Method, device and equipment for training boiler water wall temperature field prediction model

Abstract

The application discloses a boiler water-cooled wall temperature field prediction model training method, a prediction method and equipment, and is applied to a boiler monitoring system. The method acquires a water-cooled wall temperature field and operation data under different loads to construct a sample library, and divides source load working condition domains and target load working condition domains based on loads and equipment operation state parameters; operation characteristics, spatial measurement point characteristics, temperature field shape characteristics and measurement point reliability characteristics are extracted, a cross-load migration sample weight is obtained by calculating the difference degree of the source domain and the target domain in the multidimensional characteristics, and a migration training sample set is constructed accordingly; finally, the temperature field spatial gradient is constrained based on the spatial adjacency relationship of the measurement points, the measurement point reliability characteristics are taken as the prediction error weight, and the migration model is iteratively trained and verified. The application effectively blocks the negative migration interference caused by the mixed heterogeneous data, eliminates the data pollution of the poor measurement points in the industrial field, and realizes the high-precision prediction of the water-cooled wall temperature field under the small sample complex working condition.

Description

Technical Field

[0001] This invention relates to the field of boiler monitoring and temperature field modeling technology, and in particular to a training method, prediction method and equipment for a boiler water-cooled wall temperature field prediction model. Background Technology

[0002] Water-cooled walls are important heat-receiving surfaces in boiler furnaces. Their temperature field distribution can reflect the furnace heat load distribution, water-cooled wall heat exchange state, and local heat safety margin. As boiler operation modes gradually develop towards deep peak shaving, frequent load increases and decreases, and flexible operation under multiple conditions, the heat load environment, combustion organization state, and steam-water side heat exchange conditions of water-cooled walls will change significantly with changes in operating load, resulting in strong operating condition correlation and spatial variability in the water-cooled wall temperature field.

[0003] Existing water-cooled wall temperature field modeling methods typically rely on historical data from a fixed load range or relatively stable operating conditions for training. When the boiler operates under conditions such as low load, variable load, changes in burner operation combination, coal mill switching, or soot blowing disturbance, the original model is prone to problems such as increased prediction error, distortion of spatial temperature distribution, and inaccurate identification of local high-temperature areas. Especially when the number of measured samples under the target load condition is small, it is difficult to directly rely on the data of this condition to establish a reliable temperature field prediction model.

[0004] Furthermore, while there are certain commonalities among different load conditions, there are also significant differences. If historical data from different load conditions are directly mixed for model training, it can easily cause mutual interference between the data from different load conditions, making it difficult for the model to accurately learn the temperature field change patterns under the target load condition. At the same time, the temperature measurement points of the water-cooled wall on site may also be affected by factors such as installation status, long-term thermal cycling, aging of measurement points, and signal fluctuations, resulting in sudden jumps, drifts, or local discontinuities in some measurement point data, which further affects the training effect and prediction reliability of the temperature field model.

[0005] A review of publicly available technical solutions reveals that CN120542522B proposes a training method, a temperature field determination method, and a device for a temperature field determination model. The training method includes: acquiring a sample temperature data set for the target hardware module, comprising a real temperature data set, a simulated temperature data set, and a generated temperature data set; and performing transfer learning training on a pre-trained source model based on the real and simulated temperature data sets to obtain an intermediate model. The source model includes physical information neural networks constrained by the target heat conduction equation. The network; and based on real temperature data sets, simulated temperature data sets, and generated temperature data sets, the intermediate model is trained to obtain a trained temperature field determination model; however, this scheme is mainly aimed at the temperature field determination of the target hardware module, focusing on the transfer training of physical information neural networks on real, simulated and generated temperature data, without considering the differences in combustion state, steam-water side operation, temperature field morphology and measurement point reliability under different load conditions of boiler water-cooled walls, and without dividing the operating domain and controlling the transfer weights, making it difficult to solve the negative transfer problem caused by the mismatch of source operating condition data when the target load condition samples of water-cooled walls are insufficient. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a training method, prediction method, and device for a boiler water-cooled wall temperature field prediction model. Its purpose is to overcome the prediction distortion defects caused by the scarcity of measured samples under deep peak shaving and complex variable load conditions in existing models. By combining operating condition domain partitioning and multi-dimensional difference degree migration weight control, it blocks the negative migration interference caused by the blind mixing of heterogeneous historical data, enabling the model to efficiently and directionally learn from the historical experience with the highest matching degree. Simultaneously, to address the problem of sensor aging and drift in harsh industrial environments, a measurement point reliability mechanism is introduced to shield dirty data from contaminating model parameter updates. This achieves high-precision reconstruction of the global temperature field of the boiler water-cooled wall and stable identification of local high-temperature risk areas even in small-sample and high-noise actual operating environments.

[0007] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: According to a first aspect of the present invention, a method for training a boiler water-cooled wall temperature field prediction model is provided, comprising: Acquire water-cooled wall temperature field data and corresponding boiler operating condition data under different load conditions, and construct a cross-load condition sample library; Based on the load parameters and equipment operating status parameters in the boiler operating condition data, the training samples in the cross-load operating condition sample library are jointly divided into operating condition domains to determine the source load operating condition domain and the target load operating condition domain. The operating characteristics, spatial measurement point characteristics, and temperature field morphology characteristics of each load condition domain are extracted respectively, and the measurement point reliability characteristics used to characterize the quality of temperature sensor data are extracted. Based on operational characteristics and temperature field morphology characteristics, the load difference, combustion state difference, steam-water side operation difference, temperature field morphology difference, and load response difference between samples in the source load operating condition domain and samples in the target load operating condition domain are calculated respectively. The difference values ​​are then fused together to obtain the cross-load migration sample weights of the source load operating condition domain relative to the target load operating condition domain. Based on the cross-load migration sample weights, corresponding training samples are extracted from the source load condition domain to construct a migration training sample set; Based on the spatial measurement point features, the spatial adjacency relationship of the measurement points is constructed to constrain the spatial gradient of the temperature field. The measurement point credibility feature is used as the weight for calculating the single-point prediction error. The initial water-cooled wall temperature field transfer model is iteratively trained using the transfer training sample set to obtain a boiler water-cooled wall temperature field prediction model suitable for the target load operating domain. The boiler water-cooled wall temperature field prediction model is validated based on the measured temperature field data under the target load condition, and the cross-load migration sample weights, measurement point credibility features or model parameters are updated when the validation error exceeds the preset threshold.

[0008] In one possible implementation of the first aspect, the load parameters include the boiler load range and the load change rate, and the equipment operating status parameters include the burner commissioning status, the coal mill operating status, and the soot blowing status. The step of performing joint load condition domain partitioning on training samples in the cross-load condition sample library to determine the source load condition domain and the target load condition domain includes: According to the proportion of boiler load to rated load, the training samples in the cross-load condition sample library are divided into different load ranges; Based on the stability of the load change rate in the sample, the samples within the same load range are divided into sample subsets with different load change trends. The sample subsets specifically include stable load samples, load increase samples, and load decrease samples. For stable load samples, increasing load samples, and decreasing load samples, operating conditions are marked based on their corresponding burner commissioning status, coal mill operating status, and soot blowing status to form a joint operating condition domain. The joint operating condition domain with a sufficient number of historical samples and a stable temperature field distribution corresponding to the samples is defined as the source load operating condition domain, and the joint operating condition domain with insufficient measured samples and the need to train the temperature field prediction model is defined as the target load operating condition domain.

[0009] In one possible implementation of the first aspect, the operating characteristics include boiler load, load change rate, fuel quantity, air volume at each stage, flue gas oxygen content, steam-water side operating parameters, burner commissioning status, coal mill operating status, and soot blowing status. The temperature field morphology features include regional average temperature, regional extreme temperature, regional temperature gradient, spatial temperature difference, area ratio of high-temperature regions, and centroid location of high-temperature regions. The spatial measuring point characteristics include the location of the measuring point on the furnace wall, the tube panel number, the spatial coordinates, and the area affected by the equipment to which it belongs; The reliability characteristics of the measuring point are composed of the basic reliability of the measuring point, the reliability of the measuring point jump, the reliability of the measuring point drift, and the reliability of the spatial continuity of the measuring point.

[0010] In one possible implementation of the first aspect, the cross-load migration sample weights of the source load case domain relative to the target load case domain are obtained by fusing and calculating the various differences, and the calculation formula is as follows:

[0011] in, Indicates the first Cross-load migration sample weights of each source load case domain relative to the target load case domain; This indicates the number of source load case domains participating in the transfer training. Indicates the degree of load variation; Indicates the degree of difference in combustion state; Indicates the difference in operation between the steam and water sides; Indicates the degree of difference in temperature field morphology; Indicates the degree of load response difference; , , , , These represent the corresponding adjustment coefficients; This represents the zero minimum constant.

[0012] In one possible implementation of the first aspect, the calculation process for the load difference, combustion state difference, steam-water side operation difference, temperature field morphology difference, and load response difference includes: Based on the boiler load and load change rate in the operating characteristics, the load difference degree is obtained by normalization comparison; Based on the fuel quantity, air volume at each stage, flue gas oxygen content, burner commissioning status, coal mill operating status, and soot blowing status in the operating characteristics, the combustion status difference is obtained by comparing the consistency of the status. Based on the operating parameters of the steam and water sides in the operating characteristics, the degree of deviation of the heat transfer boundary conditions is compared after normalization to obtain the operating difference degree of the steam and water sides. Based on the regional average temperature, regional extreme temperature, regional temperature gradient, spatial temperature difference, area ratio of high temperature region, and centroid position of high temperature region in the temperature field morphology characteristics, the degree of deviation is compared to obtain the temperature field morphology difference. Based on the characteristics of the temperature field morphology, the consistency of the response trend of the temperature field with changes in boiler load is compared to obtain the load response difference degree.

[0013] In one possible implementation of the first aspect, the step of extracting corresponding training samples from the source load domain based on cross-load migration sample weights to construct a migration training sample set includes: The cross-load migration sample weights corresponding to each source load condition domain are compared with the preset migration weight threshold, and the source load condition domains with weights greater than the preset migration weight threshold are extracted as the migrateable source load condition domains. Normalize the weights of cross-load migration samples corresponding to all transferable source load case domains to obtain the data proportion of each transferable source load case domain in the migration training sample set. Based on the proportion of each data point, sample data of a corresponding proportion are extracted from the corresponding transferable source load condition domain and combined with the existing measured samples in the target load condition domain to form a transfer training sample set.

[0014] In one possible implementation of the first aspect, the initial water-cooled wall temperature field transfer model is iteratively trained using a transfer training sample set, and the total training loss of the water-cooled wall temperature field transfer model is... The calculation formula is:

[0015]

[0016]

[0017]

[0018] in, This represents the transfer learning loss corresponding to the source load condition domain data. This represents the correction loss corresponding to the measured data in the target load condition domain. This represents the loss of consistency of the spatial gradient of the temperature field. , , These are the adjustment coefficients corresponding to each loss term; This indicates the number of transferable source load condition types participating in the transfer training. Indicates the first The cross-load condition type migration weight corresponding to each source load condition type. Indicates the first The number of data points used in training for each source load condition type. This indicates the number of temperature measurement points on the water-cooled wall. Indicates the first In the individual load condition type, the first Group of data corresponds to the first The reliability of each measuring point This represents the predicted temperature output by the model. This indicates the corresponding measured temperature; This indicates the number of existing sample data under the target load condition type. Indicates the first in the target load condition type Group of data corresponds to the first The reliability of each measuring point The model represents the first time under the target load condition. The predicted temperature output from each measuring point. Indicates the first under the target load condition Measured temperature at each measuring point; This represents a pre-established set of adjacency relationships for measurement points based on the spatial arrangement of measurement points on the water-cooled wall, which is used to calculate the spatial gradient of the temperature field in all training samples. Indicates the first The measuring point and the first Each measuring point is a spatially adjacent measuring point. This represents the temperature difference between adjacent measuring points in the model's predicted temperature field. It represents the temperature difference between adjacent measuring points in the measured temperature field.

[0019] In one possible implementation of the first aspect, the formula for calculating the reliability feature of the measurement point is:

[0020] in, Indicates the first In the individual load condition type, the first Group of data corresponds to the first The reliability of each measuring point; To establish the reliability of the measurement points; The reliability of sudden jumps in measurement points is determined based on the temperature change amplitude between adjacent sampling times; The reliability of the measurement point drift is determined based on the degree to which it deviates from the mean of adjacent measurement points in the same area within a time window. The reliability of spatial continuity of the measurement point is determined based on the consistency of the temperature change trend between adjacent measurement points with the influence area of ​​the equipment to which the measurement point belongs in the spatial measurement point characteristics.

[0021] According to a second aspect of the present invention, a method for predicting the temperature field of a boiler water-cooled wall is provided, comprising: Real-time acquisition of boiler operating condition data under the current target load condition, and extraction of the reliability characteristics of the current measuring point; The boiler operating condition data is input into the pre-deployed boiler water-cooled wall temperature field prediction model to obtain the water-cooled wall temperature field prediction results of the boiler under the current target load condition. Based on the predicted temperature field of the water-cooled wall and the reliability characteristics of the measuring points, local high-temperature risk areas of the water-cooled wall are identified for boiler operation safety monitoring. The boiler water-cooled wall temperature field prediction model was obtained by training on a historical cross-load condition sample database using a training method.

[0022] According to a third aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the training method for a boiler water-cooled wall temperature field prediction model.

[0023] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the training method for a boiler water-cooled wall temperature field prediction model.

[0024] According to a fifth aspect of the present invention, a computer program product is provided, which, when executed by a processor, implements the training method for a boiler water-cooled wall temperature field prediction model.

[0025] Compared with the prior art, the present invention has at least the following beneficial effects: This invention, through joint operating condition domain division based on load and equipment operating status parameters, and combined with multi-dimensional feature difference degree calculation of cross-load migration sample weights, can accurately select high-quality source domain samples that highly match the target operating condition from massive historical data for targeted migration training. This effectively avoids negative migration interference caused by blind mixing of heterogeneous operating condition data and overcomes the problem of temperature field prediction distortion caused by insufficient target measured samples under complex operating conditions such as deep peak shaving and variable load. On this basis, in the model iteration training stage, this invention, on the one hand, uses the spatial adjacency relationship of measurement points to apply physical topological constraints to the spatial gradient of the temperature field, ensuring the structural rationality of the predicted temperature field spatial distribution; on the other hand, it directly uses the reliability features of measurement points as the calculation weight of single-point prediction error, dynamically weakening and shielding the pollution of model parameter updates by dirty data such as sensor aging, sudden jumps and drift in industrial sites. Ultimately, it achieves high-precision reconstruction of the global temperature field of boiler water-cooled walls and stable identification of local high-temperature risk areas. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the specific embodiments of the present invention, the drawings used in the description of the specific embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a flowchart of a training method for a boiler water-cooled wall temperature field prediction model according to the present invention.

[0028] Figure 2 This is a schematic diagram of the load condition domain division method of the present invention.

[0029] Figure 3 This is a schematic diagram of the method for constructing a migration sample set corresponding to the target load condition of the present invention.

[0030] Figure 4 This is a schematic diagram comparing the solution of the present invention with that of the traditional solution in terms of key indicators.

[0031] Figure 5 This diagram illustrates the improvement of the present invention's solution in key indicators compared to traditional solutions. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] This embodiment addresses the technical problems of insufficient measured samples of water-cooled wall temperature field under target load conditions during deep peak shaving and frequent load changes in boilers. Directly mixing multi-condition data for training can easily lead to negative transfer, and data quality defects at measurement points result in low model prediction accuracy and spatial distribution distortion. To address these issues, a training method for a boiler water-cooled wall temperature field prediction model based on cross-load transfer learning is proposed. The core idea is to accurately classify samples from different operating states through joint operating condition domain partitioning, quantify the transferability between source and target operating conditions based on multi-dimensional differences, calculate transfer sample weights, and combine measurement point credibility weighting with temperature field spatial gradient constraints to complete the transfer model training. Continuous iterative optimization using measured data ultimately yields a high-precision temperature field prediction model suitable for the target load condition. The overall flow of the boiler water-cooled wall temperature field prediction model training method provided in this embodiment is shown in Figure 1, and the specific implementation steps are as follows: S1. Obtain water-cooled wall temperature field data and corresponding boiler operating condition data under different load conditions, and construct a cross-load condition sample library.

[0034] Specifically, water-cooled wall temperature field data is acquired through temperature measuring points deployed on the outer surface of the water-cooled wall tubes, the fin area, the furnace wall area, or the hazardous heating areas of the water-cooled wall. A combination of K-type armored wall-mounted temperature elements and FBG-1550 fiber optic grating temperature elements can be used. The K-type armored wall-mounted temperature element has a measurement range of 0℃ to 800℃ and a measurement accuracy of ±1.5℃, while the FBG-1550 fiber optic grating temperature element has a measurement range of 0℃ to 650℃ and a wavelength demodulation resolution of 1pm. The temperature data sampling period can be set to 10s and connected to the edge computing station via a TMC-32 temperature acquisition terminal. In practical implementation, temperature measuring elements can also be replaced by thin-film resistance temperature elements, infrared temperature measurement windows, acoustic temperature inversion units, or distributed fiber optic temperature sensing units. The sampling period can be adjusted within the range of 5s to 60s according to the boiler load change rate.

[0035] Boiler operating condition data covers load parameters and equipment operating status parameters, specifically including boiler load, load change rate, fuel quantity, primary air volume, secondary air volume, total air volume, flue gas oxygen content, feedwater flow rate, steam drum pressure, main steam flow rate, main steam pressure, furnace outlet flue gas temperature, burner commissioning status, coal mill operating status, and soot blowing status. This data is collected through the boiler distributed control system (DCS), combustion control system, soot blowing control system, steam-water side monitoring system, or historical operating database. The operating data sampling period can be set to 5 seconds. After collection, the data is uniformly written to the storage medium through the OPC-200 data communication gateway.

[0036] The cross-load condition sample library can be deployed on a local edge server or a factory-level data center.

[0037] After data acquisition, the water-cooled wall temperature field data and the corresponding boiler operating condition data are correlated and sampled. Data alignment is completed according to a unified timestamp, and the time alignment error is controlled within 30 seconds. After removing invalid samples such as boiler shutdown, sensor communication interruption, burner start-up and shutdown transient anomalies, and missing DCS records, a cross-load condition sample library containing training sample sets corresponding to various load condition domains is finally formed.

[0038] S2. Based on the load parameters and equipment operating status parameters in the boiler operating condition data, perform joint operating condition domain division on the training samples in the cross-load operating condition sample library to determine the source load operating condition domain and the target load operating condition domain.

[0039] The load parameters include the boiler load range and load change rate, while the equipment operating status parameters include burner operation status, coal mill operation status, and soot blowing status. The process for dividing the operating condition domain is as follows: Figure 2 As shown, the specific execution follows a hierarchical progression: S21. According to the proportion of boiler load to rated load, the training samples in the cross-load condition sample library are divided into different load intervals. For example, they can be divided into low load interval, medium load interval and high load interval, or they can be subdivided into multiple load intervals according to operating requirements. In actual implementation, the interval division range can also be adjusted according to the actual peak-shaving operation strategy, or the load condition domain can be automatically formed by clustering according to the distribution of historical operating samples.

[0040] S22. Based on the stability of the load change rate in the samples, the samples within the same load range are divided into sample subsets with different load change trends. The sample subsets specifically include stable load samples, load increase samples, and load decrease samples. Among them, samples with an absolute value of load change rate less than 0.5% of rated load / min can be marked as stable load samples, samples with a load change rate greater than 0.5% of rated load / min can be marked as load increase samples, and samples with a load change rate less than -0.5% of rated load / min can be marked as load decrease samples.

[0041] S23. For stable load samples, increasing load samples, and decreasing load samples, secondary operating condition markings are performed based on their corresponding burner operation status, coal mill operation status, and soot blowing status to form a joint operating condition domain. The burner operation status can be marked into three categories: lower layer burner operation alone, lower and middle layer burner combined operation, and all layer burner combined operation. The coal mill operation status can be marked into three categories: three mills operating, four mills operating, and five mills operating. The soot blowing status can be marked into three categories: 60 minutes before soot blowing, during soot blowing, and 120 minutes after soot blowing. The time window for soot blowing status can also be set to 30 minutes, 90 minutes, or 180 minutes according to actual operating requirements.

[0042] S24. Finally, the joint operating condition domain that meets the preset quantity requirement for the number of historical samples and whose corresponding temperature field distribution is stable is determined as the source load operating condition domain. The joint operating condition domain that requires training the temperature field prediction model and has insufficient measured sample quantity, significant temperature field distribution change, or is in a new operating load range is determined as the target load operating condition domain.

[0043] S3. Extract the operating characteristics, spatial measurement point characteristics, and temperature field morphology characteristics of each load condition domain, and extract the measurement point reliability characteristics used to characterize the quality of temperature sensor data.

[0044] Specifically, the operating characteristics include boiler load, load change rate, fuel quantity, air volume at each stage, flue gas oxygen content, steam and water side operating parameters, burner commissioning status, coal mill operating status, and soot blowing status. Among these, the steam and water side operating parameters include feedwater flow rate, steam drum pressure, main steam flow rate, and main steam pressure. The spatial measurement point characteristics include the location of the measurement point on the furnace wall, the tube panel number, the spatial coordinates, and the area affected by the equipment to which it belongs. Specifically, these can be further refined into height position, lateral position, adjacent measurement point number, the area affected by the burner to which it belongs, and the area affected by the soot blower to which it belongs. The temperature field morphology characteristics include the regional average temperature, regional extreme temperature, regional temperature gradient, spatial temperature difference, the area ratio of the high-temperature region, and the centroid position of the high-temperature region. Specifically, these can be further refined into the regional maximum temperature, regional minimum temperature, temperature difference between the left and right walls, temperature difference between the front and rear walls, temperature difference between the upper and lower layers, and temperature difference between adjacent measurement points on the same tube panel.

[0045] The reliability characteristics of a measuring point are composed of the basic reliability of the measuring point, the reliability of abrupt jumps in the measuring point, the reliability of drift in the measuring point, and the reliability of spatial continuity of the measuring point. The calculation formula is as follows:

[0046] In the formula, Indicates the first In the individual load condition type, the first Group of data corresponds to the first The reliability of each measuring point; The basic reliability of the measurement point can be determined based on the installation location, historical effectiveness rate, historical missing rate, and maintenance records. For example, if the effectiveness rate of the measurement point is not less than 98% within 90 days and there are no abnormal maintenance records, the basic reliability can be set to 0.95 to 1.00. If the effectiveness rate of the measurement point is less than 95% within 90 days or there are continuous communication interruption records, the basic reliability can be set to 0.60 to 0.80. The confidence level of a sudden jump at a measurement point is determined based on the temperature change amplitude between adjacent sampling times. When the temperature change exceeds 25°C within 10 seconds and the adjacent measurement points do not change synchronously, the confidence level of the sudden jump at that measurement point can be reduced to 0.40 to 0.70. The drift confidence level is determined based on the degree of deviation from the average of adjacent measurement points in the same area within a time window. When the deviation exceeds 12°C within a 24-hour window, the drift confidence level can be reduced to 0.50 to 0.75. The reliability of spatial continuity of the measuring points is determined based on the consistency of temperature change trends between adjacent measuring points within the influence area of ​​the equipment.

[0047] In practice, the reliability of measurement points can also be achieved using rule-based thresholding, sliding window statistics, isolated forest detection, or anomaly identification methods based on neighborhood consistency.

[0048] S4. Based on the operating characteristics and temperature field morphology characteristics, calculate the load difference, combustion state difference, steam-water side operating difference, temperature field morphology difference, and load response difference between the samples in the source load operating domain and the samples in the target load operating domain, and then merge the various differences to obtain the cross-load migration sample weights of the source load operating domain relative to the target load operating domain.

[0049] Specifically, the calculation process for each difference includes: Based on the boiler load and load change rate in the operating characteristics, the load difference degree is obtained by normalization comparison. In the calculation, the average load, median load, representative steady-state load or representative load change rate of the two types of load conditions can be obtained first within the preset statistical period. Then, the difference in load level and the difference in load rise and fall trend between the two types of load conditions are normalized and compared to reflect the difference between the two types of load conditions in load level and load change trend.

[0050] Based on the fuel quantity, air volume at each stage, flue gas oxygen content, burner operation status, coal mill operation status, and soot blowing status in the operating characteristics, the combustion state difference is obtained through state consistency comparison. During the calculation, the continuous combustion operation parameters can be normalized and compared, and the state consistency comparison can be performed on the burner operation combination and the coal mill operation combination. If the combustion organization mode of the two types of load conditions is similar, the combustion state difference is small; otherwise, the difference increases, reflecting the difference in combustion organization mode between the two types of conditions.

[0051] Based on the operating parameters of the steam and water side in the operating characteristics, the deviation of the heat exchange boundary conditions is compared after normalization to obtain the steam and water side operating difference degree. In the calculation, the dimensions of each steam and water side operating parameter can be unified and normalized first, and then the deviation of the two types of load conditions on the heat exchange boundary conditions of the steam and water side can be compared to reflect the difference of the steam and water side heat exchange conditions of the two types of load conditions.

[0052] Based on the regional average temperature, regional extreme temperature, regional temperature gradient, spatial temperature difference, area ratio of high-temperature regions, and centroid position of high-temperature regions in the temperature field morphology characteristics, the degree of deviation is compared to obtain the temperature field morphology difference. During the calculation, the morphological characteristics corresponding to the typical temperature field, average temperature field, or steady-state temperature field under the two types of load conditions can be extracted and the degree of deviation can be compared. If the overall temperature level, local high-temperature region position, and spatial gradient change trend of the two types of load conditions are similar, the difference is small. If the distribution of high-temperature regions, temperature gradient, or temperature field centroid shift is significant, the difference increases, reflecting the difference in the spatial structure of the temperature field under the two types of load conditions.

[0053] Based on the characteristics of the temperature field morphology, the consistency of the response trend of the temperature field with changes in boiler load is compared to obtain the load response difference degree. During the calculation, the direction of temperature field change, the magnitude of temperature change, the response lag time, and the center of gravity migration trend of the temperature field can be compared under two types of load conditions. If the temperature response trends of the two types of load conditions are consistent with the load change, the difference degree is small. If there are inconsistent temperature response directions, large differences in response magnitude, or significantly different response lag characteristics, the difference degree increases, reflecting the difference in the response law of the temperature field with changes in load under the two types of conditions.

[0054] The above five differences are combined and calculated to obtain the cross-load migration sample weights of the source load case domain relative to the target load case domain. The calculation formula is as follows:

[0055] In the formula, Indicates the first Cross-load migration sample weights of each source load case domain relative to the target load case domain; This indicates the number of source load case domains participating in the transfer training. Indicates the degree of load variation; Indicates the degree of difference in combustion state; Indicates the difference in operation between the steam and water sides; Indicates the degree of difference in temperature field morphology; Indicates the degree of load response difference; , , , , These represent the corresponding adjustment coefficients, which can be obtained based on historical cross-load verification samples, expert experience, or model training verification error calibration. This represents the zero-limit constant, used to avoid the denominator from being zero.

[0056] S5. Based on the cross-load migration sample weights, extract the corresponding training samples from the source load condition domain to construct the migration training sample set.

[0057] The process of constructing the transfer training sample set is as follows: Figure 3 As shown, the specific process is as follows: S51. Compare the cross-load migration sample weights corresponding to each source load condition domain with the preset migration weight threshold, and extract the source load condition domains with weights greater than the preset migration weight threshold as transferable source load condition domains; the migration weight threshold can be set to 0.08 to 0.20 according to actual training needs, or the source load condition domains with the top 3 to top 6 migration weights can be selected to participate in training.

[0058] S52. Normalize the weights of cross-load migration samples corresponding to all transferable source load condition domains to obtain the data proportion of each transferable source load condition domain in the migration training sample set.

[0059] S53. Based on the proportion of each data, extract the corresponding proportion of sample data from the corresponding transferable source load condition domain, and combine it with the existing measured samples in the target load condition domain to form a transfer training sample set.

[0060] S6. Based on the spatial measurement point features, construct the spatial adjacency relationship of the measurement points to constrain the spatial gradient of the temperature field, and use the measurement point credibility features as the calculation weight of the single-point prediction error. Use the transfer training sample set to iteratively train the initial water-cooled wall temperature field transfer model to obtain a boiler water-cooled wall temperature field prediction model suitable for the target load operating domain.

[0061] Specifically, a five-layer network structure can be used to build the initial water-cooled wall temperature field migration model. The first layer is the operating condition feature encoding layer, used to process operating condition features such as boiler load, load change rate, air-coal ratio, flue gas oxygen content, steam-water side pressure, and soot blowing status. The second layer is the spatial measuring point encoding layer, used to process spatial features such as measuring point furnace wall position, tube panel number, height number, lateral number, and adjacent measuring point number. The third layer is the temperature field morphology encoding layer, used to process morphological features such as regional average temperature, regional maximum temperature, temperature difference between upper and lower layers, temperature difference between left and right walls, area ratio of high-temperature region, and centroid position of high-temperature region. The fourth layer is the feature fusion layer, used to fuse operating condition features, spatial features, and temperature field morphology features. The fifth layer is the temperature field output layer, used to output the predicted temperature corresponding to each measuring point under the target load condition. In actual implementation, the water-cooled wall temperature field migration model can also be implemented using a multilayer perceptual network, graph neural network, temporal convolutional network, or tree model integrated structure.

[0062] The model training process aims to minimize the total training loss by iteratively updating the model parameters. The total training loss of the water-cooled wall temperature field transfer model is... The calculation formula is:

[0063] in, This represents the transfer learning loss corresponding to the source load condition domain data. This represents the correction loss corresponding to the measured data in the target load condition domain. This represents the loss of consistency of the spatial gradient of the temperature field. , , These are the adjustment coefficients corresponding to each loss term, which can be set based on historical training results, validation errors, or expert experience.

[0064] The formula for calculating the transfer learning loss corresponding to the source load domain data is:

[0065] In the formula, This indicates the number of transferable source load case domains participating in the transfer training. Indicates the first Weights of cross-load migration samples corresponding to each source load condition domain. Indicates the first The number of data points used in training within each source load condition domain This indicates the number of temperature measurement points on the water-cooled wall. Indicates the first In the individual load condition domain, the first Group of data corresponds to the first The reliability of each measuring point This represents the predicted temperature output by the model. This indicates the corresponding measured temperature.

[0066] The formula for calculating the correction loss corresponding to the measured data of the target load condition domain is as follows:

[0067] In the formula, This indicates the number of existing sample data points within the target load condition domain. Indicates the first in the target load condition domain Group of data corresponds to the first The reliability of each measuring point The model represents the first time under the target load condition. The predicted temperature output from each measuring point. Indicates the first under the target load condition The measured temperature at each measuring point.

[0068] The formula for calculating the spatial gradient consistency loss of the temperature field is:

[0069] In the formula, This represents a pre-established set of adjacency relationships for measurement points based on the spatial arrangement of measurement points on the water-cooled wall, which is used to calculate the spatial gradient of the temperature field in all training samples. Indicates the first The measuring point and the first Each measuring point is a spatially adjacent measuring point. This represents the temperature difference between adjacent measuring points in the model's predicted temperature field. It represents the temperature difference between adjacent measuring points in the measured temperature field.

[0070] The model training can employ batch training, with each batch containing 64 input samples and a maximum of 120 iterations. The initial learning rate can be set to 0.001. If the validation error decreases by less than 0.2% for 10 consecutive iterations, the learning rate is reduced to 50% of its original value. If the validation error decreases by less than 0.1% for 20 consecutive iterations, training is stopped prematurely. During training, the training contribution of samples from the source load domain is determined based on the migration weight of their respective load domains. The correction weight of existing measured samples in the target load domain can be set to 1.8 times the average training weight of the source load domain. Measurement data with a confidence level below 0.60 are not used as primary training labels but only as spatial adjacency reference data for temperature field structure constraints. Training is stopped when the total training loss is less than a preset loss threshold, or when the change in validation error after multiple iterations is less than a preset convergence threshold, and the final boiler water-cooled wall temperature field prediction model is output.

[0071] S7. Verify the boiler water-cooled wall temperature field prediction model based on the measured temperature field data under the target load condition, and update the cross-load migration sample weights, measurement point credibility features, or model parameters when the verification error exceeds the preset threshold.

[0072] Specifically, model validation metrics include the average prediction error of the temperature field, the prediction error of the highest temperature point, the prediction error of the temperature gradient, the consistency of high-temperature region identification, and the centroid shift error of the temperature field. When any validation metric exceeds the corresponding preset threshold, the weights of the cross-load migration samples can be recalculated, or the adjustment coefficient of the correction loss of the target load condition samples can be increased, or the weights of samples corresponding to low-confidence measurement points can be reduced during the training process. The water-cooled wall temperature field migration model can then be retrained to achieve dynamic optimization and continuous adaptation of the model.

[0073] After completing the above model training, real-time temperature field prediction can be carried out based on the obtained boiler water-cooled wall temperature field prediction model: real-time acquisition of boiler operating condition data under the current target load condition, and extraction of the measurement point reliability characteristics of the current measurement point; inputting the boiler operating condition data into the pre-deployed boiler water-cooled wall temperature field prediction model to obtain the water-cooled wall temperature field prediction results under the current target load condition; based on the water-cooled wall temperature field prediction results and the measurement point reliability characteristics, identifying local high-temperature risk areas of the water-cooled wall for boiler operation safety monitoring.

[0074] This solution achieves accurate classification of samples from different operating states through joint operating condition domain partitioning, avoiding training interference caused by mixed data from different operating conditions at the data level. By using multi-dimensional difference quantification and cross-load migration sample weight calculation, it can select source domain data with higher similarity to the target operating condition for training, effectively suppressing negative migration problems caused by dissimilar operating condition data. Through measurement point reliability weighting and temperature field spatial gradient consistency constraints, it reduces the interference of abnormal measurement point data on model training while ensuring the rationality of the spatial distribution of the predicted temperature field. Combined with model validation and dynamic update mechanisms, it can continuously adapt to long-term changes in boiler operating conditions. Overall, this solution can fully utilize historical multi-operating condition data to complete high-precision model training in scenarios with insufficient target load operating condition samples, effectively improving the prediction accuracy, spatial distribution accuracy, and operating condition adaptability of the water-cooled wall temperature field, providing reliable data support for boiler water-cooled wall safety monitoring.

[0075] To verify the practical application effect of this scheme, an engineering verification was conducted using a subcritical natural circulation coal-fired boiler with a rated evaporation capacity of 670 t / h. Water-cooled wall panels were installed around the boiler furnace. Temperature monitoring zones were set up on the front, rear, left, and right walls. Each furnace wall was divided into a lower burner-affected zone, a central main heat exchange zone, and an upper outlet flue gas temperature-affected zone along its height. Each zone had at least 24 temperature measuring points, for a total of 384 water-cooled wall temperature measuring points throughout the boiler. These measuring points used a combination of K-type armored wall-mounted temperature elements and FBG-1550 fiber optic grating temperature elements. Temperature data was accessed via a TMC-32 temperature acquisition terminal at the edge. The computing station has a sampling period of 10 seconds. Boiler operating condition data are collected synchronously by the DCS-6000 distributed control system, BCS-300 combustion control system, SCS-200 soot blowing control system, and WSS-100 steam-water side status monitoring unit. The operating data sampling period is set to 5 seconds. After collection, the data is uniformly written into the cross-load operating condition sample library through the OPC-200 data communication gateway. Temperature data and operating data are aligned with a unified timestamp, and the time alignment error is controlled within 30 seconds. This verification continuously collected 120 days of historical operating data, with a total of 172,800 valid operating samples. After removing samples from boiler shutdowns, sensor communication interruptions, burner start-up and shutdown transient anomalies, and missing DCS records, 145,260 valid samples were retained. The samples were divided into four load ranges according to the boiler's rated load: 45% to 60% low load, 60% to 75% medium-low load, 75% to 90% medium-high load, and 90% to 100% high load. Based on the load change rate and equipment operating status, a secondary labeling was performed, forming 36 combined load condition domains. Among them, the condition domain with more than 3,000 historical samples and a temperature field standard deviation of less than 18℃ was determined as the source load condition domain, and the condition domain with less than 800 historical samples or the condition domain corresponding to recently added operating strategies was determined as the target load condition domain. This verification selected the load increase condition of 60%~75% of rated load as the target load condition domain. The burner operation status was the combined operation of the lower and middle layer burners, the coal mill operation status was four units in operation, and the soot blowing status was within 120 minutes after soot blowing. There are already 620 sets of valid measured samples in this target load condition domain, which cannot support the training of the stable model alone. Therefore, candidate source load condition domains were selected from the 45% to 60% low load increase condition domain, the 75% to 90% medium and high load increase condition domain, the 60% to 75% stable load condition domain, and the 60% to 75% load decrease condition domain.The migration weight threshold was set to 0.12. Calculations showed that the migration weight for the 45% to 60% low load increase load condition was 0.18, for the 75% to 90% medium-high load increase load condition was 0.31, for the 60% to 75% stable load condition was 0.27, and for the 60% to 75% load decrease load condition was 0.09. Therefore, the first three load conditions were considered as transferable source load conditions, while the 60% to 75% load decrease load condition was excluded from the migration training sample set. Subsequently, the migration weights of the three transferable source load case domains were normalized, and training data were extracted from the corresponding historical samples according to the normalization results. Specifically, 2,600 samples were extracted from the 45% to 60% low load increase case domain, 4,500 samples were extracted from the 75% to 90% medium-high load increase case domain, and 3,900 samples were extracted from the 60% to 75% stable load case domain. Then, the existing 620 measured samples from the target load case domain were added, resulting in a total of 11,620 migration training samples. The model training adopts a batch training method, with each batch containing 64 input samples. The maximum number of iterations is set to 120, and the initial learning rate is set to 0.001. If the validation error decreases by less than 0.2% for 10 consecutive iterations, the learning rate is reduced to 50% of the original rate. If the validation error decreases by less than 0.1% for 20 consecutive iterations, training is stopped early. The correction weights for existing measured samples in the target load domain are set to 1.8 times the average training weights of the source load domain. Measurement point data with a confidence level below 0.60 are not used as the main training labels, but only as spatial adjacency reference data.

[0076] Three control schemes were set up for comparative verification: Comparative Scheme 1 used a single target load condition training method, using only 620 sets of existing measured samples in the target load condition domain to train the temperature field prediction model, without using data from the source load condition domain; Comparative Scheme 2 used a common mixed training method, directly mixing historical samples from multiple load condition domains to train the model, without calculating cross-load migration sample weights or introducing measurement point reliability screening; this embodiment uses cross-load migration sample weight screening, constructing a migration training sample set according to weight ratios, measurement point reliability correction, and temperature field spatial gradient consistency constraints. All three schemes under the same target load condition underwent 30 days of online playback verification and 15 days of on-site tracking verification. The comparison results of various indicators are as follows: Figure 4 , Figure 5As shown. The verification results show that, in Comparative Example 1, the average prediction error of the temperature field is 9.8℃, the prediction error of the highest temperature point is 21.6℃, the average error of the temperature gradient is 6.4℃ / m, the consistency of high-temperature region identification is 72.5%, and the temperature field centroid shift error is 2.8m; in Comparative Example 2, the average prediction error of the temperature field is 7.1℃, the prediction error of the highest temperature point is 16.9℃, the average error of the temperature gradient is 4.7℃ / m, the consistency of high-temperature region identification is 81.3%, and the temperature field centroid shift error is 1.9m; ​​in this embodiment, the average prediction error of the temperature field is 4.3℃, the prediction error of the highest temperature point is 9.8℃, the average error of the temperature gradient is 2.6℃ / m, the consistency of high-temperature region identification is 91.7%, and the temperature field centroid shift error is 0.8m.

[0077] The 18 high-temperature risk areas identified during the verification period were manually verified. The verification methods included portable contact temperature measurement, infrared temperature measurement window observation, and wall thickness sampling during shutdown maintenance. Comparative Example 1 correctly identified 12 high-temperature risk areas, misidentified 5 areas, and missed 6 areas; Comparative Example 2 correctly identified 15 high-temperature risk areas, misidentified 3 areas, and missed 3 areas; This embodiment correctly identified 17 high-temperature risk areas, misidentified 1 area, and missed 1 area. Moreover, the high-temperature areas identified in this embodiment have a good correspondence with the areas of local slagging, tube screen heat absorption deviation, and delayed temperature recovery after soot blowing found during shutdown maintenance.

[0078] A specific verification was conducted to test the effectiveness of the measurement point reliability correction. In the verification sample, 42 sets of measurement point jump samples, 36 sets of measurement point drift samples, and 28 sets of local communication anomaly samples were artificially retained. Without measurement point reliability correction, the model's average prediction error increased from 4.3℃ to 6.9℃, and the prediction error of the highest temperature point increased from 9.8℃ to 14.2℃. After applying the measurement point reliability correction in this embodiment, the model's average prediction error remained within 4.6℃, and the prediction error of the highest temperature point remained within 10.5℃, proving that the measurement point reliability mechanism can effectively reduce the interference of abnormal measurement point data on model training and prediction results.

[0079] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used for the operation of a training method for a boiler water-cooled wall temperature field prediction model.

[0080] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be Random Access Memory (RAM) or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the training method for a boiler water-cooled wall temperature field prediction model in the above embodiments.

[0081] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0082] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0085] This invention also provides a computer program product for executing the training method for a boiler water-cooled wall temperature field prediction model described above. Since the computer program product provided by this invention and the training method for a boiler water-cooled wall temperature field prediction model described above belong to the same inventive concept, the computer program product provided by this invention possesses all the advantages of the training method for a boiler water-cooled wall temperature field prediction model described above. Therefore, the beneficial effects of the computer program product provided by this invention will not be elaborated upon here.

[0086] In this invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0087] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention.

Claims

1. A training method for a boiler water-cooled wall temperature field prediction model, characterized in that, include: Acquire water-cooled wall temperature field data and corresponding boiler operating condition data under different load conditions, and construct a cross-load condition sample library; Based on the load parameters and equipment operating status parameters in the boiler operating condition data, the training samples in the cross-load operating condition sample library are jointly divided into operating condition domains to determine the source load operating condition domain and the target load operating condition domain. The operating characteristics, spatial measurement point characteristics, and temperature field morphology characteristics of each load condition domain are extracted respectively, and the measurement point reliability characteristics used to characterize the quality of temperature sensor data are extracted. Based on operational characteristics and temperature field morphology characteristics, the load difference, combustion state difference, steam-water side operation difference, temperature field morphology difference, and load response difference between samples in the source load operating condition domain and samples in the target load operating condition domain are calculated respectively. The difference values ​​are then fused together to obtain the cross-load migration sample weights of the source load operating condition domain relative to the target load operating condition domain. Based on the cross-load migration sample weights, corresponding training samples are extracted from the source load condition domain to construct a migration training sample set; Based on the spatial characteristics of the measuring points, the spatial adjacency relationship of the measuring points is constructed to constrain the spatial gradient of the temperature field. The credibility characteristics of the measuring points are used as the weight for calculating the single-point prediction error. The initial water-cooled wall temperature field transfer model is iteratively trained using the transfer training sample set to obtain a boiler water-cooled wall temperature field prediction model suitable for the target load operating domain. The boiler water-cooled wall temperature field prediction model is validated based on the measured temperature field data under the target load condition, and the cross-load migration sample weights, measurement point confidence features, or model parameters are updated when the validation error exceeds the preset threshold.

2. The training method for the boiler water-cooled wall temperature field prediction model according to claim 1, characterized in that, The load parameters include the boiler load range and load change rate, and the equipment operating status parameters include the burner commissioning status, coal mill operating status, and soot blowing status. The step of performing joint load condition domain partitioning on training samples in the cross-load condition sample library to determine the source load condition domain and the target load condition domain includes: According to the proportion of boiler load to rated load, the training samples in the cross-load condition sample library are divided into different load ranges; Based on the stability of the load change rate in the sample, the samples within the same load range are divided into sample subsets with different load change trends. The sample subsets specifically include stable load samples, load increase samples, and load decrease samples. For stable load samples, increasing load samples, and decreasing load samples, operating conditions are marked based on their corresponding burner commissioning status, coal mill operating status, and soot blowing status to form a joint operating condition domain. The joint operating condition domain with a sufficient number of historical samples and a stable temperature field distribution corresponding to the samples is defined as the source load operating condition domain, and the joint operating condition domain with insufficient measured samples and the need to train the temperature field prediction model is defined as the target load operating condition domain.

3. The training method for the boiler water-cooled wall temperature field prediction model according to claim 1, characterized in that, The operating characteristics include boiler load, load change rate, fuel quantity, air volume at each stage, flue gas oxygen content, steam and water side operating parameters, burner commissioning status, coal mill operating status, and soot blowing status. The temperature field morphology features include regional average temperature, regional extreme temperature, regional temperature gradient, spatial temperature difference, area ratio of high-temperature regions, and centroid location of high-temperature regions. The spatial measuring point characteristics include the location of the measuring point on the furnace wall, the tube panel number, the spatial coordinates, and the area affected by the equipment to which it belongs; The reliability characteristics of the measuring point are composed of the basic reliability of the measuring point, the reliability of the measuring point jump, the reliability of the measuring point drift, and the reliability of the spatial continuity of the measuring point.

4. The training method for the boiler water-cooled wall temperature field prediction model according to claim 1, characterized in that, The method involves fusing and calculating the differences among various parameters to obtain the cross-load migration sample weights relative to the target load condition domain. The calculation formula is as follows: in, Indicates the first Cross-load migration sample weights of each source load case domain relative to the target load case domain; This indicates the number of source load case domains participating in the transfer training. Indicates the degree of load variation; Indicates the degree of difference in combustion state; Indicates the difference in operation between the steam and water sides; Indicates the degree of difference in temperature field morphology; Indicates the degree of load response difference; , , , , These represent the corresponding adjustment coefficients; This represents the zero minimum constant.

5. The training method for the boiler water-cooled wall temperature field prediction model according to claim 3, characterized in that, The calculation process for the load difference, combustion state difference, steam-water side operation difference, temperature field morphology difference, and load response difference includes: Based on the boiler load and load change rate in the operating characteristics, the load difference degree is obtained by normalization comparison; Based on the fuel quantity, air volume at each stage, flue gas oxygen content, burner commissioning status, coal mill operating status, and soot blowing status in the operating characteristics, the combustion status difference is obtained by comparing the consistency of the status. Based on the operating parameters of the steam and water sides in the operating characteristics, the degree of deviation of the heat transfer boundary conditions is compared after normalization to obtain the operating difference degree of the steam and water sides. Based on the regional average temperature, regional extreme temperature, regional temperature gradient, spatial temperature difference, area ratio of high temperature region, and centroid position of high temperature region in the temperature field morphology characteristics, the degree of deviation is compared to obtain the temperature field morphology difference. Based on the characteristics of the temperature field morphology, the consistency of the response trend of the temperature field with changes in boiler load is compared to obtain the load response difference degree.

6. The training method for the boiler water-cooled wall temperature field prediction model according to claim 1, characterized in that, The step of extracting corresponding training samples from the source load domain based on cross-load migration sample weights to construct a migration training sample set includes: The cross-load migration sample weights corresponding to each source load condition domain are compared with the preset migration weight threshold, and the source load condition domains with weights greater than the preset migration weight threshold are extracted as the migrateable source load condition domains. Normalize the weights of cross-load migration samples corresponding to all transferable source load case domains to obtain the data proportion of each transferable source load case domain in the migration training sample set. Based on the proportion of each data point, sample data of a corresponding proportion are extracted from the corresponding transferable source load condition domain and combined with the existing measured samples in the target load condition domain to form a transfer training sample set.

7. The training method for the boiler water-cooled wall temperature field prediction model according to claim 3, characterized in that, The initial water-cooled wall temperature field transfer model is iteratively trained using a transfer training sample set, and the total training loss of the water-cooled wall temperature field transfer model is... The calculation formula is: in, This represents the transfer learning loss corresponding to the source load condition domain data. This represents the correction loss corresponding to the measured data in the target load condition domain. This represents the loss of consistency of the spatial gradient of the temperature field. , , These are the adjustment coefficients corresponding to each loss term; This indicates the number of transferable source load condition types participating in the transfer training. Indicates the first The cross-load condition type migration weight corresponding to each source load condition type. Indicates the first The number of data points used in training for each source load condition type. This indicates the number of temperature measurement points on the water-cooled wall. Indicates the first In the individual load condition type, the first Group of data corresponds to the first The reliability of each measuring point This represents the predicted temperature output by the model. This indicates the corresponding measured temperature; This indicates the number of existing sample data under the target load condition type. Indicates the first in the target load condition type Group of data corresponds to the first The reliability of each measuring point The model represents the first time under the target load condition. The predicted temperature output from each measuring point. Indicates the first under the target load condition Measured temperature at each measuring point; This represents a pre-established set of adjacency relationships for measurement points based on the spatial arrangement of measurement points on the water-cooled wall, which is used to calculate the spatial gradient of the temperature field in all training samples. Indicates the first The measuring point and the first Each measuring point is a spatially adjacent measuring point. This represents the temperature difference between adjacent measuring points in the model's predicted temperature field. It represents the temperature difference between adjacent measuring points in the measured temperature field.

8. The training method for the boiler water-cooled wall temperature field prediction model according to claim 3, characterized in that, The formula for calculating the reliability feature of the measurement point is: in, Indicates the first In the individual load condition type, the first Group of data corresponds to the first The reliability of each measuring point; To establish the reliability of the measurement points; The reliability of sudden jumps in measurement points is determined based on the temperature change amplitude between adjacent sampling times; The reliability of the measurement point drift is determined based on the degree to which it deviates from the mean of adjacent measurement points in the same area within a time window. The reliability of spatial continuity of the measurement point is determined based on the consistency of the temperature change trend between adjacent measurement points with the influence area of ​​the equipment to which the measurement point belongs in the spatial measurement point characteristics.

9. A method for predicting the temperature field of a boiler water-cooled wall, characterized in that, include: Real-time acquisition of boiler operating condition data under the current target load condition, and extraction of the reliability characteristics of the current measuring point; The boiler operating condition data is input into the pre-deployed boiler water-cooled wall temperature field prediction model to obtain the water-cooled wall temperature field prediction results of the boiler under the current target load condition. Based on the predicted temperature field of the water-cooled wall and the reliability characteristics of the measuring points, local high-temperature risk areas of the water-cooled wall are identified for boiler operation safety monitoring. The boiler water-cooled wall temperature field prediction model is obtained by transfer learning training on a historical cross-load condition sample database using the training method described in any one of claims 1 to 8.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a training method for a boiler water-cooled wall temperature field prediction model as described in any one of claims 1 to 8.

Images