Multi-reservoir parameter prediction method of constrained multi-task deep learning model

Abstract

The invention relates to a multi-reservoir parameter prediction method based on a constrained multi-task deep learning model, and the method comprises the steps: obtaining multi-source seismic information, inputting the multi-source seismic information into a pre-trained multi-task deep learning model, and obtaining reservoir parameter prediction results of porosity, gas saturation and lithology; the multi-task deep learning model comprises a shared feature extractor and at least three task exclusive sub-networks; the process of obtaining the reservoir parameter prediction result by the multi-task deep learning model comprises the following steps: inputting multi-source seismic information into a shared feature extractor, extracting shared high-dimensional geological features suitable for a plurality of reservoir parameter prediction tasks by using the shared feature extractor, and combining one-dimensional and two-dimensional distribution constraints of data; and processing and sharing high-dimensional geologic features by using each task exclusive sub-network, and predicting reservoir parameters. Compared with the prior art, the method has the advantages that the limitation of a traditional single-task deep learning model in processing coupling geological parameters can be overcome, and collaborative prediction of a plurality of key reservoir parameters is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data interpretation, and in particular to a multi-reservoir parameter prediction method based on a constrained multi-task deep learning model. Background Art

[0002] Seismic prediction of reservoir heterogeneity based on seismic data has important application value in oil and gas resource exploration, reservoir development, and geological modeling. This is because the porosity, gas saturation, and lithologic characteristics of underground rocks are the basic physical properties that describe underground reservoirs and are crucial for fluid reserve assessment, reservoir quality evaluation, and reservoir unit division. However, in the actual process of seismic reservoir parameter prediction, it is usually necessary to find the most reasonable physical model by minimizing the error between the seismic response and the observed data. This optimization-based seismic inversion algorithm often has difficulty achieving rapid convergence, and the error propagation effect will ultimately affect the accuracy of reservoir parameter prediction, thus posing challenges to reservoir evaluation and development decisions.

[0003] Deep learning methods are widely used to explore the complex relationship between seismic data and reservoir parameters due to their powerful nonlinear mapping and feature extraction capabilities. However, most existing deep learning methods usually regard the prediction of each reservoir parameter as an independent task. For example, parameters such as lithology, porosity and gas saturation are often trained and predicted separately through different network models. Chinese patent CN111507048B discloses a method, device, equipment and system for predicting the gas content of tight sandstone reservoirs, which are used to improve the accuracy of convolutional neural networks in predicting tight sandstone reservoirs based on pre-stack seismic angle gather data. However, in fact, the seismic response is not determined by a single reservoir parameter alone, but is the result of the interaction of multiple reservoir parameters and geological characteristics. These parameters and geological characteristics influence and couple with each other, and the prediction results of certain parameters may have a chain reaction on the prediction of other parameters.

[0004] Furthermore, an ideal network model should be able to establish a mapping relationship between labels and input attributes based on limited data, and demonstrate good generalization capabilities in unlabeled three-dimensional regions. At the same time, the prediction results must be geologically reasonable and adhere to basic physical laws. Even within a multi-task learning framework, a purely data-driven approach often makes it easy for the model to ignore the constraints of reservoir geology and physical characteristics. Furthermore, due to the quality limitations of the collected data samples, simultaneous predictions of multiple tasks often result in parameter crosstalk. Therefore, within a multi-task deep learning framework, it is necessary to combine geological and data information from more sources, such as spatial information and data distribution priors, to ensure that the physical connections and distribution patterns between the multiple predicted parameters meet the existing geological priors. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a multi-reservoir parameter prediction method based on a constrained multi-task deep learning model.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A multi-reservoir parameter prediction method based on a constrained multi-task deep learning model, the method comprising:

[0008] Acquire multi-source seismic information and input it into a pre-trained multi-task deep learning model to obtain reservoir parameter prediction results for porosity, gas saturation, and lithology; the multi-task deep learning model includes a shared feature extractor and no fewer than three task-specific subnetworks;

[0009] The process of obtaining reservoir parameter prediction results using the multi-task deep learning model includes: inputting the multi-source seismic information into a shared feature extractor, using the shared feature extractor to extract shared high-dimensional geological features suitable for multiple reservoir parameter prediction tasks; using each task-specific subnetwork to process the shared high-dimensional geological features and predict the porosity, gas saturation and lithologic reservoir parameters.

[0010] Furthermore, the multi-source seismic information includes pre-stack seismic gathers, post-stack seismic records, P-wave impedance, P-wave velocity ratio, seismic sensitivity attributes, and time-space coordinate information.

[0011] Furthermore, the training process of the multi-task deep learning model includes:

[0012] Porosity, gas saturation, and lithology data were collected using well logging, combined with multi-source seismic information and preprocessed to construct a training dataset.

[0013] Building a multi-task deep learning model, and training the multi-task deep learning model based on a multi-task loss function, a two-dimensional joint distribution consistency constraint, and a physical association constraint;

[0014] After training, a blind well test is performed on the multi-task deep learning model to evaluate its prediction performance at unseen well locations, and the multi-task deep learning model is optimized based on the evaluation results.

[0015] Furthermore, the pre-processing includes:

[0016] Based on the established well-seismic relationship, the porosity, gas saturation and lithology data obtained from well logging, as well as multi-source seismic information, are converted into time and depth, and uniformly resampled to a millisecond-level time domain sampling interval.

[0017] Synchronously interpolate multi-source seismic information and target reservoir parameter labels to millisecond sampling intervals;

[0018] The porosity and gas saturation logging curves are smoothed using a millisecond sliding window;

[0019] Taking each sampling point on the logging trajectory as the center, a sliding window method is used to extract the corresponding 3D geological slice within its neighborhood, and a set of sample pairs suitable for the 3D sequence-to-sequence mapping paradigm is constructed. The input features include processed multi-source seismic information, and the output features include processed porosity, gas saturation, and lithology data.

[0020] The processed porosity, gas saturation and lithology data as well as multi-source seismic information were scaled to the range of 0-1 using the minimum-maximum normalization method.

[0021] Furthermore, the expression of the multi-task loss function is:

[0022]

[0023] Among them, N represents the total number of samples, MSE represents the mean square error loss function, CrossEntropy represents the cross entropy loss function, and P pi 、P si and P li Represent the porosity, gas saturation and lithology prediction values ​​of the i-th sample, Y pi 、Y si and Y li Represent the true labels corresponding to the porosity, gas saturation and lithology prediction values ​​of the i-th sample, respectively, W p 、W s and W l represent the loss function weights of porosity, gas saturation and lithology prediction tasks respectively.

[0024] Furthermore, the expression of the two-dimensional joint distribution consistency constraint is:

[0025]

[0026] Where χ represents the set of all predicted porosity and gas saturation values, (p p ,s p ) represents the specific predicted porosity and gas saturation value, ξ represents the set of all true porosity and gas saturation values, (p t ,s t ) represents the specific true porosity and gas saturation values, P p and P t P pred (p p ,s p ) and P true (p t ,st ), which represents the probability density distribution of porosity and gas saturation between predicted and true values, γ(p p ,s p ;p t ,s t ) is the joint distribution matrix, which describes the predicted porosity and gas saturation values ​​(p p ,s p ) to the true porosity and gas saturation value (p t ,s t ) mapping relationship, c(p p ,s p ;p t ,s t ) is the cost function that measures the square of the Euclidean distance between the predicted porosity and gas saturation values ​​and the true values, and A is the regularization term used to improve computational efficiency and alleviate numerical instability;

[0027]

[0028]

[0029] Among them, ε is the regularization weight, p p 、s p 、p t and s t Represent the predicted porosity, gas saturation, true porosity and gas saturation respectively, P pred and P true Representing the two-dimensional probability density functions of porosity and gas saturation of predicted and true values, μ pp 、μ sp , σ pp and σ sp Represents the mean and standard deviation of predicted porosity and gas saturation, μ pt 、μ st , σ pt and σ st represents the mean and standard deviation of the corresponding true porosity and gas saturation, ρ p and ρ s represent the correlation coefficients between the predicted porosity and saturation and the true porosity and saturation, respectively.

[0030] Furthermore, the expression of the physical association constraint is:

[0031] Sat=23.48×In(Por)-0.82×Por+40.06

[0032] Among them, Por and Sat represent porosity and gas saturation, respectively.

[0033] Furthermore, when using well logging to collect the porosity, gas saturation and lithology data, geological time and space constraint information is also introduced. The introduction process includes:

[0034] The absolute position information corresponding to each sampling point is extracted from the original multi-source seismic information. The absolute position information includes three-dimensional coordinates along the survey line direction, cross-survey line direction, and time dimension. The relative time information is calculated based on the three key seismic horizons interpreted by experts, and the time position or depth-related position information of each sampling point in its corresponding stratum is quantified. The quantified expression is:

[0035]

[0036] Among them, time in,x,t Indicates the current time sampling point, {inline, xline, time} is the sample location set within the study area, horizon i+1 and horizon i Respectively represent the current time sampling point time in,x,t The upper and lower seismic horizons of the stratum in which it is located.

[0037] Furthermore, when training the multi-task deep learning model, the two-dimensional joint distribution consistency constraint and the physical association constraint are integrated with the multi-task loss function to obtain a joint optimization objective function with multiple constraint mechanisms. The multi-task deep learning model is trained. The expression of the joint optimization objective function is:

[0038] L total =L Multi +W d L W λ d

[0039] Among them, W d represents the weight of the joint data distribution constraint, λ d is the scaling factor, L W Represents a constraint.

[0040] Furthermore, the UNet network is adopted as the shared feature extractor.

[0041] According to another aspect of the present invention, there is provided a

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. Based on a multi-task deep learning model, the present invention can fully decouple the complex mapping relationship between seismic response and multiple reservoir parameters, and deeply explore the statistical similarity and physical consistency between parameters, so as to achieve collaborative prediction of key reservoir parameters such as porosity, gas saturation and lithology, and overcome the limitations of traditional single-task deep learning models in processing coupled geological parameters. In terms of model architecture design, the present invention adopts the UNet network as a shared feature extractor in multi-task learning to mine the shared features between different reservoir parameters (such as porosity, gas saturation and lithology). Subsequently, the extracted shared features are input into Each task-specific sub-network realizes the collaborative seismic prediction of multiple reservoir parameters. The seismic response is essentially a comprehensive reflection of multiple reservoir physical parameters and geological characteristics, and there is a complex physical coupling relationship between these parameters. Multi-task learning can effectively improve the model's decoupling of reservoir parameters under complex geological conditions by sharing the underlying feature space and utilizing the correlation between tasks. The present invention not only improves the prediction accuracy of a single reservoir parameter, but also optimizes the prediction results from multiple dimensions such as statistical consistency, physical self-consistency, and spatial continuity, providing efficient and reliable technical support for reservoir evaluation, geological modeling, and development decisions in oil and gas exploration.

[0044] 2. The present invention uses pre-stack seismic records, post-stack seismic data, seismic inversion results and sensitive seismic attributes as multi-source input features to construct an input data volume with rich geological information.

[0045] 3. In the preprocessing stage, the present invention takes each sampling point on the logging trajectory as the center, and uses a sliding window method to extract the corresponding three-dimensional geological body slices within its neighborhood, constructing a set of sample pairs suitable for the three-dimensional sequence-to-sequence mapping paradigm. It innovatively constructs a multidimensional spatiotemporal dataset that integrates geological time, stratigraphic and spatial dimensions, more realistically simulates the spatial distribution characteristics of actual geological structures, enhances the model's ability to understand and generalize the spatial distribution characteristics of multiple reservoir parameters under complex geological conditions, and provides a new modeling idea and technical path for reservoir parameter prediction tasks.

[0046] 4. The present invention innovatively introduces two-dimensional joint distribution consistency constraints and physical correlation constraints. The two-dimensional joint distribution consistency constraints ensure the statistical consistency of the prediction results in the joint distribution space, while the physical correlation constraints construct deterministic associations between parameters. This composite constraint strategy organically combines mathematical statistical constraints with physical law constraints, and while improving the prediction accuracy of multi-reservoir parameters, it systematically improves the physical consistency between multi-parameter predictions.

[0047] 5. The model of the present invention can output the three-dimensional distribution results of porosity, gas saturation and lithology at one time. Compared with the traditional parameter-by-parameter inversion method, the calculation efficiency is improved, providing fast and comprehensive data support for the formulation of reservoir development plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Flow chart of the method of the present invention;

[0049] Figure 2 This is a flowchart of the training and application of the multi-task deep learning model in the present invention;

[0050] Figure 3 A comparison chart of cross-validation prediction results of a multi-task deep learning model and other single-task learning models on five blind well data in one embodiment of the present invention;

[0051] Figure 4 Schematic diagram of porosity prediction results based on multi-task deep learning model;

[0052] Figure 5 Schematic diagram of porosity prediction results based on single-task learning model;

[0053] Figure 6 Schematic diagram of gas saturation prediction results based on the multi-task deep learning model;

[0054] Figure 7 Schematic diagram of gas saturation prediction results based on the single-task learning model;

[0055] Figure 8 Schematic diagram of lithology prediction results based on the multi-task deep learning model;

[0056] Figure 9 Schematic diagram of lithology prediction results based on the single-task learning model. DETAILED DESCRIPTION

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0058] Example 1

[0059] This embodiment discloses a multi-reservoir parameter prediction method based on a constrained multi-task deep learning model. Figure 1 Shown, including:

[0060] Step S1, obtaining multi-source earthquake information;

[0061] Step S2: inputting the multi-source seismic information into a shared feature extractor of a multi-task deep learning model, and using the shared feature extractor to extract shared high-dimensional geological features applicable to multiple reservoir parameter prediction tasks;

[0062] Step S3: Use the task-specific sub-networks of the multi-task deep learning model to process shared high-dimensional geological features and predict porosity, gas saturation, and lithology prediction results.

[0063] The multi-task deep learning model includes a shared feature extractor and at least three task-specific subnetworks. A UNet network is used as the shared feature extractor of the multi-task deep learning model. The task-specific subnetworks of the multi-task deep learning model include a porosity subnetwork, a gas saturation subnetwork, and a lithology subnetwork. Each task-specific subnetwork comprises an input layer, a nonlinear transformation layer, a classification layer, and an output layer. The shared high-dimensional geological features are input into each task-specific subnetwork and then processed sequentially by the input layer, nonlinear transformation layer, classification layer, and output layer to output porosity, gas saturation, or lithologic reservoir parameters.

[0064] In this embodiment, multi-source seismic information is obtained through seismic exploration technology (data acquisition, inversion, and attribute extraction). The multi-source seismic information includes pre-stack seismic traces, post-stack seismic records, longitudinal wave impedance, longitudinal-to-spherical wave velocity ratio, seismic sensitivity attributes, and time-space coordinate information.

[0065] Well logging technology directly measures the physical response of the formation around the wellbore by lowering sensors (logging instruments) into the borehole to obtain logging data. The logging data includes high-precision reservoir parameters such as porosity, gas saturation, and lithology, which serve as label data for model training.

[0066] In this embodiment, the multi-task deep learning model is a multi-task learning multi-reservoir parameter prediction model based on joint data distribution and physical constraints. Its training and application process is as follows: Figure 2 As shown in the figure, it includes: seismic data preprocessing, 3D sample construction, multi-task network construction, introduction of joint data distribution and physical constraint mechanism, and final 3D reservoir parameter spatial distribution prediction. The specific steps are described as follows:

[0067] 1. 3D input of multi-source seismic information and construction of training labels

[0068] Porosity, gas saturation, and lithology data were collected using well logging technology and preprocessed with multi-source seismic information from the entire 3D work area to construct a training dataset. The multi-source seismic information includes pre-stack seismic gathers, post-stack seismic records, seismic attributes, seismic inversion results, and seismic spatiotemporal information (Xline, Inline, time, and relative time).

[0069] Preprocessing includes:

[0070] Based on the established well-seismic relationship, multi-source seismic information and porosity, gas saturation, and lithology data obtained from well logging were time-to-depth converted and uniformly resampled to a 2-millisecond time domain sampling interval to ensure that the porosity, gas saturation, and lithology data were consistent with the multi-source seismic information in the temporal dimension.

[0071] To increase the amount of sample data for deep learning training, multi-source seismic information and target reservoir parameter labels are synchronously interpolated to a 1 millisecond sampling interval;

[0072] Considering the scale difference between logging reservoir parameters and seismic main frequency, the logging curves of porosity and gas saturation are smoothed using a 20-millisecond sliding window to reduce high-frequency noise interference and achieve matching with the seismic response frequency band;

[0073] Considering that seismic response is essentially a spatial combination of underground 3D geological features, a 3D sequence-to-sequence mapping paradigm is used to construct each training sample into a small 3D volume with spatial context information. Specifically, a 3D seismic volume of size 3×3×19 is extracted within the neighborhood of each sampling point on the logging trajectory. Three traces are taken in the inline and Xline directions, and 19 sampling points are taken in the time direction, forming an input sample containing seismic trace gathers and their spatial position information.

[0074] To accelerate network convergence, the minimum-maximum normalization method is used to scale the processed porosity, gas saturation, lithology data, and multi-source seismic information to the range of 0-1.

[0075] In order to enhance the modeling capability of the temporal evolution characteristics and spatial distribution patterns of seismic data, geological temporal and spatial constraint information is introduced in the preprocessing stage. The specific process is as follows:

[0076] The absolute position information corresponding to each sampling point is extracted from the original seismic data volume, including the three-dimensional coordinates along the survey line (inline), cross-surface direction (xline), and time dimension. Relative time information is also calculated to characterize the relative depositional sequence and geological age relationship of the strata in the time domain. Relative time information is calculated based on three key seismic horizons interpreted by experts to quantify the time position or depth-related position information of each sampling point in its corresponding stratum. The quantification expression is:

[0077]

[0078] Among them, time in,x,t Indicates the current time sampling point, {inline, xline, time} is the sample location set within the study area, horizon i+1 and horizon iRespectively represent the current time sampling point time in,x,t The upper and lower seismic horizons of the stratum in which it is located.

[0079] 2. Multi-task learning network architecture construction

[0080] A multi-task deep learning model was constructed, consisting of a shared feature extractor and three task-specific sub-networks. A UNet network was used as the shared feature extractor in multi-task learning to mine shared features between different reservoir parameters (such as porosity, gas saturation, and lithology). Seismic response is essentially a comprehensive reflection of multiple reservoir physical parameters and geological characteristics, and these parameters are complexly physically coupled. By sharing the underlying feature space and leveraging inter-task correlations, multi-task learning can effectively improve the model's decoupling of reservoir parameters under complex geological conditions.

[0081] In the method of this embodiment, the multi-task deep learning model actually builds a shared-private hybrid neural network architecture based on UNet within the multi-task deep learning framework to achieve collaborative prediction of multiple reservoir parameters (such as porosity, gas saturation, and lithology). During the model training phase, the network calculates the error loss between each task's predicted output and the corresponding true label, and uses a weighted loss function to balance the training contribution of each task.

[0082] The expression of the multi-task loss function is:

[0083]

[0084] Where N represents the total number of samples, MSE represents the mean square error loss function, and CrossEntropy represents the cross entropy loss function. In the formula, for parameter prediction tasks (such as porosity and gas saturation), the mean square error (MSE) loss function is used; for the classification task of lithology identification, the cross entropy loss function is used to evaluate the error, P pi 、P si and P li Represent the porosity, gas saturation and lithology prediction values ​​of the i-th sample, Y pi 、Y si and Y li Represent the true labels corresponding to the porosity, gas saturation and lithology prediction values ​​of the i-th sample, respectively, W p 、W s and W l Represent the loss function weights of porosity, gas saturation and lithology prediction tasks respectively. To accelerate the parameter optimization process and narrow the search space, we perform a grid search on the weight of each task in the range of {0.3, 0.5, 0.7, 0.9}. The final weights are: porosity W p=0.3, gas saturation W s =0.3, lithology W l =0.7. In addition, the scaling factor λ is introduced p =0.01,λ s =1 and λ l = 0.005 is used to adjust the loss functions of porosity, gas saturation, and lithology tasks, respectively, to ensure that the three losses are at the same level, thus avoiding the loss of a certain task dominating the overall optimization process.

[0085] 3. Construction of Consistency Constraints for Two-Dimensional Joint Data Distribution

[0086] To enhance the generalization ability of multi-task deep learning models under complex geological conditions, a joint data distribution consistency constraint mechanism is introduced. The two-dimensional Wasserstein distance between the predicted output and the label in porosity-gas saturation is calculated as a distribution matching metric and incorporated into the loss function to guide the network to learn prediction results that are more consistent with geostatistical laws. Specifically, all samples are extracted from the three-parameter prediction results output by the network, and only the center point of each three-dimensional sample is selected as the predicted value and label value. Assuming that the data follows a Gaussian distribution, the two-dimensional probability density functions of the porosity and gas saturation of the predicted value and the true label are calculated using the following formulas:

[0087]

[0088] Among them, p p 、s p 、p t and s t Represent the predicted porosity, gas saturation, true porosity and gas saturation respectively, P pred and P true Representing the two-dimensional probability density functions of porosity and gas saturation of predicted and true values, μ pp 、μ sp , σ pp and σ sp Represents the mean and standard deviation of predicted porosity and gas saturation, μ pt 、μ st , σ pt and σ st represents the mean and standard deviation of the corresponding true porosity and gas saturation, ρ p and ρ s represent the correlation coefficients between the predicted porosity and saturation and the true porosity and saturation, respectively. These parameters are calculated from the predicted values ​​and well labels in each batch of data.

[0089] The Wasserstein distance is then applied to quantify the difference between the predicted results and the true labels in the two-dimensional joint distribution space of porosity and gas saturation. The core idea of ​​this metric is to measure the minimum "transportation cost" required to transform one distribution into another, thus providing an efficient and intuitive metric for evaluating the similarity between two distributions. The expression for the two-dimensional joint distribution consistency constraint is:

[0090]

[0091] Where χ represents the set of all predicted porosity and gas saturation values, (P p ,s p ) represents the specific predicted porosity and gas saturation value, ξ represents the set of all true porosity and gas saturation values, (P t ,S t ) represents the specific true porosity and gas saturation values, p p and p t P pred (P p ,S p ) and p true (p t ,s t ), which represents the probability density distribution of porosity and gas saturation between predicted and true values, γ(p p ,s p ;p t ,s t ) is the joint distribution matrix, which describes the predicted porosity and gas saturation values ​​(p p ,s p ) to the true porosity and gas saturation value (p t ,s t ) mapping relationship, c(p p ,s p ;p t ,s t ) is the cost function that measures the square of the Euclidean distance between the predicted porosity and gas saturation values ​​and the true values, and A is the regularization term used to improve computational efficiency and alleviate numerical instability;

[0092]

[0093] Among them, ε is the regularization weight, which is set to 0.01.

[0094] 4. Construction of physical correlation constraints

[0095] To enhance the physical consistency of multi-task deep learning models in predicting multiple reservoir parameters, this paper introduces a physical constraint mechanism based on the empirical relationship between porosity and gas saturation. By analyzing the cross-plot of porosity and gas saturation data in the labels, the first-order empirical relationship between the two is obtained, that is, the expression of the physical correlation constraint is:

[0096] Sat=23.48×In(Por)-0.82×Por+40.06

[0097] Among them, Por and Sat represent porosity and gas saturation, respectively.

[0098] This first-order empirical relationship is then used as a fitting operator during network training to convert the gas saturation predicted by the multi-task network into the corresponding pseudo-porosity. This pseudo-porosity is then connected as an enhanced feature to the input of the porosity prediction subnetwork. This approach allows the model to predict porosity not only based on the original input seismic information but also incorporates physical correlation features derived from the gas saturation prediction results. This enhances the coupled modeling capabilities between parameter predictions and improves the accuracy and physical rationality of porosity predictions.

[0099] 5. Joint Optimization and Application of Multi-task Learning Models Integrating Data Distribution with Physical Constraints

[0100] When training a multi-task deep learning model, the two-dimensional joint distribution consistency constraint and the physical correlation constraint are integrated with the multi-task loss function to obtain a joint optimization objective function with multiple constraint mechanisms. The expression of the joint optimization objective function is:

[0101] L total =L Multi +W d L W λ d

[0102] Among them, W d represents the weight of the joint data distribution constraint, λ d is the scaling factor, L W Represents a constraint.

[0103] In this embodiment, after hyperparameter tuning, λ d =0.001, W d =0.3. After training, the trained multi-task deep learning model with multiple constraints was used to conduct a blind well test to evaluate its prediction performance at unseen well locations. The model was then applied to 3D seismic data volumes to achieve simultaneous prediction of three reservoir parameters: porosity, gas saturation, and lithology in 3D space.

[0104] Figure 3 This is a comparison chart of the cross-validation prediction results of the multi-task deep learning model based on the method in this embodiment and other single-task learning models on five blind well data. From left to right in each well are: single-task and multi-task porosity labels and prediction results, single-task and multi-task gas saturation labels and prediction results, lithology labels, single-task lithology prediction results, and multi-task lithology prediction results. Black arrows indicate areas with better multi-task predictions, gray arrows indicate areas where multi-task performance is slightly insufficient, and black boxes indicate areas where the multiple parameters of multi-task predictions have better self-consistency. By analyzing the statistical results and prediction results of single-task and multi-task, the effectiveness of the proposed constraint-based multi-task strategy is demonstrated, that is, the multi-task learning multi-reservoir parameter prediction strategy based on joint data distribution and physical constraints shows better performance in terms of prediction accuracy, spatial continuity and physical consistency of the three reservoir parameters.

[0105] Figure 4-Figure 9 Schematic diagram of three-parameter seismic profile prediction results based on single-task learning model and multi-task deep learning model: Figure 4 is the porosity prediction result (multi-task), Figure 5 is the porosity prediction result (single task), Figure 6 For gas saturation prediction results (multi-task), Figure 7 is the gas saturation prediction result (single task), Figure 8 For lithology prediction results (multi-task), Figure 9 Figure 1 shows the lithology prediction results (single-task), where wells W29 and W27 are blind wells. Gray circles indicate that the three parameters predicted by the multi-task method are more accurate and more consistent with the well labels. The gray irregular closed shape on the porosity prediction profile indicates that the multi-task prediction results have stronger geological continuity in this area. Black circles indicate that the multi-task porosity and gas saturation predictions are more consistent than those predicted by the single-task method. That is, low porosity corresponds to low gas saturation, while high porosity does not correspond to low gas saturation as predicted by the single-task method.

[0106] Example 2

[0107] Based on Example 1, this embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, and the one or more programs include instructions for executing the multi-reservoir parameter prediction method based on the aforementioned constraint-based multi-task deep learning model.

[0108] At the hardware level, the electronic device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the above-mentioned multi-reservoir parameter prediction method based on the constraint-based multi-task deep learning model. Of course, in addition to software implementation, the present invention does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc. In other words, the execution body of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0109] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0110] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0111] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A multi-reservoir parameter prediction method based on a constrained multi-task deep learning model, characterized in that: The method comprises: Acquire multi-source seismic information and input it into a pre-trained multi-task deep learning model to obtain reservoir parameter prediction results for porosity, gas saturation, and lithology; the multi-task deep learning model includes a shared feature extractor and no fewer than three task-specific subnetworks; The process of obtaining reservoir parameter prediction results using the multi-task deep learning model includes: inputting the multi-source seismic information into a shared feature extractor, using the shared feature extractor to extract shared high-dimensional geological features suitable for multiple reservoir parameter prediction tasks; using each task-specific subnetwork to process the shared high-dimensional geological features and predict the porosity, gas saturation and lithologic reservoir parameters.

2. The multi-reservoir parameter prediction method based on a constraint-based multi-task deep learning model according to claim 1, characterized in that: The multi-source seismic information includes pre-stack seismic gathers, post-stack seismic records, P-wave impedance, P-wave velocity ratio, seismic sensitivity attributes, and time-space coordinate information.

3. The multi-reservoir parameter prediction method based on a constraint-based multi-task deep learning model according to claim 1, characterized in that: The training process of the multi-task deep learning model includes: Porosity, gas saturation, and lithology data were collected using well logging, combined with multi-source seismic information and preprocessed to construct a training dataset. Building a multi-task deep learning model, and training the multi-task deep learning model based on a multi-task loss function, a two-dimensional joint distribution consistency constraint, and a physical association constraint; After training, a blind well test is performed on the multi-task deep learning model to evaluate its prediction performance at unseen well locations, and the multi-task deep learning model is optimized based on the evaluation results.

4. The multi-reservoir parameter prediction method based on a constraint-based multi-task deep learning model according to claim 3, characterized in that: The pretreatment includes: Based on the established well-seismic relationship, the porosity, gas saturation and lithology data obtained from well logging, as well as multi-source seismic information, are converted into time and depth, and uniformly resampled to a millisecond-level time domain sampling interval. Synchronously interpolate multi-source seismic information and target reservoir parameter labels to millisecond sampling intervals; The porosity and gas saturation logging curves are smoothed using a millisecond sliding window; Taking each sampling point on the logging trajectory as the center, a sliding window method is used to extract the corresponding 3D geological slice within its neighborhood, and a set of sample pairs suitable for the 3D sequence-to-sequence mapping paradigm is constructed. The input features include processed multi-source seismic information, and the output features include processed porosity, gas saturation, and lithology data. The processed porosity, gas saturation and lithology data as well as multi-source seismic information were scaled to the range of 0-1 using the minimum-maximum normalization method.

5. The multi-reservoir parameter prediction method based on a constraint-based multi-task deep learning model according to claim 3, characterized in that: The expression of the multi-task loss function is: Among them, N represents the total number of samples, MSE represents the mean square error loss function, CrossEntropy represents the cross entropy loss function, and P pi 、P si and P li Represent the porosity, gas saturation and lithology prediction values ​​of the i-th sample, Y pi 、Y si and Y li Represent the true labels corresponding to the porosity, gas saturation and lithology prediction values ​​of the i-th sample, respectively, W p 、W s and W l represent the loss function weights of porosity, gas saturation and lithology prediction tasks respectively.

6. The method for predicting multi-reservoir parameters based on a constraint-based multi-task deep learning model according to claim 3, characterized in that: The expression of the two-dimensional joint distribution consistency constraint is: Where χ represents the set of all predicted porosity and gas saturation values, (p p ,s p ) represents the specific predicted porosity and gas saturation value, ξ represents the set of all true porosity and gas saturation values, (p t ,s t ) represents the specific true porosity and gas saturation values, P p and P t P pred (p p ,s p ) and P true (p t ,s t ), which represents the probability density distribution of porosity and gas saturation between predicted and true values, γ(p p ,s p ;p t ,s t ) is the joint distribution matrix, which describes the predicted porosity and gas saturation values ​​(p p ,s p ) to the true porosity and gas saturation value (p t ,s t ) mapping relationship, c(p p ,s p ;p t ,s t ) is the cost function that measures the square of the Euclidean distance between the predicted porosity and gas saturation values ​​and the true values, and A is the regularization term used to improve computational efficiency and alleviate numerical instability; c(p p ,s p ;p t ,s t )=(p p -p t ) 2 +(s p -s t ) 2 Among them, ε is the regularization weight, p p 、s p 、p t and s t Represent the predicted porosity, gas saturation, true porosity and gas saturation respectively, P pred and P true Representing the two-dimensional probability density functions of porosity and gas saturation of predicted and true values, μ pp 、μ sp , σ pp and σ sp Represents the mean and standard deviation of predicted porosity and gas saturation, μ pt 、μ st , σ pt and σ st represents the mean and standard deviation of the corresponding true porosity and gas saturation, ρ p and ρ s represent the correlation coefficients between the predicted porosity and saturation and the true porosity and saturation, respectively.

7. The method for predicting multi-reservoir parameters based on a constraint-based multi-task deep learning model according to claim 3, characterized in that: The expression of the physical association constraint is: Sat=23.48×In(Por)-0.82×Por+40.06 Among them, Por and Sat represent porosity and gas saturation, respectively.

8. The method for predicting multi-reservoir parameters based on a constraint-based multi-task deep learning model according to claim 3, characterized in that: When using well logging to collect the porosity, gas saturation, and lithology data, geological temporal and spatial constraint information is also introduced. The introduction process includes: The absolute position information corresponding to each sampling point is extracted from the original multi-source seismic information. The absolute position information includes three-dimensional coordinates along the survey line direction, cross-survey line direction, and time dimension. The relative time information is calculated based on the three key seismic horizons interpreted by experts, and the time position or depth-related position information of each sampling point in its corresponding stratum is quantified. The quantified expression is: Among them, time in,x,t Indicates the current time sampling point, {inline, xline, time} is the sample location set within the study area, horizon i+1 and horizon i Respectively represent the current time sampling point time in,x,t The upper and lower seismic horizons of the stratum in which it is located.

9. The method for predicting multiple reservoir parameters based on a constraint-based multi-task deep learning model according to claim 3, characterized in that: When training the multi-task deep learning model, the two-dimensional joint distribution consistency constraint and the physical association constraint are integrated with the multi-task loss function to obtain a joint optimization objective function with multiple constraint mechanisms. The multi-task deep learning model is trained. The expression of the joint optimization objective function is: L 1otal =L Multi +W d L W λ d Among them, W d represents the weight of the joint data distribution constraint, λ d is the scaling factor, L W Represents a constraint.

10. The method for predicting multiple reservoir parameters based on a constraint-based multi-task deep learning model according to claim 1, characterized in that: The UNet network is used as a shared feature extractor of the multi-task deep learning model; the task-specific subnetworks of the multi-task deep learning model include a porosity subnetwork, a gas saturation subnetwork and a lithology subnetwork, and the task-specific subnetworks all include an input layer, a nonlinear transformation layer, a classification layer and an output layer. After the shared high-dimensional geological features are input into each task-specific subnetwork, they are processed in sequence by the input layer, the nonlinear transformation layer, the classification layer and the output layer to output porosity, gas saturation or lithology reservoir parameters.

Images