Intelligent self-adaptive CO2 miscible phase flooding reservoir development method and device

Abstract

The application belongs to the technical field of oil reservoir development, and particularly relates to an intelligent self-adaptive CO2 miscible displacement oil reservoir development method and device, which comprises the following steps: constructing a cross-scale digital twin of pores-core-oil reservoir, determining the initial minimum miscibility pressure MMP0 of crude oil and CO2 in the target oil reservoir; taking the maximization of the comprehensive benefit evaluation index CEI of miscible displacement development as the target, solving and generating an initial injection-production development scheme through an NSGA-III multi-objective optimization algorithm; using standardized characteristic data, a cloud digital twin engine performs real-time dynamic updating and parameter correction on the cross-scale digital twin, and calculates the dynamic minimum miscibility pressure and full-dimensional miscibility degree factor MPF_full of the oil reservoir plane-vertical full-space grid unit; based on the full-dimensional miscibility degree factor, the whole domain of the oil reservoir is divided into five regulation and control units; based on the regulation and control units and the miscibility evolution mechanism, a multi-modal space-time fusion early prediction system is constructed, and multi-time scale prediction results and hierarchical early warning information are obtained.

Description

Technical Field

[0001] This invention belongs to the field of reservoir development technology, specifically relating to an intelligent adaptive CO2 miscible flooding reservoir development method and apparatus. Background Technology

[0002] CO2 miscible flooding is a core replacement technology for enhancing oil recovery in low-permeability tight reservoirs, and it also enables large-scale geological sequestration of CO2. It has already entered the stage of large-scale application in my country's oil and gas field development. After CO2 and crude oil become miscible, the interfacial tension between oil and water can be completely eliminated, crude oil viscosity can be significantly reduced, and light hydrocarbon components in crude oil can be efficiently extracted. The oil washing efficiency is more than 40% higher than that of immiscible flooding, making it the optimal development mode for CO2 oil recovery.

[0003] my country's continental oil reservoirs differ fundamentally from foreign marine oil reservoirs, generally exhibiting strong reservoir heterogeneity, low light hydrocarbon content in crude oil, and high minimum miscibility pressure (MMP). This leads to common industry pain points during field implementation, including "extremely heterogeneous planar-vertical miscibility, early gas channeling, difficulty in maintaining stable miscibility, and unsatisfactory recovery rates." To address these issues, existing technologies have been researched and developed. Existing technologies use empirical formulas adapted to marine reservoirs abroad for MMP calculation, miscibility characterization, and gas channeling risk quantification. However, due to the characteristics of low light hydrocarbon content and strong heterogeneity in my country's continental reservoirs, the calculation errors are large and the adaptability is poor. These technologies cannot accurately characterize the dynamic evolution of miscibility during development, resulting in a lack of theoretical foundation for the entire technical solution.

[0004] The approach remains stuck in "passive regulation centered on ground production indicators," failing to address the essential nature of oil displacement through "precise and controllable miscibility across all times and spaces." It treats miscibility merely as a post-processing analysis result rather than the core objective of proactive regulation, resulting in consistently lagging regulation and an inability to address the problem at its root.

[0005] Artificial intelligence applications are merely simple combinations of single models, failing to construct a full-link intelligent system driven by both mechanism and data, lacking a fully closed-loop self-iterative mechanism, exhibiting poor generalization, not conforming to reservoir physical laws, and unable to meet the adaptive control needs of complex on-site working conditions.

[0006] In summary, existing technologies not only lack a core mathematical model system for continental oil reservoirs in my country, but also fail to construct a fully intelligent development system that spans the entire chain from theoretical foundation to field execution. This makes it impossible to fundamentally solve the core pain points of CO2 miscible flooding in continental oil reservoirs. There is an urgent need to develop a CO2 miscible flooding development method and device with intellectual property mathematical models, deep intelligence, and full closed-loop adaptive capabilities. Summary of the Invention

[0007] To address the aforementioned shortcomings in the existing technology, this invention provides an intelligent adaptive CO2 miscible flooding reservoir development method and apparatus to solve the problems mentioned in the background technology.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A smart adaptive CO2 miscible flooding reservoir development method specifically includes the following steps: S1 constructs a digital twin of a multi-scale, multi-field coupled pore-core-reservoir system with an embedded mixed-camera physical kernel; it collects geological static data, core CT scan data, micropore structure data, long core displacement experiment data, high-pressure fluid property data, and historical development dynamic data of the target reservoir; and determines the initial minimum miscibility pressure (MMP0) of crude oil and CO2 in the target reservoir by fitting the initial MMP calculation model of the continental reservoir with fine tube experiments and interfacial tension experiments. A three-dimensional geological-numerical simulation integrated digital twin was constructed, coupling pore-level molecular dynamics, core-level seepage mechanism, and reservoir-level development dynamics across multiple scales. A real-time calculation model coupling multi-physics field, multi-component, interface dynamics, and miscibility evolution mechanism was embedded into the PINN physical information neural network kernel of the digital twin. With maximizing the comprehensive benefit evaluation index (CEI) of miscibility flooding development as the core objective, a multi-objective system was coupled to maximize the miscibility swept volume of the entire reservoir, maximize oil recovery, maximize CO2 sequestration, maximize development economic benefits, and minimize gas channeling risk. With formation safety, wellbore safety, and environmental protection thresholds as hard constraints, the initial injection-production development scheme was generated by solving the problem using the NSGA-III multi-objective optimization algorithm. S2 constructs an integrated well-to-ground-to-well cross-scale intelligent monitoring network to achieve cloud-edge-device collaborative data preprocessing and edge intelligent inference. An integrated well-to-ground-to-well intelligent monitoring network is deployed in injection wells, production wells, and observation wells of the target reservoir. This network combines distributed fiber optic monitoring in the wellbore, downhole nanosensor arrays, online analysis of inter-well tracers, and surface microseismic / magnetotelluric monitoring. It collects dynamic data of multiple physical fields, components, and scales at the micro- and macro-levels, from pore level to wellbore level to reservoir level, during reservoir development. A federated edge learning architecture is constructed through well site edge computing units to perform data denoising, spatiotemporal alignment of multi-source data, feature extraction, and standardization at the edge. At the same time, a self-supervised learning model is used to complete local anomaly early warning and emergency control, with only feature data uploaded to the cloud-based digital twin engine. S3 constructs a multi-scale intelligent representation system for mixed-phase states, enabling super-resolution reconstruction and evolution mechanism analysis of all-temporal and spatiotemporal mixed-phase distributions. The cloud-based digital twin engine, based on real-time monitoring of multi-source heterogeneous data, employs an ensemble Kalman filter combined with an adaptive history fitting algorithm to dynamically update and correct parameters of the cross-scale digital twin in real time. Through an embedded dynamic miscibility evolution mechanism coupling model, it calculates the dynamic minimum miscibility pressure of the reservoir's planar-vertical full-space grid cells in real time. By combining the full-dimensional miscibility factor MPF_full with the Transformer super-resolution reconstruction model, super-resolution reconstruction of the spatial distribution of miscibility states in sparse areas of monitoring data is achieved. Based on the MPF_full value, the calculated gas channeling risk probability CRP, the remaining oil enrichment, and the reservoir heterogeneity characteristics, the entire reservoir is finely divided into five regulatory units: a fully miscible stable zone, a near-miscible enhanced zone, an immiscible enhanced zone, a high-risk gas channel zone, and a remaining oil enrichment unaffected zone. At the same time, the dynamic evolution mechanism of CO2-crude oil multiple contact miscibility, the extraction law of light hydrocarbon components, and the dynamic change characteristics of interfacial tension are analyzed in different units. S4 constructs a multimodal spatiotemporal fusion advanced prediction system to achieve advanced early warning of multi-timescale miscibility evolution and risk classification. Based on the dynamic evolution data of cross-scale digital twins, multi-source monitoring time series data, and reservoir spatial topological connectivity data, it constructs a multimodal spatiotemporal fusion prediction model coupled with a spatiotemporal graph neural network (ST-GNN) and Bayesian uncertainty quantification (UQ) to achieve advanced prediction at three time scales: short-term 7-day accurate prediction, medium-term 1-3 month trend prediction, and long-term 6-12 month evolution prediction. Prediction indicators include the dynamic changes of MPF_full in each control unit, the evolution of miscibility zone area, CRP value, changes in remaining oil distribution, and development indicators of the entire block. At the same time, it identifies gas channeling precursor characteristics through a self-supervised anomaly detection model and, combined with CRP value, achieves a four-level classification early warning of gas channeling risk from the initiation stage to the breakthrough stage, thus locking in risk channels in advance.

[0009] Compared with the prior art, the present invention has the following beneficial effects: This invention independently constructs a core mathematical model system covering the entire CO2 miscible flooding process, which is fully adapted to the geological characteristics of my country's continental oil reservoirs. The MMP calculation error is reduced from more than 8% in the existing technology to less than 1.5%, and the accuracy of miscible state characterization is improved by one order of magnitude. This invention is the first to take the precise and controllable miscible state in all time and space as the core objective, and reconstructs the development system from the essence of oil displacement. It completely breaks the industry vicious cycle of "enhancing miscibility aggravates gas channeling and blocking gas channeling destroys miscibility". It can delay the gas channeling breakthrough time by more than 18 months and extend the stable miscible flooding cycle by more than 200%. Detailed Implementation

[0010] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below with reference to embodiments.

[0011] S1 constructs a digital twin of a multi-scale, multi-field coupled pore-core-reservoir system with an embedded mixed-camera physical kernel; it collects geological static data, core CT scan data, micropore structure data, long core displacement experiment data, high-pressure fluid property data, and historical development dynamic data of the target reservoir; and determines the initial minimum miscibility pressure (MMP0) of crude oil and CO2 in the target reservoir by fitting the initial MMP calculation model of the continental reservoir with fine tube experiments and interfacial tension experiments. A three-dimensional geological-numerical simulation integrated digital twin was constructed, coupling pore-level molecular dynamics, core-level seepage mechanism, and reservoir-level development dynamics across multiple scales. A real-time calculation model coupling multi-physics field, multi-component, interface dynamics, and miscibility evolution mechanism was embedded into the PINN physical information neural network kernel of the digital twin. With maximizing the comprehensive benefit evaluation index (CEI) of miscibility flooding development as the core objective, a multi-objective system was coupled to maximize the miscibility swept volume of the entire reservoir, maximize oil recovery, maximize CO2 sequestration, maximize development economic benefits, and minimize gas channeling risk. With formation safety, wellbore safety, and environmental protection thresholds as hard constraints, the initial injection-production development scheme was generated by solving the problem using the NSGA-III multi-objective optimization algorithm. S2 constructs an integrated well-to-ground-to-well cross-scale intelligent monitoring network to achieve cloud-edge-device collaborative data preprocessing and edge intelligent inference. An integrated well-to-ground-to-well intelligent monitoring network is deployed in injection wells, production wells, and observation wells of the target reservoir. This network combines distributed fiber optic monitoring in the wellbore, downhole nanosensor arrays, online analysis of inter-well tracers, and surface microseismic / magnetotelluric monitoring. It collects dynamic data of multiple physical fields, components, and scales at the micro- and macro-levels, from pore level to wellbore level to reservoir level, during reservoir development. A federated edge learning architecture is constructed through well site edge computing units to perform data denoising, spatiotemporal alignment of multi-source data, feature extraction, and standardization at the edge. At the same time, a self-supervised learning model is used to complete local anomaly early warning and emergency control, with only feature data uploaded to the cloud-based digital twin engine. S3 constructs a multi-scale intelligent representation system for mixed-phase states, enabling super-resolution reconstruction and evolution mechanism analysis of all-temporal and spatiotemporal mixed-phase distributions. The cloud-based digital twin engine, based on real-time monitoring of multi-source heterogeneous data, employs an ensemble Kalman filter combined with an adaptive history fitting algorithm to dynamically update and correct parameters of the cross-scale digital twin in real time. Through an embedded dynamic miscibility evolution mechanism coupling model, it calculates the dynamic minimum miscibility pressure of the reservoir's planar-vertical full-space grid cells in real time. By combining the full-dimensional miscibility factor MPF_full with the Transformer super-resolution reconstruction model, super-resolution reconstruction of the spatial distribution of miscibility states in sparse areas of monitoring data is achieved. Based on the MPF_full value, the calculated gas channeling risk probability CRP, the remaining oil enrichment, and the reservoir heterogeneity characteristics, the entire reservoir is finely divided into five regulatory units: a fully miscible stable zone, a near-miscible enhanced zone, an immiscible enhanced zone, a high-risk gas channel zone, and a remaining oil enrichment unaffected zone. At the same time, the dynamic evolution mechanism of CO2-crude oil multiple contact miscibility, the extraction law of light hydrocarbon components, and the dynamic change characteristics of interfacial tension are analyzed in different units. S4 constructs a multimodal spatiotemporal fusion advanced prediction system to achieve advanced early warning of multi-timescale hybrid evolution and risk classification. Based on dynamic evolution data from cross-scale digital twins, multi-source monitoring time-series data, and reservoir spatial topological connectivity data, a multimodal spatiotemporal fusion prediction model coupled with a spatiotemporal graph neural network (ST-GNN) and Bayesian uncertainty quantification (UQ) is constructed to achieve advanced prediction at three time scales: short-term 7-day accurate prediction, medium-term 1-3 month trend prediction, and long-term 6-12 month evolution prediction. Prediction indicators include the dynamic changes of MPF_full in each control unit, the evolution of the miscible zone area, CRP value, changes in remaining oil distribution, and development indicators of the entire block. At the same time, a self-supervised anomaly detection model is used to identify gas channeling precursor characteristics, and combined with CRP value, a four-level early warning system for gas channeling risk is implemented from the initiation stage to the breakthrough stage, thus locking in risk channels in advance. S5 constructs a multi-agent hierarchical collaborative reinforcement learning decision-making system to generate a multi-dimensional collaborative adaptive control strategy. With the continuous maximization of the Comprehensive Benefit Evaluation Index (CEI) for miscible flooding development as the core objective, and addressing the differentiated development contradictions of the five control units, corresponding unit-level control sub-objectives are matched. Combining the advanced prediction results of miscible evolution and risk warning levels, a three-level multi-agent hierarchical collaborative reinforcement learning (MADRL) decision architecture is constructed, consisting of a top-level global optimization agent, a middle-level unit collaborative agent, and a bottom-level single-well execution agent. Combined with a safety-constrained reinforcement learning mechanism, and using the single-well control optimization objective function as the solution core, an adaptive control strategy of four-dimensional collaboration is generated, which includes precise miscible enhancement, intelligent gas channeling prevention and control, global optimization of injection and production parameters, and targeted tapping of remaining oil potential. All control strategies are first verified by virtual well testing simulation and safety constraint validation in a cross-scale digital twin before being deployed for execution. The S6 constructs an integrated well-to-ground digital closed-loop intelligent execution system, enabling precise zoned control and real-time dynamic feedback of results. The validated adaptive control strategy is simultaneously distributed to the surface intelligent injection digital twin system and the downhole layered intelligent control execution unit, enabling precise targeted control operations of each control unit in collaboration between the well and the ground. At the same time, through the well-ground-well integrated cross-scale intelligent monitoring network, dynamic response data of reservoir miscibility state, multi-physics field change data, and single-well / whole block production and development data are collected in real time and fed back to the cloud cross-scale digital twin in real time, forming a fully digital closed-loop link of "decision-simulation-execution-monitoring-feedback". S7 constructs a full-chain lifelong learning self-iterative system to achieve adaptive optimization and continuous model evolution throughout the entire lifecycle. Based on the dynamic feedback data after regulation, a lifelong continuous learning framework driven by both mechanism and data is constructed. The core framework uses the model iteration error correction formula and the dynamic learning rate calculation formula, with the accuracy of the miscible state representation, the prediction deviation of development indicators, and the execution effect of the regulation strategy as core evaluation indicators. Simultaneously, the following are completed: ① parameter self-calibration and mesh reconstruction of the embedded miscible mechanism model in the cross-scale digital twin; ② online incremental learning and accuracy optimization of the multimodal spatiotemporal fusion prediction model; ③ policy iteration and dynamic optimization of the reward function of the multi-agent reinforcement learning decision model. Steps S2-S7 are repeated to achieve closed-loop adaptive development regulation and continuous model evolution throughout the entire lifecycle of CO2 miscible flooded reservoirs, from the initial gas injection stage to the development depletion stage.

[0012] In some embodiments, step S1, the initial MMP calculation model for continental oil reservoirs, taking into account the characteristics of low light hydrocarbon content and strong heterogeneity in Chinese continental oil reservoirs, couples crude oil composition, pore structure, and formation temperature and pressure characteristics based on phase equilibrium theory. The calculation formula is as follows: In the formula: The minimum miscibility pressure is the reference pressure, expressed in MPa, and is determined by the reference miscibility pressure of pure CO2-degassed crude oil measured in a capillary test. The formation temperature correction factor is calculated using the following formula: Where T is the formation temperature of the target reservoir, in °C. The experimental reference temperature is 75℃. The correction factor for the light hydrocarbon component of crude oil is calculated using the following formula: ),in The mole fraction of C2-C6 light hydrocarbon components in crude oil, expressed as %. The baseline light hydrocarbon content for marine reservoirs is 25%. The reservoir permeability heterogeneity correction coefficient is calculated using the following formula: ),in is the coefficient of variation of reservoir permeability, dimensionless, with a value range of 0-1; The pore structure correction factor is calculated using the following formula: ),in The target reservoir average porosity, expressed as a percentage. The baseline porosity is set at 15%. The geological static data includes three-dimensional reservoir structure, reservoir interlayer distribution, porosity, permeability gradient, pore structure parameters, oil saturation, lithology, fracture distribution and conductivity, and dominant seepage channel distribution data. The high-pressure fluid properties data include crude oil C1-C30 components, density, viscosity, gas-oil ratio, saturation pressure, formation temperature, formation pressure, and CO2-crude oil interfacial tension experimental data. The initial injection-production development plan includes optimized deployment of the injection-production well network, injection pressure threshold, injection rate, gas-water alternating WAG injection parameters, stratified injection scheme, slug design for the miscible enhancement system, fluid production rate and bottom hole flowing pressure control threshold, and optimization scheme for the injection-production pressure system.

[0013] In some embodiments, step S2, the well-to-ground-to-well integrated cross-scale intelligent monitoring network includes: a distributed fiber optic temperature / pressure / strain / sonic monitoring subunit deployed along the entire wellbore; a downhole chromatographic fluid component online analysis subunit deployed in the oil layer perforated section; a formation pore-level implanted nanoscale temperature, pressure, and component sensor array; an inter-well intelligent tracer online monitoring subunit; and a surface-deployed microseismic monitoring subunit and magnetotelluric monitoring subunit. The real-time acquired dynamic data includes formation temperature field, pressure field, stress field, seepage field, saturation field, CO2 concentration field, fluid full component data, fracture propagation and dominant seepage channel evolution data, and wellbore production dynamic data. The federated edge learning architecture completes local model inference and emergency control at the edge, and global model training is completed in the cloud. The original monitoring data does not leave the well site; only model update gradients and feature data are uploaded.

[0014] In some embodiments, in step S3, the dynamic minimum miscibility pressure (MMP_dynamic) calculation model is coupled with the crude oil light hydrocarbon extraction effect, CO2 dissolution effect, dynamic changes in temperature and pressure field, and pore structure contamination characteristics during the development process. The calculation formula is as follows: In the formula: Development time, in days (d). This is the initial minimum miscibility pressure, in MPa; The coefficient representing the effect of light hydrocarbon extraction is dimensionless and obtained by fitting the long core displacement experiment, with a value ranging from 0.6 to 1.2. The mole fraction of C2-C6 light hydrocarbon components in crude oil under initial conditions, expressed in % (%). The mole fraction of C2-C6 light hydrocarbon components in crude oil at time t, expressed in % %. The real-time formation temperature at time t is expressed in °C. The original formation temperature is expressed in °C. The CO2 dissolution effect correction factor is calculated using the following formula: ,in () represents the dissolved CO2 gas-oil ratio in crude oil at time t, in m³ / m³. The dynamic contamination correction factor for pore structure is calculated using the following formula: ,in The original formation permeability is expressed in mD. The real-time permeability of the formation at time t is expressed in mD; the full-dimensional miscibility factor MPF_full is a full-dimensional quantitative characterization model of the miscibility state, and its calculation formula is as follows: In the formula: The current formation fluid pressure in the grid cell, in MPa; This represents the current dynamic minimum miscibility pressure of the grid cell, in MPa. The correction factor for the contact efficiency between CO2 and crude oil is calculated using the following formula: ),in This refers to the actual contact time between CO2 and crude oil, expressed in days (d). The critical contact time required for complete miscibility, in days, is determined by core experiments. The correction factor for multiple contact miscibility time is calculated using the following formula: ),in The number of contact cycles between CO2 and crude oil is dimensionless. The reservoir heterogeneity correction factor is calculated using the following formula: ,in Permeability of the grid cell, in mD. The average permeability of the reservoir is expressed in mD. The interfacial tension correction coefficient is calculated using the following formula: ,in The real-time interfacial tension between CO2 and crude oil at time t is expressed in mN / m. The CO2-crude oil interfacial tension under the original state is expressed in mN / m. The molecular diffusion correction factor is calculated using the following formula: ,in Let be the molecular diffusion coefficient of CO2 in crude oil at time t, with units of m² / s. is the limiting diffusion coefficient of CO2 in crude oil under formation conditions, with units of m² / s; The correction factor for fracture conductivity is calculated using the following formula: ,when hour, ;in This refers to crack permeability, expressed in mD. The value represents matrix permeability, expressed in mD. The precise rules for determining the miscibility state based on MPF_full values ​​are as follows: Completely miscible stable zone: MPF_full ≥ 1.05, interfacial tension ≤ 0.001 mN / m, gas channeling risk probability (CRP) < 20%; Nearly miscible enhanced zone: 0.85 ≤ MPF_full < 1.05, 0.001 mN / m < interfacial tension ≤ 0.1 mN / m, CRP < 20%; Immiscible enhanced zone: MPF_full < 0.85, interfacial tension > 0.1 mN / m, CRP < 20%; High-risk gas channel zone: CRP ≥ 50%, regardless of MPF_full value; Remaining oil enrichment unaffected zone: Oil saturation ≥ 60% of the original oil saturation, CO2 sweep efficiency < 10%, regardless of MPF_full value.

[0015] In some embodiments, step S3, the gas channeling risk probability (CRP) calculation model, coupled with seepage mechanics characteristics, miscibility, reservoir heterogeneity, and production dynamics characteristics, is calculated using the following formula: In the formula: , , , , The weight coefficients of each influencing factor are dimensionless and determined by the coupling of the analytic hierarchy process (AHP) and the entropy weight method, satisfying the following conditions: 1; The CO2 transport rate ratio is calculated using the following formula: ,in This refers to the CO2 frontal transport rate, expressed in m / d. The migration rate at the crude oil displacement front is expressed in m / d. The permeability variation coefficient of the target area is dimensionless and ranges from 0 to 1. The normalized value of the rate of increase in the gas-oil ratio is calculated using the following formula: ,in This represents the current daily gas-oil ratio, in m³ / m³. This represents the gas-oil ratio threshold, expressed in m³ / m³. The normalized value of the spatial gradient of the degree of miscibility is calculated using the following formula: ,in The spatial gradient magnitude of MPF_full This represents the maximum value of the MPF_full gradient modulus for the entire reservoir. The crack development factor is calculated using the following formula: ,in This represents the total length of cracks within the grid cell, in meters. The grid cell side length is in meters; the time series prediction model is a bidirectional LSTM time series prediction model with a lead time of 1-3 months. The prediction indicators include the mean MPF_full value of each partition, the area ratio of the mixed zone, the CRP value, the daily oil production, the gas-oil ratio, and the crude oil recovery rate.

[0016] In some embodiments, the operating mechanism of the three-level multi-agent hierarchical collaborative reinforcement learning decision architecture in step S4 is as follows: Top-level global optimization agent: Taking the maximization of the comprehensive benefit evaluation index (CEI) of mixed-phase drive development as the global objective, it outputs the control weights and global constraints of each control unit and coordinates the development contradictions between units. Mid-level unit collaborative intelligent agent: There are 5 in total, each corresponding to 5 control units. Taking the control sub-objective of the corresponding unit as the optimization direction, based on global constraints, it outputs a unit-level control strategy framework to achieve collaborative optimization among units and resolve the inherent contradiction between hybridization enhancement and gas channeling prevention. Bottom-level single-well execution agent: For each injection well and production well, the single-well regulation optimization objective function is used as the core solution and the unit-level regulation strategy framework is used as the constraint. It outputs the precise execution parameters of single well injection pressure, injection rate, gas-water ratio, slug combination, layered injection parameters, fluid production rate, and bottom hole flowing pressure, while meeting the hard constraints of wellbore safety and formation safety. The safety constraint reinforcement learning mechanism incorporates formation fracture pressure, wellbore pressure resistance threshold, production well gas-oil ratio upper limit, and environmental emission threshold as insurmountable hard constraints, embedding them into the action space and reward function of reinforcement learning to ensure that the control strategy output by the agent always meets the requirements for safe development. The virtual well test simulation verification simulates the development dynamics of the control strategy for 3 months after its execution in a digital twin, verifying the effects of miscibility improvement, gas channeling prevention, and recovery rate enhancement. If the expected goals are not achieved, the decision model is re-optimized.

[0017] In some embodiments, the Comprehensive Benefit Evaluation Index (CEI) for mixed-phase flooding development is a normalized comprehensive evaluation model with multiple coupled objectives, and its calculation formula is as follows: In the formula: The weighting coefficients for each item are dimensionless and dynamically adjusted throughout the entire development lifecycle to meet the following requirements. Initial stage of gas injection Highest weight, mid-to-late development stage , The weight gradually increases; The total volume of the fully miscible zone and the near-miscible zone of the entire reservoir is expressed in m³. This represents the total oil-bearing volume of the reservoir, in m³. The percentage increase in oil recovery is expressed as % The percentage increase in theoretical maximum recovery rate is expressed as % CO2 geological sequestration rate, in percentages (%) The net present value is expressed in ten thousand yuan. The theoretical maximum net present value is expressed in ten thousand yuan. The average gas channeling risk probability for the entire reservoir is dimensionless; the single-well control optimization objective function, with the sub-objectives of the control unit to which the single well belongs as the core, is calculated as follows: Constraints: In the formula: -- The component weighting coefficients for single-well optimization are dimensionless and dynamically adjusted by the type of control unit to which they belong, satisfying the following requirements. ; The average MPF_full value of the single-well control area is dimensionless. This represents the daily oil production of a single well, expressed in m³ / d. This represents the theoretical maximum daily oil production of a single well, expressed in m³ / d. This represents the real-time gas-oil ratio for a single well, expressed in m³ / m³. This represents the gas-oil ratio threshold for a single well, expressed in m³ / m³. This refers to the injection pressure in a single well, expressed in MPa. The optimal miscible injection pressure for a single well is given in MPa. In the constraints: For minimum injection pressure, For formation fracture pressure, These are the upper and lower limits of the injection volume. For the bottom flow pressure of the production well, The minimum bottom hole flowing pressure is achieved. The four-dimensional synergistic adaptive control strategy performs differentiated targeted control for the five control units. Specifically, for the fully miscible stable zone: the injection and production parameters are adaptively optimized to maintain the formation pressure at 1.05-1.1 times MMP_dynamic, the WAG slug size and gas-water ratio are optimized, the uniform advancement of the displacement front is controlled, the formation of dominant seepage channels is suppressed, and the stable fully miscible displacement state is continuously maintained. For the near-miscible enhancement zone: Adaptively optimize WAG injection parameters, match the injection of low-concentration nano-enhanced miscibility strengthening additives, reduce crude oil dynamic MMP and interfacial tension, increase MPF_full value to above 1.05, and promote the transformation from near-miscible to fully miscible; For the immiscible enhancement zone: Adaptively replenish formation energy, increase injection pressure to above 90% of MMP_dynamic, and simultaneously inject a targeted miscibility strengthening system to replenish light hydrocarbon components of crude oil, significantly reduce interfacial tension and MMP, and promote the transformation of the region from immiscible to near-miscible to fully miscible; For the high-risk gas channel zone: Based on CRP early warning level, adaptively match a CO2 / pH dual-response intelligent regulation and sealing system, combined with downhole... The intelligent gas channel blocking robot achieves deep and precise sealing of advantageous channels while optimizing the injection-production pressure system. While blocking gas channel blockages, it maintains a near-miscible displacement state in the surrounding area, preventing the sealing operation from disrupting the miscible environment. For unaffected areas rich in residual oil: it adaptively optimizes the flow lines of the injection-production well network, reconstructing the formation seepage field through periodic injection-production, pulsed injection, and targeted water shut-off and profile control technologies. This forces CO2 to flow towards the remaining oil-rich area, achieving targeted tapping of the remaining oil's potential. Simultaneously, it matches miscibility enhancement measures to improve the miscibility of the affected area. The execution timing of the aforementioned control strategy is based on the advanced prediction results of miscibility evolution and CRP early warning levels, executed 1-3 injection cycles in advance, achieving proactive pre-emptive control rather than reactive post-event remediation.

[0018] In some embodiments, step S7, the model iteration error correction formula and the dynamic learning rate calculation formula are the core mathematical support for the lifelong learning self-iterative system, specifically: Iterative error correction formula: In the formula: This is the amount of correction for the model parameters. This is a correction factor, dimensionless, with a value ranging from 0.01 to 0.1; To monitor the actual values, These are the model's predicted values; For the current parameters of the model, These are the corrected model parameters; Formula for calculating dynamic learning rate: In the formula: Let be the dynamic learning rate at time t; The initial learning rate; The attenuation coefficient is dimensionless and ranges from 0.0005 to 0.002. Let t be the average prediction error of the model at time t; The initial allowable maximum error for the model; The iteration number is [number]; the twin-decision model's dual self-iteration is specifically as follows: Cross-scale digital twin self-iteration: Based on dynamic feedback data, the correction coefficients of grid attribute parameters, relative permeability curves, capillary pressure curves, and embedded mixed-phase physical models are adaptively fitted and corrected through error correction formulas. It automatically identifies areas where the model deviation exceeds the threshold, triggers local grid reconstruction and parameter refitting, and ensures that the characterization error of the mixed phase state is ≤2% and the fitting error of the core development indicators is ≤4%. The multimodal spatiotemporal fusion prediction model is self-iterative: based on the deviation between real-time collected data and prediction results, it updates model weights through online incremental learning, optimizes the learning process through a dynamic learning rate formula, and solves the catastrophic forgetting problem of the model through a continuous learning mechanism, ensuring that the prediction accuracy continues to improve with the accumulation of development data, and the average absolute percentage error of the mid-term prediction in 3 months is ≤6%. The multi-agent reinforcement learning decision model is self-iterative: using the change of the CEI index of the entire reservoir after regulation as the comprehensive reward function, and combining the deviation between the simulation results of virtual well testing and the actual execution effect, it completes the online iterative update of the strategy network, continuously optimizing the decision accuracy and strategy adaptability. The mechanism-data dual-driven lifelong continuous learning framework uses the reservoir seepage mechanism and the physical law of miscible displacement as the hard constraints of the model, avoiding the defects of poor generalization and non-compliance with physical laws of pure data-driven models. At the same time, it continuously optimizes model parameters through real-time monitoring data, realizing the deep integration and continuous evolution of the mechanism model and the data-driven model.

[0019] In some embodiments, in step S1, the PINN physical information neural network kernel uses the seepage control equation, component transport equation, energy conservation equation, and miscible state evolution equation of CO2 miscible flooding as regularization constraints of the neural network and embeds them into the loss function of the neural network to achieve rapid dynamic simulation and parameter inversion of the reservoir under physical mechanism constraints. The simulation calculation speed is improved by more than two orders of magnitude compared with traditional numerical simulation.

[0020] An intelligent adaptive CO2 miscible flooding reservoir development device is used to execute the above-mentioned intelligent adaptive CO2 miscible flooding reservoir development method. The device includes five core systems: a well-to-ground-to-well integrated cross-scale intelligent monitoring system, a cloud-edge-device collaborative edge computing system, a cloud-based cross-scale digital twin intelligent decision-making platform, a surface intelligent injection digital twin execution system, and a downhole layered intelligent control and sealing execution system. The integrated well-to-ground-to-well multi-scale intelligent monitoring system is deployed in injection wells, production wells, observation wells, and surface well sites of the target reservoir. It includes a wellbore distributed fiber optic monitoring subsystem, a downhole nanosensor array, a downhole fluid component online analysis subsystem, an inter-well intelligent tracer monitoring subsystem, and a surface microseismic / magnetotelluric monitoring subsystem. It is used to collect dynamic data of multiple scales, multiple physical fields, and multiple components at the pore level, wellbore level, and reservoir level in real time during reservoir development, and transmit the collected data to the cloud-edge-device collaborative edge computing system. The cloud-edge-device collaborative edge computing system is deployed at the explosion-proof industrial control terminal of the well site and communicates with the well-to-ground-to-well integrated cross-scale intelligent monitoring system. It has built-in federated edge learning unit, data preprocessing unit, edge intelligent inference unit, and local emergency control unit to complete multi-source data preprocessing, local anomaly early warning, edge model inference and emergency control. At the same time, it communicates with the cloud-based cross-scale digital twin intelligent decision-making platform through the 5G industrial private network to realize feature data uploading and global model update gradient distribution. The cloud-based cross-scale digital twin intelligent decision-making platform communicates with the edge computing system. Its core is a cross-scale multi-field coupled digital twin engine connecting pores, core, and reservoir. It is equipped with a multi-scale intelligent characterization module for miscible states, a multi-modal spatiotemporal fusion advanced prediction module, a three-level multi-agent hierarchical collaborative reinforcement learning decision-making module, a virtual well test simulation verification module, and a full-link lifelong learning self-iteration module. The cross-scale digital twin engine incorporates a PINN physical information neural network kernel and embeds a full-mechanism coupled calculation model for miscible evolution. The multi-scale intelligent characterization module is used to perform full-spatiotemporal miscible distribution calculation, super-resolution reconstruction, and control unit division based on formulas. The multi-modal spatiotemporal fusion advanced prediction module is used to perform multi-timescale miscible evolution prediction and gas channeling risk classification and early warning based on the CRP model. The three-level multi-agent hierarchical collaborative reinforcement learning decision-making module is used to generate a four-dimensional collaborative adaptive control strategy based on the CEI index and single-well optimization objective function. The virtual well test simulation verification module is used for simulation verification and safety checks of the control strategy. The full-link lifelong learning self-iteration module is used to perform calculations based on error correction formulas and dynamic... The state learning rate formula enables online self-correction and continuous evolution of the entire-link model. The aforementioned ground-based intelligent injection digital twin execution system, deployed at the injection well site, communicates with the cloud-based cross-scale digital twin intelligent decision-making platform. It includes a CO2 intelligent high-pressure injection subsystem, an intelligent liquid mixing subsystem, a high-pressure injection pipeline intelligent control valve group, and an injection equipment digital twin mirror subsystem. It is used to execute precise injection of CO2, the miscibility enhancement system, and the regulation and sealing system, as well as adaptive closed-loop control of injection pressure, injection rate, and slug combination, while simultaneously uploading the equipment operating status in real time. The system includes injection execution data; the downhole layered intelligent control and sealing execution system is deployed downhole in injection wells and production wells, and is connected to the cloud-based cross-scale digital twin intelligent decision-making platform and cloud-edge-device collaborative edge computing system. It includes a layered intelligent injector, a downhole electric intelligent flow control valve, a packer unit, a downhole intelligent sealing robot, and a downhole edge control unit. It is used to perform precise injection, flow control, adaptive adjustment of production regime in downhole layered sections, and precise targeted sealing operation of advantageous seepage channels, so as to realize precise targeted control of each control unit in the reservoir plane and vertical direction.

[0021] In some embodiments of the well-to-ground-to-well integrated multi-scale intelligent monitoring system, the distributed optical fiber monitoring subunit is continuously deployed along the entire wellbore to achieve distributed continuous monitoring of temperature, pressure, strain, and acoustic waves along the entire vertical section, with a spatial resolution ≤0.5m; the downhole nanosensor array is implanted in the formation pores of the perforated section of the oil layer to achieve in-situ real-time monitoring of temperature, pressure, CO2 concentration, and crude oil composition at the pore level; the detector array of the surface microseismic monitoring subunit covers the entire block to achieve real-time location and characterization of fracture propagation and microfracture initiation during CO2 injection, with a location accuracy ≤5m.

[0022] In some embodiments of the cloud-edge-device collaborative edge computing system, the local emergency control unit is preset with an upper limit for injection pressure, a formation fracture pressure threshold, an upper limit for the gas-oil ratio of the production well, a safe threshold for wellbore temperature, and a safe threshold for equipment operation. When the monitored data exceeds the preset safe threshold, the local emergency control unit immediately triggers a millisecond-level emergency shutdown command, controls the surface intelligent injection digital twin execution system and the downhole layered intelligent control and sealing execution system to perform emergency shutdown operations, and uploads early warning information to the cloud platform.

[0023] In some embodiments, the digital twin mirror subsystem of the ground intelligent injection digital twin execution system constructs a one-to-one digital twin for each injection pump, control valve, and liquid dispensing device, monitors the equipment operating status and health in real time, predicts equipment failures, and realizes predictive maintenance and adaptive operation control of the equipment. The injection pressure control accuracy is ≤ ±0.2MPa, and the injection flow control accuracy is ≤ ±1%.

[0024] In some embodiments, the layered intelligent injector of the downhole layered intelligent control and sealing execution system can realize independent flow control and injection of five or more oil-bearing sections in a single well, with a single-layer flow control accuracy of ≤±2%. The downhole intelligent sealing robot is equipped with a downhole imaging unit, a flow monitoring unit, and a targeted injection unit, and can enter the advantageous seepage channels of the target section along the wellbore to realize precise targeted injection of the sealing system and real-time monitoring of the sealing effect.

[0025] Example 1 This embodiment uses CO2 miscible flooding development of a deep, tight sandstone reservoir in the Sichuan Basin of my country as an application scenario. The core geological parameters of this reservoir are: reservoir depth 3850m, original formation temperature... =112℃, initial formation pressure 38.6MPa, average reservoir permeability =0.42mD, average porosity =8.7%, coefficient of variation of permeability =0.72, mole fraction of light hydrocarbon components in crude oil (C2-C6) =15.8%, original CO2-crude oil interfacial tension =1.8mN / m, formation fracture pressure 42MPa.

[0026] Step 1: Calculate the initial MMP using the autonomous formula and construct a cross-scale digital twin. The initial MMP0 of continental reservoirs was calculated using the independently constructed initial MMP calculation model of this invention: The measured reference miscibility pressure in the capillary tube experiment is known. =32.5MPa, =112℃, =15.8%, =0.72, =8.7%, Substitute into the formula: The final calculation yields: The relative error between the corrected initial MMP (adapted to the formation conditions) and the field-measured value of 47.8 MPa is only 0.67%, which is far better than the 8.7% calculation error of the existing Alston formula.

[0027] Based on the independently calculated initial MMP, a multi-scale digital twin of pore-core-reservoir is constructed. The independently calculated miscibility evolution model is embedded into the PINN kernel. With the goal of maximizing the independently calculated CEI index, the initial injection-production development scheme is generated by optimizing the NSGA-III algorithm: adopting the WAG injection mode, the initial gas-water ratio is 1.1:1, the slug size is 0.07PV, the single-well CO2 injection rate is 90t / d, the injection pressure is controlled at 35-39MPa, and the bottomhole flowing pressure of the production well is not less than 32MPa.

[0028] Step 2: Construct an integrated well-to-ground-to-well monitoring network and a cloud-edge-device collaborative architecture. A distributed fiber optic monitoring subsystem, a downhole nanosensor array, and an online fluid composition analysis subsystem were deployed in 8 injection wells and 11 production wells. A microseismic monitoring array was deployed on the surface to build a full-scale monitoring network. An edge computing system with a built-in federated edge learning architecture was deployed at the well site to complete data preprocessing and local emergency control. The upper limit of injection pressure was preset to 42MPa and the upper limit of gas-oil ratio to 3000m³ / m³. Exceeding the threshold would trigger a millisecond-level shutdown.

[0029] Step 3: Calculate the dynamic MMP and MPF_full using the autonomous formula to complete the miscibility characterization and partitioning. In the sixth month of development, the formation temperature of the target area was obtained through real-time monitoring data inversion. =110℃, Crude oil C2-C6 content =14.2%, CO2 dissolved gas-oil ratio =62m³ / m³, formation permeability =0.38mD, formation fluid pressure =37.2MPa, calculated using the autonomous dynamic MMP calculation model: Simultaneously calculate the MPF_full value for this region, and substitute it into the custom formula to obtain... It was determined to be a near-miscible enhancement zone; at the same time, the gas channeling risk probability CRP in this area was calculated to be 28% using the autonomous CRP model, which is low risk; based on the full reservoir calculation results, the entire region was divided into 5 characteristic control units to complete the full spatiotemporal characterization of the miscible state.

[0030] Step 4: Multimodal spatiotemporal fusion for advanced prediction and hierarchical early warning The ST-GNN model was used to predict the migratory evolution trend over three months. The results showed that the CRP of the high-permeability fracture zones corresponding to the three injection wells will rise to 68% in the next two months, entering the second-level warning (development period), thus identifying high-risk gas channeling channels in advance.

[0031] Step 5: Multi-agent collaborative decision-making generates adaptive control strategies With maximizing the autonomous CEI index as the global objective, a four-dimensional collaborative control strategy is generated through a three-level multi-agent decision-making architecture and executed one injection cycle in advance: Near-miscibility enhancement zone: Inject 3.5% concentration of nano-miscibility enhancement agent to optimize the gas-water ratio to 1.3:1, pushing MPF_full to above 1.05; High-risk gas channeling zone: Inject 0.4% concentration of a dual-response channeling control system with a slug size of 0.015PV, controlling the injection pressure to stabilize at 35MPa, maintaining a near-miscible state while blocking gas channeling; Immeasurable enhancement zone: Increase the injection pressure to 38-39MPa, inject 5.5% concentration of a miscibility enhancement system to promote near-miscibility conversion; All strategies are first verified through virtual well testing in a digital twin. The results show that the gas channeling risk can be reduced to below 18%, and the proportion of the fully miscible zone increases by 28 percentage points, meeting expectations before being implemented.

[0032] Step 6: Closed-loop execution and dynamic feedback of well-ground coordination The system precisely executes control strategies through a ground-based intelligent injection system and a downhole layered injection device. At the same time, it collects dynamic data after control in real time through a monitoring network and feeds it back to the cloud digital twin in real time, forming a fully digital closed loop.

[0033] Step 7: End-to-end model self-iteration Based on the adjusted feedback data, the parameters of the digital twin, prediction model, and decision model are simultaneously self-calibrated and iteratively updated using the autonomous error correction formula and the dynamic learning rate formula. The error in the mixed state representation is controlled within 1.8%, the prediction error over 3 months is ≤5.8%, and the model accuracy continues to improve. Repeating the above steps enables adaptive development throughout the entire lifecycle.

[0034] Compare with Example 1 This comparative example uses the exact same reservoir and development conditions as Example 1, and adopts conventional methods of existing technology: the initial MMP is calculated using the Alston empirical formula, production indicators are predicted based on the LSTM model, and injection and production parameters are optimized based on the surface gas-oil ratio and water cut. There is no autonomous mathematical model, no real-time characterization of miscibility, no advanced prediction, and no multi-agent collaborative control.

[0035] Comparison of Implementation Results The table below shows a comparison of the development effects of Example 1 and Control Example 1. The present invention has achieved unexpected and significant technical effects: Example 2 This embodiment provides an intelligent adaptive CO2 miscible flooding reservoir development device for executing the development method described in Embodiment 1. The device includes five core systems: a well-to-ground-to-well integrated cross-scale intelligent monitoring system, a cloud-edge-device collaborative edge computing system, a cloud-based cross-scale digital twin intelligent decision-making platform, a surface intelligent injection digital twin execution system, and a downhole layered intelligent control and sealing execution system.

[0036] Among them, the cloud-based cross-scale digital twin intelligent decision-making platform incorporates all the core mathematical models independently constructed by this invention, and can complete the entire process calculation from mixed-phase state characterization, risk prediction, intelligent decision-making to model self-iteration; the ground intelligent injection execution system has an injection pressure control accuracy of ≤±0.2MPa and a flow control accuracy of ≤±1%; the downhole layered intelligent injection device can realize independent control of 5 layers in a single well, and the single-layer flow control accuracy is ≤±2%; the downhole intelligent sealing robot can realize precise targeted sealing of advantageous channels.

[0037] The device in this embodiment constructs a complete fully digital closed-loop intelligent control system through the coordinated linkage of five major systems, which can fully realize the development method of this invention and ensure the implementation of the core invention objectives.

[0038] The above description is only a preferred embodiment of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart adaptive CO2 miscible flooding reservoir development method, characterized in that: Includes the following steps, S1 constructs a multi-scale digital twin of pores, core, and reservoir, collecting geological static data, core CT scan data, micropore structure data, long core displacement experiment data, full-component data of high-pressure fluid properties, and historical development dynamic data of the target reservoir. Based on the initial MMP calculation model of continental reservoirs, combined with fine tube experiments and interfacial tension experiments, the initial minimum miscibility pressure MMP0 of crude oil and CO2 in the target reservoir is determined. With the goal of maximizing the comprehensive benefit evaluation index (CEI) of miscible flooding development, and with formation safety, wellbore safety, and environmental protection thresholds as hard constraints, the initial injection and production development plan is generated by solving the problem using the NSGA-III multi-objective optimization algorithm. S2 constructs a well-to-ground-to-well monitoring network. Based on the initial injection-production development plan, the monitoring network is deployed in the injection wells, production wells, and observation wells of the target reservoir. A federated edge learning architecture is constructed through well site edge computing units. Data denoising, spatiotemporal alignment of multi-source data, feature extraction, and standardization are completed at the edge. At the same time, a self-supervised learning model is used to complete local anomaly early warning and emergency control. The standardized feature data is uploaded to the cloud digital twin engine. The S3 cloud-based digital twin engine utilizes standardized feature data and employs an ensemble Kalman filter combined with an adaptive history fitting algorithm to dynamically update and correct parameters of the cross-scale digital twin in real time, calculating the dynamic minimum miscibility pressure of the reservoir's planar-vertical full-space grid cells. With the full-dimensional miscibility factor MPF_full; based on the full-dimensional miscibility factor, the entire reservoir is divided into five control units: fully miscible stable zone, near-miscible enhanced zone, immiscible enhanced zone, high-risk gas channel zone, and unaffected zone of remaining oil enrichment. Based on the control unit and mixed-phase evolution mechanism obtained in step S3, S4 constructs a multimodal spatiotemporal fusion advanced prediction system to obtain multi-timescale prediction results and hierarchical early warning information.

2. The intelligent adaptive CO2 miscible flooding reservoir development method as described in claim 1, characterized in that: In step S1, the calculation formula for the initial MMP calculation model of the continental reservoir is as follows: In the formula: The minimum miscibility pressure is determined by the baseline miscibility pressure of pure CO2-degassed crude oil measured in a capillary test. The formation temperature correction factor is calculated using the following formula: Where T is the formation temperature of the target oil reservoir. The experimental reference temperature; The correction factor for the light hydrocarbon component of crude oil is calculated using the following formula: ),in This refers to the mole fraction of C2-C6 light hydrocarbon components in crude oil. The baseline light hydrocarbon content for marine reservoirs; The reservoir permeability heterogeneity correction coefficient is calculated using the following formula: ),in The coefficient of variation of reservoir permeability; The pore structure correction factor is calculated using the following formula: ),in The target reservoir average porosity, The baseline porosity is used.

3. The intelligent adaptive CO2 miscible flooding reservoir development method as described in claim 2, characterized in that: The geological static data includes three-dimensional reservoir structure, distribution of reservoir interlayers, porosity, permeability gradient, pore structure parameters, oil saturation, lithology, fracture distribution and conductivity, and distribution of dominant seepage channels. The fluid high-pressure physical property data includes crude oil C1-C30 full component data, density, viscosity, gas-oil ratio, saturation pressure, formation temperature, formation pressure, and CO2-crude oil interfacial tension experimental data. The initial injection-production development scheme includes injection pressure threshold, injection rate, gas-water alternating WAG injection parameters, fluid production rate, and bottom hole flowing pressure control threshold.

4. The intelligent adaptive CO2 miscible flooding reservoir development method as described in claim 1, characterized in that: In step S2, the well-to-ground-to-well monitoring network includes: a distributed fiber optic temperature / pressure / strain / acoustic monitoring subunit deployed along the entire wellbore; a downhole chromatographic fluid component online analysis subunit deployed in the oil layer perforation section; a nano-temperature-pressure / component sensor array implanted at the formation pore level; an inter-well intelligent tracer online monitoring subunit; and a microseismic monitoring subunit and a magnetotelluric monitoring subunit deployed on the surface. The dynamic data collected by the monitoring network includes formation temperature field, pressure field, stress field, seepage field, saturation field, CO2 concentration field, fluid composition data, fracture propagation and dominant seepage channel evolution data, and wellbore production dynamic data.

5. The intelligent adaptive CO2 miscible flooding reservoir development method as described in claim 1, characterized in that: In step S3, the dynamic minimum miscibility pressure The calculation formula is: In the formula: For development time; This is the initial minimum miscibility pressure; The coefficient representing the effect of light hydrocarbon extraction. This represents the mole fraction of C2-C6 light hydrocarbon components in crude oil under initial conditions. The mole fraction of C2-C6 light hydrocarbon components in crude oil at time t; Let t be the real-time temperature of the formation at time t; The original formation temperature; This is the correction factor for the CO2 dissolution effect; This is the dynamic contamination correction coefficient for pore structure.

6. The intelligent adaptive CO2 miscible flooding reservoir development method as described in claim 5, characterized in that: The formula for calculating the dissolution effect correction coefficient is as follows: ,in () represents the dissolved gas-oil ratio of CO2 in crude oil at time t; The formula for calculating the dynamic contamination correction coefficient of pore structure is: ,in The original stratum permeability, Let t be the real-time permeability of the formation.

7. The intelligent adaptive CO2 miscible flooding reservoir development method as described in claim 1, characterized in that: The full-dimensional miscibility factor MPF_full is calculated using the following formula: In the formula: The current formation fluid pressure for the grid cell; for The correction factor for contact efficiency with crude oil is calculated using the following formula: ),in This refers to the actual contact time between CO2 and crude oil. The critical contact time required for complete miscibility; The correction factor for multiple contact miscibility time is calculated using the following formula: ),in This represents the number of contact cycles between CO2 and crude oil. The reservoir heterogeneity correction factor is calculated using the following formula: ,in Permeability of the grid cell The average permeability of the reservoir; The interfacial tension correction coefficient is calculated using the following formula: ,in The real-time interfacial tension between CO2 and crude oil at time t. The interfacial tension between CO2 and crude oil in its original state; The molecular diffusion correction factor is calculated using the following formula: ,in Let be the molecular diffusion coefficient of CO2 in crude oil at time t. The limiting diffusion coefficient of CO2 in crude oil under formation conditions; The correction factor for fracture conductivity is calculated using the following formula: ,when hour, ;in For crack permeability, This represents the matrix permeability.

8. The intelligent adaptive CO2 miscible flooding reservoir development method as described in claim 1, characterized in that: In step S3, the formula for calculating the gas channeling risk probability (CRP) is as follows: In the formula: , , , , The weighting coefficients of each influencing factor satisfy the following conditions: 1; The CO2 transport rate ratio is calculated using the following formula: ,in This represents the CO2 front transport rate. For the migration rate of the crude oil displacement front; The coefficient of variation of permeability in the target area; The normalized value of the rate of increase in the gas-oil ratio is calculated using the following formula: ,in The current daily gas-oil ratio, This represents the gas-oil ratio threshold value³. The normalized value of the spatial gradient of the degree of miscibility is calculated using the following formula: ,in The spatial gradient magnitude of MPF_full This represents the maximum value of the MPF_full gradient modulus for the entire reservoir. The crack development factor is calculated using the following formula: ,in This represents the total length of the cracks within the grid cell. is the side length of the grid cell.

9. The intelligent adaptive CO2 miscible flooding reservoir development method as described in claim 1, characterized in that: The Comprehensive Benefit Evaluation Index (CEI) for mixed-phase flooding development is calculated using the following formula: In the formula: Let be the weight coefficients of each item, satisfying ; The total volume of the fully miscible zone and the near-miscible zone of the entire reservoir; This represents the total oil-bearing volume of the reservoir; This represents the increase in crude oil recovery rate; This represents the theoretical maximum recovery rate increase. CO2 geological sequestration rate; To develop net present value; This represents the theoretical maximum net present value. This represents the average gas channeling risk probability for the entire reservoir.

10. An intelligent adaptive CO2 miscible flooding reservoir development device, used to execute the intelligent adaptive CO2 miscible flooding reservoir development method as described in any one of claims 1-9, characterized in that: The device comprises five core systems: a well-to-ground-to-well integrated cross-scale intelligent monitoring system, a cloud-edge-device collaborative edge computing system, a cloud-based cross-scale digital twin intelligent decision-making platform, a surface intelligent injection digital twin execution system, and a downhole layered intelligent control and sealing execution system. The well-to-ground-to-well integrated multi-scale intelligent monitoring system is deployed in the injection wells, production wells, observation wells and surface well sites of the target reservoir. It is used to collect multi-scale, multi-physical field and multi-component dynamic data at the pore level, wellbore level and reservoir level during the reservoir development process, and transmit the collected data to the cloud-edge-device collaborative edge computing system. The cloud-edge-device collaborative edge computing system is deployed at the explosion-proof industrial control terminal of the well site and communicates with the well-ground-well integrated cross-scale intelligent monitoring system to complete multi-source data preprocessing, local anomaly early warning, edge model inference and emergency control. The cloud-based cross-scale digital twin intelligent decision-making platform is used to perform calculations of air-mixed phase distribution and division of control units, predict multi-timescale mixed phase evolution and classify and warn of gas channeling risks based on the CRP model, and generate a four-dimensional collaborative adaptive control strategy based on the CEI index and the single-well optimization objective function. The ground-based intelligent injection digital twin execution system is deployed at the injection well site and communicates with the cloud-based cross-scale digital twin intelligent decision-making platform. It is used to execute the injection of CO2, the mixed-phase enhancement system, and the regulation and sealing system, as well as the adaptive closed-loop control of injection pressure, injection rate, and slug combination. The downhole layered intelligent control and sealing execution system is deployed downhole in the injection well and production well, and is connected to the cloud-based cross-scale digital twin intelligent decision-making platform and cloud-edge-device collaborative edge computing system to execute downhole layered injection, flow control, and adaptive adjustment of production system.

Images