Circulating fluidized bed temperature prediction method based on dynamic correlation guided graph space-time learning

Abstract

The invention discloses a circulating fluidized bed temperature prediction method based on graph space-time learning guided by dynamic correlation, and belongs to the technical field of industrial process control and energy power engineering. Aiming at the problem of insufficient prediction precision and robustness caused by the fact that the space coupling relation and the time evolution characteristic between operation variables are not considered at the same time in the existing bed temperature prediction method, the method comprises the following steps of: screening key input from candidate variables by adopting multi-strategy characteristic selection; constructing a graph convolution network guided by dynamic correlation to adaptively describe a variable topological structure and extract spatial features, fusing modeling time dynamic characteristics of a long-short-term memory network, and designing a multi-objective loss function optimization model training including fitting errors, trend consistency and output smoothness; finally, high-precision and stable bed temperature prediction of the circulating fluidized bed boiler under complex working conditions is realized, and a reliable basis is provided for operation optimization and intelligent control.

Description

Technical Field

[0001] This invention belongs to the field of industrial process control and energy power engineering technology, and particularly relates to a method for predicting the bed temperature of a circulating fluidized bed based on dynamic correlation-guided graph spatiotemporal learning. Background Technology

[0002] In the field of thermal power generation, with the continuous growth of installed capacity of renewable energy sources such as wind and photovoltaic power, their inherent intermittency and volatility pose challenges to grid stability. Therefore, thermal power generation will continue to play a crucial role in base load, peak shaving, frequency regulation, and emergency backup for a considerable period of time. To achieve more efficient and cleaner thermal power conversion, circulating fluidized bed combustion technology has become the mainstream technology for small and medium-sized combined heat and power (CHP), low-calorific-value fuel utilization, and solid waste energy conversion due to its advantages such as strong fuel adaptability, high combustion efficiency, low pollutant emissions, and flexible load adjustment. Bed temperature is a key state parameter in the operation of circulating fluidized bed boilers, and its changes directly affect combustion efficiency, pollutant generation, load response rate, and equipment operating safety. In practical applications, excessively high bed temperatures may lead to coking and corrosion, while excessively low temperatures can easily cause incomplete combustion and efficiency reduction. Therefore, accurate prediction of bed temperature is of great significance for early identification of abnormal operating conditions, optimization of control strategies, and ensuring the safe and stable operation of the unit.

[0003] Currently, bed temperature modeling methods mainly fall into two categories: mechanistic modeling and data-driven modeling. Mechanistic models are typically based on thermal equilibrium and reaction kinetic equations. While possessing physical interpretability, they suffer from problems such as complex modeling, difficulty in parameter calibration, and poor adaptability in actual industrial scenarios, making them ill-suited for dynamic conditions such as fuel type switching or drastic load fluctuations. Data-driven methods, especially those based on deep learning, have been gradually introduced into the field of circulating fluidized bed temperature prediction in recent years due to their strong ability to fit complex nonlinear relationships. However, existing data-driven methods often treat various operating variables, such as primary air volume, secondary air volume, and coal feed rate, as independent inputs, failing to effectively characterize the spatial coupling relationships between variables. Furthermore, they lack sufficient modeling of the system's dynamic response and time lag characteristics, resulting in insufficient prediction accuracy and robustness under complex operating conditions. For example, in scenarios with load fluctuations or sensor noise interference, existing methods are prone to phase deviations or non-physical jitter, failing to provide stable and reliable multi-step prediction results, thus limiting their practical application in real-time control and optimization. Therefore, there is an urgent need for a bed temperature prediction method that can simultaneously capture the spatial correlation and temporal evolution characteristics between variables, in order to overcome the limitations of existing technologies in practical engineering. Summary of the Invention

[0004] To address the aforementioned technical challenges, this invention proposes a circulating fluidized bed temperature prediction method based on dynamic correlation-guided graph spatiotemporal learning. This method achieves high-precision and robust bed temperature prediction under complex scenarios such as load fluctuations, fuel switching, and sensor noise, providing reliable technical support for anomaly warning, combustion optimization, and intelligent control of circulating fluidized bed boilers.

[0005] To achieve the above objectives, this invention provides a method for predicting bed temperature in a circulating fluidized bed based on dynamic correlation-guided graph spatiotemporal learning, comprising: S1. From the historical operating data of the circulating fluidized bed boiler, the key variables related to the bed temperature and height are selected as model inputs using the multi-strategy feature selection method; S2. Construct a dynamically weighted graph structure using the key variables as nodes, and extract spatial and temporal features based on the fusion of graph convolutional networks and long short-term memory networks; S3. Train the network using a multi-objective loss function, which combines fitting error, trend consistency error, and output smoothness error. S4. Deploy the trained model to the circulating fluidized bed boiler control system and output the bed temperature prediction value based on real-time operating data.

[0006] Optionally, in S1, the process of the multi-strategy feature selection method includes: The variance thresholding method is applied to remove features whose variance is less than a preset threshold. Calculate the Pearson correlation coefficient between the retained features and the bed temperature, and filter out features whose absolute value of the correlation coefficient is greater than a preset threshold; The remaining features were scored using a random forest regression model, and the feature with the highest score was selected, along with historical bed temperature as input.

[0007] Optionally, in S2, constructing a dynamically weighted graph structure includes: Define a set of nodes, including primary air volume, secondary air volume, main steam flow rate, coal feed rate, generator power, economizer inlet flue gas temperature, left side bed pressure of the furnace, slag cooler inlet temperature, and historical bed temperature; The edge weights are initialized based on the Pearson correlation coefficient between the variables. If the absolute value of the correlation coefficient is greater than a set threshold, the edge weight is set to the absolute value of the correlation coefficient; otherwise, it is zero. During training, the edge weight matrix is ​​set as a trainable parameter, and the graph topology is dynamically updated through backpropagation.

[0008] Optionally, in S2, spatial feature extraction based on graph convolutional networks includes: Information propagation and aggregation are performed using a symmetrically normalized adjacency matrix; Node features are extracted using at least two graph convolutional layers, and the ReLU activation function is applied. The spatial features output by the graph convolutional network are fused with the original input features.

[0009] Optionally, in S2, the fusion of long short-term memory network to extract temporal features includes: The sequence after fusing spatial features is input into a long short-term memory network; Long Short-Term Memory (LSTM) networks have at least two hidden layers with a dimension of 64, and use Dropout to prevent overfitting. The hidden state of the last step of the Long Short-Term Memory network is output and mapped to the bed temperature prediction value through a fully connected layer.

[0010] Optionally, in S3, the multi-objective loss function includes: The fitting error loss is calculated by determining the mean square error between the predicted and actual values. Trend consistency loss is calculated to determine the difference between the first-order difference sign of the predicted and actual values. The output smoothness loss is calculated by summing the squares of the second-order differences of the predicted values; The total loss is a weighted sum of the three losses, and the weighting coefficients are determined by grid search.

[0011] Optionally, in S4, deploying the trained model to the circulating fluidized bed boiler control system includes: The model receives real-time operating parameter inputs, including primary air volume, secondary air volume, main steam flow rate, coal feed rate, generator power, economizer inlet flue gas temperature, left side bed pressure of the furnace, slag cooler inlet temperature, and historical bed temperature; it outputs the predicted bed temperature value for a future set time, which is used for boiler control and optimization.

[0012] Optionally, in S1, the preprocessing of historical running data includes: Outliers in the original data are removed by setting a threshold; Missing values ​​are filled using linear interpolation. The processed data is normalized to the [0,1] interval; The normalized dataset is divided into training, validation, and test sets.

[0013] Technical Effects of this Invention: This invention discloses a circulating fluidized bed (CFB) temperature prediction method based on dynamic correlation-guided graph spatiotemporal learning. Through multi-strategy feature selection, it effectively filters out key variables highly correlated with bed temperature, enhancing the representativeness of the model input. A dynamically correlated graph convolutional network is constructed, abstracting the operating variables into nodes and adaptively constructing a graph topology based on the Pearson correlation coefficient. A trainable edge weight matrix updates the spatial coupling relationship between variables in real time, accurately extracting spatial features under complex operating conditions. A long short-term memory (LSTM) network is fused to recursively model the spatiotemporal feature sequence, effectively capturing the lag effect and long-range dependency of bed temperature changes. A designed multi-objective loss function comprehensively optimizes fitting accuracy, trend consistency, and output smoothness, ensuring that the predicted curve maintains the same trend as the measured value and is continuously smooth. Finally, high-precision and robust bed temperature prediction is achieved under complex scenarios such as load fluctuations, fuel switching, and sensor noise, providing reliable technical support for anomaly early warning, combustion optimization, and intelligent control of CFB boilers. Attached Figure Description

[0014] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic flowchart of a circulating fluidized bed temperature prediction method based on dynamic correlation-guided graph spatiotemporal learning, according to an embodiment of the present invention. Detailed Implementation

[0015] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0016] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0017] like Figure 1 As shown, this embodiment provides a method for predicting the bed temperature of a circulating fluidized bed based on dynamic correlation-guided graph spatiotemporal learning, including: S1. From the historical operating data of the circulating fluidized bed boiler, the key variables related to the bed temperature and height are selected as model inputs using the multi-strategy feature selection method; S2. Construct a dynamically weighted graph structure using the key variables as nodes, and extract spatial and temporal features based on the fusion of graph convolutional networks and long short-term memory networks; S3. Train the network using a multi-objective loss function, which combines fitting error, trend consistency error, and output smoothness error. S4. Deploy the trained model to the circulating fluidized bed boiler control system and output the bed temperature prediction value based on real-time operating data.

[0018] Furthermore, in S1, the process of the multi-strategy feature selection method includes: The variance thresholding method is applied to remove features whose variance is less than a preset threshold. Calculate the Pearson correlation coefficient between the retained features and the bed temperature, and filter out features whose absolute value of the correlation coefficient is greater than a preset threshold; The remaining features were scored using a random forest regression model, and the feature with the highest score was selected, along with historical bed temperature as input.

[0019] Specifically, the implementation process of this embodiment includes: Preliminary filtering is performed using the variance thresholding method to remove features that are close to constant with a variance less than 0.1, thereby eliminating variables whose information contribution is negligible. For the retained features, the Pearson correlation coefficient is calculated to further filter out features with weaker correlations with other variables, in order to retain a more representative and relevant subset of features. The formula for calculating the Pearson correlation coefficient is as follows: ; in, This indicates the initial processing of the minimum value in the dataset; This indicates the maximum value in the initial processing of the dataset; This refers to the normalized data; This involves the initial processing of the raw, unprocessed data in the dataset; among which... and These represent the sample values ​​of the characteristic variable and the bed temperature, respectively. and This represents the corresponding average value. A threshold of 0.4 is used to select variables that have a moderate or strong correlation with bed temperature, thereby further compressing the feature space. The remaining features were scored using a random forest regression model, and the eight most important features were selected based on the scores. Considering the time-dependent nature of the modeling task, historical bed temperature was added as a dynamic input.

[0020] Furthermore, in S2, constructing a dynamically weighted graph structure includes: Define a set of nodes, including primary air volume, secondary air volume, main steam flow rate, coal feed rate, generator power, economizer inlet flue gas temperature, left side bed pressure of the furnace, slag cooler inlet temperature, and historical bed temperature; The edge weights are initialized based on the Pearson correlation coefficient between the variables. If the absolute value of the correlation coefficient is greater than a set threshold, the edge weight is set to the absolute value of the correlation coefficient; otherwise, it is zero. During training, the edge weight matrix is ​​set as a trainable parameter, and the graph topology is dynamically updated through backpropagation.

[0021] Specifically, the implementation process of this embodiment includes: A time-varying graph structure is constructed using the input variables as nodes. The nodes include primary air volume, secondary air volume, main steam flow rate, coal feed rate, generator power, economizer inlet flue gas temperature, furnace left side bed pressure, slag cooler inlet temperature, and historical bed temperature. The edge weights are determined by the Pearson correlation coefficient between the variables. If the absolute value of the correlation coefficient is greater than the set threshold, the edge weight is set to the absolute value of the correlation coefficient; otherwise, it is 0. The graph structure is input into a graph convolutional network, and information propagation and aggregation are performed using a symmetric normalized adjacency matrix to extract the spatial feature representation of each node. During model training, the edge weight matrix is ​​set as a trainable parameter, and the graph structure is dynamically updated through the backpropagation mechanism to achieve adaptive optimization of the graph topology. After fusing the spatial features output by the graph convolutional network with the original input features, the data is fed into a long short-term memory network to model the temporal evolution relationship between variables and output the predicted bed temperature value at future time.

[0022] Furthermore, in S2, the extraction of spatial features based on graph convolutional networks includes: Information propagation and aggregation are performed using a symmetrically normalized adjacency matrix; Node features are extracted using at least two graph convolutional layers, and the ReLU activation function is applied. The spatial features output by the graph convolutional network are fused with the original input features.

[0023] Furthermore, in S2, the temporal features extracted by fusing the Long Short-Term Memory network include: The sequence after fusing spatial features is input into a long short-term memory network; Long Short-Term Memory (LSTM) networks have at least two hidden layers with a dimension of 64, and use Dropout to prevent overfitting. The hidden state of the last step of the Long Short-Term Memory network is output and mapped to the bed temperature prediction value through a fully connected layer.

[0024] Specifically, the implementation process of this embodiment includes: Construction of Dynamic Graph Convolutional Networks (GCNs) To explicitly capture the spatial coupling relationships between input variables at each time step t, this embodiment constructs a dynamic weighted graph. Feature propagation is achieved using a two-layer graph convolutional network. The specific steps are as follows: (1) Node and initial edge weights: Node set Primary air volume, secondary air volume, coal feed rate, main steam flow rate, generator power, economizer inlet flue gas temperature, left side bed pressure of the furnace, slag cooler inlet temperature and historical bed temperature, total Each node.

[0025] Initial edge weight matrix : in the time window Calculation nodes within each sampling point and Pearson correlation coefficient ;like ,but ,otherwise Threshold Setting it to 0.4 preserves significant coupling while avoiding overfitting caused by dense connections.

[0026] (2) Self-loop and normalization: right Add self-loop Calculate the degree matrix Its elements Using a symmetric normalized Laplace matrix: ; Ensure consistent node feature scale during aggregation and suppress excessive influence of height nodes on their neighborhood.

[0027] (3) Two-layer graph convolution feature extraction: Node feature matrix After two layers of graph convolution: ; ; in , These are trainable weights; The ReLU activation function is used. Output This refers to the spatial characteristics that incorporate topological information at the current moment.

[0028] (4) Dynamic edge weight update mechanism: During the backpropagation process of the model, As trainable parameters, they are updated synchronously with network weights, achieving joint optimization based on both "data-driven" and "prior correlation". After training... It can automatically strengthen or weaken the corresponding edges as the working conditions change, thus completing dynamic correlation guidance.

[0029] (5) Spatiotemporal fusion output: spatial features With original features Element-by-element addition: : Then, the consecutive L=12 time steps were... As input to the LSTM, it achieves joint modeling of spatial and temporal features. The hidden layer has a dimension of 64, two layers, and Dropout = 0.2. The hidden state is shown in the last step of the LSTM process. The bed temperature prediction value at time n is obtained by mapping it to 1D through a fully connected layer. .

[0030] Furthermore, in S3, the multi-objective loss function includes: The fitting error loss is calculated by determining the mean square error between the predicted and actual values. Trend consistency loss is calculated to determine the difference between the first-order difference sign of the predicted and actual values. The output smoothness loss is calculated by summing the squares of the second-order differences of the predicted values; The total loss is a weighted sum of the three losses, and the weighting coefficients are determined by grid search.

[0031] Specifically, the implementation process of this embodiment includes: Construct a multi-objective loss function and train and optimize the network.

[0032] To ensure the model achieves engineering-ready levels in terms of fitting accuracy, trend direction, and curve smoothness, this embodiment employs a weighted multi-objective loss function for end-to-end training of the network parameters. The specific structure is as follows: ; in, The data loss for fitting; This represents a loss due to trend consistency. To compensate for the loss of output smoothness; and These are the weighting coefficients for trend consistency loss and output smoothness loss, respectively, with values ​​ranging from 0 to 1. They are determined using a grid search method to achieve a balance between fitting accuracy, trend consistency, and prediction stability.

[0033] ; ; ; in, This represents the total number of data points used in the calculation. and These represent the model at the 1st and 2nd. Predicted and actual values ​​for each time step; , ;when y When >0, sign( y )=1, when y< At time 0, sign( y )=-1, when y= At time 0, sign( y )=0.

[0034] During the training phase, the batch size was set to 256. Weights were determined from the candidate sets [0.0, 0.1, 0.3, 0.5, 1.0] and [0.0, 0.01, 0.05, 0.1, 0.2] using grid search, respectively. and The model was optimized using the Adam optimizer with an initial learning rate of 0.001, combined with a learning rate decay strategy to enhance convergence stability and prevent overfitting. If the validation loss did not improve for 20 consecutive epochs, the learning rate was reduced by a factor of 0.1.

[0035] The network's hyperparameters (such as learning rate, number of hidden layer neurons, and batch size) are optimized using a validation subset of data to improve the model's generalization ability.

[0036] The trained network was evaluated using test set data. The main evaluation metrics included root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²), which were calculated as follows: ; ; ; ; in, This indicates the number of data points for bed temperature prediction; Indicates the predicted value; This represents the target value.

[0037] Furthermore, in S4, deploying the trained model to the circulating fluidized bed boiler control system includes: The model receives real-time operating parameter inputs, including primary air volume, secondary air volume, main steam flow rate, coal feed rate, generator power, economizer inlet flue gas temperature, left side bed pressure of the furnace, slag cooler inlet temperature, and historical bed temperature; it outputs the predicted bed temperature value for a future set time, which is used for boiler control and optimization.

[0038] Specifically, the implementation process of this embodiment includes: The trained network was validated on a local computer using test set data. During validation, the model was input with parameters from the test set, including primary air volume, secondary air volume, main steam flow rate, and coal feed rate. The output was a predicted bed temperature value, which was then compared with the actual bed temperature value. The predictive performance of the model was evaluated by calculating the error between the predicted and actual values.

[0039] This invention discloses a circulating fluidized bed (CFB) temperature prediction method based on dynamic correlation-guided graph spatiotemporal learning. It effectively filters key variables highly correlated with bed temperature through multi-strategy feature selection, enhancing the representativeness of the model input. A dynamically correlated graph convolutional network is constructed, abstracting the operating variables into nodes and adaptively building a graph topology based on the Pearson correlation coefficient. A trainable edge weight matrix updates the spatial coupling relationship between variables in real time, accurately extracting spatial features under complex operating conditions. A long short-term memory (LSTM) network is fused to recursively model the spatiotemporal feature sequence, effectively capturing the lag effect and long-range dependency of bed temperature changes. A designed multi-objective loss function comprehensively optimizes fitting accuracy, trend consistency, and output smoothness, ensuring that the predicted curve maintains the same trend as the measured value and is continuously smooth. Finally, high-precision and robust bed temperature prediction is achieved under complex scenarios such as load fluctuations, fuel switching, and sensor noise, providing reliable technical support for anomaly early warning, combustion optimization, and intelligent control of CFB boilers.

[0040] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting bed temperature in a circulating fluidized bed based on dynamic correlation-guided graph spatiotemporal learning, characterized in that, include: S1. From the historical operating data of the circulating fluidized bed boiler, the key variables related to the bed temperature and height are selected as model inputs using the multi-strategy feature selection method; S2. Construct a dynamically weighted graph structure using the key variables as nodes, and extract spatial and temporal features based on the fusion of graph convolutional networks and long short-term memory networks; S3. Train the network using a multi-objective loss function, which combines fitting error, trend consistency error, and output smoothness error. S4. Deploy the trained model to the circulating fluidized bed boiler control system and output the bed temperature prediction value based on real-time operating data.

2. The method for predicting bed temperature in a circulating fluidized bed based on dynamic correlation-guided graph spatiotemporal learning as described in claim 1, characterized in that, In S1, the process of the multi-strategy feature selection method includes: The variance thresholding method is applied to remove features whose variance is less than a preset threshold. Calculate the Pearson correlation coefficient between the retained features and the bed temperature, and filter out features whose absolute value of the correlation coefficient is greater than a preset threshold; The remaining features were scored using a random forest regression model, and the feature with the highest score was selected, along with historical bed temperature as input.

3. The method for predicting bed temperature in a circulating fluidized bed based on dynamic correlation-guided graph spatiotemporal learning as described in claim 1, characterized in that, In S2, constructing a dynamically weighted graph structure includes: Define a set of nodes, including primary air volume, secondary air volume, main steam flow rate, coal feed rate, generator power, economizer inlet flue gas temperature, left side bed pressure of the furnace, slag cooler inlet temperature, and historical bed temperature; The edge weights are initialized based on the Pearson correlation coefficient between the variables. If the absolute value of the correlation coefficient is greater than a set threshold, the edge weight is set to the absolute value of the correlation coefficient; otherwise, it is zero. During training, the edge weight matrix is ​​set as a trainable parameter, and the graph topology is dynamically updated through backpropagation.

4. The method for predicting bed temperature in a circulating fluidized bed based on dynamic correlation-guided graph spatiotemporal learning as described in claim 1, characterized in that, In S2, spatial feature extraction based on graph convolutional networks includes: Information propagation and aggregation are performed using a symmetrically normalized adjacency matrix; Node features are extracted using at least two graph convolutional layers, and the ReLU activation function is applied. The spatial features output by the graph convolutional network are fused with the original input features.

5. The method for predicting bed temperature in a circulating fluidized bed based on dynamic correlation-guided graph spatiotemporal learning as described in claim 1, characterized in that, In S2, the temporal features extracted by fusing the Long Short-Term Memory network include: The sequence after fusing spatial features is input into a long short-term memory network; Long Short-Term Memory (LSTM) networks have at least two hidden layers with a dimension of 64, and use Dropout to prevent overfitting. The hidden state of the last step of the Long Short-Term Memory network is output and mapped to the bed temperature prediction value through a fully connected layer.

6. The method for predicting bed temperature in a circulating fluidized bed based on dynamic correlation-guided graph spatiotemporal learning as described in claim 1, characterized in that, In S3, the multi-objective loss function includes: The fitting error loss is calculated by determining the mean square error between the predicted and actual values. Trend consistency loss is calculated to determine the difference between the first-order difference sign of the predicted and actual values. The output smoothness loss is calculated by summing the squares of the second-order differences of the predicted values; The total loss is a weighted sum of the three losses, and the weighting coefficients are determined by grid search.

7. The method for predicting bed temperature in a circulating fluidized bed based on dynamic correlation-guided graph spatiotemporal learning as described in claim 1, characterized in that, In S4, deploying the trained model to the circulating fluidized bed boiler control system includes: The model receives real-time operating parameter inputs, including primary air volume, secondary air volume, main steam flow rate, coal feed rate, generator power, economizer inlet flue gas temperature, left side bed pressure of the furnace, slag cooler inlet temperature, and historical bed temperature; it outputs the predicted bed temperature value for a future set time, which is used for boiler control and optimization.

8. The method for predicting bed temperature in a circulating fluidized bed based on dynamic correlation-guided graph spatiotemporal learning as described in claim 1, characterized in that, In S1, the preprocessing of historical running data includes: Outliers in the original data are removed by setting a threshold; Missing values ​​are filled using linear interpolation. The processed data is normalized to the [0,1] interval; The normalized dataset is divided into training, validation, and test sets.

Images