14.4. 1 Apply grid method to approximate resource distribution.
The basic idea of this method is:
(1) According to a large number of observation data, the plane distribution map of each single parameter is compiled, which is usually a plane isoline map, such as the isothickness map of source rocks, and individual maps are zoning maps, such as the evolution degree map. Using these plane distribution maps to simplify the spatial variation of each parameter is mainly to simplify the vertical variation of each parameter to a constant value by using the average value, such as the so-called isoline map of organic matter abundance of source rocks, that is, to simplify the vertical variation of organic matter abundance at each point to a constant value. (2) Establish a fixed grid on the plane. The grid is generally set by uniform method, but non-uniform grid can also be used. The number of grids depends on the plane change speed of variables, computer speed and capacity. In principle, the finer the mesh, the more accurately it can describe the change of parameter plane.
(3) reading the specific numerical values of parameters on grid nodes (or grid midpoint) on each same grid parameter distribution map.
(4) For each grid node (or grid midpoint), according to the resource evaluation model, the hydrocarbon generation amount and hydrocarbon expulsion amount are calculated respectively, and then the isoline map of hydrocarbon generation amount and hydrocarbon expulsion amount is compiled.
(5) According to the area occupied by each isoline interval, calculate the hydrocarbon generation amount and hydrocarbon expulsion amount occupied by this interval, and then accumulate to obtain the hydrocarbon generation amount and hydrocarbon expulsion amount of the whole region. Multiplied by the corresponding migration and accumulation coefficient, the resources of the whole region can be obtained.
14.4.2 Monte Carlo method
The so-called Monte Carlo method is a numerical calculation method, which refers to taking a fixed value from the distribution curve of each parameter by random sampling, and then calculating according to the evaluation model to get a certain value as the result. The above process is repeated for thousands of times, and there are thousands of fixed values, and then these fixed values are counted to get the distribution curve of the results. This method has been widely used in the evaluation of oil and gas resources, and its advantage is that a distribution curve is used to approximate the underground evaluation object, more likely value and most likely value. This is more in line with people's understanding process, limitations and uncertainty of underground evaluation objects.
The calculation steps of this method are as follows:
(1) In principle, the more data of each parameter, the better. It is generated, collected and sorted by means of data processing and interpretation, analysis and testing, and reading drawings. At the same time, singularity is eliminated.
(2) According to the sorted data, the probability distribution curve of each parameter is established statistically. When there are many data, such as dozens or more, the statistical distribution curve is representative and reliable. However, when there are only a few or more data, the actual distribution curve can be constructed according to the distribution probability of parameters (generally an empirically known distribution model, such as normal distribution and lognormal distribution). But when there is only a small amount of data and its distribution probability is uncertain, it is best to replace its distribution with uniform distribution or triangular distribution.
(3) The simplest and most basic way to generate random numbers by computer is to distribute them evenly. Random numbers are required to undergo strict tests (such as uniformity test, independence test, combination regularity test, continuity test, etc.). ), and its properties meet the requirements before it can be put into use. The more random numbers, the better, preferably thousands. The range of random values is 0 ~ 1.
(4) Take the random value as the entrance value of the probability, and use interpolation method to get the parameter value corresponding to the probability in a parameter distribution (Figure 14- 1). Then use another random value to find the parameter value on another parameter distribution curve (Figure 14-2). And so on. Then according to the evaluation model, the obtained parameter values (each parameter has only one value) are multiplied, divided or added and subtracted to get a result (Figure 14-3). Repeat this process and get thousands of results.
Figure 14- 1 Schematic diagram of sampling calculation process
(5) Make mathematical statistics on the obtained results, and get the probability distribution diagram of the results (Figure 14-3). Generally speaking, the probability distribution of parameters used in Monte Carlo calculation can be varied, but the result distribution is generally normal or lognormal.
Figure 14-2 Schematic diagram of multi-parameter sampling calculation process
Figure 14-3 Schematic Diagram of Monte Carlo Calculation Process
14.4.3 fuzzy mathematics calculation method
In some research objects, the boundaries of different things are completely different. For example, water can have three forms: ice, water and steam, and its boundaries are generally clear. However, in some objects, the boundaries between different things are not clear. For example, in petroleum geology, "good permeability" and "poor permeability" are two completely different concepts, but sometimes it is not easy to classify a specific object as "good permeability" or "poor permeability". Fuzzy mathematics uses membership degree to describe this situation, that is, it uses numerical value to indicate the degree that an object belongs to something, and an object can "belong" to two or more things, and its degree of belonging to these two kinds of things is described by two membership degrees respectively, thus solving this kind of problem reasonably.
When fuzzy mathematics is used to evaluate the oil-gas bearing property of a trap, a vector is used to represent a trap:
Evaluation method and practice of oil and gas resources
The research object contains k traps, so the set Ui is used to represent this trap group:
Evaluation method and practice of oil and gas resources
N geological factors play different roles in hydrocarbon evaluation of traps, and each factor is represented by a weighted ai value:
Evaluation method and practice of oil and gas resources
Each geological factor is represented by M-level:
Evaluation method and practice of oil and gas resources
Ci is an attribute expressed as an integer, and its specific value varies with m.
When m=3, c = [- 1 0 1]
When m=5, c = [-2- 10 12]
When m=7, c = [-3-2- 10 123]
The geological factors of a trap are expressed by its membership degree to each attribute (such as table 14- 1).
Table 14- 1 Membership Table of Geological Factors
A trap is described by n variables, and the expression of each variable will be transformed into a vector, while a trap originally represented by a vector will be represented by a comprehensive evaluation transformation matrix r:
Evaluation method and practice of oil and gas resources
The comprehensive evaluation of each trap is calculated by the weight of each geological factor and the comprehensive evaluation transformation matrix of each trap. This calculation process is called synthesis:
Evaluation method and practice of oil and gas resources
Where H is the number of samples, Rh is the comprehensive evaluation transformation matrix of H samples, Bh is a vector composed of n (variable) numbers, and its elements are
Evaluation method and practice of oil and gas resources
Here, ○ represents an algorithm, which evolved from the following four basic algorithms (assuming that A and R are two elements in a fuzzy set).
1)a∨r=max( 1，r)
2)a∧r=min(a，r)
3)a r=ar
4)a⊕r=min(a, 1+r)
According to this synthesis, the sample vector is obtained, and then the comprehensive evaluation value (comprehensive score) d is calculated:
Evaluation method and practice of oil and gas resources
The result is a number. Traps are arranged according to their D values, which means that the advantages and disadvantages of these traps are arranged. Every time a synthesis method is adopted, there is a B, and correspondingly there is a D value and a queue. Because B is generated in different ways, variable values have different functions, and the same original data will have different queuing results.
14.4.4 neural network calculation method
Artificial neural network refers to a network composed of a large number of (artificial) neurons similar to the natural nervous system.
The structure and characteristics of neural networks are determined by the characteristics of neurons and their relationships. Artificial neurons form a network through interconnection. The way of interconnection is called connection. The connection strength between neurons is the connection weight. When the connection weight matrix of the network is determined, the connection mode of the network is also determined.
In artificial neural network, the information processing process or the change of stored knowledge is completed by modifying the connection mode between neurons. This modification process is called neural network training or learning. Different ways to adjust the weight matrix means different learning methods.
The learning of neural network and the structure of neural network are not one-to-one correspondence. Different neural networks can be trained by the same learning algorithm; The same neural network can also be trained by different learning algorithms.
Generally, multi-layer feedforward neural network and back propagation (BP) algorithm are adopted.
For the three-layer neural network model, the first layer is the input layer, the second layer is the middle layer and the third layer is the output layer. The number of neurons in the first layer is n, the middle layer is 1, and the third layer is 1.
The 1 layer is the input layer and consists of n neurons of m samples. It is agreed that the input of the kth sample (trap), namely 1 layer neurons, is xk 1, xk2, …, xkn, and the corresponding output is Tk, where k is the number of samples and k= 1.
The second layer is the hidden layer, and the number of neurons is 1, which is set by the user and calculated by multiplying x by the weight coefficient matrix W2. The middle layer of the k-th sample is
Evaluation method and practice of oil and gas resources
F(t) adopts S-type compression function:
Evaluation method and practice of oil and gas resources
To control the value of u, change the first formula to: x0=- 1, w0j=ξ, remember.
Evaluation method and practice of oil and gas resources
Then the second formula becomes
The value of t is not only related to Wij and xi, but also related to the number of variables n. In order to make the value 0 ~ 1, you need
Evaluation method and practice of oil and gas resources
Give an appropriate zeta value.
The calculation from the middle layer to the output layer is similar. Only it uses another W (matrix).
If a suitable W (two W arrays) is found, the Y value of each sample calculated from the input X of each sample should be the same as or very close to the output T of the original sample. Our task is to ask these two W arrays.
Evaluation method and practice of oil and gas resources
The initial W matrix is generated randomly. Of course, it calculates that y of each sample will not be equal to t, and we use E(W) to measure its deviation:
Evaluation method and practice of oil and gas resources
When e (w) < ε, learning is completed. When E(W) > ε, two w matrices need to be modified to make E(W) smaller gradually. As far as the current model is concerned (a * * * has three layers and the output layer has only one element), modifying W is divided into two steps. The first step is to modify W calculated by U, and the second step is to modify W calculated by X. ..
Evaluation method and practice of oil and gas resources
Evaluation method and practice of oil and gas resources
In this way, according to the calculated y, the two-layer W matrix is modified every time until e (w) < ε, and the learning is completed.
After learning is completed, two W matrices are obtained, and the X vector of the sample to be judged is calculated according to the established model, so that the Y value of each sample can be obtained, which is the evaluation of the specific object.