A boiler water wall temperature prediction system and method for a dynamic variable load process

Abstract

The application discloses a boiler water-cooled wall temperature prediction system and method for a dynamic variable load process, and belongs to the technical field of industrial intelligent computing, comprising a data acquisition and standardization module, a physical wall temperature calculation module, a sample database module, a convolutional autoencoder spatial feature extraction module, a GRU time series prediction module, a result output module and a checking module connected with the physical wall temperature calculation module and the GRU time series prediction module in sequence. The application solves the problems that the traditional measuring point monitoring in the existing coal-fired boiler variable load process is difficult to reflect the overall wall temperature change of the water-cooled wall surface, and the traditional computational fluid dynamics calculation is time-consuming and difficult to guide the power production site in real time, and realizes the rapid prediction of the water-cooled wall temperature field and the over-temperature risk under unknown dynamic conditions.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent computing technology, specifically to a boiler water-cooled wall temperature prediction system and method for dynamic load changes. Background Technology

[0002] Under the overarching context of "dual carbon" (carbon diversification and energy conservation), with the increasing proportion of renewable energy power generation, coal-fired units are undertaking more tasks related to deep peak shaving and rapid load changes. During dynamic load changes, boiler operating conditions such as fuel quantity, air supply volume, burner operation mode, and feedwater parameters change continuously, leading to unsteady changes in furnace combustion organization, flue gas temperature field, and heat flux density on the flue gas side of the water-cooled walls. As a crucial component of the boiler's evaporative heating surface, the water-cooled wall's wall temperature level directly affects tube wall strength, overheating risk, and safe unit operation.

[0003] Currently, power plant sites primarily rely on temperature monitoring points located locally on the water-cooled walls to monitor wall temperature. However, water-cooled walls have a large area, numerous tube panels, and uneven heat load distribution, making it difficult for a limited number of monitoring points to reflect the spatial distribution of the overall furnace water-cooled wall temperature and the hot spot migration process. Computational fluid dynamics (CFD) methods can simulate furnace combustion, flue gas flow, and heat flux density distribution on heated surfaces in relatively detailed terms. However, if a complete transient CFD calculation is performed for every moment of a dynamic load change process and repeatedly coupled and iterated with the water-side heat transfer model, the computational load is large and time-consuming, making it difficult to meet the needs of extensive operating condition analysis and rapid on-site prediction.

[0004] Meanwhile, existing data-driven methods can establish rapid prediction models using historical temperature fields or operating parameters. However, directly predicting high-dimensional wall temperature fields over time is difficult, and it easily overlooks physical factors such as water-cooled wall heat flux input, water-side enthalpy propagation, working fluid property changes, and tube wall thermal inertia. Therefore, it is necessary to propose a boiler water-cooled wall temperature prediction system and method for dynamic load changes, combining heat flux density field, physical heat transfer calculation, and deep learning prediction to achieve rapid prediction of water-cooled wall temperature field and over-temperature risk under unknown dynamic operating conditions. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a boiler water-cooled wall temperature prediction system and method for dynamic load change processes, so as to solve the problems that traditional measuring point monitoring is difficult to reflect the overall wall temperature change of the water-cooled wall surface during the load change process of existing coal-fired boilers, and that traditional computational fluid dynamics calculation is time-consuming and cannot provide real-time guidance for power production sites, so as to realize the rapid prediction of water-cooled wall temperature field and over-temperature risk under unknown dynamic operating conditions.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows.

[0007] A boiler water-cooled wall temperature prediction system for dynamic load changes includes a data acquisition and standardization module, a physical wall temperature calculation module, a sample database module, a convolutional autoencoder spatial feature extraction module, a GRU time series prediction module, a result output module, and a verification module connected in sequence to the physical wall temperature calculation module and the GRU time series prediction module. The data acquisition and standardization module is used to acquire the wall calculation results and operation control parameters obtained by the transient computational fluid dynamics method, and to form a dynamic heat flux density field with area conservation. The physical wall temperature calculation module is used to calculate the dynamic wall temperature field based on the heat absorption power of the heat-receiving calculation unit and the water-side conservation principle. The sample database module is used to store dynamic wall temperature fields and operating control parameters; The convolutional autoencoder spatial feature extraction module is used to encode the two-dimensional wall temperature field to obtain a low-dimensional feature vector, and to decode the low-dimensional feature vector to restore the two-dimensional wall temperature field. The GRU time-series prediction module is used to predict low-dimensional feature vectors at future times based on low-dimensional feature vectors and running control parameters. The result output module is used to output at least one of the following: maximum wall temperature, over-temperature risk, and wall temperature change trend; The verification module is used to verify the error of the physical wall temperature calculation or prediction results under representative working conditions.

[0008] A method for predicting boiler water-cooled wall temperature under dynamic load changes includes the following steps: S1. Data Acquisition and Standardization: Acquire the transient computational fluid dynamics (CFD) wall calculation results and operating control parameters of the target boiler during dynamic load change process; expand the CFD wall mesh to a two-dimensional coordinate system according to the furnace wall and geometric position, and interpolate the operating control parameters to a time series consistent with the transient CFD wall calculation results; establish a regular canvas mesh, convert the CFD wall mesh heat flux density to the regular mesh through weighted mapping, and obtain the corrected regular mesh heat flux density after area conservation correction; S2. Division of Heating Calculation Units and Calculation of Heat Absorption Power: Based on the arrangement of water-cooled wall tubes, the location of the furnace wall, and the direction of working fluid flow, the two-dimensional region of the water-cooled wall is divided into multiple water-side heating calculation units connected sequentially along the direction of working fluid flow. Each water-side heating calculation unit covers one or more regular grid units and serves as a discrete control volume for the one-dimensional finite volume water-side state progression. The heat absorption power of the heating calculation unit is calculated by integrating the area of ​​the heat flux density of the regular grid in the i-th heating calculation unit. S3. Dynamic wall temperature field calculation: Using the heat absorption power as the thermal boundary on the flue gas side, the water side state is advanced along the working fluid flow direction using the one-dimensional finite volume method. Based on the conservation of mass and energy, the working fluid enthalpy, working fluid temperature and thermophysical parameters of each heat-receiving calculation unit are updated. The convective heat transfer coefficient inside the pipe is calculated. Combined with the convective heat transfer thermal resistance on the water side and the thermal conductivity thermal resistance of the pipe wall, the quasi-steady-state target wall temperature is obtained. Then, considering the thermal inertia caused by the heat capacity of the pipe wall, the actual wall temperature is iteratively updated to obtain the dynamic wall temperature field. S4. Sample Database Construction: The dynamic wall temperature field and the operating control parameters are aligned in time to construct a dynamic wall temperature sample database, and deep learning training samples are formed according to the sliding time window; a convolutional autoencoder is used to encode the two-dimensional wall temperature field to obtain a low-dimensional feature vector representing the spatial temperature distribution and local hot spot features; the convolutional autoencoder includes an encoder and a decoder, and the encoder reduces the two-dimensional wall temperature field to a one-dimensional feature vector through a convolution algorithm; S5. Time Series Prediction and Result Output: Input the low-dimensional feature vector and operation control parameters into the GRU time series prediction model to predict the low-dimensional feature vector at future time. Then, reconstruct the water-cooled wall temperature field under future dynamic variable load conditions through the decoder. Traverse the predicted wall temperature of each regular grid to obtain the maximum wall temperature and its corresponding location. Compare the predicted wall temperature with the preset over-temperature threshold to determine the over-temperature area and over-temperature risk. Determine the wall temperature change trend based on the change in wall temperature or maximum wall temperature at adjacent future time. Output the results.

[0009] Preferably, in step S1, the transient computational fluid dynamics wall calculation results are obtained by solving the mass conservation, momentum conservation, energy conservation and component transport equations and derived from the water-cooled wall boundary after establishing a three-dimensional geometric model of the target boiler, dividing the CFD wall grid f and the volume grid, and applying time-varying fuel, air volume and heat transfer surface boundary conditions. The results include the center coordinates, area, flue gas side heat flux density and CFD time step of each CFD wall grid f. The operating control parameters are derived from the target boiler DCS historical records or actual power plant operating instructions, including unit load, load change rate, burner operation status, feedwater flow rate, feedwater temperature, and feedwater pressure. The specific method of interpolation is as follows: The sampling time of the operating control parameters is recorded as... The output time of the transient computational fluid dynamics wall calculation results is denoted as... Because the time step of the transient numerical simulation is inconsistent with the time step of the actual acquired data, the operating control parameters... When satisfied At that time, linear interpolation was used to obtain the operating control parameters. value of time This ensures that the CFD heat flux density field at each moment has the same timestamp as the corresponding operating control parameters.

[0010] Preferably, in step S1, the regular grid unit The formula for calculating heat flux density is: in, For regular grid cells Heat flux density at time t; Let f be the area of ​​the CFD wall mesh; Let f be the heat flux density of the CFD wall mesh f at time t; For CFD wall mesh f, from center to regular mesh cell Distance from the center; The effective CFD neighborhood within the same wall partition is specifically represented as:

[0011] in, R Let R be the neighborhood radius, and R > 0; This indicates that the CFD wall mesh f and the regular cell mesh g are located on the same wall. The weight kernel function is specifically expressed as: Area conservation corrections were applied to the heat flux densities of the two grids: in, This represents the dynamic heat flux correction coefficient, which is a dimensionless parameter. Regular grid cells The area; This represents the thermal power of the CFD wall mesh f; Representing a regular grid Thermal power; Revised regular grid The heat flux density is specifically: in, Represents the corrected regular grid Heat flux density, used to calculate regular meshes The local wall temperature is used as the wall temperature field.

[0012] Preferably, in step S2, the formula for calculating the heat absorption power is: in, Ωi represents the heat absorption power of the i-th heated computing unit, which is only used for water-side energy conservation and working fluid state propulsion; Ωi represents the set of regular grid cell numbers covered by the i-th heated computing unit.

[0013] Preferably, in step S3, the specific method for advancing the water-side state is as follows: Along each water-cooled wall working fluid flow path, the water-side heating calculation units are numbered i=1,2,...,N, and the outlet enthalpy of the (i-1)th unit is used as the inlet enthalpy of the i-th unit using the upwind configuration. For each time step, the working fluid mass in the unit is first updated by mass conservation, then the heat absorption power on the flue gas side is included in the working fluid enthalpy by energy conservation, and finally the working fluid temperature and thermophysical parameters are retrieved based on the updated pressure and enthalpy. The pressure is obtained by interpolation of the inlet and outlet pressures of the water-cooled wall along the working fluid flow direction; the working fluid temperature and thermophysical parameters are obtained by bilinear interpolation based on pressure-enthalpy or pressure-temperature property tables. In step S3, the convective heat transfer coefficient inside the pipe The Nusselt number correlation is obtained based on the Reynolds number (Re), Prandtl number (Pr), and Nusselt number (Nu), and the Nusselt number correlation is selected according to the interval to which the Reynolds number belongs, indicating laminar, transitional, or turbulent flow; when Re i When 2300 < 2300, the laminar flow Nusselt number correlation is used; when 2300 ≤ Re i When <10000, a transitional flow correlation is used; when Re i When the value is ≥10000, the turbulent correlation equation is used to calculate the convective heat transfer coefficient inside the pipe. The formula for calculating the Nusselt number Nu is: in, denoted by the Nusselt number of the i-th water-side heating computational unit, dimensionless, used to characterize the strength of convective heat transfer; Represents the Reynolds number of the i-th heated calculation unit, which is dimensionless and used to determine the flow state of the working fluid; The Prandtl number represents the i-th heated computational unit. It is dimensionless and is used to characterize the ratio of the working fluid's momentum diffusion to its thermal diffusion capacity. convective heat transfer coefficient The formula for calculation is: in, This represents the convective heat transfer coefficient inside the water-side pipe of the i-th heating calculation unit at time t; The thermal conductivity of the working fluid; This refers to the inner diameter of the water-cooled wall tube.

[0014] Preferably, in step S3, the formula for calculating the quasi-steady-state target wall temperature is: in, This represents the quasi-steady-state target wall temperature at time t; This represents the temperature of the working fluid inside the i-th heated calculation unit at time t; Indicates the thickness of the pipe wall metal; This indicates the thermal conductivity of the steel used in the pipe wall; Considering the thermal inertia caused by the heat capacity of the pipe wall, the formula for iteratively updating the actual wall temperature is as follows: in, express The wall temperature at any given time; This indicates the time step used in the physical wall temperature calculation; This represents the wall temperature at time t. This represents the wall temperature at time t; The dimensionless dynamic response coefficient is specifically defined as follows: in, is the thermal inertia time constant.

[0015] Preferably, in step S5, the input to the GRU time series prediction model is ,in, z t This represents the true low-dimensional feature vector at the current moment. u t The operating control parameters are set for the current moment. The GRU timing prediction model sequentially calculates the reset gate. Update door Candidate hidden state and hidden state The low-dimensional feature prediction value for the next time step is obtained from the output layer; The GRU time-series prediction model employs an autoregressive rolling prediction method, using the low-dimensional feature vector predicted at the previous time step as the historical state input for the next time step. During training, the joint loss function is constructed from the low-dimensional feature prediction error and the wall temperature field error after decoding. in, Train a joint total loss function for the GRU time series prediction model; The loss weight coefficient has a value range of [0,1]; MSE is the mean squared error. This is the predicted low-dimensional feature vector output by the GRU time series prediction model; The predicted wall temperature field is obtained after the low-dimensional features are restored by the decoder. The target wall temperature field in the sample database; During training, the GRU time-series prediction model divides the training set and validation set according to the complete dynamic load change conditions to reduce the prediction bias caused by leakage in adjacent samples under the same conditions.

[0016] Preferably, the method further includes a verification step: for some representative dynamic variable load conditions, the dynamic wall temperature field obtained in step S3 is compared with the Fluent-UDF online coupling calculation results or the wall temperature measured at the field points based on the same boundary conditions to verify the error, which is used to correct the thermal inertia time constant. And the empirical parameters for heat transfer calculation in the Nusselt number correlation; The UDF is a wall temperature calculation program based on C language programming. At each time step, the UDF reads the heat flux density on the flue gas side of the water-cooled wall, calculates the working fluid state and wall temperature on the water side according to the feed water flow rate, temperature and pressure, and then returns the wall temperature to Fluent as the wall boundary, realizing bidirectional iteration of flue gas side heat flux and water side wall temperature; the online coupling calculation result of Fluent-UDF is the wall temperature field obtained by this bidirectional iteration process. The dynamic wall temperature field obtained in step S3 is compared with the Fluent-UDF coupled calculation results under the same boundary conditions, spatial location, and time point. The resulting error is used to correct the aforementioned thermal inertia time constant. And the empirical coefficients of the Nusselt number correlation.

[0017] Preferably, in step S5, the result is output in at least one form: a numerical value, a variation curve, or a two-dimensional wall temperature distribution map.

[0018] Due to the adoption of the above technical solutions, the technical progress achieved by this invention is as follows.

[0019] This invention combines physical mechanisms with data-driven approaches, balancing accuracy and speed: This invention calculates the dynamic wall temperature field using a transient CFD thermal boundary and a one-dimensional water-side physical heat transfer model, constructs a sample database, and then utilizes a convolutional autoencoder for spatial dimensionality reduction and a GRU for temporal prediction. This architecture ensures both the physical consistency and generalization ability of the training samples and enables rapid prediction of the wall temperature field, resolving the dual contradiction between the computationally intensive nature of pure CFD and the lack of physical interpretability in purely data-driven models.

[0020] The heat flux density field normalization in this invention employs area conservation correction to ensure physical consistency. This invention employs area-weighted interpolation when mapping unstructured CFD wall meshes to regular canvas meshes, and further introduces a dynamic heat flux correction coefficient for global area conservation correction, ensuring that the total heat power before and after interpolation is consistent, thus guaranteeing the accuracy of subsequent physical wall temperature calculations.

[0021] This invention introduces a dynamic response coefficient to physically model the thermal inertia of the pipe wall, closely resembling the actual heat transfer process: This invention introduces a dynamic response coefficient to perform a first-order inertial recursive update of the wall temperature, quantifies the hysteresis effect of the pipe wall heat capacity on the wall temperature change into a thermal inertia time constant, and corrects it through a verification module, so that the wall temperature calculation is closer to the actual physical process.

[0022] The verification module of this invention implements a closed-loop optimization of "physical calculation - data prediction - verification and correction", which ensures prediction accuracy. The verification module uses high-fidelity Fluent-UDF coupled calculation results or field measurement data to correct the thermal inertia time constant and Nusselt number correlation empirical coefficients in the physical wall temperature calculation module, forming a closed-loop optimization. This ensures that the physical calculation module continuously outputs high-fidelity samples, thereby guaranteeing the prediction accuracy of the deep learning model.

[0023] This invention effectively combines a convolutional autoencoder with a GRU, reducing model complexity while maintaining prediction accuracy. The convolutional autoencoder compresses the high-dimensional wall temperature field into a low-dimensional feature vector, which significantly reduces the number of parameters and training difficulty of the GRU time series prediction model; the joint loss function considers both the low-dimensional feature prediction error and the wall temperature field reconstruction error, thus ensuring prediction accuracy.

[0024] The prediction results of this invention are intuitive and practical, facilitating on-site safety monitoring. This invention can output the maximum wall temperature and its location, overheating area and risk level, wall temperature change trend, etc., and output them in at least one form, such as numerical values, change curves or two-dimensional wall temperature distribution maps, providing an intuitive technical reference for the safe operation of power plants during load changes. Attached Figure Description

[0025] Figure 1 This is a structural diagram of the prediction system module of the present invention.

[0026] Figure 2 This is a flowchart of the prediction method of the present invention.

[0027] Figure 3 This is a schematic diagram of the wall temperature prediction process combining a convolutional autoencoder and a GRU according to the present invention.

[0028] Figure 4 This is a schematic diagram illustrating the mapping relationship between CFD wall mesh elements and regular mesh elements in this invention.

[0029] Wherein: 1. CFD wall mesh element f, 2. Regular mesh element established in the two-dimensional coordinate system of the water-cooled wall. 3. Regular grid unit 4. The geometric center of the CFD wall mesh element f. Detailed Implementation

[0030] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0031] A boiler water-cooled wall temperature prediction system for dynamic load variations, combined with Figure 1 As shown, it includes a data acquisition and standardization module, a physical wall temperature calculation module, a sample database module, a convolutional autoencoder spatial feature extraction module, a GRU temporal prediction module, a result output module, and a verification module connected in sequence to the physical wall temperature calculation module and the GRU temporal prediction module, respectively.

[0032] The data acquisition and standardization module is used to acquire the wall calculation results and operation control parameters obtained by transient computational fluid dynamics methods, and to form a dynamic heat flux density field with area conservation.

[0033] The physical wall temperature calculation module is used to calculate the dynamic wall temperature field based on the heat absorption power of the heat-receiving calculation unit and the water-side conservation.

[0034] The sample database module is used to store dynamic wall temperature fields and operating control parameters.

[0035] The convolutional autoencoder spatial feature extraction module is used to encode the two-dimensional wall temperature field to obtain a low-dimensional feature vector, and then decode the low-dimensional feature vector to restore the two-dimensional wall temperature field.

[0036] The GRU time-series prediction module is used to predict low-dimensional feature vectors at future times based on low-dimensional feature vectors and runtime control parameters.

[0037] The results output module is used to output at least one of the following: maximum wall temperature, risk of overheating, and trend of wall temperature change.

[0038] The verification module is used to verify the error of the physical wall temperature calculation or prediction results under representative working conditions.

[0039] A method for predicting boiler water-cooled wall temperature in dynamic variable load processes, such as... Figure 2 As shown, it includes the following steps: S1. Data Acquisition and Standardization.

[0040] The transient computational fluid dynamics (CFD) wall calculation results and operating control parameters of the target boiler during dynamic load change are obtained respectively. The CFD wall mesh is expanded to a two-dimensional coordinate system according to the furnace wall and geometric position, and the operating control parameters are interpolated to a time series consistent with the transient CFD wall calculation results. A regular canvas mesh is established, and the heat flux density of the CFD wall mesh is converted to the regular mesh through weighted mapping. The corrected regular mesh heat flux density is obtained after area conservation correction.

[0041] In this step, the transient computational fluid dynamics wall calculation results are obtained by solving the mass conservation, momentum conservation, energy conservation and component transport equations and derived from the water-cooled wall boundary after establishing a three-dimensional geometric model of the target boiler, dividing the CFD wall mesh f and the volume mesh, and applying time-varying fuel, air volume and heat transfer surface boundary conditions. The results include the center coordinates, area, flue gas side heat flux density and CFD time step of each CFD wall mesh f.

[0042] The operating control parameters are derived from the target boiler's DCS historical records or the power plant's actual operating instructions, including unit load, load change rate, burner operation status, feedwater flow rate, feedwater temperature, and feedwater pressure.

[0043] The specific method of interpolation is as follows: The sampling time of the operating control parameters is recorded as... The output time of the transient computational fluid dynamics wall calculation results is denoted as... Since the time step of the transient numerical simulation and the time step of the actual acquired data may not be completely consistent, time series consistency interpolation is required for the operating control parameters. When satisfied At that time, linear interpolation was used to obtain the operating control parameters. value of time This ensures that the CFD heat flux density field at each moment has the same timestamp as the corresponding operating control parameters.

[0044] In this step, since the CFD wall mesh f is mostly unstructured and does not meet the requirements for input to the subsequent convolutional autoencoder spatial feature extraction module, a regular canvas mesh is established for the unfolded two-dimensional coordinate system, and for each regular mesh unit... Select the CFD wall mesh f in its neighborhood to calculate the regular mesh element. Heat flux density, calculated for regular mesh elements The heat flux density.

[0045] The regular mesh cell g is located within a pre-selected two-dimensional effective expanded computational domain of the water-cooled wall. This effective expanded computational domain can be composed of multiple wall partitions forming a rectangular parameter domain. Each wall partition corresponds to its respective furnace wall, and the boundaries between wall partitions are known internal coordinate boundaries, not constituting undefined blank areas. The regular mesh covers the effective expanded computational domain. For a regular mesh cell g, its effective CFD neighborhood within the same wall partition is defined as... :

[0046] in, R Let R be the neighborhood radius, and R > 0; This indicates that the CFD wall mesh f and the regular cell mesh g are located on the same wall.

[0047] By pre-determining the regular grid spacing and neighborhood radius R, each regular grid cell participating in subsequent calculations has at least one effective CFD wall grid; the heat flux density is calculated only for regular grid cells g that meet the above conditions; the method for calculating the heat flux density of the regular grid cell g is as follows: in, Regular grid cells Heat flux density at time t; Let f be the area of ​​the CFD wall mesh; Let f be the heat flux density of the CFD wall mesh f at time t; For CFD wall mesh f, from center to regular mesh cell Distance from the center; The weight kernel function is specifically expressed as:

[0048] Area conservation corrections were applied to the heat flux densities of the two grids: in, This represents the dynamic heat flux correction coefficient, which is a dimensionless parameter. Regular grid cells The area; This represents the thermal power of the CFD wall mesh f; Representing a regular grid Thermal power.

[0049] Revised regular grid The heat flux density is specifically: in, Represents the corrected regular grid Heat flux density, used to calculate regular meshes The local wall temperature is used as the wall temperature field.

[0050] S2. Division of heat-receiving calculation units and calculation of heat absorption power.

[0051] Based on the arrangement of water-cooled wall tubes, the location of the furnace wall, and the direction of working fluid flow, the two-dimensional region of the water-cooled wall is divided into multiple water-side heating calculation units connected sequentially along the direction of working fluid flow. Each water-side heating calculation unit covers one or more regular grid units and serves as a discrete control volume for the one-dimensional finite volume water-side state advancement. The heat absorption power of the heating calculation unit is calculated by integrating the area of ​​the heat flux density of the regular grid in the i-th heating calculation unit.

[0052] In this step, the formula for calculating the heat absorption power is: in, Ωi represents the heat absorption power of the i-th heated computing unit, which is only used for water-side energy conservation and working fluid state propulsion; Ωi represents the set of regular grid cell numbers covered by the i-th heated computing unit.

[0053] S3. Calculation of dynamic wall temperature field.

[0054] Using the heat absorption power as the thermal boundary on the flue gas side, the water-side state is advanced along the working fluid flow direction using a one-dimensional finite volume method. Based on the conservation of mass and energy, the working fluid enthalpy, working fluid temperature, and thermophysical parameters of each heated calculation unit are updated. The convective heat transfer coefficient inside the pipe is calculated. Combined with the convective heat transfer thermal resistance on the water side and the thermal conductivity thermal resistance of the pipe wall, the quasi-steady-state target wall temperature is obtained. Then, considering the thermal inertia caused by the heat capacity of the pipe wall, the actual wall temperature is iteratively updated to obtain the dynamic wall temperature field.

[0055] In this step, the specific method for advancing the water-side state is as follows: Along each water-cooled wall working fluid flow path, the water-side heating calculation units are numbered i=1,2,...,N, and the outlet enthalpy of the (i-1)th unit is used as the inlet enthalpy of the i-th unit using the upwind configuration. For each time step, the working fluid mass in the unit is first updated by mass conservation, then the heat absorption power on the flue gas side is included in the working fluid enthalpy by energy conservation, and finally the working fluid temperature and thermophysical parameters are retrieved based on the updated pressure and enthalpy. The pressure is obtained by interpolating the inlet and outlet pressures of the water-cooled wall along the direction of the working fluid flow; the working fluid temperature and thermophysical parameters are obtained by bilinear interpolation based on the pressure-enthalpy or pressure-temperature property tables.

[0056] In this step, the convective heat transfer coefficient inside the pipe The Nusselt number correlation is obtained based on the Reynolds number (Re), Prandtl number (Pr), and Nusselt number (Nu), and the Nusselt number correlation is selected according to the interval to which the Reynolds number belongs, indicating laminar, transitional, or turbulent flow; when Re i When 2300 < 2300, the laminar flow Nusselt number correlation is used; when 2300 ≤ Re i When <10000, a transitional flow correlation is used; when Re i When the value is ≥10000, the turbulent correlation equation is used to calculate the convective heat transfer coefficient inside the pipe. The formula for calculating the Nusselt number (Nu) is: in, denoted by the Nusselt number of the i-th water-side heating computational unit, dimensionless, used to characterize the strength of convective heat transfer; Represents the Reynolds number of the i-th heated calculation unit, which is dimensionless and used to determine the flow state of the working fluid; The Prandtl number represents the i-th heated computational unit. It is dimensionless and is used to characterize the ratio of the working fluid's momentum diffusion to its thermal diffusion capacity.

[0057] convective heat transfer coefficient The formula for calculation is: in, This represents the convective heat transfer coefficient inside the water-side pipe of the i-th heating calculation unit at time t; The thermal conductivity of the working fluid; This refers to the inner diameter of the water-cooled wall tube.

[0058] In this step, the formula for calculating the quasi-steady-state target wall temperature is: in, This represents the quasi-steady-state target wall temperature at time t; This represents the temperature of the working fluid inside the i-th heated calculation unit at time t; Indicates the thickness of the pipe wall metal; This indicates the thermal conductivity of the steel used in the pipe wall.

[0059] Considering the thermal inertia caused by the heat capacity of the pipe wall, the formula for iteratively updating the actual wall temperature is as follows: in, express The wall temperature at any given time; This indicates the time step used in the physical wall temperature calculation; This represents the wall temperature at time t. This represents the wall temperature at time t; The dimensionless dynamic response coefficient is specifically defined as follows: in, is the thermal inertia time constant.

[0060] S4. Sample database construction.

[0061] The dynamic wall temperature field is aligned with the operating control parameters in time to construct a dynamic wall temperature sample database, and deep learning training samples are formed according to a sliding time window. A convolutional autoencoder is used to encode the two-dimensional wall temperature field to obtain a low-dimensional feature vector representing the spatial temperature distribution and local hot spot features. The convolutional autoencoder includes an encoder and a decoder, and the encoder reduces the two-dimensional wall temperature field to a one-dimensional feature vector through a convolution algorithm, which solves the problem of too many parameters and training difficulties caused by directly inputting the two-dimensional field into the time series prediction model.

[0062] The deep learning training samples include wall temperature field, unit load, load change rate, burner commissioning status, feedwater flow rate, feedwater temperature and pressure for L historical time steps, and the prediction label is the wall temperature field for the next K time steps.

[0063] S5. Time Series Prediction and Result Output.

[0064] The low-dimensional feature vector and operating control parameters are input into the GRU time-series prediction model to predict the low-dimensional feature vector at future time. The decoder then reconstructs the water-cooled wall temperature field under future dynamic load conditions. Specifically, the decoder restores the predicted low-dimensional feature vector to the two-dimensional wall temperature field at future time. The predicted wall temperature of each regular grid is traversed to obtain the maximum wall temperature and its corresponding location. The predicted wall temperature is compared with the preset over-temperature threshold to determine the over-temperature area and over-temperature risk. The wall temperature change trend is judged based on the change in wall temperature or maximum wall temperature at adjacent future time, and the results are output.

[0065] In this step, the input to the GRU time series prediction model is: ,in, This represents the true low-dimensional feature vector at the current moment. The operating control parameters are for the current moment.

[0066] The GRU time-series prediction model sequentially calculates the reset gate. Update door Candidate hidden state and hidden state The low-dimensional feature prediction value for the next time step is obtained from the output layer; The GRU time-series prediction model employs an autoregressive rolling prediction method, using the low-dimensional feature vector predicted at the previous time step as the historical state input for the next time step. During training, the joint loss function is constructed from the low-dimensional feature prediction error and the wall temperature field error after decoding. in, Train a joint total loss function for the GRU time series prediction model; The loss weight coefficient has a value range of [0,1]; MSE is the mean squared error. This is the predicted low-dimensional feature vector output by the GRU time series prediction model; The predicted wall temperature field is obtained after the low-dimensional features are restored by the decoder. The target wall temperature field in the sample database; During training, the GRU time-series prediction model divides the training set and validation set according to the complete dynamic load change conditions to reduce the prediction bias caused by leakage in adjacent samples under the same conditions.

[0067] In this step, the results are output in at least one form: numerical values, variation curves, or two-dimensional wall temperature distribution maps.

[0068] This method also includes a verification step: for some representative dynamic variable load conditions, the dynamic wall temperature field obtained in step S3 is compared with the Fluent-UDF online coupling calculation results or the wall temperature measured at the field points based on the same boundary conditions to verify the error, which is used to correct the thermal inertia time constant. And empirical parameters for heat transfer calculation in the Nusselt number correlation.

[0069] UDF is a wall temperature calculation program based on C language programming. At each time step, UDF reads the heat flux density on the flue gas side of the water-cooled wall and calculates the working fluid state and wall temperature on the water side according to the feed water flow rate, temperature and pressure. Then, the wall temperature is returned to Fluent as the wall boundary, realizing bidirectional iteration of flue gas side heat flux and water side wall temperature. The online coupling calculation result of Fluent-UDF is the wall temperature field obtained by this bidirectional iteration process.

[0070] Unlike the present invention, which uses a heat flux density field and a one-dimensional water-side model to quickly calculate the wall temperature, Fluent-UDF simultaneously couples furnace combustion, flue gas flow, and wall heat transfer. While this involves a larger computational load, it provides a more complete physical process and can serve as a high-fidelity verification result under representative operating conditions. The dynamic wall temperature field obtained in step S3 is compared with the Fluent-UDF coupled calculation result under the same boundary conditions, spatial location, and time point. The resulting error is used to correct the aforementioned thermal inertia time constant. The Nusselt number correlation empirical coefficient is used to improve the accuracy of the wall temperature calculation and prediction results of this invention.

[0071] During model training, a convolutional autoencoder can be trained first to accurately reconstruct the two-dimensional wall temperature field. Then, the low-dimensional features output by the encoder and the operating control parameters are input into the GRU time-series prediction model for training. Alternatively, the convolutional autoencoder and GRU can be jointly fine-tuned after initial training. The training loss includes the low-dimensional feature prediction error and the wall temperature field error after decoding, thus balancing the accuracy of spatial distribution reconstruction and the accuracy of dynamic trend prediction. Example 1

[0072] The prediction method of this invention first obtains the transient computational fluid dynamics (CFD) wall calculation results and DCS or user operating control parameters. The heat flux density field obtained from the transient CFD includes at least the water-cooled wall position, area, heat flux density, and time step information; the operating control parameters include at least the unit load, load change rate, burner operating status, feedwater flow rate, feedwater temperature, and pressure. After aligning the above data according to a unified time series, a two-dimensional unfolding process is performed according to the furnace wall position, tube panel geometry, and water-cooled wall unfolding method. The unfolded heat flux density field is then weighted and reconstructed, and area conservation correction is applied to obtain the flue gas-side dynamic thermal boundary corresponding to the two-dimensional calculation region of the water-cooled wall.

[0073] After obtaining the two-dimensional heat flux density field, the water-cooled wall is divided into multiple heat-receiving computational units i based on the arrangement of the water-cooled wall tubes and the direction of the working fluid flow. These heat-receiving computational units are discrete heat-receiving units formed along the geometric position of the water-cooled wall and the direction of the working fluid flow, used for physical calculations and water-side state progression; they differ from... Figure 1 The system's functional modules, such as the data acquisition and standardization module and the physical wall temperature calculation module, are shown. The heat flux density on the flue gas side within each heated calculation unit is integrated using an area-weighted method to obtain the heat absorption power of that calculation unit at the corresponding time step.

[0074] Subsequently, using the heat absorption power of each heated calculation unit as the thermal boundary for water-side heat transfer calculation, a one-dimensional finite volume calculation is established along the working fluid flow direction. Inlet enthalpy and water-side state variables are transferred between adjacent heated calculation units according to the flow direction, and the working fluid enthalpy, temperature, pressure, and thermophysical parameters are updated based on mass and energy conservation. Density, specific heat capacity, viscosity, and thermal conductivity are obtained by looking up or interpolating based on the working fluid state, and the convective heat transfer coefficient inside the pipe is calculated using the Reynolds number, Prandtl number, and Nusselt number correlation formula. Furthermore, combining the convective heat transfer resistance inside the pipe and the thermal conductivity resistance of the pipe wall, an iterative update method with thermal inertia is used to obtain the dynamic wall temperature field.

[0075] After obtaining the historical dynamic wall temperature field, training samples are constructed using the historical water-cooled wall temperature field and operating control parameters. The training samples are organized according to a sliding time window, where the wall temperature field and corresponding operating parameters of the past L time steps are used as input, and the wall temperature field of the next K time steps are used as prediction labels. The two-dimensional wall temperature field is spatially reduced by a convolutional autoencoder to extract low-dimensional state features that characterize the spatial temperature distribution and local hot spots.

[0076] The low-dimensional features extracted by the convolutional autoencoder are fused with control variables such as load change rate, burner operating status, feedwater flow rate, feedwater temperature, and pressure, and then input into the GRU time-series prediction model to obtain a low-dimensional feature sequence for future time periods. The decoder of the convolutional autoencoder then decodes these future low-dimensional features to reconstruct the dynamic wall temperature field under unknown dynamic load conditions, and outputs the predicted results for maximum wall temperature, over-temperature location, over-temperature risk, and wall temperature distribution. Example 2

[0077] The verification module can use the online coupled calculation results of Fluent-UDF under representative operating conditions or the wall temperature measured at field points as the verification basis. Under the same operating boundary conditions, Fluent-UDF independently calculates the high-fidelity dynamic wall temperature field. The verification module receives the Fluent-UDF wall temperature field and the wall temperature field calculated or predicted by this invention, and compares the errors according to the same spatial location and timestamp. The errors are then used to correct the thermal inertia time constant and the empirical coefficient of the heat transfer correlation. This verification process is used to improve the reliability of the model and does not require a complete online coupled calculation at every time step. Example 3

[0078] Figure 3 This is a schematic diagram of the wall temperature prediction process combining a convolutional autoencoder and a GRU according to the present invention. Figure 3 As shown, the deep learning prediction process begins with the wall temperature field and operating parameters output from the sample database. First, a time window is constructed to form the historical wall temperature state. Then, the two-dimensional wall temperature field is compressed into the current low-dimensional features by the encoder of the convolutional autoencoder. The GRU time series prediction module predicts the future low-dimensional features based on the historical low-dimensional state and operating control parameters. Finally, the future low-dimensional features are restored to the wall temperature field by the decoder of the convolutional autoencoder, and the result output module outputs the maximum wall temperature, over-temperature risk, and prediction results. Example 4

[0079] Figure 4 This is a schematic diagram illustrating the mapping relationship between CFD wall mesh elements and regular mesh elements in this invention. For example... Figure 4 As shown, for each regular grid cell Using its center as a reference, select CFD wall mesh elements f within a neighborhood radius R, based on the distance between the center of the regular mesh element and the center of each CFD wall mesh element. By combining the area and heat flux density of the CFD wall mesh cells, a weighted calculation is performed, thereby mapping the CFD wall calculation results to regular mesh cells to form a regularized two-dimensional heat flux density field.

[0080] When in use, the working principle of this invention can be summarized as the technical path of "physical calculation to generate high-fidelity samples → spatial dimensionality reduction to extract core features → temporal prediction mapping dynamic evolution → high-fidelity verification closed-loop correction".

[0081] Specifically, this invention first obtains the unstructured wall heat flux density field calculated by transient CFD, and then normalizes it into a time-aligned, spatially regular two-dimensional dynamic heat flux density field through area-weighted interpolation and global area conservation correction. This normalization process ensures the integrity and consistency of the thermal boundary data.

[0082] Secondly, based on the standardized heat flux density field, the heat-receiving calculation units divided along the working fluid flow direction of the water-cooled wall are used as discrete control volumes. The working fluid state is advanced by solving the mass conservation and energy conservation equations. Combined with the calculation of the convective heat transfer coefficient inside the pipe and the iterative update of the pipe wall thermal inertia, a physically driven dynamic wall temperature field is obtained. This physical calculation process fully considers the changes in working fluid properties with pressure and enthalpy, the heat transfer characteristics under different flow states, and the hysteresis effect of pipe wall heat capacity on wall temperature changes, ensuring the physical consistency of the sample data.

[0083] Then, a convolutional autoencoder is used to spatially reduce the dimensionality of the high-dimensional wall temperature field, extracting low-dimensional feature vectors that characterize the spatial temperature distribution and local hotspots. These low-dimensional feature vectors, along with the operating control parameters, are then input into a GRU time-series prediction model to predict the low-dimensional feature vectors for future time steps. Finally, the wall temperature field is reconstructed using a decoder. This "spatial dimensionality reduction + time-series prediction" architecture effectively solves the problems of parameter explosion and training difficulties caused by direct prediction of high-dimensional fields.

[0084] Finally, the verification module uses high-fidelity Fluent-UDF coupled calculation results or field measurement data to correct the thermal inertia time constant and heat transfer correlation empirical parameters in the physical calculation module, forming a closed-loop optimization of "physical calculation - data prediction - verification correction", which ensures the continuous accuracy and generalization ability of the system.

[0085] Based on the above principles, this invention can shorten the wall temperature field prediction time while ensuring calculation accuracy, and realize rapid and accurate prediction of the water-cooled wall temperature field under variable load conditions, providing technical reference for the safe operation of power plants in deep peak shaving and rapid load change processes.

Claims

1. A boiler water-cooled wall temperature prediction system for dynamic load changes, characterized in that: It includes a data acquisition and standardization module, a physical wall temperature calculation module, a sample database module, a convolutional autoencoder spatial feature extraction module, a GRU temporal prediction module, a result output module, and a verification module connected in sequence to the physical wall temperature calculation module and the GRU temporal prediction module respectively. The data acquisition and standardization module is used to acquire the wall calculation results and operation control parameters obtained by the transient computational fluid dynamics method, and to form a dynamic heat flux density field with area conservation. The physical wall temperature calculation module is used to calculate the dynamic wall temperature field based on the heat absorption power of the heat-receiving calculation unit and the water-side conservation principle. The sample database module is used to store dynamic wall temperature fields and operating control parameters; The convolutional autoencoder spatial feature extraction module is used to encode the two-dimensional wall temperature field to obtain a low-dimensional feature vector, and to decode the low-dimensional feature vector to restore the two-dimensional wall temperature field. The GRU time-series prediction module is used to predict low-dimensional feature vectors at future times based on low-dimensional feature vectors and running control parameters. The result output module is used to output at least one of the following: maximum wall temperature, over-temperature risk, and wall temperature change trend; The verification module is used to verify the error of the physical wall temperature calculation or prediction results under representative working conditions.

2. A method for predicting boiler water-cooled wall temperature under dynamic load variations. The system implementation based on claim 1 is characterized by comprising the following steps: S1. Data Acquisition and Standardization: Acquire the transient computational fluid dynamics (CFD) wall calculation results and operating control parameters of the target boiler during dynamic load change process; expand the CFD wall mesh to a two-dimensional coordinate system according to the furnace wall and geometric position, and interpolate the operating control parameters to a time series consistent with the transient CFD wall calculation results; establish a regular canvas mesh, convert the CFD wall mesh heat flux density to the regular mesh through weighted mapping, and obtain the corrected regular mesh heat flux density after area conservation correction; S2. Division of Heating Calculation Units and Calculation of Heat Absorption Power: Based on the arrangement of water-cooled wall tubes, the location of the furnace wall, and the direction of working fluid flow, the two-dimensional region of the water-cooled wall is divided into multiple water-side heating calculation units connected sequentially along the direction of working fluid flow. Each water-side heating calculation unit covers one or more regular grid units and serves as a discrete control volume for the one-dimensional finite volume water-side state progression. The heat absorption power of the heating calculation unit is calculated by integrating the area of ​​the heat flux density of the regular grid in the i-th heating calculation unit. S3. Dynamic wall temperature field calculation: Using the heat absorption power as the thermal boundary on the flue gas side, the water side state is advanced along the working fluid flow direction using the one-dimensional finite volume method. Based on the conservation of mass and energy, the working fluid enthalpy, working fluid temperature and thermophysical parameters of each heat-receiving calculation unit are updated. The convective heat transfer coefficient inside the pipe is calculated. Combined with the convective heat transfer thermal resistance on the water side and the thermal conductivity thermal resistance of the pipe wall, the quasi-steady-state target wall temperature is obtained. Then, considering the thermal inertia caused by the heat capacity of the pipe wall, the actual wall temperature is iteratively updated to obtain the dynamic wall temperature field. S4. Sample Database Construction: The dynamic wall temperature field and the operating control parameters are aligned in time to construct a dynamic wall temperature sample database, and deep learning training samples are formed according to the sliding time window; a convolutional autoencoder is used to encode the two-dimensional wall temperature field to obtain a low-dimensional feature vector representing the spatial temperature distribution and local hot spot features; the convolutional autoencoder includes an encoder and a decoder, and the encoder reduces the two-dimensional wall temperature field to a one-dimensional feature vector through a convolution algorithm; S5. Time Series Prediction and Result Output: Input the low-dimensional feature vector and operation control parameters into the GRU time series prediction model to predict the low-dimensional feature vector at future time. Then, reconstruct the water-cooled wall temperature field under future dynamic variable load conditions through the decoder. Traverse the predicted wall temperature of each regular grid to obtain the maximum wall temperature and its corresponding location. Compare the predicted wall temperature with the preset over-temperature threshold to determine the over-temperature area and over-temperature risk. Determine the wall temperature change trend based on the change in wall temperature or maximum wall temperature at adjacent future time. Output the results.

3. The method for predicting boiler water-cooled wall temperature for dynamic load changes according to claim 2, characterized in that: In step S1, the transient computational fluid dynamics wall calculation results are obtained by solving the mass conservation, momentum conservation, energy conservation and component transport equations and derived from the water-cooled wall boundary after establishing a three-dimensional geometric model of the target boiler, dividing the CFD wall grid f and the volume grid, and applying time-varying fuel, air volume and heat transfer surface boundary conditions. The results include the center coordinates, area, flue gas side heat flux density and CFD time step of each CFD wall grid f. The operating control parameters are derived from the target boiler DCS historical records or actual power plant operating instructions, including unit load, load change rate, burner operation status, feedwater flow rate, feedwater temperature, and feedwater pressure. The specific method of interpolation is as follows: The sampling time of the operating control parameters is recorded as... The output time of the transient computational fluid dynamics wall calculation results is denoted as... Because the time step of the transient numerical simulation is inconsistent with the time step of the actual acquired data, the operating control parameters... When satisfied At that time, linear interpolation was used to obtain the operating control parameters. value of time This ensures that the CFD heat flux density field at each moment has the same timestamp as the corresponding operating control parameters.

4. The method for predicting boiler water-cooled wall temperature for dynamic load changes according to claim 2, characterized in that: In step S1, the regular grid unit The formula for calculating heat flux density is: in, Regular grid cells Heat flux density at time t; Let f be the area of ​​the CFD wall mesh; Let f be the heat flux density of the CFD wall mesh f at time t; For CFD wall mesh f, from center to regular mesh cell Distance from the center; The effective CFD neighborhood within the same wall partition is specifically represented as: in, R Let R be the neighborhood radius, and R > 0; This indicates that the CFD wall mesh f and the regular cell mesh g are located on the same wall. The weight kernel function is specifically expressed as: Area conservation corrections were applied to the heat flux densities of the two grids: in, This represents the dynamic heat flux correction coefficient, which is a dimensionless parameter. Regular grid cells The area; This represents the thermal power of the CFD wall mesh f; Representing a regular grid Thermal power; Revised regular grid The heat flux density is specifically: in, Represents the corrected regular grid Heat flux density, used to calculate regular meshes The local wall temperature is used as the wall temperature field.

5. The method for predicting boiler water-cooled wall temperature for dynamic load changes according to claim 4, characterized in that: In step S2, the formula for calculating the heat absorption power is: in, Ωi represents the heat absorption power of the i-th heated computing unit, which is only used for water-side energy conservation and working fluid state propulsion; Ωi represents the set of regular grid cell numbers covered by the i-th heated computing unit.

6. The method for predicting boiler water-cooled wall temperature for dynamic load changes according to claim 5, characterized in that: In step S3, the specific method for advancing the water-side state is as follows: Along each water-cooled wall working fluid flow path, the water-side heating calculation units are numbered i=1,2,...,N, and the outlet enthalpy of the (i-1)th unit is used as the inlet enthalpy of the i-th unit using the upwind configuration. For each time step, the working fluid mass in the unit is first updated by mass conservation, then the heat absorption power on the flue gas side is included in the working fluid enthalpy by energy conservation, and finally the working fluid temperature and thermophysical parameters are retrieved based on the updated pressure and enthalpy. The pressure is obtained by interpolation of the inlet and outlet pressures of the water-cooled wall along the working fluid flow direction; the working fluid temperature and thermophysical parameters are obtained by bilinear interpolation based on pressure-enthalpy or pressure-temperature property tables. In step S3, the convective heat transfer coefficient inside the pipe The Nusselt number correlation is obtained based on the Reynolds number (Re), Prandtl number (Pr), and Nusselt number (Nu), and the Nusselt number correlation is selected according to the interval to which the Reynolds number belongs, indicating laminar, transitional, or turbulent flow; when Re i When 2300 < 2300, the laminar flow Nusselt number correlation is used; when 2300 ≤ Re i When <10000, a transitional flow correlation is used; when Re i When the value is ≥10000, the turbulent correlation equation is used to calculate the convective heat transfer coefficient inside the pipe. The formula for calculating the Nusselt number Nu is: in, Nu i denoted by the Nusselt number of the i-th water-side heating computational unit, dimensionless, used to characterize the strength of convective heat transfer; Represents the Reynolds number of the i-th heated calculation unit, which is dimensionless and used to determine the flow state of the working fluid; The Prandtl number represents the i-th heated computational unit. It is dimensionless and is used to characterize the ratio of the working fluid's momentum diffusion to its thermal diffusion capacity. convective heat transfer coefficient The formula for calculation is: in, This represents the convective heat transfer coefficient inside the water-side pipe of the i-th heating calculation unit at time t; The thermal conductivity of the working fluid; This refers to the inner diameter of the water-cooled wall tube.

7. The method for predicting boiler water-cooled wall temperature for dynamic load changes according to claim 6, characterized in that: In step S3, the formula for calculating the quasi-steady-state target wall temperature is: in, This represents the quasi-steady-state target wall temperature at time t; This represents the temperature of the working fluid inside the i-th heated calculation unit at time t; Indicates the thickness of the pipe wall metal; This indicates the thermal conductivity of the steel used in the pipe wall; Considering the thermal inertia caused by the heat capacity of the pipe wall, the formula for iteratively updating the actual wall temperature is as follows: in, express The wall temperature at any given time; This indicates the time step used in the physical wall temperature calculation; This represents the wall temperature at time t. This represents the wall temperature at time t; The dimensionless dynamic response coefficient is specifically defined as follows: in, is the thermal inertia time constant.

8. The method for predicting boiler water-cooled wall temperature for dynamic load changes according to claim 7, characterized in that: In step S5, the input to the GRU time series prediction model is... ,in, z t This represents the true low-dimensional feature vector at the current moment. u t The operating control parameters are set for the current moment. The GRU timing prediction model sequentially calculates the reset gate. Update door Candidate hidden state and hidden state The low-dimensional feature prediction value for the next time step is obtained from the output layer; The GRU time-series prediction model employs an autoregressive rolling prediction method, using the low-dimensional feature vector predicted at the previous time step as the historical state input for the next time step. During training, the joint loss function is constructed from the low-dimensional feature prediction error and the wall temperature field error after decoding. in, Train a joint total loss function for the GRU time series prediction model; The loss weight coefficient has a value range of [0,1]; MSE is the mean squared error. This is the predicted low-dimensional feature vector output by the GRU time series prediction model; The predicted wall temperature field is obtained after the low-dimensional features are restored by the decoder. The target wall temperature field in the sample database; During training, the GRU time-series prediction model divides the training set and validation set according to the complete dynamic load change conditions to reduce the prediction bias caused by leakage in adjacent samples under the same conditions.

9. A method for predicting boiler water-cooled wall temperature for dynamic load changes according to claim 7, characterized in that: The process also includes a verification step: for some representative dynamic variable load conditions, the dynamic wall temperature field obtained in step S3 is compared with the Fluent-UDF online coupling calculation results or the wall temperature measured at the field points based on the same boundary conditions to verify the error, which is used to correct the thermal inertia time constant. And the empirical parameters for heat transfer calculation in the Nusselt number correlation; The UDF is a wall temperature calculation program based on C language programming. At each time step, the UDF reads the heat flux density on the flue gas side of the water-cooled wall, calculates the working fluid state and wall temperature on the water side according to the feed water flow rate, temperature and pressure, and then returns the wall temperature to Fluent as the wall boundary, realizing bidirectional iteration of flue gas side heat flux and water side wall temperature; the online coupling calculation result of Fluent-UDF is the wall temperature field obtained by this bidirectional iteration process. The dynamic wall temperature field obtained in step S3 is compared with the Fluent-UDF coupled calculation results under the same boundary conditions, spatial location, and time point. The resulting error is used to correct the aforementioned thermal inertia time constant. And the empirical coefficients of the Nusselt number correlation.

10. A method for predicting boiler water-cooled wall temperature for dynamic load changes according to claim 2, characterized in that: In step S5, the result is output in at least one form, such as a numerical value, a change curve, or a two-dimensional wall temperature distribution map.

Images