CN-121999260-A - Crude oil production layer judging method based on neural network learning model

CN121999260ACN 121999260 ACN121999260 ACN 121999260ACN-121999260-A

Abstract

The invention discloses a crude oil production layer judging method based on a neural network learning model, which belongs to the technical field of oil reservoir exploration and comprises the following steps of S1, obtaining crude oil samples of different production layer sections in a research area, S2, setting gas chromatography analysis conditions, S3, obtaining a chromatography graph, S4, carrying out homogenization treatment on the chromatography graph to obtain a label data set, S5, establishing an image learning network model, inputting a training set into the image learning network model for training, S6, adjusting the trained image learning network model through a verification set, S7, extracting crude oil samples of unknown production layer sections for carrying out gas chromatography analysis test, carrying out homogenization treatment on test results, and finally inputting the image learning network model to judge the crude oil sources of the unknown production layer sections. The invention is based on the neural network, and can rapidly and accurately judge the source of the crude oil in the unknown production layer by carrying out characteristic matching on the chromatographic spectrogram of the crude oil sample in the production well in the unknown production layer section.

Inventors

CHEN MANFEI
WANG LI
WANG JIAQI
KANG QIANG
TAN JIE
YU KAI

Assignees

中国石油天然气股份有限公司

Dates

Publication Date: 20260508
Application Date: 20241106

Claims (10)

1. The crude oil production layer judging method based on the neural network learning model is characterized by comprising the following steps of: s1, obtaining crude oil samples of different oil producing intervals in a research block; S2, setting gas chromatographic analysis conditions; S3, carrying out repeated gas chromatography on crude oil samples of each layer section to obtain a chromatographic chart of a gas chromatography analysis result of each test; S4, carrying out homogenization treatment on the chromatographic curve graph, labeling the homogenized chromatographic curve graph report to obtain a label data set, and dividing the label data set into a training set, a verification set and a test set according to the ratio of 4:1:1; S5, establishing an image learning network model, inputting a training set into the image learning network model for training, and selecting a residual error network to classify a chromatographic curve graph; s6, starting verification set optimization, adjusting the trained image learning network model through the verification set, and performing performance evaluation on the adjusted image learning network model through the test set; And S7, extracting a crude oil sample of the unknown production interval, carrying out gas chromatography analysis test, carrying out homogenization treatment on a test result, and finally inputting an image learning network model to judge the crude oil source of the unknown production interval.
2. The method for judging the crude oil production layer based on the neural network learning model of claim 1, wherein in the step S2, the gas chromatography analysis condition is set, the sample inlet temperature is 320 ℃, the carrier gas flow rate is 22mL/min, the chromatographic column temperature rising program is kept at an initial temperature of 40 ℃ for 10min, the temperature rising rate is increased to 250 ℃ at 4 ℃/min, the temperature rising rate is increased to 310 ℃ for 300min, the sample injection amount is 0.5 mu L, and the sample injection amount split ratio is 50:1.
3. The method for judging a crude oil production zone based on a neural network learning model according to claim 1, wherein in S3, the repetitive gas chromatography is performed more than 20 times on each layer section of crude oil sample.
4. The method for judging a crude oil production zone based on a neural network learning model according to claim 1, wherein in the step S4, the homogenization treatment of the chromatographic graph is performed by setting an abscissa of the chromatographic graph to 0-280min, an ordinate of the chromatographic graph to 0-Xmv, and Xmv is a peak value of a highest peak in the chromatographic graph obtained by each crude oil sample.
5. The method for judging crude oil production layer based on the neural network learning model of claim 1, wherein in the step S5, selecting a residual network to classify a chromatographic graph refers to carrying out element-by-element multiplication and summation operation on a convolution kernel and an input image, extracting features in the image, and carrying out full connection operation on the extracted features by the residual network to convert the features into class probability distribution.
6. The method for judging crude oil production layer based on the neural network learning model of claim 1, wherein in S5, the classification of the chromatographic graph is completed through a loss function.
7. The method for judging a crude oil production layer based on the neural network learning model of claim 6, wherein the loss function is a cross entropy loss function; Formula 1; Wherein, the The true mark of the ith category is the prediction probability of the ith category, and n is the number of the categories.
8. The method for judging a crude oil production zone based on the neural network learning model of claim 1, wherein in S6, starting verification set optimization means starting optimization when the variation amplitude of cross entropy loss values is smaller than 10%.
9. The method for determining a crude oil production zone based on a neural network learning model according to claim 1, wherein in S6, the adjustment of the trained image learning network model by the verification set comprises increasing the network depth, changing the convolution kernel size, and increasing the attention mechanism.
10. The method for judging the crude oil production zone based on the neural network learning model of claim 1, wherein in S6, performance evaluation of the adjusted image learning network model through the test set means that the image learning network model is built when the accuracy reaches 90%, otherwise, the adjustment is continued until the accuracy reaches 90%.

Description

Crude oil production layer judging method based on neural network learning model Technical Field The invention relates to the technical field of oil reservoir exploration, in particular to a crude oil production layer judging method based on a neural network learning model. Background The current field of Sichuan basin shale oil has characteristics of multiple hydrocarbon sources and multiple producing layers, which provide challenges for molecular geochemical evaluation and crude oil producing layer comparison work. Because the maturity of shale oil source rocks in Sichuan basin is in the mature-high mature stage, part of the biomass content is low, the shale oil source rocks lose efficacy when oil source discrimination is carried out, the representative oil source discrimination basis is lacking, the hydrocarbon source rocks are all deposited in land phases, the molecular geochemical characteristics of crude oil in each set of production layers are similar, and the difference between each layer is difficult to be obtained respectively through a conventional visual inspection method and a parameter-graphic method. Meanwhile, the crude oil group component separation and chromatographic quality experiments of saturated hydrocarbon, aromatic hydrocarbon and asphaltene are required to be carried out in a matched mode, the whole experiment flow is long, the experiment cost is high, the data types are various, the time required by manual comparison is long, and the manual comparison is easily influenced by human factors. The Chinese patent document with publication number CN110987858A and publication date 2020, 04 and 10 discloses a method for rapidly detecting oil products by using a neural network data model, which comprises the following steps: the step (1) is to collect oil products; Measuring the near infrared spectrum of the oil obtained in the step (1), performing spectrogram processing on the measured spectrum data, simultaneously measuring the conventional index of the oil obtained in the step (1), dividing seed oil into a plurality of families, and performing quantitative calculation on the families by using multi-element correction, specifically, firstly taking 10 samples from new oil to fully aged lubricating oil, scanning the infrared spectrums of the 10 oil samples, obtaining a water content result in the oil sample by a Karl Fischer potentiometric titration instrument, correlating the obtained infrared spectrum with the water content result, and obtaining a curve of water content and absorbed light in the lubricating oil by multi-element linear regression calculation; Step (3), correlating the spectrogram processing data obtained in the step (2) with conventional indexes of oil products, and establishing a correction model, wherein the correction model is established by adopting a partial least square method; Step (4), performing model verification on the correction model obtained in the step (3), wherein the verified model becomes a final model; And (5) obtaining accurate analysis structure and unknown sample measurement data through a final model. According to the method for rapidly detecting the oil product by using the neural network data model, the data model is built by using a counter-propagation artificial neural network method in near infrared spectrum analysis, indexes of linear changes of viscosity, antioxidant content, hydrocarbon content and the like of the oil product can be analyzed, and the relationship between near infrared spectrum information and lubricating oil base oil can be better quantitatively studied. But the crude oil of an unknown producing zone can not be rapidly judged with high precision. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides a crude oil production layer judging method based on a neural network learning model, the invention is based on a neural network, by performing feature matching on the chromatographic spectrogram of the crude oil sample of the production well in the unknown oil production interval, the source of the crude oil in the unknown oil production interval can be rapidly and accurately identified. The invention is realized by the following technical scheme: The crude oil production layer judging method based on the neural network learning model is characterized by comprising the following steps of: s1, obtaining crude oil samples of different oil producing intervals in a research block; S2, setting gas chromatographic analysis conditions; S3, carrying out repeated gas chromatography on crude oil samples of each layer section to obtain a chromatographic chart of a gas chromatography analysis result of each test; S4, carrying out homogenization treatment on the chromatographic curve graph, labeling the homogenized chromatographic curve graph report to obtain a label data set, and dividing the label data set into a training set, a verification set and a test set according to the ratio of 4:1:1; S5,