Low-temperature economizer digital twinborn body construction method

Abstract

The invention relates to the technical field of computers, in particular to a low-temperature economizer digital twinborn body construction method, and aims to solve the problems that an existing model statically solidifies, multi-source data fusion is difficult, and degradation recognition lags. The method comprises the following steps: constructing a multi-physical field reference model covering fluid, heat transfer and corrosion mechanisms; collecting and carrying out time-space alignment on temperature, pressure, flue gas components and ash deposition data, and combining wavelet and median filtering to carry out de-noising; introducing extended Kalman filtering to correct model state variables on line, and realizing dynamic updating of parameters; a CNN-GRU hybrid network is embedded to extract time sequence degradation features, and unsupervised clustering is combined to identify working condition states; and when the detection is abnormal, triggering local high-fidelity CFD re-simulation, and completing closed-loop optimization and state synchronous mapping. Through the mechanism, the internal physical field error lt of the equipment is realized; 5% high-precision dynamic mapping, more than 95% of anomaly detection rate and rapid deployment of a new unit within 72 hours are realized, the prediction accuracy and preventive maintenance capability are remarkably improved, and the service life of equipment is prolonged by 15%-20%.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computers, and particularly relates to a low-temperature economizer digital twin construction method. BACKGROUND

[0002] With the deepening of industrial energy saving and digital transformation, as a key equipment for waste heat recovery in coal-fired power plants, the operation state monitoring and performance optimization of the low-temperature economizer are of great significance to improve the system energy efficiency. The traditional monitoring method relies on field sensor data and empirical formula estimation, and it is difficult to realize the dynamic mapping of the thermal characteristics, corrosion risk and ash deposition trend of the equipment in the whole life cycle, resulting in lagging operation and maintenance decision and weak fault warning capability. Especially under the conditions of complex load fluctuation and variable coal quality, the coupling relationship among the internal flow field distribution, wall temperature gradient and heat exchange efficiency of the equipment needs high-fidelity modeling support, which puts forward higher requirements for the real-time performance, precision and adaptive ability of the digital twin.

[0003] However, the existing digital twin construction methods are mostly based on static physical models or offline simulation data, and lack online response mechanisms for sudden changes in operating parameters and boundary condition drift, resulting in a gradual mismatch between the twin and the actual equipment state. At the same time, the multi-source heterogeneous sensor data in the high-dust and strong-corrosion environment of the low-temperature economizer has the problems of noise interference and asynchronous sampling, and the traditional data fusion strategy is difficult to support the accurate reconstruction of the high-dimensional state space. In addition, the twin update relies on artificial rule setting or fixed threshold judgment, and cannot autonomously identify the performance degradation characteristics under non-steady-state conditions, resulting in the accumulation of prediction deviation and weakening of its guiding value in preventive maintenance.

[0004] Therefore, a high-precision and adaptive digital twin construction method for low-temperature economizers is expected. SUMMARY

[0005] The purpose of the present application is to provide a low-temperature economizer digital twin construction method, which can effectively solve the problems in the background art.

[0006] To achieve the above-mentioned purpose, the technical solution adopted by the present application is: The application discloses a low-temperature economizer digital twin construction method, which comprises the following specific steps: step (1), constructing a multi-physical field coupling reference digital model: based on the geometric structure parameters, material thermal physical property parameters and typical operation conditions of the low-temperature economizer, a multi-physical field coupling reference model covering fluid dynamics, heat transfer and electrochemical corrosion mechanism is established, the model is discretely solved by using a finite volume method to solve Navier-Stokes equations, energy conservation equations and corrosion rate equations, and outputs the flow velocity field, temperature field, wall heat flux density and corrosion rate distribution in the equipment; step (2), collecting multi-source heterogeneous operation data and performing time-space alignment and denoising processing: the original data from the distributed temperature sensor, pressure transmitter, flue gas component analyzer and ash deposition monitoring device are synchronously acquired, a multi-channel signal alignment strategy based on a dynamic time warping algorithm is used to eliminate sampling time deviation, and a wavelet threshold denoising and sliding window median filtering joint method is combined to suppress random noise and pulse interference in a high dust environment; step (3), constructing an adaptive data-model fusion driving mechanism: the preprocessed operation data are used as observation input, an extended Kalman filter is introduced to modify the state variables of the reference digital model online, wherein the state vector comprises the average flow velocity, wall temperature gradient and local heat transfer coefficient of the key section, the observation equation is mapped from the physical position of the sensor to the model grid node, and dynamic updating of the model parameters and boundary conditions is realized; step (4), embedding a deep time sequence feature extraction and performance degradation identification module: a one-dimensional convolutional neural network and a gated recurrent unit hybrid architecture are used to encode the multi-dimensional operation sequence in a continuous time window, extract implicit degradation features reflecting the ash deposition growth rate, corrosion acceleration trend and heat exchange efficiency attenuation, and automatically divide the steady-state, transition state and abnormal working condition intervals through an unsupervised clustering algorithm; step (5), performing closed-loop iterative optimization and state mapping of the digital twin: the model reconstruction mechanism is triggered according to the degradation feature identification result, when it is detected that the duration of the non-steady-state working condition exceeds a preset threshold or the degradation feature vector deviates from the normal distribution by more than 3 times the standard deviation, a high-fidelity computational fluid dynamics simulation module is automatically called to finely re-simulate the local area, and the updated physical field data are injected into the reference model, so that high-precision synchronous mapping of the digital twin and the state of the entity equipment is realized.

[0007] Preferably, the geometric structure parameters in step (1) comprise tube bundle arrangement mode, tube diameter, tube length, fin spacing and flue gas flow area, the material thermal physical property parameters comprise thermal conductivity, specific heat capacity, density and electrochemical corrosion potential, and the typical operation conditions comprise flue gas temperature, flow and sulfur content combinations corresponding to five representative coal qualities in the 30%-100% load interval.

[0008] Preferably, in step (2), the distributed temperature sensor is arranged at no less than 12 measuring points along the flue gas flow direction at the economizer inlet and outlet and the middle section, the pressure transmitter is installed at the pressure difference measurement position before and after the flue, the sampling frequency of the flue gas composition analyzer is no less than 1 Hz, the ash accumulation monitoring device adopts the principle of acoustic ash measurement or microwave attenuation, and the data sampling period is uniformly set to 5 seconds.

[0009] Preferably, in step (2), wavelet threshold denoising uses the db4 wavelet basis function for 5-level decomposition, the threshold function is a soft threshold, and the threshold size is set to 1.5 times the standard deviation of noise estimation; the window length of the sliding window midpoint filter is 15 sampling points, which is used to eliminate the instantaneous data jump caused by sudden dust impact.

[0010] Preferably, in step (3), the state vector dimension of the extended Kalman filter is 64, corresponding to 64 key monitoring areas in the model grid; the observation equation maps the physical coordinates of the sensor to the four nearest grid nodes through bilinear interpolation, and the observation noise covariance matrix is ​​dynamically adjusted according to the sensor accuracy level, with the temperature sensor covariance set to 2 Kelvin square and the pressure sensor covariance set to 100 Pascal square.

[0011] Preferably, in step (4), the one-dimensional convolutional neural network contains three convolutional layers with kernel sizes of 7, 5, and 3, and the number of output channels per layer is 64, 128, and 256, followed by a max pooling layer; the gated recurrent unit has a hidden layer dimension of 512, and the input sequence length is 300 time steps, corresponding to 25 minutes of continuous running data; unsupervised clustering adopts a Gaussian mixture model, and the number of clusters is automatically determined by the Bayesian information criterion, usually 3 to 5 classes.

[0012] Preferably, the degradation feature vector in step (4) contains 128 hidden state outputs, of which the first 64 dimensions represent short-term dynamic response characteristics and the last 64 dimensions represent long-term trend cumulative effects; the abnormal working condition is determined by Mahalanobis distance metric. When the Mahalanobis distance of 10 consecutive time steps is greater than 9.21 (corresponding to 99% confidence level), it is determined to be a significant performance degradation.

[0013] Preferably, in step (5), the high-fidelity computational fluid dynamics simulation module uses the Reynolds stress turbulence model and the discrete phase model for coupled solution, with a grid number of no less than 5 million and a time step of 0.01 seconds, to accurately simulate the flow channel blockage and secondary flow effect caused by local dust accumulation; the reconstructed region is determined by the top 3 monitoring regions with the highest contribution in the degradation feature space.

[0014] Preferably, it also includes: establishing a historical operation database to store no less than 2 years of full-condition operation data and corresponding digital twin state snapshots, with a data sampling frequency of once per minute; training a transfer learning model based on the database to quickly initialize its digital twin benchmark model in the early stage of new unit commissioning, shortening the model convergence time to within 72 hours.

[0015] Preferably, it also includes: setting a digital twin health assessment index, which integrates the heat exchange efficiency deviation rate, the frequency of wall temperature exceeding the limit and the corrosion rate growth rate. When the health is below 85%, it automatically triggers preventive maintenance recommendations, including adjusting the soot blowing cycle, optimizing water quality control parameters or issuing a warning for replacing local pipe sections.

[0016] Compared with the prior art, the present invention has the following beneficial effects: High-precision dynamic mapping capability Breaking through the limitations of static models: By using a multi-physics coupled benchmark model and an extended Kalman filter online correction mechanism, dynamic high-fidelity mapping of the internal flow field, temperature field and corrosion field of the low-temperature economizer is achieved, with state variable estimation error less than 5%, which is significantly better than the traditional monitoring method based on fixed empirical formulas. Deep fusion of multi-source data: By adopting dynamic time warping and joint denoising strategies, the problem of temporal misalignment and noise interference of multi-source heterogeneous sensor data in high dust environment is effectively solved, and the data availability rate is increased to more than 98%, providing reliable input for high-dimensional state space reconstruction.

[0017] Autonomous adaptation and intelligent recognition capabilities Autonomous identification of unsteady operating conditions: Based on deep temporal feature extraction and unsupervised clustering, it can automatically identify performance degradation patterns under complex load fluctuations without the need for manual rule setting. The abnormal operating condition detection rate is greater than 95%, and the false alarm rate is less than 3%. Closed-loop iterative optimization mechanism: When significant performance degradation is detected, local high-fidelity simulation reconstruction is automatically triggered to achieve continuous synchronization between the digital twin and the physical device, avoid the accumulation of prediction bias, and keep the long-term model drift rate within 2% per month.

[0018] Preventive maintenance support capabilities Full lifecycle status tracking: By quantifying the trend of equipment performance degradation through health assessment indicators, it provides data-driven decision-making basis for soot blowing strategy optimization, water quality management and component replacement, which is expected to extend the service life of equipment by 15% to 20%; Rapid deployment and migration capabilities: Based on the historical database, the transfer learning mechanism significantly shortens the construction cycle of digital twins for new units from several weeks in the traditional method to within 72 hours, significantly improving the efficiency of technology promotion and engineering practicality. Attached Figure Description

[0019] Figure 1 This is a flowchart of the overall technical solution of the present invention; Figure 2 This is a flowchart illustrating the multi-source heterogeneous data acquisition and processing workflow of this invention. Figure 3 This is a diagram of the adaptive data-model fusion driving mechanism architecture of the present invention; Figure 4 This is a diagram of the closed-loop iterative optimization and state mapping architecture of the present invention. Detailed Implementation

[0020] Please refer to Figures 1-4 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0021] Currently, during the operation of cryogenic economizers, problems such as decreased heat exchange efficiency and shortened equipment lifespan arise due to ash accumulation, corrosion, and fluctuations in operating conditions. Traditional monitoring methods rely on empirical formulas and static models, making it difficult to achieve high-precision mapping of the dynamic evolution of internal multi-physics fields and autonomous identification of performance degradation trends. To address these technical problems, this invention proposes a method for constructing a digital twin of a cryogenic economizer and applies it to such a method.

[0022] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, step (1) involves constructing a multi-physics coupled benchmark digital model: based on the geometric structural parameters, material thermal properties, and typical operating conditions of the low-temperature economizer, a multi-physics coupled benchmark model covering fluid dynamics, heat transfer, and electrochemical corrosion mechanisms is established. This model uses the finite volume method to discretize and solve the Navier-Stokes equations, energy conservation equations, and corrosion rate equations, outputting the internal velocity field, temperature field, wall heat flux density, and corrosion rate distribution of the equipment. Specifically, the geometric structural parameters in step (1) include the tube bundle arrangement, tube diameter, tube length, fin spacing, and flue gas flow cross-sectional area; the material thermal properties include thermal conductivity, specific heat capacity, density, and electrochemical corrosion potential; and the typical operating conditions cover the flue gas temperature, flow rate, and sulfur content combinations corresponding to five representative coal types within the 30% to 100% load range. The tube bundles are arranged in a staggered pattern to enhance the turbulence effect. The tube diameter is set at 25 mm, the tube length at 6 m, and the fin spacing at 8 mm. The flue gas flow cross-sectional area is calculated based on the inlet and outlet flue dimensions to ensure flow field uniformity. The thermal conductivity ranges from 40 to 50 W / m Kelvin, the specific heat capacity is 480 J / kg Kelvin, the density is 7800 kg / m³, and the electrochemical corrosion potential is determined based on the material composition and environmental pH, initially set at -0.7 volts. Under typical operating conditions, the flue gas temperature ranges from 120°C to 200°C, the flue gas flow rate ranges from 10,000 to 30,000 standard cubic meters per hour, and the sulfur content ranges from 0.5% to 2.0%. Five sets of operating conditions are set for model training and verification. The model space is discretized using unstructured hexahedral meshes, with a total mesh count of no less than 1 million and a minimum mesh size controlled within 1 mm to ensure analytical accuracy in the boundary layer region. The time step is set to 0.1 seconds to satisfy the Courant-Friedrichs-Lewy stability condition. The Navier-Stokes equations are discretized using a second-order upwind scheme, the energy conservation equations are performed using a central difference scheme, and the corrosion rate equation is derived based on the Butler-Volmer equation, with the following expression:

[0023] in, For local current density, For exchange current density, and These are the anode and cathode transfer coefficients, respectively. It is Faraday's constant. This is the universal gas constant. Absolute temperature The equation is used to calculate the overpotential. It is solved in conjunction with the flow and temperature fields to predict the spatial distribution of corrosion rates. The model output includes a three-dimensional velocity vector field, temperature field, wall heat flux density distribution map, and corrosion rate contour map. All results are stored in VTK format, supporting subsequent visualization analysis and data retrieval.

[0024] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, step (2) involves collecting multi-source heterogeneous operating data and performing spatiotemporal alignment and denoising: synchronously acquiring raw data from distributed temperature sensors, pressure transmitters, flue gas composition analyzers, and ash accumulation monitoring devices; employing a multi-channel signal alignment strategy based on dynamic time warping algorithm to eliminate sampling timing deviations; and combining wavelet threshold denoising and sliding window midpoint filtering to suppress random noise and pulse interference in high-dust environments. Specifically, in step (2), the distributed temperature sensors are arranged with no fewer than 12 measuring points along the flue gas flow direction at the economizer inlet, outlet, and intermediate sections; the pressure transmitters are installed at the pressure difference measurement positions before and after the flue; the flue gas composition analyzer has a sampling frequency of no less than 1 Hz; and the ash accumulation monitoring device uses acoustic ash measurement or microwave attenuation principles, with a data sampling period uniformly set to 5 seconds. The temperature sensors are of platinum resistance type with an accuracy class of 0.1 degrees Celsius, and are installed at the inlet, outlet, and three intermediate sections, with four measuring points arranged at each section to form a spatial sampling array. The pressure transmitter adopts a differential pressure structure with a range of 0 to 10 kPa and an accuracy of ±0.5%. It is installed between the inlet and outlet flanges of the flue gas duct. The flue gas composition analyzer uses infrared absorption spectroscopy to detect the concentrations of carbon dioxide, sulfur dioxide, and oxygen in real time. The sampling frequency is set to 1 Hz, and the data is transmitted to the central control system via the Modbus RTU protocol. The ash accumulation monitoring device uses the ultrasonic reflection principle. The probe is installed on the outside of the tube bundle, and the transmission frequency is 40 kHz. The echo signal is amplified and analyzed by an embedded processor to measure thickness changes, with a measurement error of less than 1 mm. All sensor data is connected to the edge computing node via industrial Ethernet, and the timestamp is synchronized by a GPS timing system with millisecond-level accuracy. Multi-channel signal alignment uses a dynamic time warping algorithm to map each sensor sequence to a unified time axis, allowing a maximum time offset of ±2 seconds, and ensuring consistent data length after alignment. Wavelet thresholding denoising employs a 5-level decomposition using the db4 wavelet basis function. A soft threshold is used, with the threshold value set at 1.5 times the standard deviation of the noise estimate. A sliding window midpoint filter with a window length of 15 sampling points is used to eliminate instantaneous data jumps caused by sudden dust impacts. Specifically, the original signal is first decomposed using db4 wavelet to obtain low-frequency approximation coefficients and five levels of detail coefficients. Soft thresholding is applied to each level of detail coefficient, with the threshold estimated using the Donoho-Johnstone method. Then, an inverse transform is performed to reconstruct the signal. Subsequently, a sliding window midpoint filter is applied to the reconstructed signal, with a window length of 15 sampling points (75 seconds). The median is calculated point-by-point to replace the original value, effectively suppressing impulse noise. The final output data sequence has undergone quality checks, with a missing value rate of less than 0.5% and an outlier rate of less than 1%, making it suitable for subsequent modeling and analysis.

[0025] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, step (3) involves constructing an adaptive data-model fusion driving mechanism: using preprocessed operating data as the observation input, an extended Kalman filter is introduced to correct the state variables of the baseline digital model online. The state vector includes the average flow velocity, wall temperature gradient, and local heat transfer coefficient of the key cross-section. The observation equation is mapped from the physical location of the sensor to the model grid nodes, realizing the dynamic update of model parameters and boundary conditions. Specifically, in step (3), the state vector dimension of the extended Kalman filter is 64, corresponding to 64 key monitoring areas in the model grid. The observation equation maps the physical coordinates of the sensor to the four nearest grid nodes through bilinear interpolation. The observation noise covariance matrix is ​​dynamically adjusted according to the sensor accuracy level. The covariance of the temperature sensor is set to 2 Kelvin squared, and the covariance of the pressure sensor is set to 100 Pascal squared. The state vector is defined as:

[0026] in Indicates the first The average flow velocity in each region Indicates the wall temperature gradient. This represents the local heat transfer coefficient. The system state transition equation is provided by the numerical integration results of the baseline model, and is in the form of: ,in To control inputs (such as flue gas flow rate and temperature). For process noise, the covariance matrix Dynamically adjusted based on model uncertainties. Observation vector. It consists of sensor readings, including temperature, pressure, etc., and the observation equation is: ,in The observation matrix is ​​constructed using bilinear interpolation. Specifically, for the observation matrix located at... The coordinate sensor finds the coordinates of its four nearest neighbor grid nodes. to Calculate the weighting coefficients , making When the The corresponding grid number is the first One sensor. Observation noise. covariance matrix Based on sensor type settings, the covariance is set to 2 Kelvin squared for temperature sensors, 100 Pascal squared for pressure sensors, and other sensors are converted according to their nominal accuracy. The extended Kalman filter performs two phases: prediction and update. In the prediction phase, the state transition equation is used to calculate the prior state and covariance; in the update phase, the Kalman gain is calculated, observation data is fused, and the state estimate is corrected. The entire process is executed every 5 seconds to ensure that the model state is synchronized with actual operation. Furthermore, when the state residual exceeds 3 times the standard deviation after three consecutive updates, a model parameter reestimation process is triggered to adjust boundary conditions or material parameters, improving long-term consistency.

[0027] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, step (4) embeds a deep temporal feature extraction and performance degradation identification module: using a hybrid architecture of a one-dimensional convolutional neural network and a gated recurrent unit, the multi-dimensional operating sequence within a continuous time window is feature-encoded to extract implicit degradation features reflecting the ash accumulation growth rate, corrosion acceleration trend, and heat exchange efficiency decay, and the steady-state, transitional, and abnormal operating condition intervals are automatically divided by an unsupervised clustering algorithm. Specifically, in step (4), the one-dimensional convolutional neural network contains three convolutional layers with kernel sizes of 7, 5, and 3, and the number of output channels per layer is 64, 128, and 256, followed by a max pooling layer; the gated recurrent unit has a hidden layer dimension of 512, and the input sequence length is 300 time steps, corresponding to 25 minutes of continuous operating data; the unsupervised clustering adopts a Gaussian mixture model, and the number of clusters is automatically determined by the Bayesian information criterion, usually 3 to 5 classes. The input data is a preprocessed multidimensional time series, including temperature, pressure, flue gas composition, and ash thickness, normalized to the [0,1] interval. The first layer of the one-dimensional convolutional neural network uses 7 convolutional kernels with a stride of 1 and identical padding, outputting a 64-channel feature map; the second layer uses 5 convolutional kernels, outputting a 128-channel feature map; and the third layer uses 3 convolutional kernels, outputting a 256-channel feature map. Each layer is followed by a ReLU activation function and a max-pooling layer with a pooling window of 2 and a stride of 2 to reduce the sequence length. The flattened convolutional output is fed into a gated recurrent unit (GRU), with a hidden layer dimension of 512. The input sequence length is 300 time steps, corresponding to 25 minutes of data, with a time step of 5 seconds. The GRU unit captures long-term dependencies through update and reset gate mechanisms, outputting a sequence length of 300 and a dimension of 512. The final output is compressed into a 128-dimensional hidden state vector through a fully connected layer, serving as a degradation feature vector. The first 64 dimensions of this vector represent short-term dynamic response characteristics, such as instantaneous temperature fluctuations and sudden pressure changes; the latter 64 dimensions represent long-term cumulative effects, such as a slow increase in dust accumulation and a continuous increase in corrosion rate. Unsupervised clustering employs a Gaussian mixture model, assuming the data follows multiple Gaussian distributions, and iteratively optimizes parameters using an expectation-maximization algorithm. The number of clusters is automatically determined using the Bayesian information criterion, avoiding manual setting, and is typically 3 to 5, corresponding to steady-state, transitional, and abnormal operating conditions, respectively. The clustering results are used to divide operating condition intervals to support subsequent decision-making.

[0028] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, step (5) involves performing closed-loop iterative optimization and state mapping of the digital twin: based on the degradation feature identification results, a model reconstruction mechanism is triggered. When the duration of the unsteady working condition exceeds a preset threshold or the degradation feature vector deviates from the normal distribution by more than 3 times the standard deviation, the high-fidelity computational fluid dynamics simulation module is automatically invoked to perform a refined re-simulation of the local area, and the updated physical field data is injected into the benchmark model to complete the high-precision synchronous mapping between the digital twin and the physical equipment state. Specifically, in step (5), the high-fidelity computational fluid dynamics simulation module uses a Reynolds stress turbulence model coupled with a discrete phase model for solution, with a grid number of no less than 5 million and a time step of 0.01 seconds, to accurately simulate the flow channel blockage and secondary flow effect caused by local ash accumulation; the reconstruction area is determined by the top 3 monitoring areas with the highest contribution in the degradation feature space. When the degradation feature identification module determines that a certain area is under abnormal operating conditions for more than 10 minutes, or the Mahalanobis distance is greater than 9.21 for 10 consecutive time steps (corresponding to a 99% confidence level), the reconstruction process is initiated. The system first calculates the contribution of each monitored area in the degradation feature space, and based on its weight and change magnitude in the feature vector, selects the top three areas with the highest contribution as reconstruction targets. Subsequently, a high-fidelity CFD simulation module is invoked. This module uses the Reynolds stress turbulence model (RSM) to describe complex turbulent structures, combined with a discrete phase model (DPM) to simulate the movement and deposition behavior of fly ash particles. Structured hexahedral meshes are used, locally refined to a minimum size of 0.5 mm, with a total mesh count of no less than 5 million to ensure analytical accuracy of the boundary layer and ash accumulation zone. The time step is set to 0.01 seconds to meet numerical stability requirements. Boundary conditions are provided by real-time operating data, including inlet velocity, temperature, and particle concentration. After simulation, high-resolution flow field, temperature field, and dust accumulation distribution maps are output and mapped to the baseline model mesh using an interpolation algorithm, replacing the original data. The updated model then re-participates in the extended Kalman filter process to achieve state synchronization. The entire closed-loop process is executed hourly to ensure that the digital twin always reflects the true state of the physical equipment.

[0029] In addition, a historical operation database is established, storing at least two years of full-condition operation data and corresponding digital twin state snapshots, with a data sampling frequency of once per minute. A transfer learning model is trained based on this database to quickly initialize the digital twin baseline model during the initial commissioning of new units, shortening the model convergence time to within 72 hours. The database uses a PostgreSQL relational database management system, with table structures including fields such as timestamps, device IDs, sensor numbers, measured values, and status labels, and indexes optimize query efficiency. State snapshots are stored in JSON format, containing key information such as model parameters, state variables, and feature vectors. The transfer learning model uses a pre-trained deep neural network, with the initial operation data of the new unit as input and the initial parameter set of the baseline model as output. Domain adaptation technology reduces the difference between the source and target domains, achieving rapid convergence. After the new unit is commissioned, the system automatically loads the parameters output by the transfer learning model, starts the digital twin, and achieves stable prediction accuracy within only 72 hours, significantly better than the debugging cycle of several weeks required by traditional methods.

[0030] In addition, it includes setting a digital twin health assessment index, which comprehensively considers heat exchange efficiency deviation rate, frequency of wall temperature exceeding limits, and corrosion rate growth rate. When the health level falls below 85%, it automatically triggers preventative maintenance recommendations, including adjusting soot blowing cycles, optimizing water quality control parameters, or issuing warnings for replacing local pipe sections. The health assessment index is defined as follows:

[0031] Among them, the heat exchange efficiency deviation rate is the percentage difference between the current measured value and the design value; the wall temperature over-limit frequency is the number of times the safety threshold is exceeded per unit time; and the corrosion rate growth rate is the ratio of the current rate to the initial rate. When the system reaches less than 85% efficiency, it generates a maintenance recommendation report, suggesting adjusting the soot blowing cycle from once per shift to once every two hours, adjusting the pH value in the water quality control parameters to the range of 8.5 to 9.0, or prompting the replacement of specific pipe sections. It is recommended to push this information to maintenance personnel through the Enterprise Resource Planning (ERP) system to achieve closed-loop management.

[0032] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, the remote sensing image data in step (1) includes high-resolution optical images, synthetic aperture radar images, and hyperspectral images. The temporal resolution is no less than 15 days, and the spatial resolution is better than 2 meters. Multi-temporal image alignment is achieved through geometric correction and radiometric normalization processing based on control points. Specifically, the remote sensing image data comes from Sentinel-2, Landsat-8, and WorldView-3 satellite platforms. The optical image bands cover visible light to shortwave infrared, with a spatial resolution of 10 meters to 0.3 meters and a time interval of 5 to 15 days. The synthetic aperture radar images use the C-band, with a polarization mode of VV / VH, a spatial resolution of 10 meters, and strong cloud penetration capability. The hyperspectral images contain more than 200 continuous bands with a wavelength range of 400 to 2500 nanometers, used for fine ground feature classification. All images are first geometrically corrected, and a projection transformation model is established using ground control points (GCPs). The correction error is less than 1 pixel. Radiation normalization employs atmospheric correction models (such as the 6S model) to remove atmospheric influences and standardizes surface reflectance to the same reference. Multi-temporal images are aligned to a unified coordinate system using registration algorithms to ensure consistent ground feature locations at different time points, providing a foundation for subsequent change detection.

[0033] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, the real estate registration data in step (1) covers natural resource property rights information such as land, forest land, water area, and minerals. Entities are linked through the unified social credit code and spatial location code to ensure accurate matching between property boundaries and geographic coordinates, with a matching error of less than 0.5 meters. Specifically, the real estate registration data comes from the Ministry of Natural Resources' real estate registration system and includes fields such as parcel number, right holder, use, area, and boundary coordinates. The unified social credit code links legal entities with their owned natural resource assets, achieving cross-departmental data integration. The spatial location code uses the National Geodetic Coordinate System (CGCS2000) with centimeter-level coordinate accuracy. Property boundaries are stored as polygonal elements and analyzed by overlaying with remote sensing images, using spatial topological relationships to determine whether there are overlaps or gaps. The matching error is optimized using the least squares method to ensure that the boundary line matches the actual ground features, with the error controlled within 0.5 meters.

[0034] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, the graph neural network in step (2) adopts a graph attention mechanism. The node feature vector dimension is 128, and the edge weights are calculated by multiplying the spatial distance decay function with the functional similarity. The decay coefficient is set to 0.8, and the similarity threshold is set to 0.7. Specifically, the graph neural network is constructed with natural resource plots as nodes. The node features include area, use, vegetation index, soil type, etc., and the dimension is 128. Edge connections are based on spatial adjacency and functional similarity, and the edge weight calculation formula is:

[0035] in For spatial distance, Functional similarity is calculated based on indicators such as usage and ecological value. If the value is less than 0.7, no connection is established. The graph attention mechanism assigns different weights to neighboring nodes during message passing, increasing the influence of important nodes.

[0036] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, the spatial autocorrelation analysis in step (2) uses a combination of the Moran index and the Gillley index for discrimination. The calculation window radius is set to 5 kilometers, and the significance level is set to 0.05. This is used to identify high-high clustering areas and low-low clustering areas as the initial connection basis for the correlation map. Specifically, the Moran index measures global spatial autocorrelation, and the Gillley index measures local spatial autocorrelation. The calculation window radius is 5 kilometers, covering the scale of typical ecological units. The significance level is set to 0.05, and the significance is determined by a permutation test. High-high clustering areas refer to adjacent high-value areas, and low-low clustering areas refer to adjacent low-value areas. These areas are preferentially connected to form the initial correlation map.

[0037] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, the ecological security objective in step (3) is quantified by the ecological carrying capacity index, the economic efficiency objective by the unit resource output ratio, and the social equity objective by the reciprocal of the Gini coefficient. These three objectives constitute a three-dimensional optimization objective space. Specifically, the ecological carrying capacity index is based on ecosystem service value assessment, taking into account functions such as biodiversity, water conservation, and carbon sequestration. The calculation formula is as follows:

[0038] in As weight, For the first The value of a service. Economic efficiency is the GDP output per unit of resource input, while social equity is the reciprocal of the Gini coefficient, reflecting the degree of income distribution equilibrium. These three objectives constitute a three-dimensional objective space for multi-objective optimization.

[0039] In the above-mentioned method for constructing a digital twin of a cryogenic economizer, the improved multi-objective particle swarm optimization algorithm in step (3) introduces an elite archiving mechanism and adaptive inertia weights. The population size is set to 200, the maximum number of iterations is set to 500, and the convergence criterion is that the Pareto front change rate is less than 1% for 50 consecutive generations. Specifically, the algorithm initializes 200 particles, each representing a resource allocation scheme. The elite archiving stores non-dominated solutions, and the adaptive inertia weights are dynamically adjusted with the number of iterations, decreasing from 0.9 to 0.4 to balance exploration and development. The particle velocity and position are updated in each generation, fitness is evaluated, and individual optimality and global optimality are updated. When the Pareto front change rate is less than 1% for 50 consecutive generations, the algorithm is considered to have converged, and the optimal solution set is output.

[0040] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, the resource allocation deviation index in step (4) is calculated by weighting the spatial mismatch rate, the target deviation, and the execution lag coefficient, with weights of 0.4, 0.4, and 0.2 respectively, and the deviation index threshold is set to 0.15. Specifically, the spatial mismatch rate is the proportion of the difference between the planned allocation and the actual distribution, the target deviation is the relative value of the difference between the actual achieved value and the target value, and the execution lag coefficient is the degree of delay in policy implementation. The formula for calculating the deviation index is:

[0041] when When the value is greater than 0.15, the adjustment mechanism is triggered.

[0042] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, the reinforcement learning agent in step (4) adopts a deep deterministic policy gradient algorithm. The state space includes the current deviation index and historical adjustment records, the action space is the fine-tuning amount of the multi-objective weight vector, and the reward function is the product of the deviation index decrease and the scheme stability. Specifically, the state space has a dimension of 10 and includes the current deviation index, the past 5 adjustment records, and the trend of target weight changes. The action space is a three-dimensional vector, corresponding to the adjustment amount of the ecological, economic, and social target weights, respectively, with a range of [-0.1, 0.1]. The reward function is defined as:

[0043] in To ensure similarity between the current and historical strategies, smooth adjustments are encouraged. The agent gradually optimizes its strategy through experience replay and training with the target network.

[0044] In the above-mentioned method for constructing a digital twin of a low-temperature economizer, the visualization decision support report in step (5) is dynamically generated according to quarterly, annual, and five-year planning cycles. The spatial distribution heat map adopts a graded color scheme with 7 color levels. The target achievement radar chart includes 5 dimensions of indicators. Risk warning prompts are classified and labeled according to three levels: high, medium, and low. Specifically, the heat map color ranges from light blue to dark red, representing from low to high, with 7 color levels for easy and intuitive identification. The radar chart includes five dimensions: ecological security, economic efficiency, social equity, resource utilization efficiency, and sustainable development, with each dimension scored from 0 to 100. Risk warnings are judged based on deviation index and trend, with high risk being... A value greater than 0.2 and continuing to rise indicates a medium risk level of 0.15. <0.2, low risk <0.15, marked in red, yellow and green respectively.

[0045] This invention also includes establishing a knowledge base for the optimal allocation of natural resource assets. The knowledge base stores no fewer than 100,000 historical allocation cases, policy and regulatory provisions, and expert rules, and provides prior knowledge support for the initial setting of multi-objective weights through a semantic retrieval engine. Specifically, the knowledge base is stored using the Neo4j graph database, with nodes including cases, policies, rules, expert opinions, etc., and edges representing relationships. The semantic retrieval engine is based on the BERT model, supports natural language queries, and returns relevant cases and rules to assist in weight setting.

[0046] This invention also includes integration with the national land spatial planning "one map" system, acquiring planning control boundary data in real time through a standard geographic information service interface to ensure strict consistency between resource allocation schemes and higher-level plans, with spatial consistency verification error less than 1 pixel. Specifically, the integration adopts the WMS / WFS standard protocol to acquire layers such as planned land use nature, ecological protection red lines, and permanent basic farmland. Verification is performed through spatial overlay analysis to check whether the resource allocation scheme exceeds the control boundaries, with the error controlled within 1 pixel to ensure compliance.

[0047] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a digital twin of a cryogenic economizer, the method comprising: The method comprises the following specific steps: ​ Step (1) constructing a multi-physics coupling benchmark digital model: based on the geometric structure parameters, material thermal physical property parameters and typical operating conditions of the low-temperature economizer, a multi-physics coupling benchmark digital model covering fluid dynamics, heat transfer and electrochemical corrosion mechanism is established, the multi-physics coupling benchmark digital model is discretely solved by the finite volume method to obtain Navier-Stokes equation, energy conservation equation and corrosion rate equation, and outputs the flow velocity field, temperature field, wall heat flux density and corrosion rate distribution in the device; Step (2) collecting multi-source heterogeneous operating data and performing time-space alignment and denoising processing: synchronously obtaining the original data from the distributed temperature sensor, pressure transmitter, flue gas component analyzer and ash deposition monitoring device, adopting a multi-channel signal alignment strategy based on a dynamic time warping algorithm to eliminate sampling time deviation, and combining a wavelet threshold denoising and sliding window median filtering joint method to suppress random noise and pulse interference in a high dust environment, so as to obtain the pretreated operating data; Step (3) constructing an adaptive data-model fusion driving mechanism: taking the pretreated operating data as an observation input, introducing an extended Kalman filter to online correct the state variables of the benchmark digital model, wherein the state vector includes the average flow velocity, wall temperature gradient and local heat transfer coefficient of the key section, an observation equation is constructed to map the measured values at the physical positions of the sensors to the model grid nodes, so as to establish the corresponding relationship between the observation data and the model state, and realize the dynamic updating of the model parameters and boundary conditions; Step (4) embedding a deep time sequence feature extraction and performance degradation identification module: using a one-dimensional convolutional neural network and a gated recurrent unit hybrid architecture to encode the multi-dimensional operating sequence in a continuous time window, extract implicit degradation features reflecting the ash deposition growth rate, corrosion acceleration trend and heat exchange efficiency decay, and automatically divide the steady-state, transition-state and abnormal working condition intervals through an unsupervised clustering algorithm; Step (5) performing closed-loop iterative optimization and state mapping of the digital twin: triggering the model reconstruction mechanism according to the degradation feature identification result, when it is detected that the duration of the non-steady-state working condition exceeds the preset threshold or the degradation feature vector deviates from the normal distribution by more than 3 times the standard deviation, automatically calling a high-fidelity computational fluid dynamics simulation module to finely re-simulate the local area, and injecting the updated physical field data into the benchmark model to complete the high-precision synchronous mapping of the digital twin and the state of the physical device.

2. The low-temperature economizer digital twin construction method of claim 1, wherein: The geometric structure parameters include tube bundle arrangement, tube diameter, tube length, fin spacing and flue gas flow area, the material thermal physical property parameters include thermal conductivity, specific heat capacity, density and electrochemical corrosion potential, and the typical operating conditions include flue gas temperature, flow and sulfur content combinations corresponding to 5 representative coal qualities in the 30% to 100% load interval; wherein the tube bundle arrangement is staggered or in-line arrangement, the tube diameter ranges from 20 mm to 25 mm, the tube length ranges from 6 m to 8 m, and the fin spacing ranges from 8 mm to 10 mm.

3. The low-temperature economizer digital twin construction method of claim 1, wherein: The distributed temperature sensor is arranged at not less than 12 measuring points along the flue gas flow direction at the inlet and outlet of the economizer and the middle section, the pressure transmitter is installed at the position for measuring the pressure difference before and after the flue, the sampling frequency of the flue gas composition analyzer is not less than 1 Hz, the ash deposition monitoring device adopts the acoustic wave ash measurement or microwave attenuation principle, and the data sampling period is uniformly set to 5 seconds; the temperature sensor selects a platinum resistance or a thermocouple type, the accuracy level is better than 0.2 degrees Celsius, the pressure transmitter range is 0 to 12 kPa, and the accuracy is better than ±0.6%.

4. The low-temperature economizer digital twin construction method of claim 1, wherein: The wavelet threshold denoising adopts a db4 wavelet basis function for 5-layer decomposition, the threshold function is a soft threshold, and the threshold size is set to 1.5 times the standard deviation of noise estimation; the window length of the sliding window median filter is 15 sampling points, which is used to eliminate the instantaneous data jump caused by sudden dust impact; all sensor data are accessed to the edge computing node through the industrial Ethernet, the time stamp is synchronized by the GPS timing system, and the accuracy reaches the millisecond level.

5. The low-temperature economizer digital twin construction method of claim 1, wherein: The state vector dimension of the extended Kalman filter is 64, corresponding to 64 key monitoring areas in the model grid; the observation equation maps the sensor physical coordinates to the nearest four grid nodes through bilinear interpolation, and the observation noise covariance matrix is dynamically adjusted according to the sensor accuracy level, the temperature sensor covariance is set to 2 Kelvin square, and the pressure sensor covariance is set to 100 Pascal square; the state vector is defined as a combined vector of average flow velocity, wall temperature gradient and local heat transfer coefficient, and the prediction and update cycle is executed once every 5 seconds.

6. The low-temperature economizer digital twin construction method of claim 1, wherein: The one-dimensional convolutional neural network comprises three convolutional layers, the convolution kernel sizes are 7, 5 and 3 respectively, the output channel numbers of each layer are 64, 128 and 256 respectively, and a maximum pooling layer is connected in sequence; the gated recurrent unit hidden layer dimension is 512, the input sequence length is 300 time steps, and the corresponding continuous operation data is 25 minutes; the unsupervised clustering adopts a Gaussian mixture model, the cluster number is automatically determined through the Bayesian information criterion, and is usually 3 to 5 categories.

7. The low-temperature economizer digital twin construction method of claim 1, wherein: The degradation feature vector comprises 128-dimensional hidden state outputs, wherein the first 64 dimensions represent short-term dynamic response characteristics, and the last 64 dimensions represent long-term trend cumulative effects; the abnormal working condition determination adopts Mahalanobis distance measurement, and when the Mahalanobis distances of 10 continuous time steps are all greater than 9.21, it is determined that the performance is significantly degraded; the clustering result is used to divide the steady state, transition state and abnormal working condition intervals, and supports subsequent decision triggering.

8. The low-temperature economizer digital twin construction method of claim 1, wherein: The high-fidelity computational fluid dynamics simulation module adopts a Reynolds stress turbulence model and a discrete phase model for coupled solution, the number of grids is not less than 5 million, and the time step is 0.01 seconds; the reconstruction area is determined by the top 3 monitoring areas with the highest contribution in the degradation feature space; after the simulation is completed, high-resolution physical field data is output, which is mapped to the reference model grid through an interpolation algorithm, replaces the original data and participates in the state correction process again.

9. The cryogenic economizer digital twin construction method of claim 1, wherein: Also includes establishing a historical operation database, storing not less than 2 years of full operating condition operation data and corresponding digital twin state snapshot, data sampling frequency is 1 time per minute; based on the database, a transfer learning model is trained, which is used to quickly initialize the digital twin benchmark model of the new unit in the initial stage of operation, and the model convergence time is shortened to within 72 hours; the database adopts a relational structure, and the state snapshot stores key parameters and feature vectors in JSON format.

10. The low-temperature economizer digital twin construction method of claim 1, wherein: Also includes setting a digital twin health degree evaluation index, which comprehensively considers the heat exchange efficiency deviation rate, wall temperature overrun frequency and corrosion rate growth rate, and automatically triggers a preventive maintenance suggestion when the health degree is less than 85%, including soot blowing period adjustment, water quality control parameter optimization or local pipe section replacement warning, and pushing to the operation and maintenance personnel through the enterprise resource planning system.

Images