A method for combined modeling of secondary air temperature and material level of a vertical cooler

Abstract

The application discloses a secondary air temperature and material level combined modeling method of a vertical cooler and belongs to the vertical cooler field, and comprises the following steps: acquiring working parameter data of the vertical cooler; the working parameter data comprises air pressure distribution data, material inlet temperature data and air blower operation parameter data; the air pressure distribution data is mapped into a material bed resistance distribution matrix according to a spatial topological relationship; the material bed resistance distribution matrix, the material inlet temperature data and the air blower operation parameter data are input into a pre-set space-time mapping network model to obtain a spatial heat exchange feature map; and the spatial heat exchange feature map is associated and mapped through the space-time mapping network model, and a secondary air temperature prediction value and a corresponding material level height representation value are output.

Description

Technical Field

[0001] This invention relates to the field of vertical coolers, and in particular to a method for jointly modeling the secondary air temperature and material level of a vertical cooler. Background Technology

[0002] Vertical coolers are widely used in clinker cooling processes in the metallurgical and building materials industries. They work by blowing cold air in from the bottom, which exchanges heat with the hot material falling from the top, thereby reducing the material temperature and recovering heat.

[0003] In related technologies, the monitoring logic for the internal state of a vertical cooler is usually as follows: secondary air temperature data is collected using a thermocouple sensor installed at the air outlet, and material level height data is collected using a radar rangefinder at the top.

[0004] The related technologies have the following shortcomings: During the heat exchange process inside the vertical cooler, changes in the bed thickness directly alter the permeability of the internal cooling gas and the heat exchange contact area between the material and the gas; the independent calculation models lack a unified assessment that incorporates spatial permeation resistance into the overall heat exchange process. This independent approach means that when encountering sudden changes in feed rate, the calculation models cannot reflect the changes in local wind resistance and heat transfer lag caused by changes in material level, resulting in a discrepancy between the predicted and actual values ​​of the secondary air temperature. Summary of the Invention

[0005] This invention provides a method for jointly modeling the secondary air temperature and material level of a vertical cooler, which can solve the problems existing in the prior art.

[0006] A method for jointly modeling the secondary air temperature and material level in a vertical cooler includes: Obtain the operating parameter data of the vertical cooler; the operating parameter data includes air pressure distribution data, feed temperature data, and blower operating parameter data; The wind pressure distribution data is mapped into a bed resistance distribution matrix according to spatial topology; The bed resistance distribution matrix, the feed temperature data, and the blower operating parameter data are input into a pre-set spatiotemporal mapping network model to obtain a spatial heat exchange characteristic map. The spatial heat exchange feature map is correlated and mapped by the spatiotemporal mapping network model, and the secondary wind temperature prediction value and the corresponding material level height characterization value are output.

[0007] Based on the above scheme, the shortcomings of independent calculation models in reflecting the lag effect of material level changes on wind resistance and heat transfer are overcome, and the consistency between predicted and actual values ​​can still be maintained under non-steady-state conditions such as sudden changes in feed rate. Attached Figure Description

[0008] Figure 1 The flowchart illustrates a method for jointly modeling the secondary air temperature and material level of a vertical cooler, as provided by this invention. Detailed Implementation

[0009] The specific embodiments of the present invention will be described in detail below, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0010] The present invention provides a method for joint modeling of secondary air temperature and material level in a vertical cooler, comprising: S101. Obtain the operating parameter data of the vertical cooler; the operating parameter data includes air pressure distribution data, feed temperature data, and blower operating parameter data.

[0011] Specifically, multiple pressure sensors are arranged along different height layers in the annular section inside the vertical cooler. Each sensor transmits the detected air pressure value to the control unit in real time. A temperature sensor is installed at the material inlet to detect the feed temperature entering the cooler. An air volume detection device is installed at the air volume regulating valve of the blower to output the air volume characterization signal to the control unit.

[0012] S102. The wind pressure distribution data is mapped into a bed resistance distribution matrix according to the spatial topology.

[0013] Specifically, the internal cross-section of the vertical cooler is divided into a regular two-dimensional grid. Each pressure sensor has a fixed installation position within the vertical cooler cross-section, and its corresponding origin coordinates can be determined in the grid based on spatial topology. The grid nodes at the origin coordinates use the wind pressure value of this sensor as the initial value. For the remaining grid nodes not directly covered by the sensor, the control unit uses the wind pressure values ​​of adjacent sensors to calculate their resistance data using an interpolation algorithm, thereby expanding the wind pressure distribution data collected at each measuring point into a bed resistance distribution matrix covering the entire cross-section. The row and column directions of this bed resistance distribution matrix correspond to the width and depth directions of the vertical cooler cross-section, respectively. The value of each element in the matrix represents the resistance of the bed to the airflow at the corresponding location. The larger the value, the denser the bed at that location, and the more difficult it is for the airflow to pass through.

[0014] S103. Input the bed resistance distribution matrix, the feed temperature data, and the blower operating parameter data into a pre-set spatiotemporal mapping network model to obtain a spatial heat exchange characteristic map.

[0015] Specifically, the spatiotemporal mapping network model receives three types of input data: the bed resistance distribution matrix as spatial structural information, reflecting the airflow capacity distribution in each region of the bed; the feed temperature data as heat source boundary conditions, reflecting the heat level carried by the material entering the cooler; and the blower operating parameter data as cold source boundary conditions, reflecting the air supply parameters on the cooling air side. The spatiotemporal mapping network model internally includes a feature extraction network, which uniformly encodes and abstracts the resistance distribution pattern in the spatial structural information and the temperature and air supply parameter variation trends in the time series information, generating a spatial heat exchange feature map. Each value in this feature map encapsulates a comprehensive heat exchange state representation of the bed resistance at the corresponding spatial location and the temperature boundary and air supply conditions under the current operating conditions.

[0016] S104. The spatial heat exchange feature map is correlated and mapped through the spatiotemporal mapping network model, and the secondary wind temperature prediction value and the corresponding material level height characterization value are output.

[0017] Specifically, after obtaining the spatial heat exchange feature map, the spatiotemporal mapping network model maps the spatial heat exchange state representation in the feature map to the target output space through the association mapping layer.

[0018] The correlation mapping layer, according to a pre-defined feature-output correspondence, maps the numerical distribution reflecting heat transfer intensity in the feature map to predicted secondary air temperature values. These values ​​represent the predicted temperature of the secondary air at the outlet under the current bed resistance distribution and air supply conditions. The layer also maps the numerical values ​​reflecting the resistance distribution scale in the feature map to material level height representation values, which represent the bed thickness level corresponding to the current predicted secondary air temperature. Since both the predicted secondary air temperature and the material level height representation values ​​originate from the correlation mapping of the same spatial heat exchange feature map, a correspondence is established between them through the internal structure of the model, achieving joint modeling of secondary air temperature and material level.

[0019] As can be seen, by adopting the above scheme, the spatial permeation resistance distribution of the material bed in the vertical cooler, the feed temperature, and the blower operating parameters are uniformly input into the spatiotemporal mapping network model for joint modeling. This establishes a correspondence between the predicted secondary air temperature and the material level height based on the spatial heat exchange characteristic map. This overcomes the shortcomings of independent calculation models that cannot reflect the lag effect of material level changes on air resistance and heat transfer. It can still maintain the consistency between the predicted and actual values ​​under non-steady-state conditions such as sudden changes in feed volume.

[0020] In some embodiments, obtaining the operating parameter data of the vertical cooler includes the following methods: The measured wind pressure values ​​of multiple pressure sensors installed in the annular section inside the vertical cooler are obtained, and the wind pressure distribution data is generated by sliding sampling of the measured wind pressure values ​​within a preset time window. The measured value of the material's inlet temperature into the vertical cooler is obtained and the inlet temperature data is generated by sliding sampling according to a preset time window; The air volume characterization signal of the cooling air supply equipment is acquired and the blower operating parameter data is generated by sliding sampling according to a preset time window.

[0021] In implementation, multiple pressure sensors are arranged in layers along the circumference and height of the annular cross-section inside the vertical cooler. Each sensor transmits the measured wind pressure value to the acquisition module of the control unit in real time. The control unit maintains an annular buffer with a preset time window length. Whenever the preset time window is filled with newly acquired data, the statistical value of all data within the window, such as the mean or median, is used as the sampling output for that time window. Then, the window slides forward one step to continue receiving subsequent data. The sliding sampling of wind pressure distribution data covers all measuring points of all pressure sensors; the sliding sampling of feed temperature data is performed on the material inlet temperature sensor, using the same preset time window and step size as the wind pressure data; the sliding sampling of blower operating parameter data is extracted from the air volume characterization signal at the air volume regulating valve, also using a uniform preset time window and step size.

[0022] The above technical solution uses a unified preset time window to synchronously slide sample the three types of signals from the pressure sensor, temperature sensor, and air volume detection device, so that the air pressure distribution data, feed temperature data, and blower operating parameter data are aligned in the time direction, which is used by the spatiotemporal mapping network model to obtain time-synchronized input data.

[0023] In some embodiments, mapping wind pressure distribution data into a bed resistance distribution matrix according to spatial topological relationships includes the following methods: Obtain the mesh generation parameters for the internal cross-section of the vertical cooler; The origin coordinates of each pressure sensor in the grid are determined based on spatial topology; The bed resistance distribution matrix is ​​obtained by interpolation calculation of the grid nodes outside the origin coordinates.

[0024] During implementation, the meshing parameters include the number of rows and columns of the mesh, as well as the size of each individual mesh. These parameters are preset based on the cross-sectional dimensions of the vertical cooler and the arrangement density of the pressure sensors. The control unit reads the installation coordinates of each pressure sensor on the cross-section of the vertical cooler and converts them into row and column numbers in the mesh according to the spatial topology, using these as the origin coordinates.

[0025] For the grid nodes at the origin coordinates, the wind pressure value of this sensor is directly used as the resistance data; for the remaining grid nodes not directly covered by the sensor, the control unit performs interpolation calculations to calculate the resistance data of each node using the wind pressure values ​​at multiple origin coordinates around it, thereby generating a bed resistance distribution matrix covering the entire cross section; the interpolation method makes the resistance values ​​of adjacent grid nodes in the matrix have spatial continuity, which can reasonably reflect the gradual distribution characteristics of bed resistance on the cross section.

[0026] By using the above technical solution, the wind pressure data of each measuring point is extended into a continuous bed resistance distribution matrix through grid division and interpolation calculation based on sensor location. This enables the model to obtain the spatial resistance distribution information of the entire vertical cooler cross section, rather than relying solely on the observation values ​​of a few measuring points.

[0027] In some embodiments, the process of inputting the bed resistance distribution matrix, feed temperature data, and blower operating parameter data into a spatiotemporal mapping network model to extract spatial heat exchange feature maps includes the following methods: In the first processing branch of the spatiotemporal mapping network model, two-dimensional feature extraction is performed on the bed resistance distribution matrix to generate spatial permeability resistance features; In the second processing branch, one-dimensional feature extraction is performed on the feed temperature data and blower operating parameter data to generate time boundary features; By integrating the spatial permeability resistance characteristics and the temporal boundary characteristics, a spatial heat exchange characteristic map is obtained.

[0028] In implementation, the spatiotemporal mapping network model contains two parallel processing branches. The first branch, targeting the bed resistance distribution matrix, uses a two-dimensional convolutional structure to perform sliding window operations along the rows and columns of the matrix, extracting spatial permeation resistance features reflecting the bed resistance distribution patterns in different regions layer by layer. The second branch, targeting the feed temperature data and blower operating parameter data, uses a one-dimensional convolutional or recurrent network structure to extract temporal boundary features reflecting the changing trends of temperature and air supply parameters along the time direction. The features extracted by the two branches are combined in a preset manner in the fusion layer, comprehensively processing the corresponding positional values ​​of the spatial permeation resistance features and the temporal boundary features to generate a spatial heat exchange feature map that simultaneously contains spatial resistance distribution information and temporal boundary condition information.

[0029] The above technical solution utilizes the first and second processing branches to process spatial structure information and time series information respectively, enabling the model to simultaneously capture the spatial distribution characteristics of bed resistance and the temporal variation characteristics of temperature and air supply parameters, and unify the two types of information into a spatial heat exchange feature map through feature fusion.

[0030] In some embodiments, the spatiotemporal mapping network model is generated through the following training steps: Historical sample data of the vertical cooler is obtained, including historical operating parameter data, corresponding actual secondary air temperature data, and corresponding actual material level height data. The historical operating parameter data is input into the initial mapping network model to output sample air temperature prediction values ​​and sample material level prediction values. Constraint verification is performed based on the sample air temperature prediction values, sample material level prediction values, and historical sample data to calculate the constraint loss function. The initial mapping network model is updated according to the constraint loss function until the preset convergence condition is reached, generating a spatiotemporal mapping network model.

[0031] During implementation, historical sample data is extracted from the historical operation records of the vertical cooler. Each sample contains operating parameter data for a time window, as well as the actual secondary air temperature and actual material level height data at the corresponding moment. The actual material level height data can be obtained through actual measurement using a radar rangefinder. At the start of training, the weight parameters of the initial mapping network model are randomly assigned. The control unit inputs the historical operating parameter data into the initial mapping network model in batches. The model calculates and outputs the sample air temperature prediction value and sample material level prediction value through forward propagation. Then, the control unit performs constraint verification, comparing the sample air temperature prediction value and sample material level prediction value with the actual secondary air temperature data and actual material level height data, respectively, applying additional constraints, and calculating the value of the constraint loss function. Based on this value, the model's weight parameters are updated according to the gradient descent direction. The above process is repeated iteratively on all historical sample data until the value of the constraint loss function drops below the preset convergence condition. The network model obtained at this point is the trained spatiotemporal mapping network model.

[0032] The above technical solution uses historical operating data of vertical coolers to train the model and introduces constraint verification during the training process. This allows the model to fit historical data while meeting preset constraints, thus generating a spatiotemporal mapping network model that can more accurately reflect the actual heat exchange characteristics of vertical coolers.

[0033] In some embodiments, the process of calculating the constraint loss function based on the sample wind temperature prediction value, the sample material level prediction value and historical sample data includes the following steps: calculating the first deviation value between the sample material level prediction value and the actual material level height data; calculating the second deviation value between the sample wind temperature prediction value and the actual secondary wind temperature data; and constructing the constraint loss function based on the first deviation value and the second deviation value.

[0034] During implementation, for each batch of input sample data, the control unit compares the model's output sample material level prediction value with the actual material level height data sample by sample, calculating the deviation between the two as the first deviation value, which reflects the model's prediction accuracy for material level height. Simultaneously, it compares the sample wind temperature prediction value with the actual secondary wind temperature data, calculating the deviation between the two as the second deviation value, which reflects the model's prediction accuracy for secondary wind temperature. Then, the first and second deviation values ​​are used to construct a constrained loss function according to a preset weight ratio or combination. This loss function simultaneously measures the model's comprehensive performance across both the material level prediction and wind temperature prediction objectives.

[0035] By using the above technical solution, a constraint loss function is constructed by jointly using the material level prediction deviation and the wind temperature prediction deviation, so that the training process of the model considers two prediction objectives at the same time, avoiding the performance degradation of the other objective that may be caused by optimizing a single objective.

[0036] In some embodiments, the process of interpolating grid nodes other than the origin coordinates to obtain the bed resistance distribution matrix includes the following steps: calculating the straight-line distance from the target grid node to each origin coordinate; using the reciprocal of the straight-line distance as the distance weight allocation coefficient; multiplying the wind pressure values ​​corresponding to each pressure sensor in the wind pressure distribution data by the corresponding distance weight allocation coefficient to generate the first product data; summing the first product data and dividing by the sum of all distance weight allocation coefficients to generate the resistance pixel value of the target grid node; and traversing all grid nodes to stitch together the resistance pixel values ​​to generate the bed resistance distribution matrix.

[0037] During implementation, the control unit determines the origin coordinates of all pressure sensors in the grid and processes each target grid node not covered by the origin coordinates sequentially. For the current target grid node, the straight-line distance from it to each origin coordinate is calculated, based on the difference between the node's row and column numbers and the origin coordinates. The reciprocal of each straight-line distance is used as the distance weighting coefficient for the corresponding sensor; the closer the origin coordinate is, the larger the weighting coefficient it receives, and the greater its contribution to the wind pressure value of the target node. The wind pressure values ​​of each sensor are multiplied by the corresponding distance weighting coefficient to obtain the first product data. All first product data are summed, and the sum is divided by the sum of all distance weighting coefficients; the quotient is the resistance pixel value of the current target grid node. After the control unit traverses all target grid nodes to complete the calculation, it stitches the resistance pixel values ​​into a complete bed resistance distribution matrix according to the grid's row and column order.

[0038] By employing the above technical solution and using the distance reciprocal weighted interpolation method, the resistance value of any grid node in the bed resistance distribution matrix is ​​calculated by weighting the wind pressure values ​​of all sensors. The closer the sensor is to the target node, the greater its contribution, thus ensuring the spatial rationality and smoothness of the interpolation results.

[0039] In some embodiments, the process of fusing spatial permeability resistance features and temporal boundary features to obtain a spatial heat exchange feature map includes the following steps: performing dimensional expansion processing on the temporal boundary features to generate a feature sequence with the same dimension as the spatial permeability resistance features; calculating the element-wise product of the feature sequence and the corresponding element values ​​in the spatial permeability resistance features to generate second product data; and performing numerical compression transformation on the second product data through a normalization function to generate a spatial heat exchange feature map.

[0040] In implementation, the dimensionality of the spatial permeability resistance features generated by the first processing branch is determined by the size of the bed resistance distribution matrix and the number of feature channels; the dimensionality of the time boundary features generated by the second processing branch is usually lower. The control unit performs dimensionality expansion processing on the time boundary features, expanding the low-dimensional time boundary features into a feature sequence with the same dimension as the spatial permeability resistance features through copying or interpolation, so that each spatial location in the feature sequence has a corresponding time feature value. Element-wise multiplication is performed on the elements at corresponding positions in the feature sequence and the spatial permeability resistance features to obtain the second product data, which reflects the interaction strength between the spatial resistance features and the time boundary features at each location. The second product data is then normalized to compress the values ​​to a preset range, outputting a spatial heat exchange feature map. The normalization operation ensures that the feature maps generated by different batches of input have comparable numerical scales.

[0041] The above technical solution involves expanding the dimensionality of the time boundary features and performing element-wise product fusion with the spatial permeability resistance features. This injects the boundary condition information in the time series into the spatial resistance features at each spatial location. Then, normalization compression is used to stabilize the feature values ​​within a reasonable range, thereby enhancing the effectiveness of feature fusion and the numerical stability of the output features.

[0042] In some embodiments, the process of constructing a constraint loss function based on a first deviation value and a second deviation value includes the following steps: obtaining the specific heat capacity parameter of the material, the specific heat capacity parameter of the cooling gas, and preset reference temperatures for the blower's cold air and the material discharge; converting the predicted sample material level into the estimated mass of the material, and calculating the first temperature difference between the historical temperature data in the feed temperature data and the reference temperature of the material discharge; calculating a first thermodynamic estimation term using the estimated mass of the material, the specific heat capacity parameter of the material, and the first temperature difference; converting the historical blower airflow in the historical operating parameter data into the estimated mass of the cooling gas, and calculating the second temperature difference between the predicted sample airflow temperature and the reference temperature of the blower's cold air; calculating a second thermodynamic estimation term using the estimated mass of the cooling gas, the specific heat capacity parameter of the cooling gas, and the second temperature difference; calculating the absolute value of the difference between the first thermodynamic estimation term and the second thermodynamic estimation term to generate an energy conservation constraint term; calculating the root mean square error of the first deviation value to generate a first loss term; calculating the root mean square error of the second deviation value to generate a second loss term; and adding the first loss term, the second loss term, and the energy conservation constraint term to generate a constraint loss function.

[0043] During implementation, the specific heat capacity parameters of the material and the cooling gas are obtained from a preset parameter table based on the type of material being processed by the vertical cooler and the composition of the cooling gas. The reference temperature of the blower's cold air is taken as the nominal temperature of the cold air at the blower inlet, and the reference temperature of the material discharge is taken as the target discharge temperature of the material required by the process. The control unit converts the material level height into the estimated mass of the material based on the sample material level prediction value, the cross-sectional area of ​​the cooler, and the bulk density of the material. The difference between the historical temperature data in the feed temperature data and the reference temperature of the material discharge is calculated as the first temperature difference. The first thermodynamic estimation term is calculated using the product of the estimated mass of the material, the specific heat capacity parameter of the material, and the first temperature difference. This term represents the theoretical heat released by the material during the cooling process. Then, the control unit converts the air volume into the estimated mass of the cooling gas based on the historical air volume and gas density of the blower. The difference between the sample air temperature prediction value and the reference temperature of the blower's cold air is calculated as the second temperature difference. The second thermodynamic estimation term is calculated using the product of the estimated mass of the cooling gas, the specific heat capacity parameter of the cooling gas, and the second temperature difference. This term represents the theoretical heat absorbed by the cooling gas during the heat exchange process. Subsequently, the absolute value of the difference between the first and second thermodynamic estimation terms is calculated as the energy conservation constraint term, which reflects the degree of imbalance between the heat released by the material and the heat absorbed by the gas. Simultaneously, the root mean square errors of the first and second deviation values ​​are calculated to obtain the first and second loss terms. The first loss term, the second loss term, and the energy conservation constraint term are then added together to obtain the constraint loss function.

[0044] Through the above technical solution, the constraint loss function introduces an energy conservation constraint term based on thermodynamic principles in addition to the prediction accuracy loss term. This makes the model not only pursue the numerical consistency between the predicted value and the true value during the training process, but also subject to the constraint of the heat balance relationship. This is conducive to the model learning a mapping relationship that conforms to thermodynamic laws and reduces the generation of theoretically unreasonable prediction results.

[0045] In some embodiments, the process of updating the initial mapping network model according to the constraint loss function until a preset convergence condition is reached includes the following steps: calculating the first partial derivative of the constraint loss function with respect to each weight parameter in the initial mapping network model; multiplying each first partial derivative by a preset learning rate hyperparameter to calculate the corresponding weight adjustment amount; subtracting the corresponding weight adjustment amount from the current weight value of the initial mapping network model to generate the corresponding updated weight value; replacing the corresponding current weight value with each updated weight value until the value of the constraint loss function is less than a preset threshold.

[0046] In implementation, after each training iteration, the control unit calculates the constraint loss function value based on the current batch of sample data. It then calculates the partial derivative of this constraint loss function with respect to each weight parameter in the initial mapped network model, obtaining the first partial derivative. The magnitude and sign of the first partial derivative reflect the direction and degree of influence of the corresponding weight parameter on the constraint loss function. Next, the first partial derivative for each weight parameter is multiplied by a pre-set learning rate hyperparameter to obtain the weight adjustment amount for that weight parameter. The learning rate hyperparameter controls the step size of each weight update; a value that is too large may lead to training instability, while a value that is too small will result in slow convergence. Subsequently, the corresponding weight adjustment amount is subtracted from the current weight value of each weight parameter to obtain the updated weight value. All updated weight values ​​replace the current weight values, completing one round of weight updates. The control unit repeats the above steps, calculating the current value of the constraint loss function after each update. When this value drops below a preset threshold, it is considered to have met the preset convergence condition, and the training process ends.

[0047] The above technical solution utilizes a gradient descent-based weight update method, adjusting the model parameters round by round according to the influence direction of the constraint loss function on each weight parameter, and controlling the update step size through the learning rate hyperparameter, so that the model obtains model parameters that reflect the true heat transfer characteristics of the vertical cooler after meeting the convergence condition.

[0048] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the embodiments of the present invention are not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for jointly modeling secondary air temperature and material level in a vertical cooler, characterized in that, include: Obtain the operating parameter data of the vertical cooler; The operating parameter data includes air pressure distribution data, feed temperature data, and blower operating parameter data; The wind pressure distribution data is mapped into a bed resistance distribution matrix according to spatial topology; The bed resistance distribution matrix, the feed temperature data, and the blower operating parameter data are input into a pre-set spatiotemporal mapping network model to obtain a spatial heat exchange characteristic map. The spatial heat exchange feature map is correlated and mapped by the spatiotemporal mapping network model, and the secondary wind temperature prediction value and the corresponding material level height characterization value are output.

2. The method for joint modeling of secondary air temperature and material level in a vertical cooler as described in claim 1, characterized in that, Obtain the operating parameter data of the vertical cooler, including: Obtain the measured wind pressure values ​​from multiple pressure sensors installed in the annular section inside the vertical cooler; The wind pressure distribution data is generated by sliding sampling the measured wind pressure values ​​within a preset time window. The measured value of the material's inlet temperature into the vertical cooler is obtained and the material is sampled by sliding according to the preset time window to generate the inlet temperature data. The air volume characterization signal of the cooling air supply equipment is acquired and slid sampling is performed according to the preset time window to generate the blower operating parameter data.

3. The method for joint modeling of secondary air temperature and material level in a vertical cooler as described in claim 2, characterized in that, The wind pressure distribution data is mapped into a bed resistance distribution matrix according to spatial topological relationships, including: Obtain the mesh generation parameters of the internal cross-section of the vertical cooler; The origin coordinates of each pressure sensor in the grid are determined based on the spatial topology. The bed resistance distribution matrix is ​​obtained by interpolation calculation of the grid nodes other than the origin coordinates.

4. The method for joint modeling of secondary air temperature and material level in a vertical cooler as described in claim 3, characterized in that, The bed resistance distribution matrix, the feed temperature data, and the blower operating parameter data are input into a pre-set spatiotemporal mapping network model to obtain a spatial heat exchange characteristic map, including: In the first processing branch of the spatiotemporal mapping network model, two-dimensional feature extraction is performed on the bed resistance distribution matrix to generate spatial permeability resistance features; In the second processing branch of the spatiotemporal mapping network model, one-dimensional feature extraction is performed on the feed temperature data and the blower operating parameter data to generate time boundary features. The spatial heat exchange feature map is obtained by integrating the spatial permeability resistance feature with the temporal boundary feature.

5. The method for joint modeling of secondary air temperature and material level in a vertical cooler as described in claim 4, characterized in that, The spatiotemporal mapping network model is generated through the following training steps: Obtain historical sample data of the vertical cooler; the historical sample data includes historical operating parameter data, corresponding actual secondary air temperature data, and corresponding actual material level height data; The historical working parameter data is input into the initial mapping network model, and the sample wind temperature prediction value and sample material level prediction value are output. Based on the sample wind temperature prediction value, the sample material level prediction value and the historical sample data, constraint verification is performed, and the constraint loss function is calculated. The initial mapping network model is updated according to the constraint loss function until the preset convergence condition is reached, thereby generating the spatiotemporal mapping network model.

6. The method for joint modeling of secondary air temperature and material level in a vertical cooler as described in claim 5, characterized in that, Based on the predicted sample wind temperature, the predicted sample material level, and the historical sample data, constraint verification is performed, and the constraint loss function is calculated, including: Calculate the first deviation between the predicted sample material level and the actual material level height data; Calculate the second deviation value between the sample wind temperature prediction value and the actual secondary wind temperature data; The constraint loss function is constructed based on the first deviation value and the second deviation value.

7. The method for joint modeling of secondary air temperature and material level in a vertical cooler as described in claim 6, characterized in that, The bed resistance distribution matrix is ​​obtained by interpolating the grid nodes outside the origin coordinates, including: Calculate the straight-line distance from the target mesh node to each of the origin coordinates; The reciprocal of the straight-line distance is used as the distance weighting coefficient; Multiply the wind pressure values ​​corresponding to each pressure sensor in the wind pressure distribution data by the corresponding distance weighting coefficient to generate the first product data; The first product data is summed and divided by the sum of all the distance weight allocation coefficients to generate the resistance pixel value of the target grid node; The resistance distribution matrix of the material bed is generated by traversing all grid nodes and concatenating the resistance pixel values.

8. The method for joint modeling of secondary air temperature and material level in a vertical cooler as described in claim 7, characterized in that, The space heat exchange feature map is obtained by fusing the space permeability resistance feature with the time boundary feature, including: The time boundary features are subjected to dimensional expansion processing to generate a feature sequence with the same dimension as the spatial permeability resistance features; Calculate the element-wise product of the feature sequence and the corresponding element values ​​in the space permeability resistance feature to generate the second product data; The spatial heat exchange feature map is generated by performing a numerical compression transformation on the second product data using a normalization function.

9. The method for joint modeling of secondary air temperature and material level in a vertical cooler as described in claim 8, characterized in that, The constraint loss function is constructed based on the first deviation value and the second deviation value, including: Obtain the specific heat capacity parameters of the material, the specific heat capacity parameters of the cooling gas, and the preset reference temperature of the cold air from the blower and the reference temperature of the material discharge. The sample material level prediction value is converted into the estimated material mass, and the first temperature difference between the historical temperature data in the feed temperature data and the material discharge reference temperature is calculated. The estimated mass of the material, the specific heat capacity parameter of the material, and the first temperature difference are multiplied to calculate the first thermodynamic estimation term; The historical air volume of the blower in the historical operating parameter data is converted into the estimated mass of cooling gas, and the second temperature difference between the sample air temperature prediction value and the blower cold air reference temperature is calculated. The estimated mass of the cooling gas, the specific heat capacity parameter of the cooling gas, and the second temperature difference are multiplied to calculate the second thermodynamic estimation term; Calculate the absolute value of the difference between the first thermodynamic estimation term and the second thermodynamic estimation term to generate the energy conservation constraint term; Calculate the root mean square error of the first deviation value to generate the first loss term; The root mean square error of the second deviation value is calculated to generate the second loss term; The constraint loss function is generated by adding the first loss term, the second loss term, and the energy conservation constraint term together.

10. The method for joint modeling of secondary air temperature and material level in a vertical cooler as described in claim 9, characterized in that, Update the initial mapping network model according to the constraint loss function until a preset convergence condition is reached, including: Calculate the first partial derivative of the constraint loss function with respect to each weight parameter in the initial mapping network model; Each of the first partial derivatives is multiplied by a pre-set learning rate hyperparameter to calculate the corresponding weight adjustment. The corresponding weight adjustment amount is subtracted from the current weight value of the initial mapping network model to generate the corresponding updated weight value; Replace the corresponding current weight value with each of the updated weight values ​​until the value of the constraint loss function is less than a preset threshold.

Images