RU2571470C2 - Method of neural network analysis of telemetry data for well stock - Google Patents

Abstract

FIELD: oil and gas industry.

SUBSTANCE: invention is referred to oil production technologies, and namely to methods of production process monitoring and maintenance of formation pressure based on telemetry data processing by neural network algorithm. The method of neural network analysis of telemetry data for well stock is suggested; the above method consists in recording data on fluid consumption rate at wells, their preliminary processing and presenting as n-dimension vectors; creating m of n-dimension vectors of reference data, which are used as input data for neural network algorithm and for characterisation of wells state for a certain time interval, n-dimension vectors of consumption rate are compared with multitude of nodes of the Self Organising Map (SOM) generated in process of training. At that initial multitude m of n-dimension vectors serves as input data for SOM training algorithm and data used for number identification of the borehole belonging to the node of the trained SOM. The result of displayed borehole numbers at SOM nodes is required for identification of wells mutual influencing and interferencing.

EFFECT: suggested method allows increase in deposit development due to selection of sound operational modes based on detected phenomena of the wells mutual influencing and interferencing in process of their operation.

2 dwg

Description

The method of neural network analysis of telemetry data from a well stock relates to the oil industry, and in particular to methods for monitoring production processes and maintaining reservoir pressure (RPM) based on the processing of telemetry data by a neural network algorithm to detect the phenomenon of mutual influence of wells. The technical result is the emergence of a new method for identifying the mutual influence of wells based on a neural network analysis of data on fluid flow rates in producing wells and wells of the RPM system.

The invention relates to oil production technologies, and in particular to methods for monitoring production processes and maintaining reservoir pressure based on processing telemetry data by a neural network algorithm.

As a prototype, a method similar to the use of Kohonen’s self-organizing maps for the analysis of information “Method of neural network analysis of the state of the heart” (RF patent for the invention No. 2461877) [1] was selected [1]. By analogy with the indicated method, a sample of m n-dimensional vectors is created, where average daily measurements of fluid flow in wells are presented as reference information. This information is fed to the Kohonen self-learning card. Training is carried out, after which, in contrast to the specified method [1], the proximity of each input vector of source data to the nodes of the resulting network is determined. Because If there is an unambiguous correspondence between the number of the well and the vector of its measurements, then the numbers of the wells are located on the nodes of the Kohonen map in accordance with the similarity of their measurements. As a result of displaying the well numbers on the Kohonen map, the analyst draws conclusions about the possible effect of the mutual influence of the respective wells.

A method of neural network analysis of telemetry data from a well stock, which consists in registering data on fluid flow in wells, pre-processing them and presenting them in the form of n-dimensional vectors, creating m n-dimensional vectors of reference information that serves as input to the neural network algorithm and characterizing the state of wells for a certain time period, n-dimensional cost vectors are compared with the set of nodes of the Kohonen self-organizing map created as a result of training, This is because the initial set of m n-dimensional vectors serves as input to the Kohonen map self-learning algorithm and data to determine whether the well number belongs to the trained Kohonen map node, and the result of the display of well numbers on the Kohonen map nodes serves to detect the effects of interference and interference of wells .

, …, $$s_m=\left(q_1^m,q_2^m,\dots ,q_n^m\right)$$

, где $$q_i^j$$

- i-й замер расхода на скважине номеру. Полученные вектора служат входными данными для нейросетевого алгоритма самоорганизующейся карты Кохонена [2]. После процесса самообучения производится распределение номеров скважин 1…m в соответствии с наибольшим соответствием векторов замеров расхода (вектор веса узла на карте менее всего отличается от наблюдения) узлам на карте. В итоге получают следующее соответствие: чем ближе номера скважин на узлах карты, тем более схожи результаты замеров в векторах, тем больший эффект взаимовлияния испытывают скважины.The method comprises the steps of collecting information on fluid flow rates (injection volumes and production volumes) from oilfield telemetry systems, averaging the flow rates to daily average, generating a training sample of m n-dimensional daily flow measurement vectors for each well s ₁ ... s _m : $$s_{\mathrm{one}}=\left(q_{\mathrm{one}}^{\mathrm{one}},{\mathrm{<figure-callout\; id="2"\; label="q"\; filenames="00000007.png"\; state="\{\{state\}\}">q</figure-callout>}}_2^{\mathrm{one}},...,q_n^{\mathrm{one}}\right)$$

, ..., $$s_m=\left(q_{\mathrm{one}}^m,q_2^m,...,q_n^m\right)$$

where $$q_i^j$$

- i-th metering of the flow rate at the well number. The obtained vectors serve as input data for the neural network algorithm of the Kohonen self-organizing map [2]. After the self-learning process, the distribution of well numbers 1 ... m is carried out in accordance with the greatest correspondence of the flow measurement vectors (the weight vector of the node on the map is the least different from observation) to the nodes on the map. As a result, we obtain the following correspondence: the closer the well numbers at the map nodes, the more similar the results of measurements in the vectors, the greater the effect of mutual influence on the wells.

The figure 1 shows the algorithm of the method for detecting interference phenomena and mutual influence of oil wells according to telemetry based on the use of neural network algorithms.

A self-organizing map consists of components called nodes or neurons. Their number is set by the analyst. Each of the nodes is described by two vectors: a vector of weight m, which has the same dimension as the input data; vector r, which is the coordinates of the node on the map. Usually nodes are placed at the vertices of a regular lattice with square or hexagonal cells.

Initially, the dimension of input data is known, according to which the initial version of the map is built. In the learning process, the weight vectors of the nodes approach the input data. For each observation, the node most similar in weight vector is selected, and the value of its weight vector approaches the observation. The vectors of weight of several nodes located side by side are also approaching the observation, so if the two observations were similar in the set of input data, they would correspond to nearby nodes on the map. The cyclic learning process, sorting through the input data, ends when the map reaches an acceptable (predetermined by the analyst) error or after a specified number of iterations are completed. The self-learning algorithm is described in detail in [2].

A feature of the oil production process is the hydrodynamic interference (interference) of the wells, when a change in the operating mode (flow rate and bottomhole pressure) of one well entails a change in the operating modes of other wells. Taking this factor into account is important in the selection of rational well operation modes and increasing the efficiency of field development.

The technical result should be considered the emergence of a new type of information synthesized in the processing of telemetry data, which allows you to detect the phenomena of interference and interference of wells during the operation of oil fields. The method includes collecting information on average daily fluid flow rates (production rates and injection volumes from reservoir pressure maintenance wells) from telemetry systems of wells involved in monitoring spatially limited oil fields, organizing this information into forms suitable for processing, processing data using a neural network method, the conclusion of the processing results in a form suitable for human analysis.

The initial data for the identification are the values of the measurements of fluid flow at the producing and reservoir pressure maintenance wells (SPM) located on a certain area, so that they can conditionally be considered hydrodynamically interconnected.

Suppose that there are m wells in the observation, for each of which n daily measurements of the parameter q flow rate were performed together, i.e. available vectors:

, …, $$s_m=\left(q_1^m,q_2^m,\dots ,q_n^m\right)$$

. $$s_{\mathrm{one}}=\left(q_{\mathrm{one}}^{\mathrm{one}},{\mathrm{<figure-callout\; id="2"\; label="q"\; filenames="00000007.png"\; state="\{\{state\}\}">q</figure-callout>}}_2^{\mathrm{one}},...,q_n^{\mathrm{one}}\right)$$

, ..., $$s_m=\left(q_{\mathrm{one}}^m,q_2^m,...,q_n^m\right)$$

.

будет происходить по принципу: чем больше значение i, тем больше значение j, при котором фактор получает приращение. В обратном случае наблюдается инерционность затухания возмущения, процесс которой происходит по тому же принципу. Этот факт позволяет рассматривать s₁…s_m с точки зрения объектов, обладающих указанными свойствами $$q_j^i$$

, и решать задачу поиска сходства объектов по указанным факторам.Due to the fact that the processes of mass transfer and pressure redistribution in reservoir conditions take sufficient time and do not occur instantly, with an increase in the flow rate at the disturbing well, the RPM, the response from this disturbance in the same parameter will be detected in nearby producing wells with delay. Conditionally, if wells s ₁ ... s _m are located in one line, s ₁ is a disturbing well, and the rest are observational producers, then the increment of factors $$q_j^i$$

will occur according to the principle: the larger the value of i, the greater the value of j at which the factor increments. In the opposite case, the inertia of the attenuation of the disturbance is observed, the process of which occurs according to the same principle. This fact allows us to consider s ₁ ... s _m from the point of view of objects with the indicated properties $$q_j^i$$

, and solve the problem of finding the similarity of objects according to the specified factors.

The identification task is to find the most similar objects according to the sets of signal factors, clustering them into groups according to similar characteristics identified by statistical processing of the measurement results. Clusters can have intersections, the further the group is separated from the other by the metric, the less similarity is, and in the projection onto the technological process, the mutual influence is less. The clustering process will be performed in Euclidean space on a plane, which can be schematically represented in the following form (Figure 2).

The diagram shows how, by the example of flow measurements from nine wells, they were clustered according to their “similarity” on a plane with a hexagonal cell arrangement. As a result of clustering, wells that are most similar in process parameters are located closer to each other in the diagram. Groups of wells are adjacent to other groups with which a certain statistical relationship is also seen. The farther the well is located from the other in the resulting diagram, the less similarity they have in the indicated factors.

For clustering in our method, the method of Kohonen self-organizing maps is used, which are competitive neural networks with training without a teacher and that perform visualization and clustering tasks.

A database is being archived for storing telemetry data on fluid flow rates in wells. In an automated system, through an organized interface, a selection of wells and a time interval for observation are made. As a result of preliminary data processing, a matrix is organized consisting of vectors containing daily average values:

, …, $$s_m=\left(q_1^m,q_2^m,\dots ,q_n^m\right)$$

. $$s_{\mathrm{one}}=\left(q_{\mathrm{one}}^{\mathrm{one}},{\mathrm{<figure-callout\; id="2"\; label="q"\; filenames="00000007.png"\; state="\{\{state\}\}">q</figure-callout>}}_2^{\mathrm{one}},...,q_n^{\mathrm{one}}\right)$$

, ..., $$s_m=\left(q_{\mathrm{one}}^m,q_2^m,...,q_n^m\right)$$

.

The matrix serves as input to an algorithm that implements clustering by the method of Kohonen self-organizing maps. The results of the algorithm are displayed on the operator’s screen in hexagonal and in the form of heat maps, which are subsequently used for analysis to detect the effects of interference and interference.

Information sources

1. RF patent No. 2461877. The method of neural network analysis of the state of the heart / Bodin O.N., Volkova N.A., Loginov D.S., Ryabchikov R.V., Funtikov V.A.

2. Kohonen, T., Self-Organizing Maps, Second Edition, Berlin: Springer - Verlag, 1997.

Claims (1)

A method of neural network analysis of telemetry data from a well stock, which consists in registering data on fluid flow in wells, pre-processing them and presenting them in the form of n-dimensional vectors, creating m n-dimensional vectors of reference information that serves as input to the neural network algorithm and characterizing the state of wells for a certain time period, n-dimensional cost vectors are compared with the set of nodes of the Kohonen self-organizing map created as a result of training, This is because the initial set of m n-dimensional vectors serves as input to the Kohonen map self-learning algorithm and data for determining the number of the well belonging to the node of the trained Kohonen map, and the result of mapping the number of wells on the nodes of the Kohonen map serves to detect the effects of interference and interference of wells .

Images