Crude oil detection digital twin system and application method

Abstract

The invention relates to the technical field of crude oil processing and digital twinning, and discloses a crude oil detection digital twinning system and an application method, and the method comprises the steps: constructing a digital twinning model integrated with an AI correction sub-model and a performance degradation prediction sub-model; collecting production data in real time to update a model state; virtual simulation and multi-target optimization screening are carried out on the multiple machining schemes based on the updated model; issuing the optimal scheme to guide production; and model closed-loop self-learning is driven by using production feedback data. The system comprises a physical entity layer, a data acquisition and transmission layer, a digital twin model layer, a simulation optimization layer and a production guidance layer. Through deep fusion of a mechanism and artificial intelligence and dynamic adaptive evolution of a model, high-fidelity real-time mapping, predictive maintenance support and autonomous optimization decision making of the whole crude oil processing flow are realized, and the technical problems that traditional offline simulation is disjointed with actual working conditions, the optimization period is long and the trial and error cost is high are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of crude oil processing and digital twin technology, specifically to a crude oil detection digital twin system and its application method. Background Technology

[0002] Crude oil processing is the core process of the refining industry, and its production efficiency and economic benefits highly depend on the accurate understanding of crude oil properties and the optimized control of the entire process. However, crude oil itself is complex and highly volatile, and the processing flow is long, involves many pieces of equipment, and is highly coupled, making the achievement of global precision optimization a continuous challenge. In existing technologies, refinery operation optimization mainly relies on two methods: one is qualitative adjustments based on operator experience; the other is offline simulation calculations using process simulation software. However, there are still significant limitations in practical applications. For example, existing methods mostly use static mechanistic models or purely data-driven models for simulation. Static mechanistic models are difficult to dynamically update their key parameters based on real-time production data, causing the model's prediction accuracy to become out of sync with actual operating conditions over time; while purely data-driven models lack physical constraints, resulting in low reliability and poor interpretability of prediction results when operating conditions fluctuate significantly, failing to provide a reliable basis for fine-grained optimization.

[0003] On the other hand, existing optimization technologies rely on testing and adjustments on physical equipment. This physical trial-and-error approach is risky, resource-intensive, and has a long optimization iteration cycle, failing to meet the demands of modern refining and chemical production for agile market response and real-time process adjustments.

[0004] To address the above issues, we propose a digital twin system for crude oil detection and its application method. Summary of the Invention

[0005] This invention provides a digital twin system and application method for crude oil detection, which helps to solve the problems mentioned in the background art.

[0006] On the one hand, the present invention provides the following technical solution: a method for applying digital twins in crude oil detection, comprising the following steps: S1. Construct and calibrate a digital twin model that maps the entire process of a physical oil refinery. The digital twin model includes a geometric model, a mechanism and process model, an equipment model, and a material model. The mechanism and process model has an embedded AI correction sub-model, and the equipment model integrates a performance degradation prediction sub-model. S2. Collect crude oil property data, equipment operating status data and material flow data of the physical refinery in real time, and transmit the processed data to the digital twin model in real time to update the operating status of the digital twin model; S3. Based on the updated digital twin model, perform virtual simulation of multiple processing schemes, and select processing schemes according to preset optimization objectives; S4. The selected processing scheme parameters are converted into production instructions and sent to the control system of the physical refinery to guide production; S5. Collect feedback data after the production instruction is executed, compare the feedback data with the corresponding simulation prediction value, and trigger self-learning and parameter update of relevant models in the digital twin model based on the comparison result.

[0007] By adopting the above technical solutions, a digital twin model integrating an AI correction sub-model and a performance degradation prediction sub-model is constructed. A closed-loop process from data acquisition and simulation optimization to instruction issuance and self-learning is established, realizing dynamic, high-fidelity virtual mapping and autonomous optimization of the entire oil refining process. This fundamentally solves the industry problems of the disconnect between traditional offline simulation and actual production, and the high cost and long cycle of optimization trial and error.

[0008] Furthermore, the mechanism-process model is constructed as follows: A basic process model is established based on the reaction kinetics equation, material balance equation, energy balance equation and phase equilibrium equation of crude oil refining. The AI ​​correction sub-model is constructed based on a physical information neural network, and its training loss function includes the physical equation constraint terms of the basic process model, which is used to dynamically correct the parameters of the basic process model based on real-time data.

[0009] By adopting the above technical solution, an AI correction sub-model is constructed using a physical information neural network, and its loss function is deeply integrated with the physical constraint terms of the mechanistic equation. This enables the data-driven learning process to strictly follow physical and chemical laws, thereby ensuring the rationality of the correction direction and the accuracy of the model's long-term prediction when dynamically correcting model parameters using real-time data. This overcomes the deficiency of pure data-driven models in lacking physical interpretability.

[0010] Furthermore, the performance degradation prediction sub-model is constructed based on a long short-term memory network to process historical operating sequence data of the equipment and output the efficiency decay coefficient of the equipment in the next planned maintenance cycle. The attenuation coefficient is used to dynamically adjust the operating efficiency parameters of the device model.

[0011] By adopting the above technical solution, a performance degradation prediction sub-model is constructed using a long short-term memory network, which can effectively learn the time series characteristics of equipment operating status, accurately predict future efficiency decay trends, and dynamically feed the predicted value back to the equipment model. This enables the full-process simulation to truly reflect the time-varying degradation impact of equipment performance, providing a key decision-making basis for achieving predictive maintenance and refined production scheduling based on equipment health status.

[0012] Furthermore, the crude oil property data includes the crude oil's density, viscosity, sulfur content, acid value, and true boiling point distillation data; The equipment operating status data includes temperature, pressure, flow rate, rotational speed, and energy consumption data.

[0013] Furthermore, the specific logic for updating the operating state of the digital twin model using data in S2 is as follows: Real-time collected data, predicted data within the model, and feedback data are associated and managed using a unified time-series identifier to form a continuous data trajectory. The real-time data in the data trajectory is compared in real time with the corresponding predicted values ​​of the mechanism-process model to generate a parameter deviation vector. The parameter deviation vector is used as the online learning signal of the AI ​​correction sub-model to adjust its network weights in real time, thereby correcting the output of the mechanism-process model.

[0014] By adopting the above technical solution, a digital thread based on a unified temporal identifier is proposed to manage data trajectories. The deviation vector between real-time data and model predictions is used to drive the online learning of the AI ​​model, realizing smooth, continuous, and automated updates of the digital twin model's state. This ensures that the virtual model responds agilely and synchronizes accurately to minor changes and dynamic processes in physical entities, thereby improving the model's real-time performance and fidelity.

[0015] Furthermore, the logic of the virtual simulation is as follows: The digital twin model performs a full-process simulation of multiple input processing schemes. The simulation simultaneously calls the mechanism process model, equipment model, and material model to calculate the material balance, product yield distribution, key equipment load, and comprehensive economic benefit indicators under each scheme, which serve as the basis for scheme comparison.

[0016] By adopting the above technical solutions, the virtual simulation is limited to simultaneously calling multiple models such as mechanisms, equipment, and materials to conduct full-process simulations and calculate comprehensive indicators such as material balance, product yield, equipment load, and economic benefits. This enables a comprehensive evaluation of the multi-dimensional and integrated impact of processing schemes, providing a unified, quantitative, and production management-oriented evaluation basis for scheme comparison and supporting global collaborative optimization.

[0017] Furthermore, the preset optimization objective is a multi-objective joint optimization function, whose input variables are the material balance, product yield distribution, key equipment load, and comprehensive economic benefit indicators output by the virtual simulation. The various processing schemes are comprehensively evaluated and screened through a weighted method.

[0018] By adopting the above technical solution, the optimization objective is concretized into a multi-objective joint optimization function that integrates multiple indicators such as product yield, energy consumption, and economic benefits. A weighted method is used for comprehensive evaluation, realizing scientific trade-offs and quantitative decision-making between conflicting objectives. This enables the system to automatically select the processing scheme that best suits the current comprehensive production strategy, thereby improving the intelligence and systematization of decision-making.

[0019] Furthermore, the logic for self-learning and parameter updating of the relevant model in S5 is as follows: An incremental learning algorithm with an integrated forgetting prevention mechanism is used to fine-tune the parameters of the AI ​​correction sub-model and the performance degradation prediction sub-model by comparing the deviation between the feedback data and the simulation prediction values, and the learning trajectory is recorded in the model knowledge base.

[0020] By adopting the above technical solution and using an incremental learning algorithm with an integrated forgetting prevention mechanism for model self-updating, it is ensured that while the model learns to adapt to new working conditions using new feedback data, it can effectively retain the ability to model historical working conditions, preventing catastrophic forgetting. This enables the continuous and stable accumulation and evolution of experience in the long-term operation of the digital twin model, ensuring the robustness of system performance.

[0021] On the other hand, a crude oil detection digital twin system, used to implement the above-mentioned crude oil detection digital twin application method, includes: The physical entity layer consists of the crude oil conveying equipment, reaction unit, separation equipment, storage equipment, and materials flowing through the above equipment in the oil refinery; The data acquisition and transmission layer is used to acquire crude oil property data, equipment operating status data, and material flow data in real time, and to preprocess and transmit the acquired data. The digital twin model layer includes: A geometric model is used to map the three-dimensional spatial structure of the physical entity layer; The mechanism and process model is constructed based on the reaction kinetics equation, material balance equation and energy balance equation of crude oil refining. Equipment models are used to simulate the operating status and performance degradation patterns of equipment. Material models are used to simulate changes in the composition of materials during processing. The simulation optimization layer is used to drive the digital twin model to perform working condition simulation, simulate the effects of multiple input processing schemes, and select processing schemes based on optimization objectives. The production guidance layer is used for visual presentation and issuing production instructions.

[0022] By adopting the above technical solutions, a complete system architecture is constructed, which includes a physical entity layer, a data acquisition and transmission layer, a digital twin model layer, a simulation optimization layer, and a production guidance layer. With the clear division and collaboration of hardware and software modules, a stable, efficient, and engineerable physical and logical carrier is provided for realizing the fully closed-loop intelligent application method, ensuring the effective implementation of the method innovation.

[0023] Furthermore, the digital twin model layer also includes a rapid adaptation module built on a meta-learning framework. When a change in crude oil type is detected, the rapid adaptation module uses historical learning experience to drive the digital twin model to quickly adjust to a state that matches the characteristics of the new crude oil under a small amount of new data.

[0024] By adopting the above technical solution, a rapid adaptation module based on the meta-learning framework is added to the system model, which gives the digital twin system the ability to quickly learn new tasks using historical experience. When faced with major changes in operating conditions such as switching of crude oil types, it can drive the model to quickly converge to a new state with a small amount of new data, which greatly shortens the model re-adaptation cycle and significantly improves the system's agility and practicality in dealing with production fluctuations.

[0025] The technical effects and advantages of this invention are as follows: 1. This invention achieves self-correction of the data-driven model under strict adherence to physicochemical laws by deeply integrating the training loss function of the physical information neural network with the physical equation constraint terms of the mechanism process model, thereby obtaining a high-fidelity, long-term accurate dynamic process simulation capability. 2. This invention integrates a performance degradation prediction sub-model based on a long short-term memory network into the equipment model to achieve accurate prediction of the future efficiency decline trend of the equipment, and feeds back the decline coefficient to the simulation process in real time, thereby realizing predictive maintenance support and refined production scheduling based on the health status of the equipment. 3. This invention performs full-process simulation and multi-objective optimization evaluation of various processing schemes in a digital twin space, and automatically converts the optimal scheme parameters into executable production instructions, thereby achieving a safe and efficient closed loop from virtual simulation to physical execution, and thus completing autonomous production optimization decision-making. 4. This invention employs an incremental learning algorithm with an integrated forgetting prevention mechanism for model self-updating, and combines it with a rapid adaptation module based on a meta-learning framework to achieve continuous and stable accumulation of experience in the long-term operation of the digital twin model, as well as rapid adaptation in the face of significant changes in operating conditions, thereby ensuring the long-term evolution and robustness of the system's intelligence. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0027] Figure 2 This is a schematic diagram of the system flow of the present invention.

[0028] In the diagram: 1. Physical entity layer; 2. Data acquisition and transmission layer; 3. Digital twin model layer; 4. Simulation optimization layer; 5. Production guidance layer. Detailed Implementation

[0029] The technical solutions of 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. Example

[0030] Reference Figure 1 A method for applying digital twins to crude oil detection includes the following steps: S1. Construct and calibrate a digital twin model that maps the entire process of a physical refinery. The digital twin model includes a geometric model, a mechanism and process model, an equipment model, and a material model. The mechanism and process model has an embedded AI correction sub-model, and the equipment model integrates a performance degradation prediction sub-model. It should be further explained that the mechanism-process model is constructed as follows: A basic process model is established based on the reaction kinetics equation, material balance equation, energy balance equation and phase equilibrium equation of crude oil refining. The AI ​​correction sub-model is built on a physical information neural network. Its training loss function includes physical equation constraint terms of the basic process model, which are used to dynamically correct the parameters of the basic process model based on real-time data.

[0031] The performance degradation prediction sub-model is built on a long short-term memory network to process historical operating sequence data of the equipment and output the efficiency decay coefficient of the equipment in the next planned maintenance cycle. The attenuation coefficient is used to dynamically adjust the operating efficiency parameters of the equipment model.

[0032] Specifically, the core of this step is to build a high-fidelity, updatable cluster of hybrid intelligent models; First, based on the factory's CAD drawings and P&ID diagrams, a geometric model with the same layout and dimensions as the physical factory is constructed using 3D modeling software, with the error controlled within 0.5%, for 3D visualization display.

[0033] Secondly, a core mechanistic process model is constructed. Taking the atmospheric and vacuum distillation unit as an example, its basic process model is built based on the following core mechanistic equations: Material balance equation: ,in, Represents all logistics Moore flows entering the system. Represents all Moore flows of logistics leaving the system. This represents the cumulative amount of material within the system; for steady-state simulations, the cumulative amount... Zero; Energy balance equation: ,in, The molar enthalpy of the feed stream. Let Q be the molar enthalpy of the output material, Q be the heat absorbed by the system from the outside (heat absorption is positive), and W be the shaft work done by the system on the outside (output work is positive). Phase equilibrium equation: The gas-liquid equilibrium constant is calculated using the Peng-Robinson equation of state. ; in, Let i be the mole fraction of component i in the gas phase. Let i be the mole fraction of component i in the liquid phase. Let be the fugacity coefficient of component i in the liquid phase. Let i be the fugacity coefficient of component i in the gas phase; Reaction kinetic equations: Taking catalytic cracking as an example, using a lumped kinetic model, such as a three-lumped model, the reaction rate follows the Arrhenius equation: Where r is the reaction rate, denoted as the pre-exponential factor, E as the activation energy of the reaction, R as the ideal gas constant, T as the reaction temperature, C as the reactant concentration, and n as the reaction order.

[0034] It should be further explained that, in terms of enhancing the basic process model, we have embedded an AI correction sub-model into the basic process model of each key unit. This AI correction sub-model is implemented using a physical information neural network, which is essentially a feedforward neural network, but its training process is strongly constrained by physical rules.

[0035] Specifically, its total loss function Designed as follows: ; The total loss value during the training of the physical information neural network; Data error term: Calculates the mean square error between the process parameters predicted by the neural network and the actual values ​​measured by the sensor; Physical constraint term: This term uses the above mechanistic equations as constraints; for example, it substitutes the predicted flow rates of each material stream output by the neural network into the material balance equations. Calculate its residuals. That is, the sum of squares of the residuals of these physical equations; , To balance the hyperparameters, they are real numbers greater than zero, used to adjust the balance between the accuracy of data fitting and the degree to which physical laws are satisfied.

[0036] By minimizing During training, the AI ​​correction sub-model not only learns to fit data, but its internal weight adjustments also spontaneously conform to physical laws. When running online, it receives real-time data and outputs dynamic correction values ​​Δ for key parameters in the basic mechanism model, thereby enabling the model to adaptively approximate the current real working conditions.

[0037] Third, a device model is constructed. In addition to integrating the device's mechanical parameters and design thresholds, a performance degradation prediction sub-model is integrated into the model of key rotating equipment. This sub-model is built based on a long short-term memory network; its input is the historical operating state sequence data of the device. ,in, Let be the outlet pressure at time t. Let be the bearing temperature at time t. Let be the vibration amplitude at time t; the Long Short-Term Memory (LSTM) network learns long-term dependencies in the sequence through its gating mechanism, and finally outputs an efficiency decay coefficient η (0 < η < 1) for the next planned maintenance cycle; this coefficient will be used as a dynamic parameter input to the device model to calculate the actual efficiency of the device in its current health state in real time. ; in, For the actual operating efficiency of the equipment, The rated efficiency of the equipment.

[0038] Finally, a material model is constructed based on the crude oil evaluation database and real-time online analysis data. A set of attribute vectors is maintained for each stream, including density, average molecular weight, hydrocarbon composition, etc., and is updated in real time during process simulation as the material migrates and reacts.

[0039] S2. Real-time acquisition of crude oil property data, equipment operating status data, and material flow data from physical refineries, and real-time transmission of the processed data to the digital twin model to update the operating status of the digital twin model; It should be further noted that the crude oil property data includes the crude oil's density, viscosity, sulfur content, acid value, and true boiling point distillation data; Equipment operating status data includes temperature, pressure, flow rate, speed, and energy consumption data.

[0040] In S2, the running status of the digital twin model is updated using data. The specific logic is as follows: Real-time collected data, predicted data within the model, and feedback data are associated and managed through a unified time-series identifier to form a continuous data trajectory. The real-time data in the data trajectory is compared with the corresponding predicted values ​​of the mechanism and process model in real time to generate a parameter deviation vector. The parameter deviation vector is used as an online learning signal for the AI ​​correction sub-model to adjust its network weights in real time to correct the output of the mechanism and process model.

[0041] Specifically, the data acquisition and transmission layer 2 operates continuously. The online analyzer provides a set of key crude oil property data every 5 minutes, including the crude oil's density, viscosity, sulfur content, acid value, and true boiling point distillation curve. The sensors collect temperature, pressure, flow rate, rotation speed, and energy consumption data once per second. The edge gateway uses median filtering to remove impulse noise and linear interpolation to complete the data points that are momentarily lost, forming a regular time-series data stream. Furthermore, a digital thread based on a unified time-series identifier is established to manage the data. Every data packet, whether it is real-time acquired data, model-internal predicted data, or subsequent production feedback data, is marked with a unified timestamp and material / equipment identifier, forming a traceable data trajectory.

[0042] Specifically, the model update process is as follows: when a new batch of real-time data arrives, the system first compares it with the predicted value of the mechanism-process model for the same measuring point at the previous moment, based on the timestamp and identifier, and generates a parameter deviation vector. ,For example, This represents the deviation between the actual tower top temperature and the predicted value; subsequently, this deviation vector... Instead of being used directly to correct the mechanistic equations, it is used as an online learning signal input to the AI ​​correction sub-model. This sub-model fine-tunes its network weights through a round of forward and backward propagation, thereby changing the amount of correction Δ to the parameters of the basic mechanistic model. In this way, the model uses real-time data streams for continuous, incremental learning, achieving smooth and adaptive updates to its operating state.

[0043] S3. Based on the updated digital twin model, perform virtual simulation of multiple processing schemes, and select processing schemes according to preset optimization objectives; It should be further explained that the logic of virtual simulation is as follows: The digital twin model performs a full-process simulation of multiple input processing schemes. The simulation simultaneously calls the mechanism process model, equipment model and material model to calculate the material balance, product yield distribution, key equipment load and comprehensive economic benefit indicators under each scheme, which serve as the basis for scheme comparison.

[0044] The preset optimization objective is a multi-objective joint optimization function, whose input variables are the material balance, product yield distribution, key equipment load and comprehensive economic benefit indicators output by virtual simulation. Multiple processing schemes are comprehensively evaluated and screened through a weighted method.

[0045] Specifically, operators can set or modify processing parameters on the visualization model through the production guidance layer 5 interface. For example, they can adjust the reaction temperature of the catalytic cracking unit and the side stream extraction rate of the vacuum tower. Multiple alternative schemes can be entered at the same time. Simulation optimization layer 4 initiates the simulation. The simulation logic is as follows: it calls the updated entire digital twin model cluster to perform a full-process, unsteady-state deduction for each input processing scheme. During the deduction, the mechanism process model is responsible for calculating the material and energy balance of each unit; the equipment model is responsible for calculating the actual power consumption and processing capacity limitations under the current healthy state; the material model is responsible for tracking the evolution of the properties of each stream of logistics. After the simulation, the system outputs a complete set of performance indicators for each scheme, including: material balance sheet, yield distribution of target product, load rate of key equipment, and comprehensive economic benefit indicators calculated based on current market prices and energy consumption.

[0046] Specifically, a multi-objective joint optimization function F is used to comprehensively evaluate the above performance indicators. This function is expressed as follows: ; F: Overall evaluation score of the processing plan; : This is the normalized value of the target product yield, i.e., the ratio of the actual yield to the maximum designed yield; : The comprehensive energy consumption per unit of product; : This is the normalized value of the hourly net return, which is the ratio of the actual return to the historical highest return; : Pollutant emissions per unit of product; , , , : A weighting coefficient greater than 0, pre-set according to the production strategy; for example, in the stage of pursuing high efficiency, The setup is relatively large; during periods of stringent environmental regulations... The setting is relatively large.

[0047] The system calculates the score of each solution under this evaluation function F and automatically selects the solution with the highest score as the recommended solution.

[0048] S4. Convert the selected processing scheme parameters into production instructions and send them to the control system of the physical refinery to guide production; Specifically, the production guidance layer 5 automatically encapsulates all executable parameters in the recommended scheme into a standard instruction sequence that conforms to the communication protocol of the refinery's distributed control system, and sends it out through a secure network link. After receiving the instruction, the DCS system executes it automatically or after operator confirmation, thereby applying the optimization results of the virtual space to the physical entity.

[0049] S5. Collect feedback data after the production command is executed, compare the feedback data with the corresponding simulation prediction values, and trigger self-learning and parameter updates of relevant models in the digital twin model based on the comparison results.

[0050] It should be further explained that the logic of self-learning and parameter update of the relevant model in S5 is as follows: An incremental learning algorithm with an integrated forgetting prevention mechanism is adopted. By comparing the deviation between feedback data and simulation prediction values, the parameters of the AI ​​correction sub-model and the performance degradation prediction sub-model are fine-tuned, and the learning trajectory is recorded in the model knowledge base.

[0051] Specifically, after the instruction is executed, the system will continuously collect actual production results as feedback data during a subsequent evaluation cycle, including the output, quality, and actual energy consumption of the final product. The system will accurately compare the feedback data with the simulation prediction values ​​corresponding to the simulation, and calculate the deviation of the key performance indicators. This deviation is not only used to evaluate the effectiveness of the scheme, but more importantly, it will trigger the deep self-learning of the model. Specifically, an incremental learning algorithm (EWC) with an integrated forgetting prevention mechanism is used to update the AI ​​correction sub-model and the performance degradation prediction sub-model. Taking the updating of the AI ​​corrector sub-model as an example, the EWC algorithm, while minimizing the loss caused by the new data, adds a penalty term to the network weights that are crucial to the previously learned task. Its loss function is expanded as follows: ; The total loss value of the EWC algorithm; Standard loss calculated based on new feedback data; The current value of the i-th weight parameter in the neural network; This weight parameter represents the important values ​​saved from previous training sessions, i.e., parameter values ​​that were important for historical tasks. The diagonal values ​​of the historical Fisher information matrix for this weight parameter are used to quantitatively measure the importance of this parameter to past tasks. : The hyperparameter that controls the strength of preventing forgetting is a real number greater than 0.

[0052] By minimizing The model can absorb new knowledge while retaining old knowledge to the greatest extent, preventing catastrophic forgetting caused by continuous learning. All weight adjustment trajectories in the learning process are recorded in the model knowledge base, forming a systematic accumulation of experience.

[0053] Example 2 Reference Figure 2 A crude oil detection digital twin system, used to implement the aforementioned digital twin application method, includes: Physical entity layer 1 consists of the crude oil conveying equipment, reaction unit, separation equipment, storage equipment of the oil refinery, and the materials flowing through the above equipment; Data acquisition and transmission layer 2 is used to acquire crude oil property data, equipment operating status data and material flow data in real time, and to preprocess and transmit the acquired data. Digital twin model layer 3 includes: Geometric model, used to map the three-dimensional spatial structure of physical entity layer 1; The mechanism and process model is constructed based on the reaction kinetics equation, material balance equation and energy balance equation of crude oil refining. Equipment models are used to simulate the operating status and performance degradation patterns of equipment. Material models are used to simulate changes in the composition of materials during processing. Simulation optimization layer 4 is used to drive the digital twin model to perform working condition simulation, simulate the effects of multiple input processing schemes, and select processing schemes based on optimization objectives. Production guidance layer 5 is used for visual display and issuing production instructions.

[0054] It should be further explained that the digital twin model layer 3 also includes a rapid adaptation module built on a meta-learning framework. When a change in crude oil type is detected, the rapid adaptation module uses historical learning experience to drive the digital twin model to quickly adjust to a state that matches the characteristics of the new crude oil with a small amount of new data.

[0055] Specifically, the crude oil detection digital twin system in this embodiment is deployed in a physical refinery. The physical entity layer 1 is a refinery with a complete process including atmospheric and vacuum distillation, catalytic cracking, and hydrorefining. In the data acquisition and transmission layer 2, more than 300 measuring points, such as density sensors, online near-infrared analyzers, temperature sensors, pressure sensors, mass flow meters, and electricity meters, are deployed at locations such as the crude oil inlet pipeline, the inlet and outlet of each key reactor, and the side line of the fractionation tower. These sensors aggregate the data to the edge computing gateway for preprocessing via industrial Ethernet and 5G network before transmitting it to the central server. The digital twin model layer 3, simulation optimization layer 4, and production guidance layer 5 run in a high-performance computing cluster on a central server. The visualization interface of the production guidance layer 5 is accessible to the workstation in the control room via a web interface.

[0056] Specifically, in the third layer of the digital twin model, a rapid adaptation module based on a meta-learning framework is also deployed. This module runs continuously in the background, and its goal is to enable the system to learn how to learn quickly. When the online detection module identifies a significant switch in crude oil type, this module is activated. Instead of training the model from scratch, it calls upon the experience of learning multiple crude oils in the past from the model knowledge base, and based on a small amount of new data collected in a short period of time after the switch, it drives the AI ​​correction sub-model and other models to quickly adjust to an initial state that fits the characteristics of the new crude oil well within a very small number of iteration steps, thereby shortening the model adaptation cycle for new oil types from the traditional several days to several hours.

[0057] It should be further explained that the online detection module is configured to perform automatic detection of crude oil type switching. Its detection logic is based on real-time trend analysis and abrupt change identification of key crude oil property parameters. The specific steps are as follows: The system continuously acquires key crude oil property data, including density ρ and sulfur content S, provided in real time by the online analyzer. The system maintains a sliding time window for density and sulfur content respectively, and calculates the average value μ and standard deviation σ of the parameter within each data update cycle. When the collected data points Upon arrival, the system calculates its standardized deviation Z-score relative to the current trend; for example, for density: ; Similarly, calculate the sulfur content. ; A joint judgment threshold is established. When both of the following conditions are met simultaneously, the system automatically determines that a change in crude oil type has occurred: Condition 1: The Z-score of density or sulfur content exceeds the preset significant mutation threshold, indicating that the parameter has undergone a statistically significant jump; Condition 2: This abrupt change state is maintained for more than one stable time period, such as three consecutive sampling periods, to eliminate single measurement noise or transient interference from the pipeline oil mixing interface; Once the judgment conditions are met, the online detection module immediately generates a crude oil type switching event signal. This signal is marked with a unified timestamp and published to the system's event bus through the data acquisition and transmission layer 2.

[0058] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments of this disclosure. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit 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 crude oil detection digital twin application method, characterized by, The method comprises the following steps: S1. Constructing and calibrating a digital twin model mapping a physical refinery process, the digital twin model comprising a geometric model, a mechanism process model, a device model and a material model; the mechanism process model has an AI correction sub-model embedded therein, and the device model has a performance degradation prediction sub-model integrated therein; S2. Collecting real-time crude oil property data, device operating state data and material flow data of the physical refinery, and transmitting the processed data to the digital twin model in real time to update the operating state of the digital twin model; S3. Based on the updated digital twin model, a plurality of processing schemes are virtually simulated, and a processing scheme is selected according to a preset optimization target; S4. The parameters of the selected processing scheme are converted into production instructions and sent to the control system of the physical refinery to guide production; S5. Collecting feedback data after the production instructions are executed, comparing the feedback data with the corresponding simulation prediction values, and triggering self-learning and parameter updating of related models in the digital twin model according to the comparison result.

2. A crude oil detection digital twin application method according to claim 1, characterized in that, The mechanism process model is constructed in the following manner: A basic process model is established based on the reaction kinetics equation, material balance equation, energy balance equation and phase equilibrium equation of crude oil refining; The AI correction sub-model is constructed based on a physical information neural network, and its training loss function contains a physical equation constraint term of the basic process model, which is used to dynamically correct the parameters of the basic process model according to real-time data.

3. The crude oil detection digital twin application method of claim 1, wherein, The performance degradation prediction sub-model is constructed based on a long short-term memory network, which is used to process historical operating sequence data of the device and output an efficiency decay coefficient of the device in the next planned maintenance period; The decay coefficient is used to dynamically adjust the operating efficiency parameters of the device model.

4. The crude oil detection digital twin application method of claim 1, wherein, The crude oil property data includes the density, viscosity, sulfur content, acid value and true boiling point distillation data of crude oil; The device operating state data includes temperature, pressure, flow rate, rotational speed and energy consumption data.

5. The crude oil detection digital twin application method of claim 1, wherein, In S2, the operating state of the digital twin model is updated using data, and the specific logic is as follows: The real-time collected data, predicted data in the model and feedback data are associated and managed through a unified time sequence identifier to form a continuous data track; the real-time data in the data track is compared with the corresponding predicted values of the mechanism process model in real time to generate a parameter deviation vector; the parameter deviation vector is used as an online learning signal of the AI correction sub-model to adjust the network weight in real time to correct the output of the mechanism process model.

6. The crude oil detection digital twin application method of claim 1, wherein, The logic of virtual simulation is as follows: The inputted plurality of processing schemes are deduced in the digital twin model, the mechanism process model, the device model and the material model are called synchronously, the material balance, product yield distribution, key device load and comprehensive economic benefit index under each scheme are calculated, and the calculation results are used as the basis for scheme comparison.

7. The crude oil detection digital twin application method of claim 1, wherein, The preset optimization target is a multi-objective joint optimization function, input variables of which are material balance, product yield distribution, key equipment load and comprehensive economic benefit index of the virtual simulation output, and the multiple processing schemes are comprehensively evaluated and screened through a weighting method.

8. The crude oil detection digital twin application method of claim 1, wherein, The logic of self-learning and parameter updating of the relevant model in S5 is as follows: An incremental learning algorithm with integrated forgetting prevention mechanism is adopted, the comparison deviation between the feedback data and the simulation prediction value is used to fine-tune the parameters of the AI correction sub-model and the performance degradation prediction sub-model, and the learning trajectory is recorded to the model knowledge base.

9. A crude oil detection digital twin system for implementing the crude oil detection digital twin application method of any one of claims 1-8, wherein, Comprise: A physical entity layer (1) composed of crude oil transportation equipment, reaction devices, separation equipment, storage equipment of a refinery and materials flowing through the above-mentioned equipment; A data acquisition and transmission layer (2) for real-time acquisition of crude oil property data, equipment operating state data and material flow data, and preprocessing and transmission of the acquired data; A digital twin model layer (3) comprising: A geometric model for mapping the three-dimensional spatial structure of the physical entity layer (1); A mechanism process model constructed based on the reaction kinetics equation, material balance equation and energy balance equation of crude oil refining; An equipment model for simulating the operating state and performance degradation law of the equipment; A material model for simulating the component change of the material in the processing process; An analog optimization layer (4) for driving the digital twin model to simulate the working condition, simulating the effect of the input multiple processing schemes, and selecting the processing schemes based on the optimization target; A production guidance layer (5) for visual display and issuing production instructions.

10. The crude oil detection digital twin system of claim 9, wherein, The digital twin model layer (3) further comprises a rapid adaptation module constructed based on a meta-learning framework, which utilizes historical learning experience to drive the digital twin model to quickly adjust to a state matched with the characteristics of new crude oil under a small amount of new data when detecting a crude oil type switching.

Images