Continuous reforming regenerator temperature peak value positioning method and system

Abstract

The invention discloses a temperature peak positioning method and system for a continuous reforming regenerator, and solves the problems of poor combustion temperature prediction precision and insufficient stability of the continuous reforming regenerator in the prior art, and the method comprises the steps: collecting process variables from a DCS system, building a three-dimensional model of the continuous reforming regenerator, and carrying out the prediction of the combustion temperature of the continuous reforming regenerator; obtaining temperature field and temperature peak position coordinates under each different working condition; establishing a mechanism model based on mass-energy balance and reaction kinetics in the regeneration process, adopting an XGBoost algorithm as data driving compensation, constructing a hybrid model of mechanism constraint and XGBoost data driving compensation, and performing training; and based on the trained hybrid model, batch processing reasoning is adopted, and current temperature peak coordinates and prediction credibility are output. A hybrid model with physical interpretability and high precision is constructed, prediction errors under extreme working conditions are reduced, and model stability is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of petrochemical process monitoring and optimization technology, and in particular to a method and system for locating the temperature peak of a continuous reforming regenerator. Background Technology

[0002] Continuous reforming units are core equipment in the petroleum refining industry for producing high-octane gasoline components and aromatic feedstocks. They restore catalyst activity by burning coke deposits on the catalyst surface. The internal environment is a high-temperature, high-pressure, and highly corrosive gas-solid two-phase flow environment, and the product quality and operational stability directly determine the economic benefits of refineries. As a "core consumable" in continuous reforming units, the catalyst gradually accumulates coke deposits on its surface during the reaction process, leading to a decrease in catalytic activity. Periodic regeneration is required through a regenerator—oxygen-containing gas is introduced into the regenerator to react with the coke deposits on the catalyst surface (the temperature is typically controlled at 480-520℃) to restore catalyst activity. Traditional methods mostly focus on modifying the regeneration reactor of the catalyst regeneration system, aiming to optimize the combustion temperature in the coke burning zone of the regenerator and extend the catalyst's lifespan.

[0003] For example, patent CN110931085A describes an online prediction and optimization method for coke content in continuous reforming catalysts. Based on data-driven analysis, it uses regeneration system parameters as input to predict the coke content of the catalyst online, partially linking it with the APC controller to adjust the oxygen content in the regenerator, thus improving prediction real-time performance and operational optimization efficiency. However, some of its models rely on specific parameters (such as regeneration system data), making them susceptible to interference under extreme conditions. Furthermore, some aspects do not fully address the issues of multivariate coupling and dynamic condition adaptability, and some deviations from actual on-site conditions require frequent correction, resulting in poor prediction accuracy. Summary of the Invention

[0004] The purpose of this invention is to solve the problems of poor accuracy and insufficient stability in the prediction of combustion temperature in continuous reforming regenerators in the prior art. It provides a method and system for locating the peak temperature of continuous reforming regenerators, constructs a hybrid model that combines physical interpretability and high accuracy, reduces prediction errors under extreme operating conditions, improves the accuracy of peak temperature prediction, and ensures model stability.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for locating the temperature peak in a continuous reforming regenerator includes the following steps: Process variables are collected from the DCS system to establish a three-dimensional model of the continuous reforming regenerator and obtain the temperature field and temperature peak location coordinates under each different operating condition. A mechanistic model was established based on the mass-energy balance and reaction kinetics of the regeneration process. The XGBoost algorithm was used as data-driven compensation. A hybrid model of mechanistic constraints and XGBoost data-driven compensation was constructed and trained. Based on the trained hybrid model, batch inference is used to output the current temperature peak coordinates and prediction confidence.

[0006] This application collects process variables of catalyst regeneration in a continuous reforming regenerator from a DCS system, and then uses CFD simulation to obtain the preliminary temperature peak coordinates and corresponding temperature values ​​under different operating parameters. These are used as a dataset for subsequent hybrid model training. The XGBoost algorithm is used as data-driven compensation, with actual operating parameters and the output of a simplified mechanistic model as features, and the peak grid numbers obtained from simulation as labels, for hybrid model training. The trained hybrid model is deployed on an industrial control server to collect real-time process data, enabling real-time prediction of the coordinates of the current temperature peak. Under extreme conditions (such as sudden changes in raw material sulfur content), the prediction error is ≤±2℃, which is more accurate than pure mechanistic models (error ±5-8℃) and pure BP neural networks (error ±3-4℃), and there are no abnormal results of "low temperature predicting high temperature".

[0007] Preferably, the mechanistic model outputs preliminary temperature peak coordinates and corresponding temperature values; the XGBoost algorithm is used as data-driven compensation, with process variables and the output of the mechanistic model as inputs, and the residual between the actual / simulated temperature peak coordinates and the output of the mechanistic model as output; and a hybrid model is trained based on the process variables and the output data of the three-dimensional model.

[0008] As a preferred method, the continuous reforming regenerator is meshed, and the temperature peak coordinates are represented by the mesh number; the axial coordinates of the temperature peak are corrected in combination with the catalyst circulation rate trend: when the circulation rate increases by 20%, the axial position of the temperature peak is shifted upward by 0.5-1 meshes.

[0009] Preferably, the mechanism model includes: obtaining the coke combustion rate based on oxygen concentration and activation energy, and establishing a kinetic equation for the coking reaction; establishing a micro-element mass-energy balance equation based on the coke combustion rate; solving the two equations using the finite difference method, and outputting the preliminary temperature peak coordinates.

[0010] As a preferred approach, a three-dimensional model of the continuous reforming regenerator is established using CFD simulation. n sets of parameters are set for each of the 12 operating conditions, and temperature distribution data under each operating condition are output. The peak coordinates of the steady-state stage are extracted as labels and time-synchronized with DCS data to form a complete dataset. The dataset is divided into training set, validation set and test set according to the proportion, and a hybrid model is trained using iterative training and early shutdown mechanism.

[0011] Preferably, the process variables include regeneration recirculation gas flow rate, regenerator top pressure, catalyst circulation rate, and recirculation gas oxygen concentration.

[0012] As a preferred approach, based on the temperature peak location results, two levels of early warning thresholds are set: if the temperature peak temperature is greater than 530℃ or the temperature peak coordinate is less than 0.3 meters from the device wall, a level one early warning is triggered, and operation suggestions are displayed simultaneously; if the temperature peak temperature is greater than 550℃ or the temperature peak coordinate remains in the danger zone for more than 30 seconds, a level two early warning is triggered, and early warning information is pushed. If there is no manual intervention within a preset time, an adjustment instruction is automatically sent to the DCS system.

[0013] As a preferred approach, the collected process variables are preprocessed: outliers are removed using the 3σ criterion, and missing data are supplemented by linear interpolation between adjacent time points; high-frequency noise is eliminated using Kalman filtering; and the filtered data is standardized using Z-score.

[0014] A continuous reforming regenerator temperature peak positioning system includes: The DCS data acquisition layer acquires DCS data in real time via the OPC UA protocol to obtain raw training data and perform preprocessing. The model computation layer constructs a hybrid model of mechanistic constraints and XGBoost data-driven compensation to perform real-time prediction of temperature peaks. The visualization and early warning layer displays a two-dimensional grid diagram of the regenerator and the location of temperature peaks, shows real-time values ​​and trend curves of process variables, and issues early warnings based on real-time temperature peaks.

[0015] Preferably, the prediction model includes a mechanism module and an XGBoost module. The mechanism module provides the physical boundary of the continuous reforming regenerator and reconstructs the global temperature field. The XGBoost module learns the residuals and captures the nonlinear relationships in the catalyst regeneration process of the continuous reforming regenerator.

[0016] Therefore, the present invention has the following beneficial effects: it adopts a hybrid model architecture of "mechanism constraint + XGBoost", which provides physical boundaries through the mechanism model (avoiding predictions that violate physical laws) and the XGBoost module learns residuals (capturing nonlinear relationships), thus solving the problems of low accuracy of pure mechanism models and poor stability of pure data-driven models. Attached Figure Description

[0017] Figure 1 This is a flowchart of the steps in the method for locating the temperature peak of the continuous reforming regenerator in this invention.

[0018] Figure 2 This is a schematic diagram of the virtual mesh division of the continuous reforming regenerator in this invention.

[0019] Figure 3 This is a schematic diagram of the architecture for locating the temperature peak of the continuous reforming regenerator in this invention.

[0020] In the diagram: L1: Axial first layer; L5: Axial fifth layer; L10: Axial tenth layer; L15: Axial fifteenth layer; L20: Axial twentieth layer; L25: Axial twenty-fifth layer; L30: Axial thirtieth layer; R1: Radial first zone; R2: Radial second zone; R3: Radial third zone; R4: Radial fourth zone; R5: Radial fifth zone; 1. Continuous reforming device; 2. DCS data acquisition layer; 3. Real-time database; 4. Model calculation layer; 5. Visualization and early warning layer; 6. Mechanism module; 7. XGBoost module. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this application will be further described in detail below through embodiments and in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only one preferred embodiment of this application and are only used to explain this application. They do not limit the scope of protection of this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0022] According to industry statistics, the catalyst replacement cost for a 1.2 Mt / a continuous reforming unit is as high as 5-8 million RMB per cycle, and the daily loss from unplanned shutdowns for maintenance is 300,000-500,000 RMB. Temperature control is crucial during regeneration: too low a temperature leads to incomplete combustion of coke deposits, resulting in a catalyst activity recovery rate of less than 70%, which in turn reduces the reforming reaction conversion rate by 5%-8%; too high a temperature (especially localized "hot spots" exceeding 550℃) can cause catalyst crystal structure sintering, resulting in permanent deactivation and directly leading to unplanned shutdowns of the unit.

[0023] Modifying the regeneration reactor requires different design specifications for different scale units, leading to high exploration and maintenance costs. Relying on computer software modeling for prediction faces complex model training and testing issues, and its applicability is relatively limited to a single operating condition. Therefore, this application uses temperature prediction based on the combustion temperature distribution in the coke burning zone to adjust the temperature during the regeneration process and ensure catalyst activity.

[0024] Currently, the mainstream method for monitoring regenerator temperature in industry is "thermocouple monitoring." This involves placing 8-12 high-temperature, corrosion-resistant thermocouples (mostly made of Inconel alloy) on the inner wall or internal components of the regenerator to collect local temperature data in real time and transmit it to the DCS system for display. To improve monitoring accuracy, some technicians add distributed fiber optic temperature measurement equipment, which uses optical fibers laid along the regenerator axis to achieve continuous axial temperature monitoring.

[0025] In addition, there are two auxiliary technical solutions in the academic and industrial fields: one is temperature prediction based on mechanism models, which calculates the temperature distribution trend in the regenerator by establishing the mass-energy balance equation and reaction kinetic model of the regeneration process; the other is prediction based on pure data-driven methods, which uses machine learning algorithms such as BP neural networks and random forests, with historical DCS data as input, to predict whether the local temperature exceeds the standard.

[0026] However, existing technologies mainly have the following drawbacks: (1) Incomplete measurement and monitoring blind spots: Due to high installation costs and limited erosion resistance, thermocouples cannot be densely arranged (usually only covering 3-5 axial sections and 1-2 radial areas of the regenerator), making it impossible to capture the full picture of the three-dimensional non-uniform temperature field. For example, when the temperature peak appears at the radial edge or axial transition section between two adjacent thermocouples, the existing measuring points cannot detect it, forming a "blind spot". A case from a refinery in 2023 showed that 5% of the catalyst sintered because hot spots appeared in areas not covered by thermocouples, and the problem was only discovered afterward through unloading inspection.

[0027] (2) Delayed early warning and irreversible losses: Both thermocouple and fiber optic temperature measurement rely on the "heat conduction-signal conversion" process, with a response delay of 0.5-1 minute; and after the hot spot is formed, it needs to be transferred to the measuring point through heat diffusion, with the actual early warning delay time as long as 3-5 minutes. When the measuring point detects an abnormal temperature, the high temperature has already caused irreversible damage to the catalyst - experimental data shows that if the high temperature above 550℃ is maintained for 1 minute, the catalyst activity will decrease by 20%; if it is maintained for 5 minutes, it will be completely deactivated. At this time, the control can only prevent the hot spot from expanding, but cannot recover the losses.

[0028] (3) Insufficient model accuracy and stability: Existing mechanistic models simplify complex factors such as gas-solid two-phase flow mixing and local flow field disturbance, resulting in temperature prediction errors of ±5-8℃, making it impossible to accurately locate peak values; Pure data-driven models (such as BP neural networks) can fit historical data, but lack physical constraints, and are prone to "violation of physical laws" prediction results under extreme conditions (such as sudden changes in raw material sulfur content and fluctuations in circulating gas flow rate) (such as predicting high-temperature hot spots under low-temperature conditions), resulting in insufficient stability.

[0029] (4) Weak operational guidance: Existing technology can only provide an alarm signal for "temperature exceeding the standard" and cannot inform the operator of the specific location of the hot spot (such as "5 meters in the axial direction and 0.6-0.9 meters in the radial direction"), which leads to a large degree of blindness when the operator adjusts the operating parameters (such as secondary air volume and catalyst circulation rate). About 30% of hot spot accidents are aggravated by untimely operation adjustments or misoperation.

[0030] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0031] Example 1: This embodiment provides a method for locating the temperature peak of a continuous reforming regenerator, such as... Figure 1 As shown, it includes the following steps: Step 1: Collect process variables from the DCS system, establish a three-dimensional model of the continuous reforming regenerator, and obtain the temperature field and temperature peak location coordinates under each different operating condition as target variables.

[0032] Key process variables of the catalytic process in the continuous reforming regenerator are collected from the DCS system, and the collected process variables are preprocessed and standardized; the target variable is the temperature peak position coordinates (X_peak, Y_peak), which is determined by CFD simulation.

[0033] Specifically, this involves: using ANSYS Fluent software to build a 3D model of the regenerator; inputting parameters for different operating conditions; setting n sets of parameters for each of the 12 operating conditions; outputting temperature distribution data for each operating condition; simulating the global temperature distribution; extracting the virtual grid coordinates of the peak values ​​as labels; and synchronizing and matching them with DCS data to form a complete dataset. This complete dataset is then used as the dataset for training the hybrid model in step two, employing iterative training and an early stopping mechanism to train the hybrid model.

[0034] Temperature peak (hot spot): The area where the local temperature inside the regenerator rises abnormally. When the temperature exceeds 550℃, it is easy to cause the catalyst crystal structure to sinter, resulting in permanent deactivation.

[0035] In this embodiment, the continuous reforming regenerator is divided into grids, and the temperature peak coordinates are represented by grid numbers.

[0036] Step 2: Establish a mechanistic model based on the mass-energy balance and reaction kinetics of the regeneration process, use the XGBoost algorithm as data-driven compensation, construct a hybrid model of mechanistic constraints and XGBoost data-driven compensation, and train it.

[0037] XGBoost, or Extreme Gradient Boosting, is a high-efficiency machine learning algorithm that uses multiple decision trees trained sequentially to fit the residuals, resulting in high accuracy and resistance to overfitting.

[0038] A simplified model (mechanistic model) is established based on the mass-energy balance and reaction kinetics of the regeneration process: the coke combustion rate is obtained based on oxygen concentration and activation energy, and the coking reaction kinetic equation is established; the micro-element mass-energy balance equation is established based on the coke combustion rate; the two equations are solved using the finite difference method, and the preliminary temperature peak coordinates are output.

[0039] Using process variables and mechanistic model outputs as inputs, and the residual between actual / simulated peak coordinates and mechanistic model outputs as outputs, the model learns nonlinear relationships (such as local flow field disturbances) that the mechanistic model cannot capture. Based on the process variables and 3D model output data obtained in step one, the model is trained using a hybrid model.

[0040] Step 3: Based on the trained hybrid model, batch inference is used to output the current temperature peak coordinates and prediction confidence.

[0041] The trained hybrid model is deployed on an industrial control server, and batch inference is used to output the current temperature peak coordinates and prediction confidence. Simultaneously, correction coefficients are established to adjust the prediction results.

[0042] The continuous reforming regenerator temperature peak location method provided in this embodiment has the following beneficial effects: Based on a hybrid "mechanism + data" model, it achieves an integrated solution with "full-range blind-spot-free location, ultra-short response early warning, physical stability and reliability, zero hardware cost, and closed-loop operation guidance." This solves the high cost problem of pure hardware solutions and overcomes the shortcomings of pure data-driven or simplified mechanism models, such as "insufficient accuracy, physical anomalies, and functional deficiencies." It better meets the actual needs of industrial scenarios for "accuracy, reliability, low cost, and ease of implementation." It is applicable to the reconstruction of the internal temperature field, real-time location and early warning of temperature peaks in continuous reforming regenerators, ensuring catalyst activity and extending the unit's operating cycle.

[0043] Example 2: This embodiment provides a method for locating the temperature peak of a continuous reforming regenerator, including: Step 1, data acquisition and preprocessing; Step 2, construction and training of a hybrid prediction model; Step 3, real-time location of the temperature peak; and Step 4, visualization and early warning.

[0044] Specifically: Step one provides high-quality data support for the hybrid model, which includes three parts: Step (1.1): Determine the data collection range.

[0045] Eighteen key process variables were collected from the DCS system, covering: regeneration circulating gas flow rate (5000-8000 Nm³). 3 / h, accuracy ±1%), regenerator top pressure (0.3-0.5 MPa, accuracy ±0.01 MPa), circulating gas oxygen concentration (1%-5%, accuracy ±0.1%), catalyst circulation rate (50-100 t / h, accuracy ±2 t / h), temperature of 12 thermocouple measuring points (450-550℃, accuracy ±1℃), coking zone inlet temperature, chlorination zone temperature, etc.

[0046] The coordinates of the temperature peak location (X_peak, Y_peak) were determined using CFD simulation: A 3D model of the regenerator was established with 2 million mesh elements, employing the RNG k-ε turbulence model and the finite rate / eddy dissipation reaction model. Parameters under different operating conditions were input, including 12 conditions such as normal operation, load fluctuation, and changes in feedstock properties. The global temperature distribution was simulated, and the virtual mesh coordinates of the peak location were extracted as labels. (See attached image) Figure 2 As shown, the axis is divided into 30 layers and the radial division into 5 zones. The coordinates "10,3" represent the 10th layer in the axis and the 3rd zone in the radial direction.

[0047] Step (1.2): Data collection implementation.

[0048] DCS data was acquired in real time via the OPC UA protocol at a frequency of once per minute for three years, resulting in approximately 1.6 million raw data points. CFD simulations were performed for 12 operating conditions, with 8-10 sets of parameters set for each condition, for a total of 100 simulations. Each simulation lasted for 2 hours and output temperature distribution data every 10 seconds. The peak coordinates of the steady-state phase were extracted as labels and synchronized with the DCS data to form a complete dataset.

[0049] Step (1.3): Data preprocessing.

[0050] This includes: (a) Data cleaning: outliers (such as jump data caused by sensor failure) are removed using the 3σ criterion, and missing data is supplemented by linear interpolation of adjacent time points to ensure that the missing rate is <0.5%.

[0051] (b) Data filtering: Kalman filtering is used to eliminate high-frequency noise. The filtering window is set to 5 data points to balance the filtering effect and response speed.

[0052] (c) Data normalization: Z-score standardization is used to eliminate the influence of dimensions, so that the mean of all variables is 0 and the standard deviation is 1.

[0053] Step two is used to address the shortcomings of purely mechanistic or purely data-driven models.

[0054] Step (2.1): Establish a mechanism model.

[0055] A simplified model is established based on the mass-energy balance and reaction kinetics of the regeneration process. The core equations include the char reaction kinetic equation and the micro-element mass-energy balance equation.

[0056] Specifically, the kinetic equation for the coking reaction is expressed as: the rate of coke combustion is equal to the pre-exponential factor k0 and the oxygen concentration C. O2 The product of the product of the natural number e and the gas constant R and temperature T, and the exponent of the product of the activation energy Ea, is the product of the gas constant R and temperature T.

[0057] Among them, the pre-exponential factor k0 is 8.2 × 10⁻⁶. 5 kg / (m 3 The gas constant is 8.314 J / (mol·K), and the activation energy is 1.2 × 10⁻⁶ s. 5 J / mol.

[0058] The energy balance equation of a micro-element is expressed as the product of the density, specific heat capacity, and rate of temperature change of the gas-solid mixture, which is equal to the product of the carbon deposition combustion rate and the combustion enthalpy minus the product of the thermal conductivity and the diffusion rate of the temperature field, minus the product of the temperature difference between the fluid and the convective heat transfer coefficient.

[0059] In this embodiment, the density of the gas-solid mixture is 1200 kg / m³. 3 Its specific heat capacity is 1.2 kJ / (kg·K), and its combustion enthalpy is -4.18×10⁻⁶. 4 kJ / kg, thermal conductivity is 0.3 W / (m·K), convective heat transfer coefficient is 50 W / (m·K). 2 ·K).

[0060] By solving the kinetic equations of the coking reaction and the mass-energy balance equations of the infinitesimal element using the linear difference method, the three-dimensional problem is simplified into an axial-radial two-dimensional problem, and the preliminary peak coordinates (X1, Y1) are output with an error of approximately ±1.5 grids.

[0061] Step (2.2): XGBoost data-driven compensation.

[0062] The system takes "18 preprocessed process variables + mechanism model output (X1, Y1 and corresponding temperature values)" as input and "the residual between the actual / simulated peak coordinates and the mechanism model output (ΔX=X_peak-X1, ΔY=Y_peak-Y1)" as output to learn nonlinear relationships (such as local flow field disturbances) that the mechanism model cannot capture.

[0063] The XGBoost parameters are set as follows: number of decision trees (n_estimators) = 100, maximum tree depth (max_depth) = 6, learning rate (learning_rate) = 0.1, L1 regularization (reg_alpha) = 0.1, L2 regularization (reg_lambda) = 1.0, and sampling ratio (subsample) = 0.8.

[0064] Step (2.3): Model training.

[0065] The dataset obtained in step one is divided into a training set (70,000 records), a validation set (20,000 records), and a test set (10,000 records) in a 7:2:1 ratio. An iterative training + early stopping mechanism is adopted (if the error of the validation set does not decrease after 3 consecutive iterations, the training is stopped). After training, the peak localization accuracy of the test set is >95%, with an error of ±0.5 grids.

[0066] Step 3 involves deploying the hybrid model online to ensure the real-time accuracy of temperature peak predictions.

[0067] Step (3.1): Real-time data acquisition.

[0068] Communicating with the DCS system via industrial Ethernet, the OPC UA protocol reads 18 process variables at a frequency of once per second; the data is verified in real time, and if it exceeds a reasonable range (such as temperature > 550℃ or < 450℃), the average of the previous 3 readings is used to replace it to ensure data reliability.

[0069] Step (3.2): Online preprocessing.

[0070] The real-time data is processed using the normalization method in step one to ensure that the input format is consistent with the training of the hybrid model.

[0071] Step (3.3): Hybrid model inference.

[0072] The hybrid model is deployed on an industrial control server, using batch inference (calculated once every 10 data points), with a single inference time of <0.1 seconds; it outputs the current peak coordinates (X_peak, Y_peak) and prediction confidence (0-1, the higher the confidence, the more reliable the result).

[0073] Step (3.4): Correction of prediction results.

[0074] The axial coordinates are corrected by combining the trend of catalyst circulation rate. When the circulation rate increases by 20%, the catalyst residence time is shortened and the peak axial position shifts upward by 0.5-1 grids. Based on this rule, a correction coefficient is established, and the final positioning accuracy reaches ±0.3 meters.

[0075] Step four provides operators with an intuitive interface and intervention guidance.

[0076] Step (4.1): Visual presentation.

[0077] The development interface is based on WinCC configuration software. The left side shows the two-dimensional mesh diagram of the regenerator (as attached). Figure 2The peak position is indicated by red flashing dots (flashing frequency associated with confidence level: ≥0.9 fast flashing, 0.7-0.9 medium flashing, <0.7 slow flashing); the middle is the data panel, displaying the real-time values ​​and 1-hour trend curves of 18 process variables; the right side is the operation guidance area, recommending adjustment strategies based on the peak position (e.g., "The peak is located in Zone 5 on the 10th floor, it is recommended to increase the secondary air volume in this area by 0.5 Nm³"). 3 / h”).

[0078] Step (4.2): Multi-level early warning settings.

[0079] Two warning thresholds are set: (a) Level 1 warning (yellow): peak temperature > 530℃ or distance from the vessel wall < 0.3 meters (radial zone 5); (b) Level 2 warning (red): peak temperature > 550℃ or stay in the danger zone for > 30 seconds.

[0080] Step (4.3): Early warning response.

[0081] A Level 1 warning triggers a yellow light and a 1Hz buzzer, simultaneously displaying operational suggestions; a Level 2 warning triggers a flashing red light and a 3Hz buzzer, pushing warning information to the operator's mobile app. If there is no manual intervention within 30 seconds, it automatically sends suggested adjustment instructions to the DCS system (such as reducing the regeneration circulating gas flow rate by 5%).

[0082] The following examples and specific application scenarios further illustrate the technical solutions and effects of the present invention. The examples below are explanations of the present invention, but the present invention is not limited to the examples below.

[0083] A 1.2 Mt / a continuous reforming unit in an oil refinery has a regenerator that is approximately 15 meters high and 3 meters in diameter.

[0084] Step (1): Data preparation.

[0085] Historical operational data of the device over the past three years was collected, with one data point per hour. Simultaneously, high-precision simulations of the regenerator under 100 different operating conditions were performed using computational fluid dynamics software, obtaining detailed temperature fields and precise peak position coordinates for each condition. These 100 sets of simulation data were then matched with corresponding actual operational data to form a training dataset.

[0086] Step (2): Hybrid model training.

[0087] like Figure 2As shown, the regenerator is axially divided into 30 layers (0.5 meters / layer), including layers L1-L30. For ease of representation, only the layer numbers L1, L5, L10, L15, L20, L25, and L30 are labeled in the figure. Radially, it is divided into 5 regions: R1, R2, R3, R4, and R5 (0.3 meters / region; the radial division is circular, so 0.3 meters refers to the radius), totaling 150 virtual grids. The location of the temperature peak is indicated by the grid number (e.g., "Layer 10, Region 3"). High, medium, and low temperatures are represented by a red, yellow, and blue gradient.

[0088] The XGBoost algorithm was used as data-driven compensation, with actual operating parameters and the output of a simplified mechanistic model as features, and the peak grid numbers obtained from simulation as labels for model training. The trained model achieved a peak grid prediction accuracy of over 95% on the test set.

[0089] Step (3): Online application.

[0090] The trained model is deployed to the device's real-time optimization system. The system reads DCS data every minute and predicts the grid location of the current temperature peak in real time.

[0091] Results: By applying the temperature peak location method provided in this embodiment, the real-time optimization system successfully predicted a trend of peak temperature moving towards the unit wall due to partial blockage of the main air distributor, issuing an early warning 30 minutes in advance. Operators adjusted their operations in time, avoiding a potential unplanned shutdown, resulting in significant economic benefits.

[0092] The continuous reforming regenerator temperature peak location method provided in this embodiment addresses the shortcomings of existing technologies, such as "measuring blind spots, delayed early warning, low model accuracy, and weak operational guidance." Specific objectives include: achieving full-domain reconstruction of the three-dimensional temperature field inside the regenerator, eliminating monitoring blind spots, and achieving a temperature peak location accuracy of ±0.3 meters (corresponding to ±0.5 virtual grid units); reducing the temperature peak identification response time from 3-5 minutes in existing technologies to less than 1 minute, providing operators with at least 2 minutes of intervention time; constructing a hybrid model that combines physical interpretability and high accuracy, with a prediction error ≤±2℃ under extreme operating conditions, ensuring model stability; and providing a visual interface and operational guidance, intuitively displaying the hotspot location and movement trend, reducing reliance on personal experience for operation, and achieving precise control.

[0093] The method for locating the peak temperature of a continuous reforming regenerator provided in this embodiment has the following beneficial effects: (1) “Mechanism constraint + XGBoost” hybrid model architecture: The physical boundary is provided by the mechanism model (to avoid predictions that violate physical laws), and the XGBoost module learns the residual (to capture nonlinear relationships), which solves the problems of low accuracy of pure mechanism model and poor stability of pure data-driven model.

[0094] (2) Fine positioning of 150 virtual grids: axial 0.5m / layer and radial 0.3m / area grid division, combined with real-time correction algorithm, to achieve positioning accuracy of ±0.3m and eliminate monitoring blind spots.

[0095] (3) Closed-loop real-time positioning process: The entire process response time from DCS real-time data acquisition to early warning response is less than 1 minute, including data verification, batch processing inference and other optimization methods to meet industrial real-time requirements.

[0096] (4) Visualization of confidence correlation and multi-level early warning: The reliability of prediction is reflected by the flashing frequency. Two-level early warning + operation guidance reduces the blindness of operation and realizes the "prediction-intervention" closed loop.

[0097] Example 3: This embodiment provides a continuous reforming regenerator temperature peak location system, such as... Figure 3 As shown, it includes a three-layer architecture: DCS data acquisition layer 2 → model calculation layer 4 (mechanism module + XGBoost module) → visualization and early warning layer 5.

[0098] The DCS data acquisition layer is connected to the continuous reforming unit 1 and is used to collect process variables during the catalyst regeneration process of the continuous reforming regenerator and process the collected process variables.

[0099] The DCS data acquisition layer connects to the real-time database 3. The real-time database is used to preprocess the acquired process variables. At the same time, a three-dimensional model of the continuous reforming regenerator is established using CFD simulation method to obtain the temperature peak position coordinates of the continuous reforming regenerator under different operating conditions. The obtained temperature peak position coordinates are time-synchronized and matched with the preprocessed process variables to form a set of test data. All the acquired data are processed in the aforementioned manner to obtain a complete dataset, which is then sent to the model calculation layer.

[0100] The model computation layer connects to a real-time database to build a hybrid model combining mechanistic constraints and XGBoost data-driven compensation. Simultaneously, it trains the hybrid model using a complete dataset sent from the real-time database. The model computation layer includes a mechanistic module (6) and an XGBoost module (7). The mechanistic module provides the physical boundaries of the continuous reforming regenerator and reconstructs the global temperature field; the XGBoost module learns the residuals and captures the nonlinear relationships in the catalyst regeneration process of the continuous reforming regenerator.

[0101] The visualization and early warning layer is connected to the model calculation layer to display the prediction results of the model calculation layer and send corresponding alarm signals based on the prediction results. If the peak temperature exceeds 530℃ or the distance between the peak temperature coordinates and the device wall is less than 0.3 meters, a level one early warning is triggered, and operation suggestions are displayed simultaneously. If the peak temperature exceeds 550℃ or the peak temperature coordinates remain in the danger zone for more than 30 seconds, a level two early warning is triggered, and early warning information is pushed. If there is no manual intervention within a preset time, an adjustment instruction is automatically sent to the DCS system.

[0102] When making real-time predictions of temperature peaks, the DCS data acquisition layer collects real-time process variables and sends them to the real-time database for data standardization. The real-time database then sends the standardized real-time process variable data to the model prediction layer. The model prediction layer predicts the real-time temperature peak coordinates based on the real-time process variable data and outputs the real-time temperature peak coordinates and the corresponding temperature values ​​to the visualization and early warning layer for display. The visualization layer then issues corresponding early warnings based on the real-time temperature peak coordinates.

[0103] In other embodiments, a hybrid model can be pre-trained, installed in the model computation layer, and then the three-layer architecture can be used directly for real-time temperature peak prediction, thus eliminating the model building and training process in the three-layer architecture.

[0104] Compared with existing thermocouple monitoring, pure mechanism models, and pure data-driven models, the continuous reforming regenerator temperature peak location method and system provided in this application have the following advantages: (1) From “point monitoring” to “field reconstruction”: No need to add hardware measurement points, the whole temperature field is reconstructed through algorithms, covering all spatial locations of the regenerator, solving the monitoring blind spot problem of existing technology, and the positioning accuracy reaches ±0.3 meters, which is much higher than the monitoring error of thermocouple ±1-2 meters.

[0105] (2) Significant early warning capability: The response time of the entire process is less than 1 minute, compared with the 3-5 minute delay of existing technologies, providing operators with at least 2 minutes of intervention time, which can prevent catalyst thermal deactivation from the source. In a pilot application at a refinery, a hot spot caused by blockage of the main air distributor was successfully warned 30 minutes in advance, avoiding unplanned shutdowns.

[0106] (3) Balance between model stability and accuracy: The hybrid model combines the advantages of mechanism constraints and XGBoost. Under extreme conditions (such as sudden changes in the sulfur content of raw materials), the prediction error is ≤ ±2℃, which is more accurate than the pure mechanism model (error ±5-8℃) and the pure BP neural network (error ±3-4℃). There are no abnormal results of "predicting high temperature from low temperature".

[0107] (4) Strong engineering practicality: Based on existing DCS system data, no additional hardware investment is required. The visual interface and operation guidance reduce the dependence on the operator's experience. A case study shows that the operator's adjustment accuracy increased from 60% to 95%, and the catalyst activity recovery rate increased from 70% to 90%.

[0108] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Other variations and modifications may be made without departing from the technical solutions described in the claims.

Claims

1. A method for locating the peak temperature in a continuous reforming regenerator, characterized in that, include: Process variables are collected from the DCS system to establish a three-dimensional model of the continuous reforming regenerator and obtain the temperature field and temperature peak location coordinates under each different operating condition. A mechanistic model was established based on the mass-energy balance and reaction kinetics of the regeneration process. The XGBoost algorithm was used as data-driven compensation. A hybrid model of mechanistic constraints and XGBoost data-driven compensation was constructed and trained. Based on the trained hybrid model, batch inference is used to output the current temperature peak coordinates and prediction confidence.

2. The method for locating the peak temperature of a continuous reforming regenerator according to claim 1, characterized in that, The mechanism model outputs the initial temperature peak coordinates and the corresponding temperature values; the XGBoost algorithm is used as data-driven compensation, with process variables and the output of the mechanism model as inputs, and the residual between the actual / simulated temperature peak coordinates and the output of the mechanism model as the output; Hybrid model training is performed based on process variables and 3D model output data.

3. A method for locating the peak temperature of a continuous reforming regenerator according to claim 1 or 2, characterized in that, The continuous reforming regenerator is meshed, and the temperature peak coordinates are represented by the mesh number. The axial coordinates of the temperature peak are corrected in combination with the catalyst circulation rate trend: when the circulation rate increases by 20%, the axial position of the temperature peak shifts upward by 0.5-1 meshes.

4. A method for locating the peak temperature of a continuous reforming regenerator according to claim 1 or 2, characterized in that, The mechanism model includes: obtaining the coke combustion rate based on oxygen concentration and activation energy, and establishing a kinetic equation for the coking reaction; establishing a micro-element mass-energy balance equation based on the coke combustion rate; solving the two equations using the finite difference method, and outputting the preliminary temperature peak coordinates.

5. The method for locating the peak temperature of a continuous reforming regenerator according to claim 1, characterized in that, A three-dimensional model of a continuous reforming regenerator was established using CFD simulation. n sets of parameters were set for each of the 12 operating conditions. Temperature distribution data under each operating condition was output, and the peak coordinates of the steady-state stage were extracted as labels. The data were synchronized with DCS data to form a complete dataset. The dataset was divided into training set, validation set and test set according to the proportion. A hybrid model was trained using iterative training and early shutdown mechanism.

6. A method for locating the peak temperature of a continuous reforming regenerator according to claim 1, 2, or 5, characterized in that, The process variables include regeneration recirculation gas flow rate, regenerator top pressure, catalyst circulation rate, and recirculation gas oxygen concentration.

7. A method for locating the peak temperature of a continuous reforming regenerator according to claim 1, 2, or 5, characterized in that, Based on the temperature peak location results, two levels of early warning thresholds are set: if the temperature peak is greater than 530℃ or the temperature peak coordinate is less than 0.3 meters away from the device wall, a level one early warning is triggered, and operation suggestions are displayed simultaneously. If the temperature peak exceeds 550℃ or the temperature peak coordinates remain in the danger zone for more than 30 seconds, a level two warning will be triggered, and a warning message will be pushed. If there is no manual intervention within the preset time, a suggested adjustment instruction will be automatically sent to the DCS system.

8. A method for locating the peak temperature of a continuous reforming regenerator according to claim 1, 2, or 5, characterized in that, The collected process variables were preprocessed as follows: outliers were removed using the 3σ criterion, and missing data were supplemented by linear interpolation between adjacent time points; high-frequency noise was eliminated using Kalman filtering; and the filtered data were standardized using Z-score.

9. A continuous reforming regenerator temperature peak location system, employing the continuous reforming regenerator temperature peak location method according to any one of claims 1-8, characterized in that, include: The DCS data acquisition layer acquires DCS data in real time via the OPC UA protocol to obtain raw training data and perform preprocessing. The model computation layer constructs a hybrid model of mechanistic constraints and XGBoost data-driven compensation to perform real-time prediction of temperature peaks. The visualization and early warning layer displays a two-dimensional grid diagram of the regenerator and the location of temperature peaks, shows real-time values ​​and trend curves of process variables, and issues early warnings based on real-time temperature peaks.

10. A continuous reforming regenerator temperature peak positioning system according to claim 9, characterized in that, The prediction model includes a mechanism module and an XGBoost module. The mechanism module provides the physical boundary of the continuous reforming regenerator and reconstructs the global temperature field. The XGBoost module learns the residuals and captures the nonlinear relationship in the catalyst regeneration process of the continuous reforming regenerator.

Images