Oil and gas reservoir optimized mining method based on multi-modal data

Abstract

The invention relates to the technical field of petroleum and natural gas engineering, and discloses an oil and gas reservoir optimized mining method based on multi-modal data, which comprises the following steps: constructing an oil and gas reservoir multi-modal data acquisition system, seismic wave field data, logging interpretation data, production dynamic data, micro-seismic monitoring data, underground temperature and pressure time sequence data and shaft structure parameter data are obtained through the acquisition system; and performing space-time alignment processing on the acquired multi-modal data, establishing a unified geological coordinate system and a time reference, and eliminating data isomerism caused by different acquisition frequencies and spatial resolutions. A six-dimensional heterogeneous data acquisition system covering a seismic wave field, well logging interpretation, production dynamics, micro-seismic monitoring, an underground temperature and pressure time sequence and shaft structure parameters is constructed, so that the depiction precision of a reservoir porosity field, a permeability field, a saturation field and a pressure field is essentially improved.

Description

Technical Field

[0001] This invention relates to the field of oil and gas engineering technology, specifically to an optimized oil and gas reservoir development method based on multimodal data. Background Technology

[0002] Currently, the technical means used in the industry for reservoir characterization and development optimization are still mainly at a rudimentary stage, relying on the independent application or simple overlay of single-modal data. Specifically, existing technologies generally rely on seismic wavefield data for structural interpretation and reservoir delineation, obtain porosity, permeability, and saturation parameters at well points through well logging data, and analyze production decline patterns using production dynamic data. However, microseismic monitoring, downhole temperature and pressure time series data, and wellbore structural parameters are mostly used separately as independent monitoring methods. Although some studies in recent years have attempted to incorporate multi-source data into a unified analysis framework, the implementation methods are still limited to data extraction and format conversion based on standard templates, lacking in-depth exploration of the inherent physical correlation between seismic response, well logging sequence, and production dynamics. Even when machine learning algorithms are introduced, they are mostly supervised learning predictions for single data types, failing to establish a cross-modal collaborative characterization mechanism.

[0003] More critically, existing oil and gas reservoir models generally employ a static geological modeling approach, with seismic inversion and production history fitting performed independently and in separate steps. This results in long model update cycles and significant response lags, preventing key parameters such as reservoir porosity and permeability from reflecting dynamic evolution processes like water injection displacement, pressure deficits, and fracture propagation in real time. Such static models are severely inaccurate in depicting highly heterogeneous reservoirs such as thin interbedded layers and fractured reservoirs, leading to generally high recovery rate prediction errors. This directly causes problems such as unreasonable injection-production ratios in well groups and inaccurate decision-making regarding the on / off states of multi-stage fracturing stages.

[0004] Furthermore, the optimization process relies heavily on the experience and judgment of engineers. The optimization objectives are singular and lack quantitative coordination, making it difficult to achieve a Pareto optimal trade-off between maximizing recovery, minimizing extraction costs, and maximizing equipment lifespan. The decision-making process, from identifying geological changes to implementing engineering adjustments, often takes weeks or even months, resulting in significant production losses and inefficient energy cycles. Simultaneously, data silos are severe across different business units, with geological, engineering, economic, and environmental risk assessments being fragmented. The optimization schemes lack closed-loop feedback verification mechanisms, making it difficult to translate AI-generated control commands into executable well group production plans. Environmental risks such as surface subsidence, produced water backpressure, and methane leakage cannot be predicted, ultimately leading to a systemic defect of multidimensional imbalance between recovery benefits, economic costs, and environmental protection.

[0005] To address these issues, those skilled in the art have proposed an optimized oil and gas reservoir development method based on multimodal data. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an optimized oil and gas reservoir development method based on multimodal data, which solves the problems mentioned in the background section.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an optimized oil and gas reservoir development method based on multimodal data, comprising the following steps:

[0008] A multimodal data acquisition system for oil and gas reservoirs is constructed, through which seismic wavefield data, well logging interpretation data, production dynamic data, microseismic monitoring data, downhole temperature and pressure time series data, and wellbore structural parameter data are acquired;

[0009] The acquired multimodal data is spatiotemporally aligned to establish a unified geological coordinate system and time reference, thereby eliminating data heterogeneity caused by different acquisition frequencies and spatial resolutions.

[0010] The aligned multimodal data is input into a pre-trained multi-scale feature extraction network. The network adopts a three-branch parallel architecture to extract waveform features of seismic wavefield, vertical layering features of well logging data, and time-varying features of production dynamic data, respectively.

[0011] The extracted multi-scale features are mapped to a multimodal fusion transformer model, and a nonlinear correlation between seismic response and production dynamics is established through a cross-modal attention mechanism to generate digital feature vectors for oil and gas reservoirs.

[0012] Based on the aforementioned digital feature vectors, a parameterized digital twin of an oil and gas reservoir is constructed. The digital twin includes dynamic evolution models of the porosity field, permeability field, saturation field, and pressure field.

[0013] Multi-objective optimization calculations are performed in the digital twin. The optimization objectives include maximizing recovery rate, minimizing mining cost, and maximizing equipment life. The constraints cover bottom hole flowing pressure limit, water cut threshold, and formation sand production critical value.

[0014] The optimization results are converted into well group production parameter control commands, which include single well stroke frequency, pump hanging depth, injection-production well production ratio, and multi-stage fracturing section switching status.

[0015] Establish a closed-loop feedback mechanism to continuously input actual production data into the digital twin for assimilation and updates, thereby achieving rolling optimization of mining parameters.

[0016] Preferably, the multimodal data acquisition system further includes:

[0017] Deploy distributed fiber optic acoustic sensors in injection and production wellbores to acquire continuous inter-well seismic wave field data;

[0018] A logging-while-drilling toolchain was deployed in the horizontal well section to collect real-time resistivity, neutron porosity and gamma-ray energy spectrum data.

[0019] A multiphase flow meter and a high-frequency pressure transmitter are installed at the wellhead to collect second-level production dynamic data.

[0020] A microseismic monitoring array was deployed in the formation around the well to collect source mechanism solutions for microseismic events induced by fracturing and production.

[0021] Preferably, the spatiotemporal alignment process specifically includes:

[0022] A dynamic time warping algorithm is used to align the time series of seismic data and production dynamics data.

[0023] Kriging spatial interpolation was used to reconstruct well logging data into three-dimensional volume data with a resolution that matched the voxel size of the seismic data.

[0024] A geomechanical finite element model was established to calculate the geometric correction of the formation stress field to the wellbore measuring point location.

[0025] Preferably, the multi-scale feature extraction network is further configured as follows:

[0026] The seismic wavefield feature branch uses a three-dimensional convolution kernel to capture local patterns of the stratigraphic reflection structure, with the kernel size increasing along the depth direction;

[0027] The feature branch of well logging data uses a bidirectional long short-term memory network to capture the vertical dependencies of sequence stratigraphic units;

[0028] The production dynamics feature branch uses time-frequency convolution to extract the composite features of the output decline curve and the pressure derivative curve.

[0029] Preferably, the multimodal fusion transformer model is specifically implemented as follows:

[0030] Seismic features, well logging features, and production features are encoded into query vectors, key vectors, and value vectors, respectively.

[0031] Introduce a geological prior knowledge mask into the self-attention layer to block feature interaction paths that do not conform to sedimentary patterns;

[0032] Enhance the physical consistency representation between different modalities by using a cross-modal contrastive learning loss function;

[0033] The output fusion feature dimensions correspond to the number of oil and gas reservoir geological units, and each dimension characterizes the reservoir performance parameters of the corresponding unit.

[0034] Preferably, the parameterized digital twin of the oil and gas reservoir includes:

[0035] The digital feature vector is input into the physical information neural network, and the network is embedded with Darcy's law, the mass balance equation and the crack seepage equation as partial differential constraints.

[0036] A generative adversarial network architecture is adopted, in which the generator outputs porosity field and permeability field tensors, and the discriminator distinguishes the statistical differences between the generated field and the historical inversion field;

[0037] A discrete crack network model is implanted in a digital twin, and the crack aperture field is driven by the density field of microseismic events to evolve.

[0038] The pressure field solver employs an adaptive mesh refinement strategy, densifying the computation nodes in the well perimeter and fracture tip regions.

[0039] Preferably, the multi-objective optimization calculation adopts the following steps:

[0040] The recovery rate objective function is defined as the time integral of the cumulative oil production, the cost objective function includes electricity consumption, chemical injection and well workover costs, and the equipment life objective function is based on the rod and tubing combination fatigue cumulative damage model.

[0041] A non-dominated sorting genetic algorithm is used to search for the Pareto optimal solution set. The population size is proportional to the number of wells in the well group, and the crossover operator simulates the linear superposition effect of inter-well disturbances.

[0042] Uncertainty quantification is performed on the optimized solution set, and a Bayesian neural network is used to predict the success probability of each optimization scheme. The confidence interval is determined by multiple random simulations using a digital twin.

[0043] The objective solution is selected from the Pareto frontier based on the decision preference weight vector, which is dynamically adjusted by oil prices, policy subsidies, and carbon emission costs.

[0044] Preferably, the closed-loop feedback mechanism is specifically implemented as follows:

[0045] Lightweight digital twin replicas are deployed on edge computing nodes, and these replicas receive real-time wellhead data hourly for partial updates.

[0046] When the deviation between the predicted output and the actual output exceeds a preset threshold, the digital twin is triggered to undergo global retraining. The retraining process uses an incremental learning approach to retain historical knowledge.

[0047] Establish an expert review module for adjusting mining parameters, and transform the control commands generated by artificial intelligence into well group production allocation schemes that engineers can understand;

[0048] A blockchain architecture is used to store key optimization decisions, ensuring the traceability and tamper-proof nature of the mining adjustment process.

[0049] Preferably, the method further includes:

[0050] Before optimization, the quality of multimodal data is assessed, and the assessment indicators include the signal-to-noise ratio of seismic data, the repetition error of well logging data, and the integrity of production dynamic data. Modal data that are below the quality threshold are repaired through an adversarial data augmentation network.

[0051] An environmental risk pre-assessment was conducted on the optimized mining plan. The assessment indicators included surface subsidence, produced water backflow pressure, and methane leakage probability. If the risk exceeded the standard, the plan would automatically trigger an early warning and return to be re-optimized.

[0052] Generate knowledge increments for this optimization cycle. These knowledge increments include newly identified geological patterns, optimized hyperparameter combinations, and failed mining strategies. The knowledge increments are stored in an enterprise-level knowledge graph for subsequent well group calls.

[0053] This invention provides an optimized oil and gas reservoir development method based on multimodal data. It has the following beneficial effects:

[0054] 1. This invention constructs a six-dimensional heterogeneous data acquisition system encompassing seismic wavefield, well logging interpretation, production dynamics, microseismic monitoring, downhole temperature and pressure time series, and wellbore structural parameters. It employs a dynamic time warping algorithm and Kriging spatial interpolation to achieve spatiotemporal benchmark unification. Simultaneously, a three-branch parallel architecture multi-scale feature extraction network captures composite features of formation reflection structure, sequence stratigraphic unit vertical dependence, and production decline curves. Furthermore, a cross-modal attention mechanism (transformer model) incorporating geological prior knowledge masks establishes a nonlinear correlation between seismic response and production dynamics, ultimately generating digital feature vectors corresponding to the number of geological units in the oil and gas reservoir. This significantly improves the accuracy of reservoir porosity, permeability, saturation, and pressure fields, particularly enhancing the identification of thin interbedded and fractured reservoirs. Consequently, the recovery rate prediction error is controlled to a low level unattainable by traditional single-modal methods, effectively avoiding deviations in exploitation strategies due to inaccurate geological models.

[0055] 2. This invention employs a physical information neural network with embedded Darcy's law, mass balance equation, and fracture seepage equation as partial differential constraints to construct a parameterized digital twin. A discrete fracture network model driven by the density field of microseismic events is embedded within this model. An adaptive mesh refinement strategy is used to densify computational nodes around the well and at the fracture tip. Simultaneously, a non-dominated sorting genetic algorithm is deployed to perform Pareto front search for three objectives: maximizing recovery, minimizing extraction costs, and maximizing equipment lifespan. Furthermore, a Bayesian neural network is used to quantify the uncertainty of the optimized solution set. This allows extraction parameter adjustments to no longer rely on static geological models and human experience, but rather on real-time rolling optimization based on dynamic formation evolution. Especially in key decisions such as well group injection-production ratio and the on / off status of multi-stage fracturing sections, hourly responses are achieved, significantly shortening the decision-making cycle from geological change identification to engineering adjustment implementation. This solves the problems of production loss and ineffective energy circulation caused by the lag in traditional methods.

[0056] 3. This invention establishes a two-layer digital twin architecture between the cloud-based optimization decision-making layer and edge computing nodes. The lightweight edge replica undertakes millisecond-level anomaly diagnosis and high-frequency data assimilation tasks. When the deviation between predicted and actual production exceeds a preset threshold, global retraining in the cloud is triggered. The retraining process uses incremental learning to retain historical knowledge, while a blockchain architecture is used to store key optimization decisions to ensure traceability and tamper-proofing. Furthermore, environmental risk pre-assessments are conducted on the optimization scheme regarding surface subsidence, produced water backtracking pressure, and methane leakage probability. This creates a complete closed loop of data perception, intelligent decision-making, execution feedback, and knowledge accumulation throughout the entire mining process. This not only effectively extends the service life of downhole equipment and reduces the frequency of well repairs caused by fatigue-accumulated damage to the rod and pipe assembly, but also significantly reduces the amount of chemical injection and power consumption. It achieves a multi-dimensional balance between recovery benefits, economic costs, and environmental protection, overcoming the systemic defects of existing technologies, such as isolated operation of various business units, difficulty in implementing optimization results, and inability to predict environmental risks. Attached Figure Description

[0057] Figure 1 This is the overall flowchart of the present invention;

[0058] Figure 2 This is a flowchart of the closed-loop control optimization of the digital twin according to the present invention. Detailed Implementation

[0059] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Please see the appendix Figure 1 - Appendix Figure 2 This invention provides a method for optimized oil and gas reservoir development based on multimodal data, comprising the following steps:

[0061] Construct a multimodal data acquisition system for oil and gas reservoirs, and acquire seismic wavefield data, well logging interpretation data, production dynamic data, microseismic monitoring data, downhole temperature and pressure time series data, and wellbore structural parameter data through the acquisition system;

[0062] The multimodal data acquisition system further includes:

[0063] Deploy distributed fiber optic acoustic sensors in injection and production wellbores to acquire continuous inter-well seismic wave field data;

[0064] A logging-while-drilling toolchain was deployed in the horizontal well section to collect real-time resistivity, neutron porosity and gamma-ray energy spectrum data.

[0065] A multiphase flow meter and a high-frequency pressure transmitter are installed at the wellhead to collect second-level production dynamic data.

[0066] A microseismic monitoring array was deployed in the formation around the well to collect source mechanism solutions for microseismic events induced by fracturing and production.

[0067] Specifically, the backscattering Rayleigh effect of optical fibers is used to convert the seismic wave field between wells into optical phase changes, realizing continuous distributed acoustic field monitoring throughout the well section. This overcomes the limitation of sparse spatial sampling of traditional point geophones, enabling high-density capture of the propagation path of stress waves between wells and formation response characteristics. A logging-while-drilling toolchain is embedded in the horizontal well section, and resistivity, neutron porosity, and gamma-ray energy spectrum are acquired in real time based on the principle of electromagnetic induction. Geological steering and parameter measurement are synchronized with the drilling process, avoiding the lag of post-logging and enabling the horizontal well trajectory to be dynamically adjusted to accurately drill the target reservoir. A multiphase flow meter and a high-frequency pressure transmitter are deployed at the wellhead, and the Venturi and capacitance tomography fusion principle is used to achieve second-level resolution of the three-phase flow of oil, gas, and water. Combined with the piezoresistive effect, the instantaneous pulsation of the flowing pressure is captured, thus completely recording the dynamic fingerprint of the entire process from formation seepage to wellbore lifting.

[0068] Spatiotemporal alignment processing specifically includes:

[0069] A dynamic time warping algorithm is used to align the time series of seismic data and production dynamics data.

[0070] Kriging spatial interpolation was used to reconstruct well logging data into three-dimensional volume data with a resolution that matched the voxel size of the seismic data.

[0071] A geomechanical finite element model was established to calculate the geometric correction of the formation stress field to the wellbore measuring point location.

[0072] Specifically, a dynamic time warping algorithm is employed, based on dynamic programming, to nonlinearly and elastically match the event-driven sequence of seismic data sampled at millisecond levels with the continuous process sequence of production dynamics sampled at seconds. By finding the optimal time mapping path with the minimum cumulative distance, the fracturing-induced seismic events and the production pressure response are precisely aligned on the physical causal chain, thus overcoming the limitation of traditional linear interpolation in capturing asynchronous time delay effects. Subsequently, Kriging spatial interpolation is used to quantitatively characterize the spatial correlation of well logging data using the variogram function, reconstructing discrete logging points on the one-dimensional well trajectory to perfectly match the voxel size of the seismic data. The three-dimensional regular grid data enables a resolution scale transition from local wellbore measurements to reservoir-level macroscopic characterization. Simultaneously, a geomechanical finite element model is established to solve the deformation effect of the formation stress field on the wellbore structure based on the principle of virtual work. The geometric displacement correction amount caused by factors such as casing bending and wellbore collapse at the measuring points is calculated, and the spatial coordinates of the logging data are corrected in reverse to ensure that the data of each mode still share a unified absolute geographic reference frame under the disturbance of geostress. The three progressively connect to finally generate a multimodal dataset with strict spatiotemporal reference alignment and physical self-consistency, which eliminates the pseudo-correlation noise caused by spatiotemporal misalignment for subsequent cross-modal feature fusion.

[0073] The acquired multimodal data is spatiotemporally aligned to establish a unified geological coordinate system and time reference, thereby eliminating data heterogeneity caused by different acquisition frequencies and spatial resolutions.

[0074] The aligned multimodal data is input into a pre-trained multi-scale feature extraction network. The network adopts a three-branch parallel architecture to extract waveform features of seismic wavefield, vertical layering features of well logging data, and time-varying features of production dynamic data, respectively.

[0075] The extracted multi-scale features are mapped to a multimodal fusion transformer model, and a nonlinear correlation between seismic response and production dynamics is established through a cross-modal attention mechanism to generate digital feature vectors for oil and gas reservoirs.

[0076] Specifically, the acquired multimodal data is first spatiotemporally aligned. By establishing a unified geological coordinate system, the spatial sampling grid of the seismic wavefield, the well trajectory coordinates of the logging data, and the wellhead nodes of the production dynamics are incorporated into the same three-dimensional geographic analysis framework. Simultaneously, a time reference synchronization mechanism is used to map millisecond-level seismic events, second-level production time series, and hourly logging operations to a unified time axis. Dynamic time warping and spatial interpolation algorithms are used to eliminate data heterogeneity caused by different acquisition frequencies and spatial resolutions, ensuring comparability and superposition of the various modal data in the spatiotemporal dimensions. Then, the aligned data is input into a pre-trained multi-scale feature extraction network. This network employs a three-branch parallel architecture. The seismic branch captures waveform amplitude, frequency, and phase reflection features through three-dimensional convolution to characterize the stratigraphic structure, while the logging branch uses a bidirectional recurrent structure to extract features. The lithological cycles and gradual changes in physical properties of vertical sequence units are analyzed. For the production branch, time-frequency analysis is used to extract the time-varying laws of production decline, pressure derivative, and water cut increase, achieving independent decoupling and deep encoding of the physical meaning of each mode. Finally, the extracted multi-scale features are mapped to a multi-modal fusion transformer model. The mutual information weight between seismic features and production features is calculated through a cross-modal attention mechanism, and a nonlinear causal relationship between seismic response anomalies and production mutations is dynamically established. At the same time, geological prior knowledge is introduced to mask out false associations that do not conform to sedimentary laws, and finally, a digital feature vector of the oil and gas reservoir is generated. Each dimension of the vector corresponds to a geological unit, comprehensively representing the static physical parameters and dynamic production response characteristics of the reservoir, providing a high-dimensional abstract representation with physical consistency for the subsequent construction of digital twins.

[0077] The multimodal fusion transformer model is specifically implemented as follows:

[0078] Seismic features, well logging features, and production features are encoded into query vectors, key vectors, and value vectors, respectively.

[0079] Introduce a geological prior knowledge mask into the self-attention layer to block feature interaction paths that do not conform to sedimentary patterns;

[0080] Enhance the physical consistency representation between different modalities by using a cross-modal contrastive learning loss function;

[0081] The output fusion feature dimensions correspond to the number of oil and gas reservoir geological units, and each dimension characterizes the reservoir performance parameters of the corresponding unit.

[0082] Specifically, seismic features, well logging features, and production features are encoded into query vectors, key vectors, and value vectors, respectively. The query vector carries the geological target features currently of interest, the key vector represents the feature index of each modal data source, and the value vector stores specific physical parameter information. The similarity score between the query and key vectors is used to adaptively allocate the weight of the impact of seismic response anomalies on production dynamics. Simultaneously, a geological prior knowledge mask is introduced into the self-attention layer. A binary mask matrix is ​​constructed based on sedimentary facies distribution patterns and tectonic constraints to forcibly block false feature interaction paths that do not conform to geological genetic logic, thus ensuring that attention... The force weights focus only on the intermodal correlations with physical causal relationships. Then, by using a cross-modal contrastive learning loss function, the mutual information of positive sample pairs (different modal features of the same geological unit) is maximized, and the similarity of negative sample pairs (cross-modal features of different geological units) is minimized. This aligns heterogeneous features such as seismic impedance, logging porosity, and production pressure derivative to a unified physical consistency representation space. Finally, the output fused feature dimensions strictly correspond to the number of oil and gas reservoir geological units, and each dimension independently represents the composite parameters of the reservoir performance of the corresponding unit, realizing a precise mapping from multi-source heterogeneous data to the geological semantic vector space.

[0083] The multi-scale feature extraction network is further configured as follows:

[0084] The seismic wavefield feature branch uses a three-dimensional convolution kernel to capture local patterns of the stratigraphic reflection structure, with the kernel size increasing along the depth direction;

[0085] The feature branch of well logging data uses a bidirectional long short-term memory network to capture the vertical dependencies of sequence stratigraphic units;

[0086] The production dynamics feature branch uses time-frequency convolution to extract the composite features of the output decline curve and the pressure derivative curve.

[0087] Specifically, the seismic wavefield feature branch uses a three-dimensional convolution kernel to capture the spatial local patterns of formation reflection structures. The kernel size increases along the depth direction to match the larger receptive field requirements of longer wavelength seismic signals from deep strata, thereby effectively identifying gradient changes in structural anomalies such as faults, pinch-outs, and thin interbedded layers. The well logging data feature branch applies a bidirectional long short-term memory network, using forward and reverse cyclic structures to simultaneously capture the vertical sedimentary cycle dependence and diagenetic evolution correlation of sequence stratigraphic units, enabling the identification of geological boundaries of interbedded sandstone and mudstone and physical property rhythm segments to have bidirectional contextual memory capabilities. The production dynamics feature branch uses time-frequency convolution, and through wavelet transform, it simultaneously extracts the long-term trend characteristics of the production decline curve and the instantaneous spectral characteristics of the pressure derivative curve in the joint domain, realizing a multi-scale composite characterization from macroscopic production laws to microscopic seepage response. The parallel output of these three features constitutes a multi-dimensional feature base that characterizes the static geological structure and dynamic production behavior of oil and gas reservoirs.

[0088] A parameterized digital twin of an oil and gas reservoir is constructed based on digital feature vectors. The digital twin includes dynamic evolution models of porosity field, permeability field, saturation field and pressure field.

[0089] The parameterized digital twin of an oil and gas reservoir includes:

[0090] The digital feature vector is input into the physical information neural network, and Darcy's law, the mass balance equation and the crack seepage equation are embedded in the network as partial differential constraints.

[0091] A generative adversarial network architecture is adopted, in which the generator outputs porosity field and permeability field tensors, and the discriminator distinguishes the statistical differences between the generated field and the historical inversion field;

[0092] A discrete crack network model is implanted in a digital twin, and the crack aperture field is driven by the density field of microseismic events to evolve.

[0093] The pressure field solver employs an adaptive mesh refinement strategy, densifying the computation nodes in the well perimeter and fracture tip regions.

[0094] Specifically, firstly, the digitized feature vectors are input into a physical information neural network. This network embeds Darcy's law, the mass balance equation, and the fracture seepage equation as partial differential constraints in its loss function. This ensures that the network training process not only fits the observed data but also enforces the fundamental laws of seepage mechanics, thus preventing purely data-driven models from making predictions that violate physical principles in sparse data regions. Next, a generative adversarial network architecture is employed. The generator decodes and outputs porosity and permeability tensors based on the feature vectors, while the discriminator distinguishes between the generated field and the historical inversion field based on statistical characteristics such as the variogram and histogram distribution using the Wasserstein distance. Through adversarial training, the generated field incorporates geological statistical characteristics. Consistency and spatial continuity are achieved. A discrete fracture network model is embedded in the digital twin. The fracture aperture field is no longer statically set, but dynamically evolved by the microseismic event density field through tensor product operations, realizing real-time coupling between the fracturing effect and production-induced fracture propagation. The pressure field solver adopts an adaptive mesh refinement strategy, which automatically densifies the calculation nodes in the near-wellbore zone and the high pressure gradient region at the fracture tip based on posterior error estimation, while maintaining a coarse mesh in the homogeneous region. This reduces the computational complexity by an order of magnitude while ensuring the accuracy of the solution, ultimately forming a high-fidelity digital twin that can assimilate production data in real time, dynamically update reservoir status, and accurately predict development indicators.

[0095] Multi-objective optimization calculations are performed in the digital twin. The optimization objectives include maximizing recovery rate, minimizing extraction cost, and maximizing equipment life. The constraints cover bottom hole flowing pressure limit, water cut threshold, and formation sand production critical value.

[0096] The optimization results are converted into well group production parameter control commands, which include single well stroke frequency, pump hanging depth, injection-production well production ratio and multi-stage fracturing section switching status.

[0097] Establish a closed-loop feedback mechanism to continuously input actual production data into the digital twin for assimilation and updates, thereby achieving rolling optimization of mining parameters.

[0098] Specifically, when performing multi-objective optimization calculations in the digital twin, the objective function for maximizing oil recovery is designed as the time integral of cumulative oil production to characterize development benefits. The objective for minimizing extraction costs covers electricity consumption, chemical injection, and well workover costs to quantify economic input. The objective for maximizing equipment lifespan is based on the fatigue cumulative damage model of the rod and tubing assembly to predict pump inspection cycles. These three objectives constitute a mutually restraining Pareto front. At the same time, bottomhole flowing pressure limits are introduced to prevent formation energy depletion, water cut thresholds are introduced to avoid water channeling risks, and formation sand production thresholds are introduced to protect wellbore integrity, ensuring the engineering feasibility of the optimal solution. Subsequently, the Pareto optimal solution set is transformed into well group production parameter control commands through a parameter mapping engine. The stroke frequency is jointly solved by the production target and pump efficiency curve, the pump hanger depth is determined based on the dynamic fluid level and submersion constraints, the injection-production well production ratio is optimized and allocated through the inter-well interference coefficient, and the multi-stage fracturing stage switching status is dynamically sorted according to the fracture conductivity, realizing the semantic conversion from numerical solutions to engineering language.

[0099] The closed-loop feedback mechanism is specifically implemented as follows:

[0100] Lightweight digital twin replicas are deployed on edge computing nodes, and the replicas receive real-time wellhead data hourly for partial updates;

[0101] When the deviation between the predicted output and the actual output exceeds a preset threshold, the digital twin is triggered to undergo global retraining. The retraining process uses an incremental learning approach to retain historical knowledge.

[0102] Establish an expert review module for adjusting mining parameters, and transform the control commands generated by artificial intelligence into well group production allocation schemes that engineers can understand;

[0103] A blockchain architecture is used to store key optimization decisions, ensuring the traceability and tamper-proof nature of the mining adjustment process.

[0104] Multi-objective optimization calculations employ the following steps:

[0105] The recovery rate objective function is defined as the time integral of the cumulative oil production, the cost objective function includes electricity consumption, chemical injection and well workover costs, and the equipment life objective function is based on the rod and tubing combination fatigue cumulative damage model.

[0106] A non-dominated sorting genetic algorithm is used to search for the Pareto optimal solution set. The population size is proportional to the number of wells in the well group, and the crossover operator simulates the linear superposition effect of inter-well disturbances.

[0107] Uncertainty quantification is performed on the optimized solution set, and a Bayesian neural network is used to predict the success probability of each optimization scheme. The confidence interval is determined by multiple random simulations using a digital twin.

[0108] The objective solution is selected from the Pareto frontier based on the decision preference weight vector, which is dynamically adjusted by oil prices, policy subsidies, and carbon emission costs.

[0109] Specifically, the recovery rate objective function is defined as the time integral of the cumulative oil production, balancing short-term and long-term development benefits through an exponential discounting mechanism. The cost objective function integrates electricity consumption, chemical injection, and well workover costs to form a full life-cycle economic assessment. The equipment life objective function quantifies the reliable service life of the pumping equipment based on a rod-tube assembly fatigue cumulative damage model. These three factors constitute a conflicting Pareto front space. When using a non-dominated sorting genetic algorithm to search for the optimal solution set, the population size is proportional to the number of wells in the well group to ensure sufficient sampling of the decision variable space. The crossover operator uses a linear superposition strategy to simulate the physical coupling effect of pressure interference between injection and production wells, while the mutation operator randomly adjusts the fracturing section switching state to explore the nonlinear solution space, thereby quickly converging to the Pareto front that takes into account the three objectives. A Bayesian neural network is used to quantify the uncertainty of each optimization scheme, predicting the success probability of the scheme through the posterior distribution of network weights, and generating confidence intervals through multiple random simulations using a digital twin, making the reliability of the optimized solution set explicit.

[0110] The method also includes:

[0111] Before optimization, the quality of multimodal data is assessed, and the assessment indicators include the signal-to-noise ratio of seismic data, the repetition error of well logging data, and the integrity of production dynamic data. Modal data that are below the quality threshold are repaired through an adversarial data augmentation network.

[0112] An environmental risk pre-assessment was conducted on the optimized mining plan. The assessment indicators included surface subsidence, produced water backflow pressure, and methane leakage probability. If the risk exceeded the standard, the plan would automatically trigger an early warning and return to be re-optimized.

[0113] Generate knowledge increments for this optimization cycle. These knowledge increments include newly identified geological models, optimized hyperparameter combinations, and failed mining strategies. The knowledge increments are stored in an enterprise-level knowledge graph for subsequent well group access.

[0114] Specifically, when assessing the quality of multimodal data before optimization, a modal differentiation index is used. The seismic data signal-to-noise ratio is used to measure the recognizability of reflected signals, the repeatability error of well logging data is used to assess instrument stability, and the integrity of production dynamic data is used to check the packet loss rate during transmission. When any index falls below a preset quality threshold, an adversarial data augmentation network is triggered for repair. This network learns the spatiotemporal correlation distribution of high-quality data through a generator, and the discriminator distinguishes the statistical differences between real data and generated data. The game-like iteration between the two ensures that the repaired data retains the original physical trend while eliminating random noise, thereby avoiding low-quality data from misleading the optimization direction. When conducting environmental risk pre-assessment of the optimized mining scheme, a digital twin is used to simulate the surface subsidence, produced water backflow pressure, and methane leakage probability under different production schemes. Once the risk index exceeds the environmental carrying capacity threshold, an early warning mechanism is automatically triggered and the process returns to re-optimization. Rigid constraints are embedded into the decision-making closed loop to ensure a balance between economic benefits and ecological safety. When generating knowledge increments for this optimization cycle, newly identified geological patterns, hyperparameter combinations verified by Bayesian optimization, and mining strategies that fail due to water channeling or sand blockage are extracted from massive optimization logs.

[0115] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimized oil and gas reservoir development based on multimodal data, characterized in that, Includes the following steps: A multimodal data acquisition system for oil and gas reservoirs is constructed, through which seismic wavefield data, well logging interpretation data, production dynamic data, microseismic monitoring data, downhole temperature and pressure time series data, and wellbore structural parameter data are acquired; The acquired multimodal data is spatiotemporally aligned to establish a unified geological coordinate system and time reference, thereby eliminating data heterogeneity caused by different acquisition frequencies and spatial resolutions. The aligned multimodal data is input into a pre-trained multi-scale feature extraction network. The network adopts a three-branch parallel architecture to extract waveform features of seismic wavefield, vertical layering features of well logging data, and time-varying features of production dynamic data, respectively. The extracted multi-scale features are mapped to a multimodal fusion transformer model, and a nonlinear correlation between seismic response and production dynamics is established through a cross-modal attention mechanism to generate digital feature vectors for oil and gas reservoirs. Based on the aforementioned digital feature vectors, a parameterized digital twin of an oil and gas reservoir is constructed. The digital twin includes dynamic evolution models of the porosity field, permeability field, saturation field, and pressure field. Multi-objective optimization calculations are performed in the digital twin. The optimization objectives include maximizing recovery rate, minimizing mining cost, and maximizing equipment life. The constraints cover bottom hole flowing pressure limit, water cut threshold, and formation sand production critical value. The optimization results are converted into well group production parameter control commands, which include single well stroke frequency, pump hanging depth, injection-production well production ratio, and multi-stage fracturing section switching status. Establish a closed-loop feedback mechanism to continuously input actual production data into the digital twin for assimilation and updates, thereby achieving rolling optimization of mining parameters.

2. The method for optimized oil and gas reservoir development based on multimodal data according to claim 1, characterized in that, The multimodal data acquisition system further includes: Deploy distributed fiber optic acoustic sensors in injection and production wellbores to acquire continuous inter-well seismic wave field data; A logging-while-drilling toolchain was deployed in the horizontal well section to collect real-time resistivity, neutron porosity and gamma-ray energy spectrum data. A multiphase flow meter and a high-frequency pressure transmitter are installed at the wellhead to collect second-level production dynamic data. A microseismic monitoring array was deployed in the formation around the well to collect source mechanism solutions for microseismic events induced by fracturing and production.

3. The method for optimized oil and gas reservoir development based on multimodal data according to claim 1, characterized in that, The spatiotemporal alignment process specifically includes: A dynamic time warping algorithm is used to align the time series of seismic data and production dynamics data. Kriging spatial interpolation was used to reconstruct well logging data into three-dimensional volume data with a resolution that matched the voxel size of the seismic data. A geomechanical finite element model was established to calculate the geometric correction of the formation stress field to the wellbore measuring point location.

4. The method for optimized oil and gas reservoir development based on multimodal data according to claim 1, characterized in that, The multi-scale feature extraction network is further configured as follows: The seismic wavefield feature branch uses a three-dimensional convolution kernel to capture local patterns of the stratigraphic reflection structure, with the kernel size increasing along the depth direction; The feature branch of well logging data uses a bidirectional long short-term memory network to capture the vertical dependencies of sequence stratigraphic units; The production dynamics feature branch uses time-frequency convolution to extract the composite features of the output decline curve and the pressure derivative curve.

5. The method for optimized oil and gas reservoir development based on multimodal data according to claim 1, characterized in that, The multimodal fusion transformer model is specifically implemented as follows: Seismic features, well logging features, and production features are encoded into query vectors, key vectors, and value vectors, respectively. Introduce a geological prior knowledge mask into the self-attention layer to block feature interaction paths that do not conform to sedimentary patterns; Enhance the physical consistency representation between different modalities by using a cross-modal contrastive learning loss function; The output fusion feature dimensions correspond to the number of oil and gas reservoir geological units, and each dimension characterizes the reservoir performance parameters of the corresponding unit.

6. The method for optimized oil and gas reservoir development based on multimodal data according to claim 1, characterized in that, The parameterized digital twin of the oil and gas reservoir includes: The digital feature vector is input into the physical information neural network, and the network is embedded with Darcy's law, the mass balance equation and the crack seepage equation as partial differential constraints. A generative adversarial network architecture is adopted, in which the generator outputs porosity field and permeability field tensors, and the discriminator distinguishes the statistical differences between the generated field and the historical inversion field; A discrete crack network model is implanted into the digital twin, and the crack aperture field is driven by the density field of microseismic events to evolve. The pressure field solver employs an adaptive mesh refinement strategy, densifying the computation nodes in the well perimeter and fracture tip regions.

7. The method for optimized oil and gas reservoir development based on multimodal data according to claim 1, characterized in that, The multi-objective optimization calculation adopts the following steps: The recovery rate objective function is defined as the time integral of the cumulative oil production, the cost objective function includes electricity consumption, chemical injection and well workover costs, and the equipment life objective function is based on the rod and tubing combination fatigue cumulative damage model. A non-dominated sorting genetic algorithm is used to search for the Pareto optimal solution set. The population size is proportional to the number of wells in the well group, and the crossover operator simulates the linear superposition effect of inter-well disturbances. Uncertainty quantification is performed on the optimized solution set, and a Bayesian neural network is used to predict the success probability of each optimization scheme. The confidence interval is determined by multiple random simulations using a digital twin. The objective solution is selected from the Pareto frontier based on the decision preference weight vector, which is dynamically adjusted by oil prices, policy subsidies, and carbon emission costs.

8. The method for optimized oil and gas reservoir development based on multimodal data according to claim 1, characterized in that, The closed-loop feedback mechanism is specifically implemented as follows: Lightweight digital twin replicas are deployed on edge computing nodes, and these replicas receive real-time wellhead data hourly for partial updates. When the deviation between the predicted output and the actual output exceeds a preset threshold, the digital twin is triggered to undergo global retraining. The retraining process uses an incremental learning approach to retain historical knowledge. Establish an expert review module for adjusting mining parameters, and transform the control commands generated by artificial intelligence into well group production allocation schemes that engineers can understand; A blockchain architecture is used to store key optimization decisions, ensuring the traceability and tamper-proof nature of the mining adjustment process.

9. The method for optimized oil and gas reservoir development based on multimodal data according to claim 1, characterized in that, The method further includes: Before optimization, the quality of multimodal data is assessed, and the assessment indicators include the signal-to-noise ratio of seismic data, the repetition error of well logging data, and the integrity of production dynamic data. Modal data that are below the quality threshold are repaired through an adversarial data augmentation network. An environmental risk pre-assessment was conducted on the optimized mining plan. The assessment indicators included surface subsidence, produced water backflow pressure, and methane leakage probability. If the risk exceeded the standard, an early warning was automatically triggered and the plan was returned for re-optimization. Generate knowledge increments for this optimization cycle. These knowledge increments include newly identified geological models, optimized hyperparameter combinations, and failed mining strategies. The knowledge increments are stored in an enterprise-level knowledge graph for subsequent well group calls.

Images