A temperature field prediction system and method for dynamic variable load process of coal-fired boiler

Abstract

This invention discloses a temperature field prediction system and method for dynamic load change processes in coal-fired boilers, belonging to the field of industrial intelligent computing and digital twin technology for coal-fired boilers. The system includes the following steps: reading real-time boiler operation control parameters; obtaining unsteady-state numerical simulation temperature field data through computational fluid dynamics methods, and constructing a spatially continuous distribution matrix with boundary constraints by combining geometric boundary constraints; constructing a spatial feature dimensionality reduction and restoration model, encoding the spatially continuous distribution matrix to obtain low-dimensional implicit vectors; performing multi-dimensional fusion of boiler operation control parameters and historical low-dimensional implicit vectors; constructing a time-series evolution prediction model, performing adaptive closed-loop rolling iterative deduction, and predicting future low-dimensional implicit vector sets; and inversely mapping the predicted low-dimensional implicit vector sets to a temperature field matrix through a spatial restoration model and returning it to the user terminal. This invention provides real-time and reliable technical support for temperature monitoring and over-temperature early warning of power plant boilers under variable load.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent computing and digital twin technology of coal-fired boilers, specifically to a system and method for predicting the temperature field during dynamic load changes in a coal-fired boiler. Background Technology

[0002] The installed capacity of existing new energy power generation continues to climb. To address the intermittent and unstable nature of new energy power generation, large-capacity coal-fired boilers need to be capable of low-load operation and rapid load changes. These processes cause rapid changes in the furnace temperature field and local combustion conditions. If the evolution trend of the temperature field cannot be grasped in time, it can easily affect the boiler's safety, economy, and pollutant control effectiveness.

[0003] Computational fluid dynamics (CFD) methods can describe the flow, heat transfer, and combustion processes inside the furnace with relatively high precision, and are an important means of obtaining temperature field distribution. However, unsteady-state numerical simulations of ultra-large coal-fired boilers typically require high computational resources and long computation times, making it difficult to directly meet the real-time requirements of online monitoring, rapid simulation, and operational decision-making. Therefore, it is necessary to provide a temperature field prediction system and method that combines unsteady-state numerical simulation data, geometric boundary constraints, spatial feature dimensionality reduction, and temporal state prediction, so as to achieve rapid prediction and visualization of the temperature field during dynamic load changes in coal-fired boilers without excessively increasing the online computational burden. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a system and method for predicting the temperature field during dynamic load changes in coal-fired boilers, in order to solve the problems of long unsteady-state solution time in existing computational fluid dynamics, difficulty in reflecting the overall temperature field distribution of the furnace by traditional measurement point monitoring, and insufficient continuous extrapolation capability of existing data-driven methods in dynamic load change scenarios, so as to provide real-time and reliable technical support for temperature monitoring and over-temperature early warning of power plant boilers under load changes.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows.

[0006] A method for predicting the temperature field during dynamic load changes in a coal-fired boiler includes the following steps: S1. Read the boiler's real-time operation control parameters based on user-side input or DCS operation data to achieve online prediction; S2. Obtain the unsteady-state numerical simulation temperature field data of the target coal-fired boiler during dynamic load change process, and construct a spatially continuous distribution matrix with boundary constraints in combination with geometric boundary constraints; S3. Construct a spatial feature dimensionality reduction and restoration model, encode the spatial continuous distribution matrix with boundary constraints in step S2, obtain a low-dimensional implicit vector representing the spatial distribution of the temperature field, and the spatial feature dimensionality reduction and restoration model can also restore the spatial continuous distribution matrix with boundary constraints based on the low-dimensional implicit vector. S4. Extract the real-time operation control parameters of the boiler during the dynamic load change process and perform multi-dimensional fusion with the low-dimensional implicit vector of historical time. S5. Construct a time-series evolution prediction model, input the fused features into the time-series evolution prediction model, and predict the low-dimensional implicit vector of the next target time. Use the predicted low-dimensional implicit vector as the state input for subsequent time periods, and perform closed-loop rolling deduction according to the preset time step to obtain the low-dimensional implicit vector group for future time periods. S6. Perform spatial reconstruction on the low-dimensional implicit vector group and output the prediction results of the unsteady temperature field during the future dynamic load change process; S7. Return the prediction results to the user.

[0007] Preferably, in step S1, the real-time boiler operation control parameters include the boiler load change rate, the current boiler load, and the burner operating status.

[0008] Preferably, in step S2, the unsteady-state numerical simulation is based on computational fluid dynamics. Initial conditions for typical operating conditions are obtained from the power plant's DCS operating data for offline training sample generation. The initial conditions include coal type, primary air velocity, secondary air velocity, burnout air velocity, primary air temperature, secondary air temperature, burnout air temperature, coal feed rate, and burner tilt angle. The internal temperature field of the furnace is obtained through unsteady-state numerical simulation using a coupled turbulence model, radiation model, discrete phase model, and coke combustion model, resulting in an unsteady-state temperature field sequence that is continuous before and after the initial state, retains thermal inertia, and exhibits response hysteresis. In step S2, the unsteady-state numerical simulation specifically involves: establishing a computational domain and grid consistent with the target coal-fired boiler; inputting the initial conditions of typical operating conditions into a solver based on the finite volume method; solving to obtain the convergent temperature field at the initial moment; then enabling unsteady-state solution; setting the physical time step; and updating the load, coal feed rate, wind speed, wind temperature, and burner swing angle conditions at each time step according to the variable load trajectory using a time series table or a user-defined function; at the nth time step, using the convergent flow field of the previous time step as the initial value, implicit time discretization is adopted to iteratively solve the continuity equation, momentum equation, energy equation, and component transport equation, and coupling the turbulence model, radiation model, discrete phase model, and coke combustion model; after satisfying the convergence requirements of residuals and temperature monitoring quantities, the temperature field at that moment is saved, and then the process is advanced to the next time step until the entire variable load process is completed; In step S2, the geometric boundary constraint is a masking matrix formed based on the actual boiler's geometric parameters. Its constraint effect on the complete mesh matrix after interpolation of the original continuous flow field is as follows: in, This is the complete grid matrix after interpolation of the original continuous flow field; The spatially continuous distribution matrix with boundary constraints is used to train the spatial feature dimensionality reduction and restoration model in step S3 and the temporal evolution prediction model in step S5. It represents the Hadamardi (or Hadama) stack; Let be the masking matrix, which is defined as: in, y For the row index of the grid matrix; x This is the column index of the grid matrix.

[0009] Preferably, in step S2, the geometric boundary constraint includes: generating a boundary mask based on the true geometric contour of the target coal-fired boiler furnace or the target cross section, masking or zeroing invalid grid points outside the boundary, and retaining the valid area inside the boundary to participate in model training, inference, and result display.

[0010] Preferably, in step S3, the low-dimensional implicit vector specifically refers to the core representation vector that can completely reproduce the original continuous physical field spatial topology after projecting a high-dimensional, highly redundant spatially continuous distribution matrix onto a low-dimensional implicit feature subspace through a nonlinear transformation; the spatial feature dimensionality reduction and restoration model includes an encoder. and decoder The encoder will convert the spatially continuous distribution matrix with boundary constraints from step S2. Mapped to low-dimensional implicit vectors The decoder reconstructs the spatially continuous distribution matrix with boundary constraints based on this implicit vector, expressed as: in, The spatial continuous distribution matrix is ​​represented by the decoder; the spatial feature dimensionality reduction and restoration model reduces the size of the matrix. and The reconstruction error between the two is used for training to establish a bidirectional mapping relationship between the spatially continuous distribution matrix and the low-dimensional implicit vector.

[0011] Preferably, in step S5, the expression for the time-series evolution prediction model is: in, For time-series evolution prediction models, These are model parameters; This is the low-dimensional implicit vector predicted by the model at the previous time step; These are the real-time operating control parameters for the boiler. This is the low-dimensional implicit vector of the current time t predicted by the model; The loss function of the time-series evolution prediction model is defined as: in, The loss value of the time series evolution prediction model at time t; is the dimension of the low-dimensional implicit feature subspace; Characterizing the first in low-dimensional space The index of each feature component; The encoder of the spatial feature dimensionality reduction and restoration model at time t is the implicit vector true value obtained after spatial dimensionality reduction of the temperature field in the unsteady numerical simulation. The implicit vector prediction value at time t is generated by the rolling prediction of the time series evolution prediction model.

[0012] Preferably, in step S6, the predicted low-dimensional implicit vector at the current time t The method for evaluating the error between the spatially continuous distribution matrix obtained after decoder reconstruction and the spatially continuous distribution matrix obtained after geometric boundary constraints on the unsteady numerical simulation temperature field at time t is as follows: in, Let be the transient average relative error at time t; Number of columns in the image; The number of rows in the image; This indicates that at time t, the coordinates are... Unsteady-state numerical simulation temperature at the location; This indicates that at time t, the coordinates are... Deep learning predicts temperature at a location.

[0013] Preferably, the spatial feature dimensionality reduction and restoration model is a nonlinear dimensionality reduction model with spatial feature extraction capability, which maps a high-dimensional spatial continuous distribution matrix to a low-dimensional state space to form a low-dimensional implicit vector, and inversely maps the low-dimensional implicit vector to a spatial continuous distribution matrix. The time-series evolution prediction model includes a control parameter input branch and a historical low-dimensional implicit vector input branch, which are used to learn the correspondence between the running control parameters and the low-dimensional implicit vector evolution of the temperature field.

[0014] Preferably, when training the spatial feature dimensionality reduction and restoration model, at least one of temperature field reconstruction consistency constraints, spatial gradient consistency constraints, and local structure consistency constraints is comprehensively adopted to ensure that the restored spatial continuous distribution matrix maintains boundary changes and main spatial distribution features. When training the time-series evolution prediction model, the complete variable load process is used as the sample organization unit, and at least one of low-dimensional implicit vector error and consistency of state changes between adjacent time steps is used to constrain the model. In step S6, the unsteady temperature field prediction results are output in at least one of the following forms: temperature cloud map, time series matrix, dynamic image, key area temperature index, or over-temperature warning information; when the predicted temperature exceeds the preset threshold, a prompt message for boiler operation adjustment is generated.

[0015] A temperature field prediction system for dynamic load change process of coal-fired boiler includes an online prediction link and an offline training and model support link; The online prediction link includes a data reading module, a data preprocessing module, a time series prediction module, and a result output module connected in sequence; The data reading module is used to receive user input or automatically read power plant DCS operation data, obtain real-time boiler operation control parameters, and transmit the real-time boiler operation control parameters to the data preprocessing module. The data preprocessing module is used to perform gridding, normalization and geometric boundary masking constraint processing on the input data to form input data suitable for subsequent model inference, and then pass the processed data to the time series prediction module. The time series prediction module has a built-in trained time series evolution prediction model, which is used to predict the future low-dimensional implicit vector group based on the real-time operation control parameters of the boiler and the low-dimensional implicit vector of the previous moment through multi-dimensional feature fusion and closed-loop rolling inference, and transmit the prediction results to the result output module. The result output module has a built-in trained decoder, which is used to restore the future low-dimensional implicit vector group into the unsteady temperature field prediction result, and display or warn. The offline training and model support link includes a nonsteady-state numerical simulation data generation module, a geometric boundary definition module, a training sample library, and a model library connected in sequence. The unsteady-state numerical simulation data generation module is used to obtain the unsteady-state temperature field sequence and corresponding boundary conditions of the target coal-fired boiler during dynamic load change process based on computational fluid dynamics methods. The geometric boundary definition module is used to generate the furnace cross section and masking matrix based on the actual geometric parameters of the target coal-fired boiler. The training sample library is used to store training samples, which are used to train spatial feature dimensionality reduction and restoration models and temporal evolution prediction models. The model parameters obtained from the training are stored in the model library. The model library is used to store the parameters of the trained models; The output of the geometric boundary definition module is also connected to the input of the data preprocessing module in the online prediction link, and is used to provide a masking matrix to the data preprocessing module so that the data preprocessing module can perform geometric boundary masking constraint processing on the input data. The output of the model library is also connected to the input of the time series prediction module and the result output module in the online prediction link, respectively, to provide model support to the online prediction link.

[0016] Due to the adoption of the above technical solutions, the technical progress achieved by this invention is as follows.

[0017] This invention maps a high-dimensional, highly redundant spatially continuous distribution matrix to a low-dimensional implicit feature subspace through a spatial feature dimensionality reduction and restoration model encoder. It extracts a low-dimensional implicit vector that can completely reproduce the original continuous physical field spatial topology and performs temporal evolution prediction in the low-dimensional implicit feature subspace. This significantly reduces the online computational load of dynamic temperature field prediction and realizes second-level real-time dynamic prediction of the temperature field during dynamic load change processes of coal-fired boilers, which can meet the real-time requirements of online monitoring and operation decision-making.

[0018] This invention forms a masking matrix based on the geometric parameters of a real boiler and applies geometric boundary masking constraints to the complete grid matrix after interpolation of the original continuous flow field. This concentrates the prediction results in the actual physical region of the boiler furnace or target section and masks or zeros invalid grid points outside the boundary. This reduces the interference of physically meaningless regions on model training, inference and result display, and improves prediction accuracy and visualization effect.

[0019] This invention integrates boiler load, load change rate, burner operating status, and low-dimensional implicit vector from the previous moment into a multi-dimensional system. This enables the time-series evolution prediction model to not only perceive the current operating control conditions but also inherit the low-dimensional implicit vector of the historical temperature field. This enhances the ability to express the continuous evolution law of dynamic load change processes and improves the prediction accuracy under varying operating conditions.

[0020] The prediction system of this invention has an online prediction link and an offline training and model support link. The data reading module supports two data acquisition methods: user input and automatic reading of power plant DCS operation data, which facilitates offline scheme evaluation and online digital twin prediction applications, and has good flexibility and engineering applicability. The system uses the unsteady temperature field sequence and boundary conditions provided by the unsteady numerical simulation data generation module and the masking matrix provided by the geometric boundary definition module to form a training sample set with geometric boundary constraints. The training sample set is used to train the spatial feature dimensionality reduction and restoration model and the temporal evolution prediction model, and the trained model parameters are stored in the model library to provide model support for the online prediction link, so that the online prediction stage does not need to repeat the high-cost unsteady numerical simulation.

[0021] This invention establishes a computational domain and computational grid consistent with the target coal-fired boiler, adopts implicit time discretization, and uses the convergent flow field of the previous time step as the initial value to advance the solution step by step. This results in a non-steady-state temperature field sequence that is continuous between the previous and subsequent states, retains thermal inertia and response lag. This provides high-quality time-series training samples for spatial feature dimensionality reduction and restoration models and time-series evolution prediction models, enabling the prediction results to reflect the real physical characteristics of the boiler during dynamic load changes.

[0022] When training the spatial feature dimensionality reduction and restoration model, this invention comprehensively adopts at least one of temperature field reconstruction consistency constraints, spatial gradient consistency constraints, and local structure consistency constraints to ensure that the restored spatial continuous distribution matrix maintains boundary changes and main spatial distribution characteristics. When training the time series evolution prediction model, the complete variable load process is used as the sample organization unit, and at least one of low-dimensional implicit vector error and consistency of state changes at adjacent time moments is used to constrain the model, thereby improving the stability of continuous prediction under dynamic variable operating conditions. Attached Figure Description

[0023] Figure 1 This is a flowchart of a method for predicting the temperature field during dynamic load changes in a coal-fired boiler, as described in this invention. Figure 2 This is a schematic diagram illustrating how the predicted low-dimensional implicit vector is used as the state input for subsequent time steps in the method of the present invention, and how closed-loop rolling deduction and temperature field reconstruction are performed according to a preset time step. Figure 3 This is a schematic diagram of the structure of a dynamic variable load process temperature field prediction system for a coal-fired boiler according to the present invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0025] A method for predicting the temperature field during dynamic load changes in a coal-fired boiler can be executed in the following order: user-side input or DCS operation data reading, acquisition of unsteady-state numerical simulation temperature field data, construction of a spatially continuous distribution matrix with geometrical boundary constraints, spatial feature dimensionality reduction and restoration, fusion of operating control parameters and historical low-dimensional implicit vectors, closed-loop prediction of the time-series evolution model, spatial restoration of the temperature field, and return of the prediction results. Combined with... Figure 1 As shown, it includes the following steps: S1. Read real-time boiler operation control parameters based on user input or DCS operation data. It is used to achieve online prediction.

[0026] In this step, the real-time operation control parameters of the boiler are read based on user input or DCS operation data, including two data acquisition methods: (1) Obtain operation control parameters based on user input data; (2) Automatically read the power plant DCS operation data to obtain operation control parameters.

[0027] Boiler real-time operation control parameters Including boiler load change rate Boiler current load and burner operating status ,Right now .

[0028] S2. Obtain the unsteady-state numerical simulation temperature field data of the target coal-fired boiler during dynamic load change process, and construct a spatially continuous distribution matrix with boundary constraints by combining geometric boundary constraints.

[0029] In this step, the unsteady-state numerical simulation is based on computational fluid dynamics methods, using initial conditions for typical operating conditions obtained from the power plant's DCS operating data. Used for offline training sample generation, where The unsteady-state time in CFD is represented by the initial conditions, which include coal type, primary air velocity, secondary air velocity, burnout air velocity, primary air temperature, secondary air temperature, burnout air temperature, coal feed rate, and burner tilt angle. The internal temperature field of the furnace is obtained by unsteady-state numerical simulation method that couples turbulence model, radiation model, discrete phase model, and coke combustion model, and a sequence of unsteady-state temperature fields with continuous front and rear states, retaining thermal inertia and response hysteresis is obtained.

[0030] The unsteady-state numerical simulation specifically involves: establishing a computational domain and computational grid consistent with the target coal-fired boiler, and setting initial conditions for typical operating conditions. Input a solver based on the finite volume method to obtain the convergent temperature field at the initial moment, then enable unsteady-state solution and set the physical time step. The method updates the load, coal feed rate, wind speed, wind temperature, and burner swing angle conditions at each time step according to the variable load trajectory using a time series table or user-defined functions. At the nth time step, the converged flow field of the previous time step is used as the initial value. Implicit time discretization is employed to iteratively solve the continuity equation, momentum equation, energy equation, and component transport equation, coupled with turbulence model, radiation model, discrete phase model, and coke combustion model. After the convergence requirements of residuals and temperature monitoring are met, the temperature field at that moment is saved, and the process is then advanced to the next time step until the entire variable load process is completed. This method differs from performing steady-state or weakly unsteady-state calculations separately for multiple stable time windows.

[0031] In this step, the geometric boundary constraint is a masking matrix formed based on the actual boiler's geometric parameters. Its constraint effect on the complete mesh matrix after interpolation of the original continuous flow field is as follows: in, This is the complete grid matrix after interpolation of the original continuous flow field; The spatially continuous distribution matrix with boundary constraints is used to train the spatial feature dimensionality reduction and restoration model in step S3 and the temporal evolution prediction model in step S5. It represents the Hadamardi (or Hadama) stack; Let be the masking matrix, which is defined as: in, y For the row index of the grid matrix; x This is the column index of the grid matrix.

[0032] Geometric boundary constraints include: generating a boundary mask based on the true geometric contour of the target coal-fired boiler furnace or target cross-section, masking or zeroing invalid mesh points outside the boundary, and retaining the valid region inside the boundary for model training, inference, and result display.

[0033] S3. Construct a spatial feature dimensionality reduction and restoration model, encode the spatial continuous distribution matrix with boundary constraints in step S2, obtain a low-dimensional implicit vector representing the spatial distribution of the temperature field, and the spatial feature dimensionality reduction and restoration model can also restore the spatial continuous distribution matrix with boundary constraints based on the low-dimensional implicit vector.

[0034] In this step, the low-dimensional implicit vector specifically refers to the core representation vector that can completely reproduce the original continuous physical field spatial topology, obtained by projecting a high-dimensional, highly redundant spatially continuous distribution matrix onto a low-dimensional implicit feature subspace through a nonlinear transformation; the spatial feature dimensionality reduction and restoration model includes an encoder. and decoder The encoder will convert the spatially continuous distribution matrix with boundary constraints from step S2. Mapped to low-dimensional implicit vectors The decoder reconstructs the spatially continuous distribution matrix with boundary constraints based on this implicit vector, expressed as: in, The spatial continuous distribution matrix is ​​represented by the decoder; the spatial feature dimensionality reduction and restoration model reduces the size of the matrix. and The reconstruction error between the two is used for training to establish a bidirectional mapping relationship between the spatially continuous distribution matrix and the low-dimensional implicit vector.

[0035] S4. Extract the real-time operation control parameters of the boiler during the dynamic load change process and perform multi-dimensional fusion with the low-dimensional implicit vectors of historical time.

[0036] In this step, multidimensional fusion refers to fusing boiler load, load change rate, burner operating status, and the low-dimensional implicit vector from the previous moment.

[0037] S5. Construct a temporal evolution prediction model, input the fused features into the temporal evolution prediction model, and predict the low-dimensional implicit vector of the next target time. Use the predicted low-dimensional implicit vector as the state input for subsequent time periods, and perform closed-loop rolling deduction according to the preset time step to obtain the low-dimensional implicit vector group for future time periods.

[0038] In this step, the expression for the time-series evolution prediction model is: in, For time-series evolution prediction models, These are model parameters; This is the low-dimensional implicit vector predicted by the model at the previous time step; These are the real-time operating control parameters for the boiler. This is the low-dimensional implicit vector of the current time t predicted by the model.

[0039] Using the predicted low-dimensional implicit vectors as the state input for subsequent time periods, the low-dimensional implicit vector set for future time periods is obtained by closed-loop rolling deduction according to the preset time step.

[0040] The loss function of the time-series evolution prediction model is defined as: in, The loss value of the time series evolution prediction model at time t; is the dimension of the low-dimensional implicit feature subspace; Characterizing the first in low-dimensional space The index of each feature component; The encoder of the spatial feature dimensionality reduction and restoration model at time t is the implicit vector true value obtained after spatial dimensionality reduction of the temperature field in the unsteady numerical simulation. The implicit vector prediction value at time t is generated by the rolling prediction of the time series evolution prediction model.

[0041] S6. Perform spatial reconstruction on the low-dimensional implicit vector group and output the prediction results of the unsteady temperature field during the future dynamic load change process.

[0042] In this step, the low-dimensional implicit vector at the current time t is predicted. The method for evaluating the error between the spatially continuous distribution matrix obtained after decoder reconstruction and the spatially continuous distribution matrix obtained after geometric boundary constraints on the unsteady numerical simulation temperature field at time t is as follows: in, Let be the transient average relative error at time t; Number of columns in the image; The number of rows in the image; This indicates that at time t, the coordinates are... Unsteady-state numerical simulation temperature at the location; This indicates that at time t, the coordinates are... Deep learning predicts temperature at a location.

[0043] S7. Return the prediction results to the user.

[0044] In this step, the unsteady temperature field prediction results are output in at least one of the following forms: temperature cloud map, time series matrix, dynamic image, key area temperature index, or over-temperature warning information; when the predicted temperature exceeds the preset threshold, a prompt message is generated for boiler operation adjustment.

[0045] Specifically, the spatial feature dimensionality reduction and restoration model is a nonlinear dimensionality reduction model with spatial feature extraction capabilities. It maps a high-dimensional spatial continuous distribution matrix to a low-dimensional state space to form a low-dimensional implicit vector, and then inversely maps the low-dimensional implicit vector to a spatial continuous distribution matrix.

[0046] The time-series evolution prediction model includes a control parameter input branch and a historical low-dimensional implicit vector input branch, which are used to learn the correspondence between the running control parameters and the low-dimensional implicit vector evolution of the temperature field.

[0047] When training the spatial feature dimensionality reduction and restoration model, at least one of the following is used in combination: temperature field reconstruction consistency constraint, spatial gradient consistency constraint, and local structure consistency constraint, so that the restored spatial continuous distribution matrix maintains the boundary changes and main spatial distribution features.

[0048] When training the time-series evolution prediction model, the complete variable load process is used as the sample organization unit, and at least one of low-dimensional implicit vector error and consistency of state changes between adjacent time moments is used to constrain the model in order to improve the continuous prediction stability under dynamic variable operating conditions.

[0049] Example 1: First, execute step S1, which involves user-side input or DCS operation data reading. User-side input may include boiler load change rate, current boiler load, and boiler burner operating status; DCS operation data reading may include obtaining the current boiler load, boiler load change rate, boiler burner operating status, and related initial operating conditions from the power plant's distributed control system. Prediction results can be provided based on manually input data, or the power plant's DCS operation data can be automatically read to predict the temperature field for the next moment or future period.

[0050] When executing S2, the unsteady-state numerical simulation temperature field data of the target coal-fired boiler during dynamic load change process is obtained, and a spatially continuous distribution matrix with boundary constraints is constructed by combining geometric boundary constraints.

[0051] Unsteady-state numerical simulation temperature field data can be obtained based on computational fluid dynamics methods, with initial conditions derived from historical operating data of the power plant's DCS under typical variable load conditions. By comprehensively considering the furnace heat transfer and combustion processes, the internal temperature field of the furnace can be obtained, providing a data foundation for subsequent spatial feature dimensionality reduction and restoration models, as well as for training time-series evolution prediction models.

[0052] Geometric boundary constraints are used to transform the complete grid matrix after interpolation of the original continuous flow field into a spatially continuous distribution matrix with boundary constraints, which is actually used in the training of spatial feature dimensionality reduction and restoration models and temporal evolution prediction models. Specifically, a masking matrix is ​​formed based on the geometric parameters of the actual boiler, and this masking matrix is ​​used to constrain the complete grid matrix. Invalid grid points outside the boundary are masked or zeroed out during training, prediction, and visualization, while valid regions inside the boundary are retained and participate in model training, inference, and result display. This reduces the interference of physically meaningless regions on the prediction results.

[0053] When executing S3, a spatial feature dimensionality reduction and restoration model is constructed. The spatially continuous distribution matrix with geometric boundary constraints is encoded to obtain a low-dimensional implicit vector representing the spatial distribution of the temperature field. Furthermore, the spatially continuous distribution matrix with boundary constraints can be restored from the low-dimensional implicit vector. The low-dimensional implicit vector refers to the core representation vector obtained by projecting the high-dimensional, highly redundant spatially continuous distribution matrix onto a low-dimensional implicit feature subspace through a nonlinear transformation. It is used to represent the main spatial topology and evolutionary state of the temperature field.

[0054] When executing S4, the real-time operation control parameters of the boiler during the dynamic load change process are extracted and fused with the low-dimensional implicit vectors from historical moments. The multi-dimensional fusion includes fusing the boiler load, boiler load change rate, burner operating status, and the low-dimensional implicit vectors from the previous moment, so that the time-series evolution prediction model can both perceive the current operation control conditions and inherit the low-dimensional implicit vectors from historical moments.

[0055] When executing S5, a temporal evolution prediction model is constructed. The fused features are input into the model to predict the low-dimensional implicit vector at the next target time step. Using the predicted low-dimensional implicit vector as the state input for subsequent time steps, a closed-loop rolling derivation is performed according to a preset time step to obtain the low-dimensional implicit vector set for future time periods. The temporal evolution prediction model can include a control parameter input branch and a historical state input branch, used to learn the correspondence between the operating control parameters and the evolution of the low-dimensional implicit vector.

[0056] When executing S6, spatial reconstruction is performed on the low-dimensional implicit vector group, and the prediction results of the unsteady temperature field during the future dynamic load change process are output. The prediction results may include the unsteady temperature field matrix, temperature contour map, dynamic temperature field image, temperature index of key areas, or over-temperature warning results.

[0057] When S7 is executed, the prediction results obtained by the above prediction method are returned to the user side, thus completing a complete temperature field prediction process from input to prediction output.

[0058] Example 2: As Figure 2 As shown, the present invention uses the predicted low-dimensional implicit vector as the state input for subsequent time steps, and performs closed-loop rolling deduction according to a preset time step. The process includes input of running control parameters, input of the low-dimensional implicit vector of the previous time step, fusion of multi-dimensional features, prediction by the temporal evolution prediction model, output of the low-dimensional implicit vector of the target time step, spatial reconstruction by the decoder, and return of temperature field results. Figure 2 Operating control parameters This includes boiler load, boiler load change rate, and burner operating status, and the low-dimensional implicit vector of the previous time step. This is the previous predicted state.

[0059] like Figure 2 As shown, at the target time t, the running control parameters will be... Compared with the low-dimensional implicit vector of the previous time step The features are spliced ​​or fused to obtain a fused feature; this fused feature is then input into the temporal evolution prediction model, and the output is a low-dimensional implicit vector of the target time. Due to the predicted It can be used as the input of the historical state for the next moment, thus forming a closed-loop rolling deduction, thereby gradually deducing the low-dimensional implicit vector group of the future period according to the preset time step.

[0060] like Figure 2 As shown, the low-dimensional implicit vector at the target time. After inputting the decoder (with a built-in spatial reconstruction model), a spatially continuous distribution matrix (temperature field matrix) is obtained. A geometric mask M is used to impose geometric boundary constraints on the spatial reconstruction result, shielding the mesh outside the boundary and forming a displayable and evaluable unsteady-state temperature field matrix within the effective region. This temperature field matrix can be further converted into temperature contour maps, time-series matrices, dynamic images, or key area indicators, and returned to the user terminal, digital twin dashboard, or over-temperature warning module.

[0061] A temperature field prediction system for dynamic load change process of coal-fired boiler, such as Figure 3 As shown, it includes the online prediction link and the offline training and model support link.

[0062] The online prediction pipeline consists of a data reading module, a data preprocessing module, a time series prediction module, and a result output module, which are connected in sequence.

[0063] The data reading module is used to receive user input or automatically read power plant DCS operation data, obtain real-time boiler operation control parameters, and transmit the real-time boiler operation control parameters to the data preprocessing module.

[0064] The data preprocessing module is used to perform gridding, normalization, and geometric boundary masking constraints on the input data to form input data suitable for subsequent model inference, and then passes the processed data to the time series prediction module.

[0065] The time series prediction module has a built-in pre-trained time series evolution prediction model, which is used to predict the future low-dimensional implicit vector group based on the real-time operation control parameters of the boiler and the low-dimensional implicit vector of the previous moment through multi-dimensional feature fusion and closed-loop rolling inference, and then transmits the prediction results to the result output module.

[0066] The output module has a built-in trained decoder, which is used to restore the future low-dimensional implicit vector group into the unsteady temperature field prediction result, and then display or issue an early warning.

[0067] The offline training and model support chain includes a nonsteady-state numerical simulation data generation module, a geometric boundary definition module, a training sample library, and a model library, which are connected in sequence.

[0068] The unsteady-state numerical simulation data generation module is used to obtain the unsteady-state temperature field sequence and corresponding boundary conditions of the target coal-fired boiler during dynamic load change process based on computational fluid dynamics methods.

[0069] The geometric boundary definition module is used to generate the furnace cross section and masking matrix based on the actual geometric parameters of the target coal-fired boiler.

[0070] The training sample library is used to train spatial feature dimensionality reduction and restoration models and temporal evolution prediction models, and the trained model parameters are stored in the model library.

[0071] The model library is used to store the parameters of the trained model.

[0072] Specifically, the output of the geometric boundary definition module is also connected to the input of the data preprocessing module in the online prediction link, which is used to provide a masking matrix to the data preprocessing module so that the data preprocessing module can perform geometric boundary masking constraint processing on the input data.

[0073] The output of the model library is also connected to the input of the time series prediction module and the result output module in the online prediction link, respectively, to provide model support to the online prediction link.

[0074] The prediction system has two operating modes: (1) It provides prediction results based on user input data. (2) The prediction system automatically reads the power plant's DCS operating data and completes the prediction of the temperature field at the next moment. This facilitates offline scheme evaluation and online digital twin prediction applications.

[0075] The working principle of this invention is as follows: First, a high-fidelity temperature field sequence of the target coal-fired boiler under typical variable load conditions is obtained through computational fluid dynamics unsteady-state numerical simulation. Then, a masking matrix is ​​generated by combining the actual boiler geometric parameters, and geometric boundary constraints are applied to the temperature field matrix to form a spatially continuous distribution matrix with boundary constraints.

[0076] Then, a spatial feature dimensionality reduction and restoration model (encoder-decoder structure) is constructed. The high-dimensional, highly redundant spatial continuous distribution matrix is ​​mapped to a low-dimensional implicit feature subspace through the encoder, extracting low-dimensional implicit vectors that can completely reproduce the spatial topology of the original physical field. At the same time, the model establishes an inverse mapping relationship from the low-dimensional implicit vectors to the spatial continuous distribution matrix through the decoder, realizing the spatial restoration of the temperature field.

[0077] Next, the real-time operating control parameters of the boiler (load, variable load rate, burner status) are fused with the low-dimensional implicit vectors of historical moments in a multi-dimensional manner, so that the time-series evolution prediction model can not only perceive the current operating control conditions, but also inherit the historical temperature field state information.

[0078] Subsequently, the temporal evolution prediction model uses the fused features as input to predict the low-dimensional implicit vector of the next moment, and uses the prediction result as the state input for subsequent moments. It performs closed-loop rolling deduction according to the preset time step to realize continuous prediction of the temperature field evolution trend in future periods.

[0079] Finally, the predicted low-dimensional implicit vector group is inversely mapped back to the high-dimensional space through the decoder, and after geometric boundary masking constraint processing, the unsteady temperature field prediction result with boundary constraints is output and returned to the user terminal in the form of temperature cloud map, time series matrix, dynamic image, temperature index of key area or over-temperature warning information.

[0080] Throughout the entire prediction process, the spatial feature dimensionality reduction and restoration model and the temporal evolution prediction model are trained offline based on unsteady-state numerical simulation data. After training, they are deployed in the online prediction link. In the online prediction stage, there is no need to re-perform high-cost unsteady-state numerical simulations, thereby achieving second-level real-time dynamic prediction of the temperature field during the dynamic load change process of coal-fired boilers.

Claims

1. A method for predicting the temperature field during dynamic load changes in a coal-fired boiler, characterized in that: Includes the following steps: S1. Read the boiler's real-time operation control parameters based on user-side input or DCS operation data to achieve online prediction; S2. Obtain the unsteady-state numerical simulation temperature field data of the target coal-fired boiler during dynamic load change process, and construct a spatially continuous distribution matrix with boundary constraints in combination with geometric boundary constraints; S3. Construct a spatial feature dimensionality reduction and restoration model, encode the spatial continuous distribution matrix with boundary constraints in step S2, obtain a low-dimensional implicit vector representing the spatial distribution of the temperature field, and the spatial feature dimensionality reduction and restoration model can also restore the spatial continuous distribution matrix with boundary constraints based on the low-dimensional implicit vector. S4. Extract the real-time operation control parameters of the boiler during the dynamic load change process and perform multi-dimensional fusion with the low-dimensional implicit vector of historical time. S5. Construct a time-series evolution prediction model, input the fused features into the time-series evolution prediction model, and predict the low-dimensional implicit vector of the next target time. Use the predicted low-dimensional implicit vector as the state input for subsequent time periods, and perform closed-loop rolling deduction according to the preset time step to obtain the low-dimensional implicit vector group for future time periods. S6. Perform spatial reconstruction on the low-dimensional implicit vector group and output the prediction results of the unsteady temperature field during the future dynamic load change process; S7. Return the prediction results to the user.

2. The method for predicting the temperature field during dynamic load change in a coal-fired boiler according to claim 1, characterized in that: In step S1, the real-time operation control parameters of the boiler include the boiler load change rate, the current load of the boiler, and the burner operating status.

3. The method for predicting the temperature field during dynamic load change in a coal-fired boiler according to claim 1, characterized in that: In step S2, the unsteady-state numerical simulation is based on computational fluid dynamics. Initial conditions for typical operating conditions are obtained from the power plant's DCS operating data and used to generate offline training samples. The initial conditions include coal type, primary air velocity, secondary air velocity, burnout air velocity, primary air temperature, secondary air temperature, burnout air temperature, coal feed rate, and burner tilt angle. The internal temperature field of the furnace is obtained through unsteady-state numerical simulation using a coupled turbulence model, radiation model, discrete phase model, and coke combustion model, resulting in an unsteady-state temperature field sequence that is continuous before and after the initial state, retains thermal inertia, and exhibits response hysteresis. In step S2, the unsteady-state numerical simulation specifically involves: establishing a computational domain and grid consistent with the target coal-fired boiler; inputting the initial conditions of typical operating conditions into a solver based on the finite volume method; solving to obtain the convergent temperature field at the initial moment; then enabling unsteady-state solution; setting the physical time step; and updating the load, coal feed rate, wind speed, wind temperature, and burner swing angle conditions at each time step according to the variable load trajectory using a time series table or a user-defined function; at the nth time step, using the convergent flow field of the previous time step as the initial value, implicit time discretization is adopted to iteratively solve the continuity equation, momentum equation, energy equation, and component transport equation, and coupling the turbulence model, radiation model, discrete phase model, and coke combustion model; after satisfying the convergence requirements of residuals and temperature monitoring quantities, the temperature field at that moment is saved, and then the process is advanced to the next time step until the entire variable load process is completed; In step S2, the geometric boundary constraint is a masking matrix formed based on the actual boiler's geometric parameters. Its constraint effect on the complete mesh matrix after interpolation of the original continuous flow field is as follows: in, This is the complete grid matrix after interpolation of the original continuous flow field; The spatially continuous distribution matrix with boundary constraints is used to train the spatial feature dimensionality reduction and restoration model in step S3 and the temporal evolution prediction model in step S5. It represents the Hadamardi (or Hadama) stack; Let be the masking matrix, which is defined as: in, y This is the row index of the grid matrix; x This is the column index of the grid matrix.

4. The method for predicting the temperature field during dynamic load change in a coal-fired boiler according to claim 3, characterized in that: In step S2, the geometric boundary constraint includes: generating a boundary mask based on the true geometric contour of the target coal-fired boiler furnace or the target cross section, masking or zeroing invalid grid points outside the boundary, and retaining the valid area inside the boundary to participate in model training, inference and result display.

5. The method for predicting the temperature field during dynamic load change in a coal-fired boiler according to claim 1, characterized in that: In step S3, the low-dimensional implicit vector specifically refers to the core representation vector that can completely reproduce the original continuous physical field spatial topology, obtained by projecting a high-dimensional, highly redundant spatially continuous distribution matrix onto a low-dimensional implicit feature subspace through a nonlinear transformation; the spatial feature dimensionality reduction and restoration model includes an encoder. and decoder The encoder will convert the spatially continuous distribution matrix with boundary constraints from step S2. Mapped to low-dimensional implicit vectors The decoder reconstructs the spatially continuous distribution matrix with boundary constraints based on this implicit vector, expressed as: in, The spatial continuous distribution matrix is ​​represented by the decoder; the spatial feature dimensionality reduction and restoration model reduces the size of the matrix. and The reconstruction error between the two is used for training to establish a bidirectional mapping relationship between the spatially continuous distribution matrix and the low-dimensional implicit vector.

6. The method for predicting the temperature field during dynamic load change in a coal-fired boiler according to claim 5, characterized in that: In step S5, the expression for the time-series evolution prediction model is: in, For time-series evolution prediction models, These are model parameters; This is the low-dimensional implicit vector predicted by the model at the previous time step; These are the real-time operating control parameters for the boiler. This is the low-dimensional implicit vector of the current time t predicted by the model; The loss function of the time-series evolution prediction model is defined as: in, The loss value of the time series evolution prediction model at time t; is the dimension of the low-dimensional implicit feature subspace; Characterizing the first in low-dimensional space The index of each feature component; The encoder of the spatial feature dimensionality reduction and restoration model at time t is the implicit vector true value obtained after spatial dimensionality reduction of the temperature field in the unsteady numerical simulation. The implicit vector prediction value at time t is generated by the rolling prediction of the time series evolution prediction model.

7. The method for predicting the temperature field during dynamic load change in a coal-fired boiler according to claim 6, characterized in that: In step S6, the predicted low-dimensional implicit vector at the current time t is... The method for evaluating the error between the spatially continuous distribution matrix obtained after decoder reconstruction and the spatially continuous distribution matrix obtained after geometric boundary constraints on the unsteady numerical simulation temperature field at time t is as follows: in, Let be the transient average relative error at time t; Number of columns in the image; The number of rows in the image; This indicates that at time t, the coordinates are... Unsteady-state numerical simulation temperature at the location; This indicates that at time t, the coordinates are... Deep learning is used to predict temperature at a location.

8. The method for predicting the temperature field during dynamic load change in a coal-fired boiler according to claim 1, characterized in that: The spatial feature dimensionality reduction and restoration model is a nonlinear dimensionality reduction model with spatial feature extraction capability. It maps a high-dimensional spatial continuous distribution matrix to a low-dimensional state space to form a low-dimensional implicit vector, and then inversely maps the low-dimensional implicit vector to a spatial continuous distribution matrix. The time-series evolution prediction model includes a control parameter input branch and a historical low-dimensional implicit vector input branch, which are used to learn the correspondence between the running control parameters and the low-dimensional implicit vector evolution of the temperature field.

9. The method for predicting the temperature field during dynamic load change in a coal-fired boiler according to claim 1, characterized in that: When training the spatial feature dimensionality reduction and restoration model, at least one of temperature field reconstruction consistency constraints, spatial gradient consistency constraints, and local structure consistency constraints is comprehensively adopted to ensure that the restored spatial continuous distribution matrix maintains boundary changes and main spatial distribution features. When training the time-series evolution prediction model, the complete variable load process is used as the sample organization unit, and at least one of low-dimensional implicit vector error and consistency of state changes between adjacent time steps is used to constrain the model. In step S6, the unsteady temperature field prediction results are output in at least one of the following forms: temperature cloud map, time series matrix, dynamic image, key area temperature index, or over-temperature warning information; when the predicted temperature exceeds the preset threshold, a prompt message for boiler operation adjustment is generated.

10. A temperature field prediction system for a dynamic load change process of a coal-fired boiler, used to implement the method as described in any one of claims 1 to 9, characterized in that, This includes online prediction pipelines and offline training and model support pipelines; The online prediction link includes a data reading module, a data preprocessing module, a time series prediction module, and a result output module connected in sequence; The data reading module is used to receive user input or automatically read power plant DCS operation data, obtain real-time boiler operation control parameters, and transmit the real-time boiler operation control parameters to the data preprocessing module. The data preprocessing module is used to perform gridding, normalization and geometric boundary masking constraints on the input data to form input data suitable for subsequent model inference, and then pass the processed data to the time series prediction module. The time series prediction module has a built-in trained time series evolution prediction model, which is used to predict the future low-dimensional implicit vector group based on the real-time operation control parameters of the boiler and the low-dimensional implicit vector of the previous moment through multi-dimensional feature fusion and closed-loop rolling inference, and transmit the prediction results to the result output module. The result output module has a built-in trained decoder, which is used to restore the future low-dimensional implicit vector group into the unsteady temperature field prediction result, and display or warn. The offline training and model support link includes a nonsteady-state numerical simulation data generation module, a geometric boundary definition module, a training sample library, and a model library connected in sequence. The unsteady-state numerical simulation data generation module is used to obtain the unsteady-state temperature field sequence and corresponding boundary conditions of the target coal-fired boiler during dynamic load change process based on computational fluid dynamics methods. The geometric boundary definition module is used to generate the furnace cross section and masking matrix based on the actual geometric parameters of the target coal-fired boiler. The training sample library is used to store training samples, which are used to train spatial feature dimensionality reduction and restoration models and temporal evolution prediction models. The model parameters obtained from the training are stored in the model library. The model library is used to store the parameters of the trained models; The output of the geometric boundary definition module is also connected to the input of the data preprocessing module in the online prediction link, and is used to provide a masking matrix to the data preprocessing module so that the data preprocessing module can perform geometric boundary masking constraint processing on the input data. The output of the model library is also connected to the input of the time series prediction module and the result output module in the online prediction link, respectively, to provide model support to the online prediction link.

Images