A method and system for predicting the heat exchange coefficient of a heat exchanger of a desulfurization tower of a thermal power plant

Abstract

The present application belongs to the field of desulfurization tower heat exchange performance monitoring and optimization control, and discloses a kind of power plant desulfurization tower heat exchanger heat exchange coefficient prediction method and system, the multivariate operating data of desulfurization tower heat exchanger system is collected by multiple sensors, and the graph structure data set containing the space relationship of measuring point is constructed after pretreatment;Design time decoupling module, decompose operating sequence into trend and periodic component, respectively extract long-term trend feature and periodic fluctuation feature and carry out fusion;Fusion operating condition mode, measuring point space attention and graph structure information, construct space-time feature matrix representing time correlation and multi-measuring point space correlation simultaneously;Through heat exchange coefficient prediction module, generate future multi-step heat exchange coefficient prediction sequence.The present application can effectively fuse time and space dependence and operating condition mode information in system operation, realize accurate, multi-step prediction of heat exchange coefficient, and provide reliable basis for desulfurization tower heat exchange performance evaluation and operation optimization control.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of monitoring and optimizing the heat exchange performance of desulfurization towers, and particularly relates to a method and system for predicting the heat exchange coefficient of a desulfurization tower heat exchanger in a thermal power plant. Background Technology

[0002] The heat exchanger in the desulfurization tower is a key piece of equipment in the flue gas desulfurization system of thermal power plants. Its core function is to reduce the flue gas temperature to adapt to subsequent processes through heat exchange between the slurry and the high-temperature flue gas, while simultaneously recovering heat to improve system energy efficiency. The performance of the heat exchanger is directly reflected in its heat transfer coefficient, and its changes can sensitively reflect degradation conditions such as scaling, corrosion, and blockage. It is a key indicator for evaluating heat exchange efficiency, predicting maintenance needs, and optimizing operational control. Therefore, achieving accurate and online prediction of the heat transfer coefficient of the desulfurization tower is of great significance for ensuring the stable operation of the desulfurization system, reducing energy consumption, and enabling preventative maintenance.

[0003] Currently, the mainstream methods for performance prediction and condition assessment of industrial process equipment can be broadly categorized into two types: mechanistic model-based methods and data-driven methods. Mechanism-based methods rely on precise physicochemical equations to describe the heat and mass transfer processes. However, in the complex and variable operating environment of desulfurization towers (such as flue gas composition fluctuations, slurry characteristic changes, and non-uniform scaling), the models often require numerous simplification assumptions, resulting in poor adaptability and limited prediction accuracy, making them difficult to widely apply in actual operation.

[0004] Existing technologies for predicting the heat transfer coefficient of desulfurization towers generally suffer from several drawbacks, including insufficient utilization of spatial correlation information among multiple measurement points, inadequate decoupling and extraction of temporal features from time series data, lack of adaptive modeling for multiple operating conditions, and insufficient accuracy in multi-step prediction. Therefore, there is an urgent need for an intelligent prediction method for the heat transfer coefficient of desulfurization towers that can deeply integrate spatiotemporal features, adapt to different operating conditions, and achieve high-precision multi-step prediction.

[0005] Based on the above analysis, the urgent technical problems that need to be solved in the existing technology are:

[0006] Existing prediction methods suffer from insufficient accuracy and weak adaptability.

[0007] The internal structure of desulfurization towers is complex. Existing methods for predicting heat transfer coefficients simply splice together data from multiple measurement points or model them independently, failing to adequately characterize the spatial coupling relationships between measurement points and their correlation mechanisms as they change with operating conditions. The internal parameters of desulfurization towers fluctuate dramatically, making it difficult to extract stable and effective non-stationary characteristics that distinguish between trends and cycles. Furthermore, it is difficult to account for both long-term degradation and short-term fluctuations, resulting in weak generalization ability across operating conditions and making it difficult to accurately and effectively predict heat transfer coefficients. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a method and system for predicting the heat transfer coefficient of a desulfurization tower heat exchanger in a thermal power plant.

[0009] This invention is implemented as follows: a method for predicting the heat transfer coefficient of a heat exchanger in a desulfurization tower of a thermal power plant, the method specifically including:

[0010] Step S1: Use multiple sensors arranged on the heat exchanger body of the desulfurization tower and its inlet and outlet pipelines to collect multivariate operating data. Perform noise reduction, missing data interpolation and standardization on the collected raw data, and divide it according to the preset time window and sliding step size to obtain a multivariate operating sample sequence. At the same time, construct the measurement point relationship diagram and normalized adjacency matrix according to the heat exchanger structure and measurement point location to form the basic dataset for subsequent network modeling.

[0011] Step S2: Input the multivariate running sample sequence into the time decoupling module, decompose it into a trend part and a periodic part through moving average, and then superimpose time embedding and spatial embedding based on the measurement point relationship graph through linear mapping, respectively. In the trend branch, the long-term running trend features are extracted using a multilayer perceptron, and in the periodic branch, the periodic fluctuation features are extracted using frequency domain multi-head self-attention. The two outputs are fused to obtain a time feature matrix that represents the time correlation of long-term operating conditions.

[0012] Step S3: Divide the time feature matrix obtained in S2 into segments according to preset time intervals and extract the operating condition features. Obtain typical operating condition patterns through clustering, and count the frequency of pattern consistency of each pair of measuring points in each time period to construct the operating condition pattern correlation matrix. In the spatial multi-head self-attention module, calculate the spatial attention weight between measuring points based on the time features, and fuse it with the operating condition pattern correlation matrix and the normalized adjacency matrix to obtain a spatiotemporal feature matrix that has both time correlation and multi-measuring point spatial correlation.

[0013] Step S4: Input the spatiotemporal feature matrix obtained in step S3 into the heat transfer coefficient prediction module, and map it into a multi-step initial prediction sequence of heat transfer coefficients for several future times through a fully connected layer or a one-dimensional convolutional layer. Apply self-attention with Gaussian time weights to the prediction sequence to model the correlation between different prediction steps, and obtain the final multi-step heat exchanger heat transfer coefficient prediction result, which is used for desulfurization tower heat transfer performance evaluation and operation optimization control.

[0014] Furthermore, step S1 includes the following specific steps:

[0015] Step S11: Collect operational data using multiple sensors installed on the heat exchanger body and inlet / outlet pipelines of the desulfurization tower in the thermal power plant, and process the raw time series collected by the sensors according to a uniform sampling period. Time synchronization and resampling are performed to construct a multivariate runtime data matrix. The specific calculation method is as follows:

[0016] ;

[0017] in, For the first One sample, Indicates a time step. Indicates the sensor dimension.

[0018] Step S12: Clean the multivariate running data matrix: Impute missing values, remove or replace outliers to obtain a cleaned and standardized data matrix. Then, use a sliding window (window length L) to segment the data matrix along the time axis to obtain the nth sample segment of length L. The sliding window segmentation described above yields a set of multivariate running sample fragments, which serve as the foundation dataset for network training and prediction.

[0019] Step S13: Construct a measurement point relationship diagram: For the set of measurement points, The connection relationship between measuring points; the connection between any two measuring points is determined based on the fluid connectivity within the heat exchanger. and Determine whether they are adjacent, and construct an adjacency matrix accordingly. ,like and If directly connected or the distance is less than a preset threshold, then ,otherwise Normalize the adjacency matrix to obtain the normalized adjacency matrix. The specific calculation method is as follows:

[0020] ;

[0021] in, Represents the identity matrix. This is the angle matrix.

[0022] Furthermore, step S2 includes the following steps:

[0023] Step S21: Take each sample fragment obtained in step S1 The input time decoupling module uses a moving average (window L) to smooth the sequence of each measurement point to obtain the trend component. Then, subtract the trend component from the sequence to obtain the periodic component. The specific calculation method is as follows:

[0024] ;

[0025] Through the above processing, each sample segment is decomposed into a trend component sequence. and periodic component sequences This serves as the input for step S22.

[0026] Step S22: The trend component sequence obtained in step S21... With periodic component sequences The physical quantities are input into a linear mapping layer, and the feature dimensions of each physical quantity are unified to obtain an initial trend feature tensor with the same dimensions as the original sample fragment. and periodic initial feature tensor Using the normalized adjacency matrix obtained in step S1 Feature propagation is performed using a graph convolutional neural network to obtain spatial embedding vectors representing the internal structural relationships of the heat exchanger. Combined with the positional information at each time step, trend feature tensors with spatiotemporal embedding are formed. and periodic feature tensors with spatiotemporal embedding .

[0027] Step S23, The trend-time feature vector is obtained by inputting the multilayer perceptron. At the same time, for In the frequency domain, multi-head self-attention is used to extract periodic features, and then an inverse Fourier transform is performed to restore them to the time domain, resulting in a periodic time feature tensor. .

[0028] Step S24: Convert the trend time feature tensor generated in step S23 into a multi-dimensional vector. With periodic time feature tensor Aligning each measurement point according to its dimension and time step, the feature matrix is ​​fused by element-wise addition or linear mapping after concatenation on the feature channels to obtain a time feature matrix that simultaneously contains both long-term trend information and cyclical fluctuation information. .

[0029] Furthermore, step S3 includes the following specific steps:

[0030] Step S31: The time feature matrix obtained in step 2 The time feature vectors are used as input to the operating condition mode extraction module and segmented along the time feature dimension of the matrix to form the operating condition feature description vectors for that time period. The obtained operating condition feature description vectors are then used as input for unsupervised clustering analysis to obtain the operating condition mode numbers representing the operating conditions. ,in This is a time period number.

[0031] Step S32: Statistical analysis of any two measurement points and The frequency of operating under the same mode of operation within the same time period is calculated using the following method:

[0032] ;

[0033] in, Represents the total time period. and Indicates the first Measurement points within a time period and The number assigned to the operating mode. This is an indicator function.

[0034] Based on the consistent frequency of the aforementioned operating mode The operating condition correlation matrix is ​​constructed, and the specific calculation method is as follows:

[0035] ;

[0036] in, Set a preset threshold for the weighting coefficient of the operating mode. ,when Greater than season Take the maximum value; when Less than season Take the minimum value. This yields the operating condition correlation matrix. This is used for subsequent spatial association modeling.

[0037] Step 33: Obtain the time feature matrix from step S24. The input spatial multi-head attention module calculates the correlation of all measurement point features at each time step, and obtains the spatial attention weight matrix through the softmax activation function. .Will and Element-wise multiplication yields the fusion space weight matrix. . use right Perform spatial weighted aggregation and combine it with a normalized adjacency matrix. The features of each measurement point are updated by an M-layer graph convolutional neural network to obtain a spatiotemporal feature matrix that simultaneously reflects temporal dependence and spatial correlation. The specific calculation method is as follows:

[0038] ;

[0039] in, This represents an M-layer graph convolutional neural network.

[0040] Furthermore, step S4 includes the following specific steps:

[0041] Step S41: Define the heat exchanger coefficient prediction as a regression prediction task.

[0042] Step S42: Use the set of time feature matrices obtained in step S3 as the input to the prediction layer, and map it through a linear layer to form the initial heat transfer coefficient prediction matrix for the next P time steps. The specific calculation method is as follows:

[0043] ;

[0044] in, , These are the weight matrix and the bias coefficients, respectively. For feature dimensions.

[0045] Step S43: The initial prediction matrix obtained in step S42 The prediction sequence features are rearranged along the prediction step dimension to obtain a prediction part feature of length P, and attention is obtained through linear mapping. The specific calculation method is as follows:

[0046] ;

[0047] in, , , For linear attention, the query vector, key vector, and value vector are given. , , These are the weight matrices for the query vector, key vector, and value vector of linear attention, respectively.

[0048] Step S44: Introduce Gaussian time modeling to construct the Gaussian time feature matrix. The specific calculation method is as follows:

[0049] ;

[0050] in, and For the prediction step index, and These are the mean and standard deviation of the predicted sequence values, respectively.

[0051] The obtained Gaussian time feature matrix is ​​multiplied element-wise with the linear attention matrix obtained in step S43, then normalized using softmax, and then... After weighting, the corrected Gaussian time prediction feature matrix is ​​obtained, and the specific calculation method is as follows:

[0052] ;

[0053] in, This is the attention scaling factor.

[0054] The obtained modified Gaussian time prediction feature matrix is ​​restored to the heat transfer coefficient space through inverse linear mapping to obtain multi-step prediction results. The specific calculation method is as follows:

[0055] ;

[0056] in, and These are the weight matrix and bias coefficients for the multi-step prediction results, respectively.

[0057] Step S45: The initial prediction matrix obtained in step S43 is... The heat exchanger coefficient is predicted using the trained model in the input regression layer with root mean square error loss. The root mean square error loss is calculated as follows:

[0058] ;

[0059] in, This is the predicted heat transfer coefficient sequence after the regression layer expansion. This is the actual heat transfer coefficient sequence.

[0060] Another objective of this invention is to provide a system for predicting the heat transfer coefficient of a desulfurization tower heat exchanger in a thermal power plant, the system specifically comprising:

[0061] The data acquisition and preprocessing module is used to collect multivariate operational data using multiple sensors, perform noise reduction, missing data interpolation and standardization on the collected raw data, and segment it according to a preset time window and sliding step size to obtain a multivariate operational sample sequence. At the same time, it constructs a measurement point relationship diagram and a normalized adjacency matrix based on the heat exchanger structure layout and measurement point location to form a basic dataset for subsequent network modeling.

[0062] The time decoupling feature extraction module decomposes the data into a trend part and a periodic part through a moving average. The trend part is then subjected to linear mapping and time embedding and spatial embedding based on the measurement point relationship graph. In the trend branch, a multilayer perceptron is used to extract long-term operating trend features, and in the periodic branch, a frequency domain multi-head self-attention is used to extract periodic fluctuation features. The two outputs are fused to obtain a time feature matrix that represents the time correlation of long-term operating conditions.

[0063] The spatial relationship feature mixing module is used to segment the time feature matrix according to a preset time interval and extract the operating condition features. Typical operating condition patterns are obtained through clustering. The frequency of pattern consistency of each pair of measuring points in each time period is counted to construct the operating condition pattern correlation matrix. Based on the time features, the spatial attention weight between measuring points is calculated and fused with the operating condition pattern correlation matrix and the normalized adjacency matrix to obtain a spatiotemporal feature matrix that has both time correlation and multi-measuring point spatial correlation.

[0064] The prediction step correlation modeling module is used to input the obtained spatiotemporal feature matrix into the heat transfer coefficient prediction module. It is then mapped to a multi-step heat transfer coefficient initial prediction sequence for several future times through a fully connected layer or a one-dimensional convolutional layer. Self-attention with Gaussian time weights is applied to this prediction sequence to model the correlation between different prediction steps, thus obtaining the final multi-step heat transfer coefficient prediction result of the heat exchanger.

[0065] The heat transfer coefficient prediction output module is used to output the predicted heat transfer coefficient of the multi-step heat exchanger, which is used for the heat transfer performance evaluation and operation optimization control of the desulfurization tower.

[0066] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0067] The proposed method for predicting the heat transfer coefficient of a desulfurization tower heat exchanger in a thermal power plant utilizes the spatiotemporal correlation between multi-sensor, multi-variable operating data. It designs a hierarchical fusion network with temporal decoupling and operating condition pattern recognition, meticulously extracting temporal features by separating and extracting trend and periodic components. Furthermore, it constructs a unified spatiotemporal feature representation by fusing spatial attention, operating condition correlation matrices, and graph structures, thereby applying temporal correlation modeling to the multi-step prediction sequence. These techniques integrate temporal decomposition of multivariable sequences, clustering and identification of typical operating conditions, and spatial modeling using graph structures and attention. Therefore, this invention can more fully utilize the system's spatiotemporal dependence and operating condition information, significantly improving the accuracy and adaptability of multi-step heat transfer coefficient prediction, and providing a new solution for desulfurization tower performance evaluation and optimization. Attached Figure Description

[0068] Figure 1 This is a flowchart of the method for predicting the heat transfer coefficient of a desulfurization tower heat exchanger in a thermal power plant, provided in an embodiment of the present invention.

[0069] Figure 2 This is a flowchart of signal acquisition provided in an embodiment of the present invention;

[0070] Figure 3 This is a module diagram of the heat transfer coefficient prediction system for the desulfurization tower heat exchanger in a thermal power plant provided in an embodiment of the present invention;

[0071] Figure 4 A comparison of error indices of different prediction models in the task of predicting heat transfer coefficient. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0073] like Figure 1As shown in the embodiment of the present invention, the method for predicting the heat transfer coefficient of a desulfurization tower heat exchanger in a thermal power plant specifically includes:

[0074] Step S1 as follows Figure 2 The signal acquisition flowchart is shown below:

[0075] This is a flowchart illustrating the signal acquisition process for the desulfurization tower heat exchanger system operating parameters involved in this invention. The system uses a desulfurization wastewater evaporation tower and a desulfurization wastewater heat exchanger as core equipment, and employs multiple types of sensors and transmitters to collect and monitor key operating parameters in real time. Specifically, a level sensor is installed inside the desulfurization wastewater evaporation tower to detect changes in the liquid level in real time and transmits the corresponding level signal to the signal acquisition system to reflect system load changes and operating status. A circulation pump is installed in the circulation pipeline between the evaporation tower and the heat exchanger to maintain wastewater circulation. Pressure transmitters are also installed at key locations in the pipeline to continuously monitor the system pipeline pressure and obtain pressure operating data.

[0076] Pressure measuring points (PTs) are installed at both ends of the desulfurization wastewater heat exchanger, and the pressure difference between the inlet and outlet of the heat exchanger is measured using a differential pressure transmitter to reflect changes in the internal flow resistance and heat exchange state. Simultaneously, a flow transmitter (FT) is installed on the heat medium return pipeline to measure the mass flow rate of the heat medium, thereby characterizing the heat transfer capacity of the heat exchanger and the system operating load. The aforementioned level, pressure, differential pressure, and mass flow signals are all uniformly collected and processed through an industrial signal acquisition system to form a multivariate operational dataset. This signal acquisition process can comprehensively obtain key operating parameters of the desulfurization tower heat exchange system, providing a reliable data foundation for subsequent heat transfer coefficient prediction.

[0077] Multivariate operational data was collected using multiple sensors deployed on the heat exchanger body and its inlet and outlet pipelines of the desulfurization tower. The raw data was then denoised, interpolated for missing values, and standardized. It was then segmented according to a preset time window and sliding step size to obtain a multivariate operational sample sequence. Simultaneously, a measurement point relationship graph and a normalized adjacency matrix were constructed based on the heat exchanger structure and measurement point locations, forming the foundational dataset for subsequent network modeling. The specific method is as follows:

[0078] Step S11: Collect operational data using multiple sensors installed on the heat exchanger body and inlet / outlet pipelines of the desulfurization tower in the thermal power plant, and process the raw time series collected by the sensors according to a uniform sampling period. Time synchronization and resampling are performed to construct a multivariate runtime data matrix. The specific calculation method is as follows:

[0079] ;

[0080] in, For the first One sample, Indicates a time step. Indicates the sensor dimension.

[0081] Step S12: Clean the multivariate running data matrix: Impute missing values, remove or replace outliers to obtain a cleaned and standardized data matrix. Then, use a sliding window of length L to segment the data matrix along the time axis to obtain the nth sample segment of length L. The sliding window segmentation described above yields a set of multivariate running sample fragments, which serve as the foundation dataset for network training and prediction.

[0082] Step S13: Construct a measurement point relationship diagram: For the set of measurement points, The connection relationship between measuring points; the connection between any two measuring points is determined based on the fluid connectivity within the heat exchanger. and Determine whether they are adjacent, and construct an adjacency matrix accordingly. ,like and If directly connected or the distance is less than a preset threshold, then ,otherwise Normalize the adjacency matrix to obtain the normalized adjacency matrix. The specific calculation method is as follows:

[0083] ;

[0084] in, Represents the identity matrix. This is the angle matrix.

[0085] Step S2: Input the multivariate running sample sequence into the time decoupling module. It decomposes the sequence into a trend component and a periodic component using a moving average. These components are then subjected to linear mapping and superimposed with time embedding and spatial embedding based on a measurement point relationship graph. In the trend branch, a multilayer perceptron is used to extract long-term operating trend features, and in the periodic branch, a frequency domain multi-head self-attention model is used to extract periodic fluctuation features. The two outputs are then fused to obtain a time feature matrix representing the time correlation of long-term operating conditions. The specific steps are as follows:

[0086] Step S21: Take each sample fragment obtained in step S1 The input time decoupling module uses a moving average of window L to smooth the sequence of each measurement point to obtain the trend component. Then, subtract the trend component from the sequence to obtain the periodic component. The specific calculation method is as follows:

[0087] ;

[0088] Through the above processing, each sample segment is decomposed into a trend component sequence. and periodic component sequences This serves as the input for step S22.

[0089] Step S22: The trend component sequence obtained in step S21... With periodic component sequences The physical quantities are input into a linear mapping layer, and the feature dimensions of each physical quantity are unified to obtain an initial trend feature tensor with the same dimensions as the original sample fragment. and periodic initial feature tensor Using the normalized adjacency matrix obtained in step S1 Feature propagation is performed using a graph convolutional neural network to obtain spatial embedding vectors representing the internal structural relationships of the heat exchanger. Combined with the positional information at each time step, trend feature tensors with spatiotemporal embedding are formed. and periodic feature tensors with spatiotemporal embedding .

[0090] Step S23, The trend-time feature vector is obtained by inputting the multilayer perceptron. At the same time, In the frequency domain, multi-head self-attention is used to extract periodic features. The specific calculation method is as follows:

[0091] ;

[0092] in, , and These are the query vector, key vector, and value vector, respectively, representing the frequency domain features. , and These are the weight matrices for the frequency domain feature query vector, key vector, and value vector, respectively.

[0093] Then, an inverse Fourier transform is performed to restore the time domain, resulting in the periodic time feature tensor. The specific calculation method is as follows:

[0094] ;

[0095] Step S24: Convert the trend time feature tensor generated in step S23 into a multi-dimensional vector. With periodic time feature tensor Aligning each measurement point according to its dimension and time step, the feature matrix is ​​fused by element-wise addition or linear mapping after concatenation on the feature channels to obtain a time feature matrix that simultaneously contains both long-term trend information and cyclical fluctuation information. .

[0096] Step S3: Segment the time feature matrix obtained in Step S2 according to a preset time interval and extract the operating condition features. Obtain typical operating condition patterns through clustering, and count the frequency of pattern consistency for each pair of measuring points in each time period to construct an operating condition pattern correlation matrix. In the spatial multi-head self-attention module, calculate the spatial attention weights between measuring points based on the time features, and fuse them with the operating condition pattern correlation matrix and the normalized adjacency matrix to obtain a spatiotemporal feature matrix that combines time correlation and multi-measuring point spatial correlation. The specific steps are as follows:

[0097] Step S31: The time feature matrix obtained in step S2 The time feature vectors are used as input to the operating condition mode extraction module and segmented along the time feature dimension of the matrix to form the operating condition feature description vectors for that time period. The obtained operating condition feature description vectors are then used as input for unsupervised clustering analysis to obtain the operating condition mode numbers representing the operating conditions. ,in This is a time period number.

[0098] Step S32: Statistical analysis of any two measurement points and The frequency of operating under the same mode of operation within the same time period is calculated using the following method:

[0099] ;

[0100] in, Represents the total time period. and Indicates the first Measurement points within a time period and The number assigned to the operating mode. This is an indicator function.

[0101] Based on the consistent frequency of the aforementioned operating mode The operating condition correlation matrix is ​​constructed, and the specific calculation method is as follows:

[0102] ;

[0103] in, Set a preset threshold for the weighting coefficient of the operating mode. ,when Greater than season Take the maximum value; when Less than season Take the minimum value. This yields the operating condition correlation matrix. This is used for subsequent spatial association modeling.

[0104] Step 33: Obtain the time feature matrix from step S24. The input spatial multi-head attention module calculates the correlation of all measurement point features at each time step, and obtains the spatial attention weight matrix through the softmax activation function. .Will and Element-wise multiplication yields the fusion space weight matrix. . use right Perform spatial weighted aggregation and combine it with a normalized adjacency matrix. The features of each measurement point are updated by an M-layer graph convolutional neural network to obtain a spatiotemporal feature matrix that simultaneously reflects temporal dependence and spatial correlation. The specific calculation method is as follows:

[0105] ;

[0106] in, This represents an M-layer graph convolutional neural network.

[0107] Step S4: Input the spatiotemporal feature matrix obtained in Step S3 into the heat transfer coefficient prediction module. Map it through a fully connected layer or a one-dimensional convolutional layer to a multi-step initial prediction sequence of heat transfer coefficients for several future time steps. Apply self-attention with Gaussian time weights to this prediction sequence to model the correlation between different prediction steps, obtaining the final multi-step heat exchanger heat transfer coefficient prediction result, which is used for desulfurization tower heat transfer performance evaluation and operation optimization control. Specific steps are as follows:

[0108] Step S41: Define the heat exchanger coefficient prediction as a regression prediction task.

[0109] Step S42: Use the set of time feature matrices obtained in step S3 as the input to the prediction layer, and map it through a linear layer to form the initial heat transfer coefficient prediction matrix for the next P time steps. The specific calculation method is as follows:

[0110] ;

[0111] in, , These are the weight matrix and the bias coefficients, respectively. For feature dimensions.

[0112] Step S43: The initial prediction matrix obtained in step S42 The prediction sequence features are rearranged along the prediction step dimension to obtain a prediction part feature of length P, and attention is obtained through linear mapping. The specific calculation method is as follows:

[0113] ;

[0114] in, , , For linear attention, the query vector, key vector, and value vector are given. , , These are the weight matrices for the query vector, key vector, and value vector of linear attention, respectively.

[0115] Step S44: Introduce Gaussian time modeling to construct the Gaussian time feature matrix. The specific calculation method is as follows:

[0116] ;

[0117] in, and For the prediction step index, and These are the mean and standard deviation of the predicted sequence values, respectively.

[0118] The obtained Gaussian time feature matrix is ​​multiplied element-wise with the linear attention matrix obtained in step S43, then normalized using softmax, and then... After weighting, the corrected Gaussian time prediction feature matrix is ​​obtained, and the specific calculation method is as follows:

[0119] ;

[0120] in, This is the attention scaling factor.

[0121] The obtained modified Gaussian time prediction feature matrix is ​​restored to the heat transfer coefficient space through inverse linear mapping to obtain multi-step prediction results. The specific calculation method is as follows:

[0122] ;

[0123] in, and These are the weight matrix and bias coefficients for the multi-step prediction results, respectively.

[0124] Step S45: The initial prediction matrix obtained in step S43 is... The heat exchanger coefficient is predicted using the trained model in the input regression layer with root mean square error loss. The root mean square error loss is calculated as follows:

[0125] ;

[0126] in, This is the predicted heat transfer coefficient sequence after the regression layer expansion. This is the actual heat transfer coefficient sequence.

[0127] like Figure 3 As shown in the embodiment of the present invention, the heat transfer coefficient prediction system for the desulfurization tower heat exchanger in a thermal power plant specifically includes:

[0128] The data acquisition and preprocessing module is used to collect multivariate operational data using multiple sensors, perform noise reduction, missing data interpolation and standardization on the collected raw data, and segment it according to a preset time window and sliding step size to obtain a multivariate operational sample sequence. At the same time, it constructs a measurement point relationship diagram and a normalized adjacency matrix based on the heat exchanger structure layout and measurement point location to form a basic dataset for subsequent network modeling.

[0129] The time decoupling feature extraction module decomposes the data into a trend part and a periodic part through a moving average. The trend part is then subjected to linear mapping and time embedding and spatial embedding based on the measurement point relationship graph. In the trend branch, a multilayer perceptron is used to extract long-term operating trend features, and in the periodic branch, a frequency domain multi-head self-attention is used to extract periodic fluctuation features. The two outputs are fused to obtain a time feature matrix that represents the time correlation of long-term operating conditions.

[0130] The spatial relationship feature mixing module is used to segment the time feature matrix according to a preset time interval and extract the operating condition features. Typical operating condition patterns are obtained through clustering. The frequency of pattern consistency of each pair of measuring points in each time period is counted to construct the operating condition pattern correlation matrix. Based on the time features, the spatial attention weight between measuring points is calculated and fused with the operating condition pattern correlation matrix and the normalized adjacency matrix to obtain a spatiotemporal feature matrix that has both time correlation and multi-measuring point spatial correlation.

[0131] The prediction step correlation modeling module is used to input the obtained spatiotemporal feature matrix into the heat transfer coefficient prediction module. It is then mapped to a multi-step heat transfer coefficient initial prediction sequence for several future times through a fully connected layer or a one-dimensional convolutional layer. Self-attention with Gaussian time weights is applied to this prediction sequence to model the correlation between different prediction steps, thus obtaining the final multi-step heat transfer coefficient prediction result of the heat exchanger.

[0132] The heat transfer coefficient prediction output module is used to output the predicted heat transfer coefficient of the multi-step heat exchanger, which is used for the heat transfer performance evaluation and operation optimization control of the desulfurization tower.

[0133] Example 1: Heat transfer coefficient prediction based on desulfurization tower steady-state operating data:

[0134] In this embodiment, historical operating data of a desulfurization tower heat exchanger in a thermal power plant during its stable operation phase is selected as the input data source. The operating data is collected from multiple measuring points located on the heat exchanger body and inlet / outlet pipelines. The collected data includes multivariate time series data such as medium temperature, flow rate, and pressure. First, the data from each measuring point undergoes unified time synchronization processing, and missing data is repaired using interpolation. Outliers are removed using statistical rules, and the data is then standardized. Subsequently, the processed multivariate data is slidably divided along the time axis according to a preset time window length, forming a multivariate operating sample sequence. Simultaneously, a measuring point relationship diagram is constructed based on the connectivity of the fluid channels inside the heat exchanger, and a corresponding normalized measuring point adjacency matrix is ​​generated.

[0135] Building upon the above, the multivariate operating sample sequence is input into the time decoupling module. A moving average method is used to separate the trend component sequence and the periodic component sequence. Long-term operating trend features and periodic fluctuation features are extracted and fused to form a time feature matrix. Subsequently, spatiotemporal features are extracted by combining the measurement point relationship diagram and the operating condition correlation matrix, and mapped to a heat transfer coefficient prediction sequence for multiple future time steps. Experimental results show that this method can accurately reflect the changing trend of the heat exchanger's heat transfer coefficient under stable operating conditions, and the prediction results exhibit good smoothness and consistency.

[0136] Example 2: Time-Decoupled Prediction Based on Load Fluctuation Conditions

[0137] In this embodiment, the operating data of the desulfurization system under conditions of frequent load changes are selected as the research object. Since load adjustments cause periodic changes in flow and temperature, direct modeling can easily lead to oscillations in the prediction results. Therefore, the collected multi-measuring point operating data is first synchronized and cleaned, and then divided into multiple operating sample segments according to a sliding time window. In the sample construction stage, data from multiple consecutive time periods before and after load changes are retained to ensure that the time series simultaneously includes both slow changing trends and significant periodic fluctuations.

[0138] In the feature extraction stage, a time decoupling module decomposes the sample sequence into trend and periodic components. The trend component characterizes the long-term effects of load changes, while the periodic component characterizes the short-term changes caused by flow regulation and temperature fluctuations. By modeling and fusing the two types of time features separately, the interference of periodic fluctuations on long-term trend prediction is effectively avoided. Prediction results show that under load fluctuation conditions, this method can maintain the stability of heat transfer coefficient prediction and significantly reduce the impact of short-term oscillations on prediction accuracy.

[0139] Example 3: Heat transfer coefficient prediction based on enhanced spatial correlation of measurement points.

[0140] In this embodiment, considering the complex internal structure of the desulfurization tower heat exchanger and the strong spatial correlation between different measuring points, the focus is on using a measuring point relationship modeling mechanism for prediction. By analyzing the fluid flow direction inside the heat exchanger and the location of the measuring points, the connection relationship between the measuring points is constructed, and a normalized measuring point adjacency matrix is ​​generated accordingly. In the data preprocessing stage, the multivariate operating data is divided into sample sequences according to a unified time window, ensuring that data from multiple interrelated measuring points are included within the same time window.

[0141] In the spatiotemporal feature extraction process, a temporal feature matrix is ​​first obtained, and then spatial feature propagation is performed by combining it with a measurement point relationship diagram, enabling the operating states of adjacent measurement points to influence each other. This spatial modeling method effectively characterizes the coordinated change patterns between different regions within the heat exchanger. Experimental results show that, compared to methods that do not consider the spatial relationships of measurement points, this implementation maintains overall prediction consistency even when local operating conditions of the heat exchanger change, thus improving the reliability of heat transfer coefficient prediction.

[0142] Example 4: Prediction of Coordination Relationships Based on Operating Condition Modes

[0143] In this embodiment, the time feature matrix is ​​segmented according to a preset time interval, and a description vector of operating conditions is extracted for each time interval. Unsupervised clustering is used to perform cluster analysis on the operating condition features to obtain multiple typical operating condition patterns. Subsequently, the frequency with which any two measuring points are classified into the same operating condition pattern within each time interval is counted, and an operating condition pattern correlation matrix is ​​constructed based on this to characterize the synergy between measuring points in their operating states.

[0144] In the subsequent spatiotemporal feature modeling process, the operating condition correlation matrix is ​​fused with the measurement point relationship diagram and the time feature matrix, so that spatial correlation not only depends on physical structure but also reflects the consistency of operating conditions. By introducing this operating condition coordination mechanism, the model can distinguish changes in the measurement point correlation under different operating conditions. Prediction results show that this method can maintain high prediction accuracy in operating scenarios with frequent operating condition switching, avoiding the error accumulation caused by single spatial relationship modeling.

[0145] Example 5: Application example of optimizing control based on actual operation scenario:

[0146] In this embodiment, the predicted multi-step heat transfer coefficient results are used for the heat transfer performance evaluation and operational optimization control of the desulfurization tower. By predicting the trend of heat transfer coefficient changes over a future period, operators can anticipate the risk of decreased heat transfer efficiency and adjust flow allocation and operating parameters based on the prediction results. This implementation method uses an online update approach, continuously inputting the latest operating data into the model to ensure that the prediction results reflect the equipment's operating status in real time.

[0147] Actual operation results show that by obtaining heat transfer coefficient prediction information in advance, the system efficiency loss caused by a sudden drop in heat transfer performance can be effectively avoided, providing a reliable decision-making basis for desulfurization tower operation optimization. This embodiment demonstrates that this method can not only predict heat transfer coefficients but also directly serve operation optimization control in engineering applications, exhibiting significant practical engineering value.

[0148] like Figure 4 The figure shows a performance comparison of different prediction models in predicting the heat transfer coefficient of a heat exchanger in a desulfurization tower of a thermal power plant. Five typical prediction models were selected for experimental comparison: Convolutional Neural Network (CNN), Temporal Convolutional Network (TCN), Long Short-Term Memory Network (LSTM), Bidirectional Long Short-Term Memory Network (BiLSTM), and the prediction method proposed in this invention. Evaluation metrics include Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Symmetric Mean Absolute Percentage Error (SMAPE), and Mean Absolute Percentage Error (MAPE), used to comprehensively measure the accuracy and stability of each model in the heat transfer coefficient prediction task.

[0149] As shown in the figure, the method of this invention outperforms other comparative models in all evaluation metrics. Specifically, the method achieves MAE, MSE, and RMSE values ​​of 0.233531, 0.101013, and 0.317825, respectively, demonstrating significantly lower prediction errors compared to traditional CNN, TCN, and recurrent neural network models. Furthermore, the method also achieves the lowest values ​​for SMAPE and MAPE, indicating better prediction stability and generalization ability in terms of relative error control. These results demonstrate that by introducing a time decoupling mechanism, spatial correlation modeling of measurement points, and fusion of operating conditions, this invention can more effectively characterize the complex spatiotemporal dependencies of the desulfurization tower heat exchanger system, thereby significantly improving the prediction accuracy of the heat transfer coefficient and providing more reliable data support for subsequent heat transfer performance evaluation and operational optimization.

[0150] A comparative verification was conducted on the heat transfer coefficient prediction method for desulfurization tower heat exchangers in thermal power plants constructed based on embodiments of the present invention. The experimental results are shown in the figure. Under the same dataset and evaluation criteria, the method of the present invention achieves the best prediction performance compared with CNN, TCN, LSTM, and BILSTM, with MAE of 0.233531, MSE of 0.101013, RMSE of 0.317825, SMAPE of 0.075067, and MAPE of 0.073271. Taking MSE as an example, compared with C... NN decreased from 0.160318 to 0.101013, a reduction of approximately 37.0%, and RMSE decreased from 0.400397 to 0.317825, a reduction of approximately 20.6%. Compared to BILSTM, MAE decreased from 0.266621 to 0.233531, a reduction of approximately 12.4%, MSE decreased from 0.129294 to 0.101013, a reduction of approximately 21.9%, and RMSE decreased from 0.359575 to 0.317825, a reduction of approximately 11.6%. The above results demonstrate that the present invention enhances the ability to express non-stationary features through trend-period temporal decoupling and frequency domain attention enhancement. It also achieves multi-point coupled modeling by combining the working condition correlation obtained from working condition pattern clustering with spatial attention / graph convolution. Furthermore, it utilizes multi-step attention with Gaussian time weights to characterize the correlation between different prediction steps, thereby effectively reducing prediction errors and suppressing the propagation of multi-step prediction errors. This verifies the effectiveness and superiority of the method of the present invention in the task of heat transfer coefficient prediction.

[0151] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the heat exchange coefficient of a heat exchanger of a desulfurization tower of a thermal power plant, characterized in that, The method includes the following steps: Multi-point, multi-variable operating data are acquired from the heat exchanger body of the desulfurization tower and its inlet and outlet pipelines. The operating data is then synchronized in time, missing values ​​are repaired, outliers are processed and standardized. Based on a preset time window, the data is slidably segmented along the time axis to form a multi-variable operating sample sequence. A measurement point relationship diagram is constructed based on the fluid connectivity within the heat exchanger, and a normalized measurement point adjacency matrix is ​​generated accordingly. The multivariate operating sample sequence is input into the time decoupling module, and decomposed into trend component sequence and periodic component sequence by moving average. The long-term operating trend time features and periodic fluctuation time features are extracted and then fused to obtain a time feature matrix that characterizes the time correlation of operating conditions. The time feature matrix is ​​processed by time period to extract operating condition features. The operating condition mode is obtained by unsupervised clustering. The operating condition mode correlation matrix is ​​constructed based on the consistency of the operating condition mode of the measurement points in each time period. The time feature matrix, the working condition mode correlation matrix, and the normalized measurement point adjacency matrix are used in synergy to extract a spatiotemporal feature matrix that simultaneously reflects time dependence and spatial correlation of multiple measurement points. The spatiotemporal feature matrix is ​​mapped to a heat transfer coefficient prediction sequence for multiple future time steps, and a time correlation modeling mechanism is introduced in the prediction step dimension to output the multi-step heat transfer coefficient prediction results of the heat exchanger.

2. The method of claim 1, wherein, The measurement point relationship diagram is constructed based on the fluid flow path inside the heat exchanger and the physical location relationship of the measurement points. When there is a direct fluid connection between any two measurement points or the spatial distance is less than a preset threshold, it is determined that there is a connection between the two.

3. The method of claim 1, wherein, The normalized measurement point adjacency matrix is ​​obtained by introducing unit connections into the measurement point connection matrix and performing symmetric normalization processing in combination with measurement point degree information, so as to ensure the numerical stability of features during the propagation process in multi-layer space.

4. A time decoupling feature extraction method for modeling the operation data of a heat exchanger of a desulfurization tower of a thermal power plant based on the heat exchanger coefficient prediction method of the desulfurization tower of the thermal power plant according to claim 1, characterized in that, Includes the following steps: For multi-point, multi-variable running sample sequences, the moving average method is used to extract the trend component sequence reflecting the long-term change pattern; Subtract the corresponding trend component sequence from the original sample sequence to obtain the periodic component sequence that reflects short-term cyclical fluctuations. The trend component sequence and the periodic component sequence are processed in a unified manner according to their feature dimensions, and spatiotemporal embedding is performed by combining the spatial embedding of measurement points and the temporal location information. For the trend component sequence, a multilayer perceptron is used to extract long-term time-dependent features, and for the periodic component sequence, a multi-head self-attention method is used to extract periodic variation features in the frequency domain and then recovers them in the time domain. The trend time features and periodic time features are fused together to output a unified time feature representation.

5. The method of claim 4, wherein, The trend component is obtained by smoothing the calculation by using a fixed-length sliding window over the time dimension for each measurement point.

6. The method of claim 4, wherein, After the periodic component is modeled in the frequency domain, it is restored to a periodic time feature representation aligned with the original time step through inverse frequency domain transformation.

7. A time-step correlation modeling method for multi-step prediction of heat exchanger heat transfer coefficient based on the heat transfer coefficient prediction method for desulfurization tower heat exchangers in thermal power plants according to claim 1, characterized in that, Includes the following steps: The spatiotemporal feature matrix is ​​mapped into a sequence of initial heat transfer coefficient predictions for multiple future time steps; Reconstruct the feature representation of the predicted sequence according to the predicted time step order; Correlation weights between prediction steps are constructed along the prediction time step dimension, and the prediction step features are weighted in combination with the time weight distribution; The final prediction result of the multi-step heat exchanger heat transfer coefficient is generated based on the weighted prediction step features.

8. The method as described in claim 7, characterized in that, The time weight distribution is constructed based on the distance relationship between prediction time steps, so that prediction steps that are closer to the current time step receive higher weights.

9. The method as described in claim 7, characterized in that, The prediction step correlation weights are obtained by linearly mapping and normalizing the prediction step features to reflect the dependence strength between different prediction steps.

10. A system for predicting the heat transfer coefficient of a desulfurization tower heat exchanger in a thermal power plant, implementing the method as described in claim 1, characterized in that, include: The data acquisition and preprocessing module is used to generate multivariate running sample sequences and normalized measurement point adjacency matrices. The time decoupling feature extraction module is used to extract and fuse trend time features and periodic time features. The working condition mode and spatial relationship modeling module is used to construct the working condition mode correlation matrix and extract the spatiotemporal feature matrix; The prediction step correlation modeling module is used to generate prediction results for the heat transfer coefficient of multi-step heat exchangers.

Images