method for calculating a global solution using an artificial neural network will be described in more detail.

(Application example)

As an application, the above method is applied in the matter of determining the drilling position which can maximize the oil gas production in the reservoir where oil gas is buried. The storage layer has a structure as shown in FIG. 9, and FIG. 9 is gridded into 4800 grids in the vertical direction, the left and right directions, and the rear direction for simulation by the computational model (I direction × J direction × K direction = 40 X 40 x 3).

The existing production wells for collecting oil gas are set in the middle of the reservoir, and it is a problem of calculating one drilling position for horizontal wells in which oil gas production is maximized in addition to the existing production wells. At this time, since the horizontal tablet can be placed in two directions, one in the vertical direction and the other in the left and right directions, the data to be analyzed are data about 3200 drilling positions (= 40X40X2 direction).

First, the computer model was used for 3200 data, and the actual output data was obtained in advance, and then the wide-range solution was verified.

In the initial learning data generation step (S100), the data on the drilling positions forming the reservoir layer, that is, 20 drilling positions are selected as the input data from the analysis target data, and each data for the 20 drilling positions selected as the input data. Is inputted into the instrumented computer model to perform the simulation. At this time, the computational model is a computational model for oil gas production according to the drilling position information. The 20 real output data outputted by the computational model are classified into the training data along with the input data.

In the initial search group generation step (S200), the analysis target data are classified into search groups, and 2858 of the reservoirs are classified as search groups except for the external boundary part and the existing production well installation part.

Next, build an artificial neural network model based on the training material through the model building step (S300). In this case, the model building step (S300) applied to the embodiment will be described in detail as follows.

First, an artificial neural network model is constructed through 20 input data and 20 actual output data through a first modeling step S310, and 2858 data constituting a search group are artificially created through a second modeling step S320. Input to neural network model to obtain 2858 output prediction data.

10 illustrates a graph comparing output prediction data output through an initial artificial neural network model with actual output data obtained by inputting 2858 data forming a search group into a computer model. Here, the X axis represents the actual output data values obtained by inputting 2858 data forming the search group into the computational model, and the Y axis represents the output prediction data values outputted from the initial neural network model with 2858 data. will be. In addition, the triangle on the upper right is the real-world solution, that is, the data with the maximum yield, the squares in the middle correspond to the training data used to construct the neural network model, and the dotted line is the baseline for representing the output ratio of 30%. That is, points corresponding to the upper side of the dotted line are data included in the calculation rate, and points corresponding to the lower side of the dotted line are data not included in the calculation rate. In this case, assuming that y = x virtual line (not shown) is added to FIG. 10, the points located above the y = x virtual line are output through the artificial neural network model earlier than the data output through the computational model. The larger the data, the lower points of the imaginary line with y = x indicate that the data output through the initial neural network model is smaller than the data output through the computational model.

Next, through the first selection step S331, the calculated output prediction data are ranked in descending order. At this time, the calculation rate is set to 30% to select output prediction data from the 1st rank to the 857th rank in the determined rank.

Next, through the screening data extraction step (S332), the input data corresponding to each of the 857 output prediction data is extracted, and the learning material used to build the artificial neural network of the 857 output prediction data through the without data extraction step (S333) Extract input data not included in.

In this case, as a result of determining the initial neural network model in the re-establishment determination step, it is determined that the initial neural network model needs to be rebuilt because the number of input data selected in the non-data extraction step (S333) is 20 or more, which is a predetermined number of determination criteria.

Next, a new learning material and a new search group are determined through the fourth modeling step (S340). That is, 20 input data among the input data selected in the non-data extraction step S333 are selected, and the 20 input data are input to the computational model to calculate actual output data. At this time, the new learning materials are determined by including 20 newly selected input data and actual output data in the existing learning materials. In addition, the input data included in the non-data extraction step (S333) and the training data used to construct the initial artificial neural network model are set as a new search group.

Next, repeating the first modeling step (S310) to the fourth modeling step (S340) using a new learning material and a new search group, the neural network model reconstruction is not necessary in the third modeling step (S330). Repeat until you determine no.

11 illustrates a graph comparing output prediction data output through a second artificial neural network model and actual output data obtained by inputting a search group into a computer model, and FIG. 12 illustrates a third artificial neural network model. A graph comparing the output prediction data output through the search group with the actual output data obtained by inputting the search group into the computational model is disclosed. In FIG. 13, the output prediction data and the search group output through the fourth artificial neural network model are disclosed. The graph comparing the actual output data obtained by inputting into the computational model is disclosed. The markers shown in FIGS. 11 to 13 have the same meanings as the markers shown in FIG. 10. 10 to 13, the number of search groups decreases as the number of reconstructions of the artificial neural network model increases, and the output prediction data value output through the artificial neural network model is similar to the actual output data value output through the computational model. It can be seen that an artificial neural network can be used to calculate a real global solution, that is, a highly accurate global solution for data having an actual maximum value.

On the other hand, in the case of the above application, the total number of input data selected in the data extraction step (S333) at the time when the construction of a total of five artificial neural network model is completed is less than 20, which is a predetermined number of discrimination criteria, the reconstruction of the artificial neural network model It is determined that this is not necessary, and the eight input data are inputted to the computational model to output the actual output data. The input data corresponding to the actual output data having the largest value among the actual output data and the actual output data of all learning materials is determined as a wide area solution. This concludes the process of constructing an artificial neural network to calculate the global optimal solution.

Returning to the present invention will be described again.

In the present invention, as shown in Figure 1, while reducing the search space gradually increasing the resolution, it is derived through the artificial neural network for the optimal solution for a plurality of drilling estimates. In other words, based on the basic lattice system, the stages are constructed sequentially from the first-stage grid system to the n-th grid system, and the artificial neural network is constructed for each stage according to the above process, and the optimal solution for the plural drilling estimates is obtained. Find. Each time the order (i) is increased by one, the step is made, using the grid on the new grid system to create a learning material, using it to build an artificial neural network model, and reduce the artificial neural network by reducing the search group of the constructed neural network Rebuild Finally, the constructed neural network is used to calculate the solution for the optimal position at each stage.

Now, both the configuration and the flow of the present invention have been described, and specific embodiments will be described.

(Example)

An embodiment of the present invention is to determine two drilling locations that can maximize the oil gas production in the oil gas reservoir shown in FIG.

The storage layer has the same structure as shown in Fig. 2, and the storage layer is lattice-shaped in the I direction × J direction × K direction = 61 × 37 × 3 for the computational model simulation. There are 6 existing production wells (hatched parts, w) in the reservoir, and it is a problem to select the optimum position when adding two production wells that can maximize the oil production of the entire reservoir.

For the addition of two vertical production wells, the total possible combination is the number of 2,545,896 cases (planar). Of course, the above input data can be applied to the artificial neural network algorithm shown in FIG. 8 as it is to find two optimal positions we want to find. However, when the number of input data is so large, when the algorithm shown in FIG. 8 is applied as it is, the probability of not finding an actual optimal position may increase. The present invention has been made to solve this problem. That is, while using the algorithm of Figure 8 developed by the researchers of the present invention, it is applied in a sequential step. In other words, the multi-stage grid system was applied to reduce the number of cases. As a design variable, n1 = 5, n2 = 3, n3 = 1, and r1 = r2 = 5. Here, n is the size of the grid in each stage, n1 means that in the first stage grid system, the basic grid is grouped by 5 * 5 to form one grid, and n2 = 3 is based on the size of the grid in stage 2. N 3 = 1 means that the lattice in the third stage is the same as the basic lattice. r1 = r2 = 5 means that the upper ranks of the first and second stages are calculated only to the fifth position, respectively. Of course, the above values may vary depending on the embodiment, and the steps may be further increased.

First, n1 = 5 was set to construct a one-stage grid system, and the 5x5 grid of the basic grid system was composed of one grid. A one-stage grid system is constructed in FIG. Here, it is a one-stage grating system in which dark solid lines are constructed, and a black grating indicates a position where a drilling estimate can be placed in this system. Thin solid line means basic grid system. In the first-stage grid system, the number of cases where two drilling estimates can be located, that is, the search space, has been dramatically reduced to a total of 3,486. The algorithm for finding the global solution shown in FIG. 8 is applied to this grid system. FIG. 15 illustrates a process of constructing and rebuilding an artificial neural network through the process shown in FIG. 8. 15 (a) to 15 (d) correspond to FIGS. 10 to 13.

The data indicated by a triangle in the upper right of FIG. 15 (d) is the global solution in the first step, that is, the optimal position of the plurality of drilling estimates.

Then, a set of r1 = 5 and the solutions of the top 5 ranks in the final result of FIG. 15 (d) are the hatched gratings of FIG. 16. Since there are many grids overlapping among the top ranks, only six grids were actually selected.

The second stage grating system was constructed by further expanding the upper 5 rank region of FIG. 16. The two-stage grid system is shown in dark solid lines in FIG. 17 and the drillable positions are marked in black. The size of the search space in the second-stage grid system is 280. The results of reconstructing the artificial neural network model shown in FIG. The grating hatched in FIG. 18 corresponds to a top 5 rank solution, and this area is further expanded to set a search space, and then a third stage grating system is constructed as shown in FIG. Referring to FIG. 19, the third stage grid system is the same as the grid size of the basic grid system, and the search space has 2,800 cases (input data).

Again, a wide area solution is derived using the artificial neural network model shown in FIG. The process of building and reconstructing the artificial neural network is shown sequentially in FIGS. 20 (a) to 20 (d). The point shown by the triangular shape at the upper right of FIG. 20 (d) is the final wide area solution. The input data of the wide area solution is shown in FIG. 21 as shown in FIG.

Since the grid size in the third stage grid system is the same as the existing grid size, the solution derived in the third stage is now determined as the final solution without updating the grid system. The two gratings denoted by r in FIG. 21 are optimal positions where two drilling estimates will be respectively installed.

For reference, FIG. 22 compares six production profiles of existing production wells and production profiles including two optimal positions of additional production wells through a computerized model.

As described above, when the neural network model is applied once using the basic grid system to find two optimal positions, there are difficult situations in which 2,545,896 cases must be covered. However, in the present invention, if the multi-stage grid system approach is used, only 6,566 (= 3,486 + 280 + 2,800) cases need to be considered in three steps.

That is, according to the present invention, the algorithm for finding a global solution using a multi-stage grid system has an advantage of finding the global solution very easily even when the number of cases to be considered increases exponentially. Of course, the neural network must be newly constructed at each stage, but the construction rate of the artificial neural network is very fast in recent years, so it does not matter in applying the present invention.

The “method of calculating a wide area solution using an artificial neural network” developed by the researchers of the present invention has a practical limitation in finding an actual wide area solution by directly applying the input data at a time when there are a lot of input data. However, not only is the process of finding the optimal solution very fast by using the multi-step approach through the present invention, but also the possibility that the found wide-area coincides with the actual wide-area solution is increased. Therefore, the present invention is expected to be able to replace the simulation using a conventional computer model.

As described above, the present invention is implemented as a program on a computer. Accordingly, the present invention provides a recording medium containing a program for executing the above method on a computer.

Meanwhile, the artificial neural network model used in the present invention has been described with reference to the form shown in FIG. 8 as an example, but the present invention is not limited thereto, and other wide area search models, such as genetic algorithms, which have been previously developed, as well as Another artificial neural network model could be used.

In addition, although the present invention has been described as using a plurality of locations of the drilling estimates, it can be used to find one drilling estimate.

In addition, although the number of criteria for determining whether to reconstruct the artificial neural network model is described and illustrated as N, the number of criteria is not limited thereto. For example, it is extended to a case where T is smaller than a predetermined value between N and 3N. can do. It can be further extended to apply for values above 3N.

The protection scope of the present invention is not limited to the description and expression of the embodiments explicitly described above. In addition, it is again noted that the scope of protection of the present invention may not be limited due to obvious changes or substitutions in the technical field to which the present invention pertains.

Claims (12)

1. (a) setting an entire area of the basic grid system formed by dividing an oil or gas storage layer into a plurality of basic grids as a search space, and subdividing the search space into a grid having a size larger than the basic grid;

   (b) applying an artificial neural network evaluation model to determine which one of the plurality of gratings in the search space is located in one of the gratings, and indicating the goodness of fit of each grating as a rank;

   (c) If the current grid size used in the evaluation step is larger than the base lattice size by a predetermined range or more, proceed to the next step to rebuild the search space, if within a certain range, the result of the evaluation step to the optimal position Determining;

   (d) reconstructing the lattice recorded in a higher rank up to a predetermined rank in the goodness-of-fit of the lattice in the evaluation step into a new search space;

   (e) re-segmenting the reconstructed search space into grids of size greater than or equal to the basic grid size but smaller than the grid size used in the preceding evaluation step;

   The grid size in the search space is gradually reduced to be equal to the size of the basic grid within a predetermined range, so that the optimal position of the drilling estimate is determined in step (c), in step (e) to step (e). Method of optimizing a plurality of drilling position using the artificial neural network in the oil gas reservoir characterized in that it performs repeatedly until.
2. The method of claim 1,

   In the step (a) or (e), when the search space is partitioned into a grid having a size larger than that of the basic grid, the neural network in the oil gas reservoir layer is formed by tying a plurality of the basic grids together. A plurality of drilling estimation position optimization method using.
3. The method of claim 1,

   When rebuilding the search space in step (d),

   The plurality of drilling positions using the artificial neural network in the oil gas reservoir layer, wherein the base area occupied by the grids of the upper rank and the additional area extended within a predetermined range from the base area are reconstructed into the search space. Optimization method.
4. The method of claim 3,

   And the additional region extends only to a range that does not include a central portion of a grid adjacent immediately adjacent to the grids of the upper rank, using a neural network in the oil gas reservoir.
5. The method of claim 4, wherein

   The method of optimizing a plurality of drilling positions using artificial neural network in the oil gas reservoir layer, characterized in that for expanding the additional region, the basic grid in the grid.
6. The method of claim 1,

   Evaluation method using the artificial neural network model,

   An initial learning material generation step of classifying input data selected from a plurality of analysis object data and actual output data outputted by inputting the input data to a computational model constructed to analyze the analysis object data as learning data;

   An initial search group generation step of classifying the analysis target data into a search group;

   Constructing an artificial neural network model for predicting data output from the computational model using the training data, and determining whether to reconstruct the artificial neural network model based on the data outputted when the search group is input to the artificial neural network model. And rebuilding the artificial neural network model by resetting training data when it is determined that reconstruction of the artificial neural network model is necessary; And

   A wide area calculation step of calculating a wide area solution when it is determined that reconstruction of the artificial neural network model is not necessary through the model building step;

   The analysis target data is a plurality of drilling position optimization method using the artificial neural network in the oil gas reservoir, characterized in that the data on the position of the grid on which the plurality of drilling wells are installed.
7. The method of claim 6,

   The model building step,

   A first modeling step of constructing an artificial neural network model for predicting data output from the computational model using the training material;

   A second modeling step of obtaining output prediction data by inputting respective data included in the search group to the artificial neural network model;

   A third modeling step of determining whether to reconstruct the artificial neural network model based on the output prediction data; And

   If it is determined that the neural network model needs to be reconstructed through the third modeling step, a fourth modeling step of determining a new learning material and the new search group including the learning material,

   The method of optimizing a plurality of drilling positions using the artificial neural network in the oil gas reservoir layer, characterized in that it is repeated from the first modeling step to the fourth modeling step using the new learning material and the new search group.
8. The method of claim 7, wherein

   The third modeling step

   A first selecting step of selecting output prediction data corresponding to a predetermined selection criterion among the output prediction data calculated through the second modeling step;

   A selection data extraction step of extracting input data corresponding to the output prediction data selected through the first selection step, respectively;

   A non-data extraction step of extracting input data not included in the learning data used for constructing the artificial neural network among the input data extracted through the selection data extraction step; And

   If the number of input data extracted in the non-data extraction step is less than the predetermined reference number, it is determined that reconstruction of the artificial neural network model is not necessary, and the number of input data selected in the non-data extraction step is preset If it is more than a number, re-establishment determination step of determining that it is necessary to rebuild the artificial neural network model; comprising a plurality of drilling estimation position optimization method using the artificial neural network in the oil gas reservoir.
9. The method of claim 8,

   The third modeling step,

   The determination criterion number is a plurality of drilling position optimization method using an artificial neural network in the oil gas reservoir, characterized in that the number of data selected as the input data of the analysis target data in the initial learning data generation step.
10. The method of claim 8,

    The fourth modeling step,

    A second selection step of selecting some of the input data extracted in the non-data extraction step;

    A material addition step of generating new learning material by adding the input data selected in the second selection step to the learning material; And

    Optimizing a plurality of drilling positions using an artificial neural network in the oil gas reservoir, characterized in that it comprises a group addition step of setting the input data selected in the non-data extraction step and the input data included in the learning material as a new search group; Way.
11. The method of claim 10,

    In the second selection step,

    Selecting data from the input data extracted in the extracting data extraction step by less than the number of input data selected as the learning material or the number of input data selected as the learning material in the initial learning material generation step. A Method for Optimizing Multiple Drilling Positions Using Artificial Neural Networks in Oil Gas Reservoir.
12. The method of claim 10,

    The data addition step,

    An additional data acquiring step of acquiring the actual output data outputted by inputting the input data selected in the second selecting step to the computational model; And

    And a data generation step of generating new learning data by adding the input data selected in the second selection step and the actual output data acquired in the additional data acquisition step to the learning data. A plurality of drilling location optimization methods using neural networks.

Images