SYSTEM FOR INTELLIGENTLY RECOVERING WATER FLOODED LAYER OF OIL RESERVOIR BASED ON SPIKING NEURAL NETWORK

TECHNICAL FIELD

[0001]The present disclosure relates to the field of petroleum exploration and development technology, particularly to a system for intelligently recovering water-flooded layer of oil reservoir based on spiking neural network.

BACKGROUND ART

[0002]With the gradual deepening of global oil field development, many oil fields have successively entered the stage of high water cut development and production decline. Therefore, water-flooded layer interpretation technology has become a core link in improving recovery rate and optimizing development strategies in the middle and later stages of oil reservoir development. However, traditional well logging interpreting models are often constructed based on assumptions of static stratum parameters, which makes it difficult to fully adapt to the significant increase in reservoir heterogeneity and highly nonlinear characteristics presented by resistivity response during water flooding processes. This non-adaptability leads to a deviation between the resistivity curve and the true water saturation, resulting in significant ambiguity in the interpretation results, making it difficult to accurately reflect the actual dynamics of the oil reservoir.

[0003]It is particularly noteworthy that when the water-flooded layer presents a complex geological phenomenon of “low-resistance oil layer” and “high-resistance water layer” coexisting, the traditional static training mode appears inadequate. It cannot effectively capture the temporal correlation characteristics of dynamic evolution of the resistivity during the water flooding processes, and exhibits low robustness to small sample and high noise well logging data, further limiting its accuracy and reliability in water-flooded layer interpretation.

[0004]In recent years, the rise of machine learning algorithms has brought new hope to the field of well logging interpretation. These algorithms, through nonlinear mapping techniques, have to some extent broken through the parameterization limitations of traditional models and improved their ability to interpret complex geological relationships. However, their ability to handle dynamic process modeling is still limited, and there are still shortcomings in explaining complex nonlinear relationships, making it difficult to meet the high-precision requirements for dynamic recovery and interpretation of resistivity of the water-flooded layer.

[0005]Therefore, developing an intelligent interpretation method with powerful dynamic characteristic extraction capabilities has become an urgent need in the current research field of resistivity recovery and interpretation of the water-flooded layer.

SUMMARY

[0006]The purpose of the present disclosure is to provide a system for intelligently recovering water-flooded layer of oil reservoir based on spiking neural network, which can accurately recover the resistivity of the water-flooded layer, quantify the impact of water flooding accurately, and provide scientific and reliable basis for oil reservoir development and adjustment.

[0013]In some embodiments, the conventional well logging curves include natural gamma-ray, acoustic interval transit time, compensated neutron log, density, and self potential; and the electrical well logging curves include deep laterolog resistivity and shallow laterolog resistivity.

[0014]In some embodiments the range normalization is expressed as:

$x_{\mathrm{norm}}=\frac{x-x_{\min }}{x_{\max }-x_{\min }}$

$x_{\mathrm{std}}=\frac{x-\mu }{\sigma }$

where x_std is a standardized data, μ is a mean of the original data, and σ is a standard deviation of the original data.

[0016]In some embodiments, a dynamic expression of the Leaky Integrate-and-Fire neuron model is:

${\tau }_m\frac{\mathrm{dV}}{\mathrm{dt}}=-\left(V-V_{\mathrm{rest}}\right)+I_{\mathrm{syn}}\left(t\right);$

[0018]In some embodiments, an expression for adjusting weights of the synapses is:

$\Delta w_{\mathrm{ij}}=\eta ·{\sum }_t{\delta }_j\left(t\right)·S_i\left(t-\Delta t\right);$

$\begin{array}{c}\mathrm{RMSE}=\sqrt{\frac{1}{2}{\sum }_{i=1}^n{\left(R_{\mathrm{pred}}^{\left(i\right)}-R_{\mathrm{true}}^{\left(i\right)}\right)}^2};\\ \mathrm{MAE}=\frac{1}{n}{\sum }_{i=1}^n❘R_{\mathrm{pred}}^{\left(i\right)}-R_{\mathrm{true}}^{\left(i\right)}❘;\\ R^2=1-\frac{{\sum }_{i=1}^n{\left(R_{\mathrm{pred}}^{\left(i\right)}-R_{\mathrm{true}}^{\left(i\right)}\right)}^2}{{\sum }_{i=1}^n{\left(R_{\mathrm{true}}^{\left(i\right)}-\stackrel{\_}{R_{\mathrm{true}}}\right)}^2};\end{array}$

where RMSE is a root mean square error, MAE is a mean absolute error, R² is a coefficient of determination,

$R_{\mathrm{pred}}^{\left(i\right)}$

is a predicted value of an i-th resistivity, and

$R_{\mathrm{true}}^{\left(i\right)}$

is a true value of the i-th resistivity.

[0020]In some embodiments, an expression of the oil saturation is:

$S_o=1-{\left(\frac{{\mathrm{RT}}_{\mathrm{original}}}{{\mathrm{RT}}_{\mathrm{current}}}·\frac{a·{\varphi }^{-m}}{{\sigma }_w}\right)}^{\frac{1}{n}};$

$F_w=\frac{\frac{K_{\mathrm{rw}}}{{\mu }_w}}{\frac{K_{\mathrm{rw}}}{{\mu }_w}+\frac{K_{\mathrm{ro}}}{{\mu }_o}};$

where K_rw is a relative permeability of an aqueous phase, K_ro is a relative permeability of an oil phase, μ_w is a viscosity of the aqueous phase, and μ_o is a viscosity of the oil phase.

[0022]In some embodiments, the quantitatively classify includes steps that the water-flooded layer is quantitatively classified and interpreted based on an oil saturation-change value ΔS_o and a resistivity-change value ΔRT, and the water-flooded layer is quantitatively classified to be four levels: unflooded, weakly flooded, moderately-strongly flooded, and strongly flooded.

[0023]Therefore, the present disclosure adopts the above-mentioned system for intelligently recovering water-flooded layer of oil reservoir based on spiking neural network, and the technical effects are as follows.

[0024]1. High-precision recovery of resistivity: By constructing a dynamic resistivity recovering model based on spiking neural network, high-precision recovery of resistivity in water-flooded layer has been achieved. This model combines spatiotemporal dynamic characteristics of pulse neurons with rock physical constraints to effectively capture dynamic changes in resistivity during water flooding, improving the accuracy of recovery of the resistivity.

[0025]2. Solution of electrical confusion issues: In response to the electrical confusion issues caused by the coexistence of “low-resistance oil layer” and “high-resistance water layer” in the water-flooded layer, the SNN model effectively distinguishes between low-resistance oil layer and high-resistance water layer through high-precision modeling and dynamic feature extraction, improving the accuracy of interpretation of the water-flooded layer.

[0026]3. Enhanced accuracy and reliability of interpretation: By combining the data of multi-modal well logging curves, the SNN model demonstrates high accuracy and strong interpretability in resistivity recovery tasks. The evaluation of the model combines multiple indicators such as RMSE, MAE, and R² to verify its reliability and accuracy in interpreting water-flooded layers.

[0027]4. Improved robustness of the interpreting model: The SNN model exhibits strong robustness to small sample and high noise well logging data, and can maintain stable interpretation performance under complex geological conditions.

BRIEF DESCRIPTION OF THE DRAWINGS

[0028]FIG. 1 is a schematic diagram of the process of recovering resistivity of the original oil layer based on machine learning according to the present disclosure.

[0029]FIGS. 2A-2B show schematic diagrams of training results of SVR algorithm according to the present disclosure.

[0030]FIGS. 3A-3B show schematic diagrams of training results of LightGBM algorithm according to the present disclosure.

[0031]FIGS. 4A-4B show schematic diagrams of training results of SNN algorithm according to the present disclosure.

[0032]FIG. 5 is a diagram of evaluation indicators REMS, MAE, and R² when resistance is recovered by using SNN algorithm, SVR algorithm, and LightGBM algorithm and training set and validation set in the present disclosure.

[0033]FIGS. 6A-6H show schematic diagrams of prediction results of test well RT according to the present disclosure.

[0034]FIG. 7 is a flowchart for constructing a quantitative interpreting model for water-flooded layer of well logging according to the present disclosure.

[0035]FIG. 8 is a schematic diagram of comprehensive interpretation results of well logging LUD3245 in an experimental area according to the present disclosure.

[0036]FIG. 9 is a schematic diagram of comprehensive interpretation results of well logging LUD3264 in an experimental area according to the present disclosure.

[0037]FIGS. 10A-10H show schematic diagrams of resistivity prediction curves for 8 encrypted wells according to the present disclosure.

[0038]FIG. 11 shows a view of plane thickness of strong water-flooded layer in an experimental area according to the present disclosure.

[0039]FIG. 12 is a plane view of degree of strong water flooding in an experimental area of an oil reservoir according to the present disclosure.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040]The following provides further explanation of the technical solution of the present disclosure through the accompanying drawings and embodiments.

[0041]Unless otherwise defined, technical or scientific terms used in the present disclosure shall have the usual meanings as understood by those skilled in the art to which the present disclosure belongs.

Embodiment 1

[0042]As shown in FIG. 1, the present disclosure provides a system for intelligently recovering water-flooded layer of oil reservoir based on spiking neural network. The system includes: a data acquiring module, a data preprocessing module, a model training module, a resistivity recovering module, and a water-flooded layer interpreting module.

[0043]The data acquiring module is configured to acquire data of multi-modal well logging curves to characterize dynamic evolution laws of resistivity in water-flooded layer. This study selects the following multi-modal well logging curves to construct a dataset.

[0044]The conventional well logging curves include: natural gamma-ray (GR), acoustic interval transit time (AC), compensated neutron log (CNL), density (DEN), and self potential (SP), they are used to characterize static physical characteristics such as reservoir lithology, porosity, and shale content.

[0045]The electrical well logging curves include deep laterolog resistivity (R_D) and shallow laterolog resistivity (Rs), they are used to directly reflect dynamic changes in resistivity of the stratum affected by water flooding.

[0046]The data preprocessing module is configured to perform range normalization and Z-Score standardization on the acquired data. These two normalization methods eliminate dimensional differences and numerical range deviations of different well logging curves, improving the model's ability to capture physical meanings.

[0047]For values that are evenly distributed and have a relatively stable range in well logging curves, such as self potential (SP), gamma-ray (GR), etc. A range normalization method is adopted to preserve its original distribution form, making it more suitable for threshold based classification tasks. The range normalization is expressed as:

$x_{\mathrm{norm}}=\frac{x-x_{\min }}{x_{\max }-x_{\min }};$

where X_norm is a normalized data, x_min is a minimum value in original data, x_max is a maximum value in the original data, and x_max−x_min is a range of data.

[0048]For values in well logging curves that approximate Gaussian distribution, such as acoustic interval transit time (HAC), density (DEN) curves, etc. The Z-Score standardization method has strong robustness to outliers, and the standardized data distribution is more stable, making it more suitable for statistical inference scenarios. The Z-Score standardization is expressed as:

$x_{\mathrm{std}}=\frac{x-\mu }{\sigma };$

where x_std is a standardized data, μ is a mean of the original data, and σ is a standard deviation of the original data.

[0049]The model training module is configured to construct a resistivity recovering model based on a spiking neural network. The resistivity recovering model adopts a Leaky Integrate-and-Fire neuron model as a basic unit and receives the acquired data via a feedforward-loop hybrid structure to generate a prediction value of resistivity. The model training module is further configured to combine plasticity of synapses and global error backpropagation to adjust weights of the synapses to optimize performance of the resistivity recovering model, and evaluate and verify reliability of the resistivity recovering model.

[0050]A dynamic expression of the aforementioned Leaky Integrate-and-Fire neuron model is:

${\tau }_m\frac{\mathrm{dV}}{\mathrm{dt}}=-\left(V-V_{\mathrm{rest}}\right)+I_{\mathrm{syn}}\left(t\right);$

where V is a membrane potential, τ_m is a membrane time constant, V_rest is a resting potential, and I_syn(t) is an input current of the synapses.

[0051]Prediction of a resistivity is that when the membrane potential exceeds a threshold V_th, neurons emit a pulse and reset to V_reset. The feedforward-loop hybrid structure is used. An input layer of the resistivity recovering model receives the acquired data. A hidden layer of the resistivity recovering model contains pulse neuron nodes. An output layer of the resistivity recovering model generates the prediction value of resistivity via pulse frequency encoding.

[0052]For resistivity regression tasks, a supervised learning algorithm based on inner product of pulse sequences is adopted, and a multi-pulse error function is constructed according to the inner product of pulse sequences. The multi-pulse error function is expressed as:

$E={\sum }_i\left({\int }_0^T{\left(S_{\mathrm{target}}^{\left(i\right)}\left(t\right)-S_{\mathrm{output}}^{\left(i\right)}\left(t\right)\right)}^2\mathrm{dt}\right);$

where S_target and S_output are target pulse sequence and output pulse sequence, respectively.

[0053]Weights of the synapses are adjusted by combining plasticity of synapses and global error backpropagation. A specific expression for adjusting weights of the synapses is:

$\Delta w_{ij}=\eta ·{\sum }_t{\delta }_j\left(t\right)·S_i\left(t-\Delta t\right);$

where η is a learning rate, δ_j(t) is an error gradient, Δw_ij is a change in weights of the synapses, and S_i(t−Δt) is a pulse state before time Δt.

[0054]A sandstone oil reservoir block Lis taken as a research object, and well logging data of 11 wells are selected for the study. 80% of the well logging data are served as a training set, and remaining 20% of the well logging data are served as a validation set.

[0055]The spiking neural network demonstrates high accuracy and strong interpretability in the task of recovering resistivity in water-flooded layers through spatiotemporal information encoding and biologically inspired optimization strategies. Model evaluation should combine with multiple indicators such as RMSE, MAE, R², and rely on geological dynamic data to verify its reliability.

[0056]The mean deviation between the predicted resistivity and the measured value is evaluated using root mean square error (RMSE), and the formula is as follows:

$\mathrm{RMSE}=\sqrt{\frac{1}{n}{\sum }_{i=1}^n{\left(R_{\mathrm{pred}}^{\left(i\right)}-R_{\mathrm{true}}^{\left(i\right)}\right)}^2}.$

[0057]To reflect the presence of heterogeneous reservoirs or local water flooding anomalies, mean absolute error (MAE) is used, and the formula is as follows:

$\mathrm{MAE}=\frac{1}{n}{\sum }_{i=1}^n❘R_{\mathrm{pred}}^{\left(i\right)}-R_{\mathrm{true}}^{\left(i\right)}❘.$

[0058]The formula of coefficient of determination (R²) for measuring explanatory power of the model on changes in stratum resistivity is as follows:

$R^2=1-\frac{{\sum }_{i=1}^n{\left(R_{\mathrm{pred}}^{\left(i\right)}-R_{\mathrm{true}}^{\left(i\right)}\right)}^2}{{\sum }_{i=1}^n{\left(R_{\mathrm{pred}}^{\left(i\right)}-\stackrel{\_}{R_{\mathrm{true}}}\right)}^2}.$

[0059]The closer the value is to 1, the higher the fit of the model to key factors such as lithology and pore fluid.

[0060]RMSE is a root mean square error, which measures the deviation between the predicted value and the true value and is sensitive to outliers. MAE is a mean absolute error, which represents an average of absolute values of the prediction errors, and can intuitively reflect the actual situation of prediction error. R² is a coefficient of determination, representing the explanatory proportion of the model on the variance of the target variables.

$R_{\mathrm{pred}}^{\left(i\right)}$

is a predicted value of an i-th resistivity, and

$R_{\mathrm{true}}^{\left(i\right)}$

is a true value of the i-th resistivity.

[0061]As shown in FIGS. 2A-2B, 3A-3B, and 4A-4B, it is known that the Jurassic oil reservoir densification test area L9 in the study area is a low amplitude anticline structure. The reservoir X1 is in low porosity and low permeability, and the reservoir X2 is in medium porosity and low permeability. The average porosity of the reservoir X1 is 14.5%, and the average porosity of the reservoir X2 is 16.6%. The average permeability of the reservoir X1 is 2.1 mD, and the average permeability of the reservoir X2 is 29.9 mD. The reservoir X3 is in medium porosity and medium permeability, with the average porosity of 18.8% and the average permeability of 73.3 mD. The oil layer is controlled by structure and lithology, and has a relatively large thickness, averaging between 6 m and 10 m. The oiliness is influenced by physical properties, with X1 and X2 mainly being in the same layer of oil and water, and X3 mainly being in the oil layer. The oil reservoir is developed with edge water and bottom water. Different machine learning algorithms are designed to predict and verify the resistivity recovery of the group of wells in the experimental area.

[0062]As shown in FIG. 5, the SNN model has a prediction accuracy of R²=0.98, REMS=0.011, and MAE=0.08 in the training set, and the best performance in the validation set with R²=0.58, REMS=0.41, and MAE=0.40. The accuracy of recovering the original resistivity of the oil layer has been improved by 24% with the SVR regression model, and by 7% with the LightGBM algorithm model.

[0063]The resistivity prediction curves and true fitting results of resistivity of the SNN and LightGBM algorithm models for 8 test wells are shown in FIGS. 6A-6H. It can be seen from FIGS. 6A-6H that the SNN algorithm model achieves an average accuracy of 82.7% in resistivity recovery for multiple wells. This further confirms the significant advantages of spiking neural networks in dealing with complex and nonlinear issues. Especially in the application scenario of resistivity recovery, compared with other machine learning methods such as support vector machines, LightGBM algorithm, etc., spiking neural network has higher computational efficiency and stronger generalization ability, thereby providing more accurate prediction and recovery effects.

[0064]As shown in FIG. 7, a spiking neural network (SNN) is used to construct a resistivity recovering model, which outputs the recovered original resistivity curve (RT_original), providing basic data for subsequent oil saturation calculations. Based on the recovered original resistivity, the oil saturation (So) and water production rate (Fw) are further calculated, and the water-flooded layer is comprehensively determined through changes in oil saturation and resistivity, and water production rate of the oil layer section. The current oil saturation is calculated using an improved Archie formula and combining with the recovered original resistivity (RT_original) and the current resistivity (RT_current).

[0065]An expression of the oil saturation is:

$S_o=1-{\left(\frac{{\mathrm{RT}}_{\mathrm{original}}}{{\mathrm{RT}}_{\mathrm{current}}}·\frac{a·{\varphi }^{-m}}{{\sigma }_w}\right)}^{\frac{1}{n}};$

where S_o is a current oil saturation of a reservoir body; RT_original is the original resistivity; RT_current is a current resistivity; a, m, n are coefficients related to rock porosity structure and saturation, usually used to describe oil-water flow characteristics; ϕ is an effective porosity of stratum; and σ_w is a viscosity of water, and is measured in pascal seconds (Pa s).

[0066]The oil saturation evaluation model is:

$S_o=1-\sqrt{\frac{a·b·R_w}{{\phi }^m·R_t}};$

where a, b, m, n are dynamic correction coefficients, and σ_w is conductivity of the mixed liquid.

[0067]An expression for the water production rate is

$F_w=\frac{\frac{K_{\mathrm{rw}}}{{\mu }_w}}{\frac{K_{\mathrm{rw}}}{{\mu }_w}+\frac{K_{\mathrm{ro}}}{{\mu }_o}};$

where K_rw is a relative permeability of an aqueous phase, K_ro is a relative permeability of an oil phase, μ_w is a viscosity of the aqueous phase, and μ_o is a viscosity of the oil phase.

[0068]The water production rate evaluation model is

$f_w=1-\frac{1.0542}{\left(1+e^{\frac{S_w-0.4742}{0.049}}\right)}.$

[0069]As shown in Table 1, based on the results of water saturation and water production rate, ΔSo and ΔRT were used to quantitatively classify and interpret the water-flooded layer: the water-flooded layer is classified to be four levels: unflooded, weakly flooded, moderately-strongly flooded, and strongly flooded. By combining the resistivity curve shape with production dynamic data, water flooding patterns (such as bottom flooding, top flooding, and intra layer flooding) are identified, providing a basis for analyzing the waterflood front advance behavior.

[0070]To verify the generalization ability of the model, two new wells are selected to test and validate its generalization. The well logging data and dynamic production data of the new wells are input into the model to recovery the original resistivity and calculate the oil saturation and water production rate.

[0071]As shown in FIGS. 8-9, the water-flooded layers of wells LU3245 and LU3264 in the experimental area are explained and their results are verified. The specific error analysis is shown in Table 2.

[0072]By comparing the well logging interpretation results with the rock core analysis results, the relative error of porosity is between −3.26% and 3.70%, and the relative error of oil saturation is between 1.04% and 4.25%. This proves that the water flooding interpreting model has a high degree of conformity, and the rock core oil saturation analysis results are basically consistent with the water flooding interpretation.

[0073]Based on the constructed quantitative interpreting model of water-flooded layer, water-flooded layer interpretation is carried out for 8 encrypted wells in the encrypted adjustment area. The resistivity prediction curves and water-flooded layer interpretation results are shown in FIGS. 10A-10H.

[0074]As shown in FIGS. 11-12, the results indicate that the average thickness of the strong water-flooded layer in the encrypted test area is 2.8 m. The top is mainly composed of no water flooding and weak water flooding, with a remaining oil layer thickness of 1.0 m-3.5 m, gradually transitioning downwards to medium strong water flooding and strong water flooding, with a total thickness of 1.0 m-6.0 m. The bottom is the original water layer. The thickness range of the remaining no water flooding to medium water flooding at the top is 1.5 m-3.5 m, with an average of 2.8 m. The current oil saturation range above the original oil-water interface is 39.0%-63.0%, with an average of 47.3%.

[0075]The planar distribution of the thickness of the medium strong water-flooded layer is controlled by multiple mechanisms such as permeability differences in microfacies, structural driven dominant channels, and reservoir heterogeneity. The weak correlation between porosity and permeability leads to local high permeability bands, which preferentially forms dominant flow channels for injected water in high structural areas. The resistivity decreases by 30% to 50% (2.0<ΔRT≤4.0), and scattered plaques of the medium strong water flooding are presented on the plane view.

[0076]Therefore, the present disclosure adopts the aforementioned system for intelligently recovering water-flooded layer of oil reservoir based on spiking neural network. The system based on spiking neural network (SNN) is adopted to achieve high-precision recovery of the original resistivity of the water-flooded layer and effective interpretation of the water-flooded layer. The system obtains multi-modal well logging curves through the data acquiring module, normalizes and standardizes the multi-modal well logging curves through the preprocessing module, and constructs and optimizes the resistivity recovery model through the model training module. The resistivity recovering module recoveries the original resistivity of the oil layer based on this, while the water-flooded layer interpreting module further calculates the oil saturation and water production rate, achieving quantitative classification and interpretation of the water-flooded layer.

[0077]Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present disclosure and not to limit it. Although the present disclosure has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solution of the present disclosure, and these modifications or equivalent substitutions cannot make the modified technical solution deviate from the spirit and scope of the technical solution of the present disclosure.