A coal-fired boiler main steam temperature prediction modeling method based on a multi-branch convolutional neural network

Abstract

This invention discloses a method for predicting and modeling the main steam temperature of a coal-fired boiler based on a multi-branch convolutional neural network, belonging to the field of coal-fired boiler operation optimization and control technology. The method includes: extracting historical data from the boiler monitoring system and using a random forest algorithm to select core features; constructing the core features into a two-dimensional time-series tensor, which is then sequentially input into a first standard convolutional layer, a selective kernel convolutional layer, and a second standard convolutional layer. Through the splitting, fusion, and selection operations of the selective kernel convolutional layer, features at multiple time scales with different receptive fields are adaptively fused, and hidden features of furnace-side combustion and boiler-side steam are automatically identified. Finally, the predicted value of the main steam temperature is output through a fully connected network. This invention, by introducing a selective kernel mechanism, achieves dynamic decoupling between high-frequency disturbances in the desuperheating water and low-frequency inertia on the combustion side, solving the problem that traditional convolutional neural networks cannot simultaneously handle fast and slow time-scale feature extraction, and improving the prediction accuracy and stability of the main steam temperature under variable load conditions.

Description

Technical Field

[0001] This invention belongs to the field of coal-fired boiler operation optimization and control technology, specifically involving a modeling method for predicting the main steam temperature of a coal-fired boiler based on a multi-branch convolutional neural network, which is used for predicting the main steam temperature of a coal-fired boiler. Background Technology

[0002] The continued growth of fossil fuel emissions has severely compressed the window period for the 1.5°C temperature control target. The penetration rate of renewable energy sources, such as wind and solar power, in the power grid has increased dramatically, and the intermittent and fluctuating nature of their output has led to significant high-frequency fluctuations in the grid's net load. To ensure the safe and stable operation of the power grid, large coal-fired power units are transitioning from baseload power sources to flexible load-regulating power sources. However, under ultra-low load conditions, the ratio of combustion to heat transfer in the furnace changes significantly, increasing the risk of thermal imbalance in the unit. Main steam temperature (MST) is a key parameter for measuring the thermal efficiency and operational safety of power plants, and its dynamic stability has become a significant bottleneck restricting the flexible peak-shaving of units. Existing research indicates that frequent fluctuations in thermal parameters can induce flow-accelerated corrosion (FAC) in the header system, severely shortening equipment lifespan.

[0003] For a 1000 MW double reheat tower boiler, the main steam temperature (MST) control faces extremely complex dynamic challenges. On the one hand, the introduction of the two-stage reheat system creates strong dynamic coupling of heat on the flue gas side between multiple heating surfaces. On the other hand, the dynamic evolution of the MST exhibits significant multi-timescale characteristics: desuperheating water regulation is a rapid phase change heat transfer process with relatively low inertia, while coal-water ratio regulation involves combustion and hydrodynamic circulation, exhibiting longer response inertia, especially in the coal mill pulverization and transportation processes. Traditional constant-parameter PID control strategies are prone to control hysteresis and overshoot in such strongly coupled, nonlinear systems, limiting the unit's regulation performance. This intertwining of fast and slow timescales makes it difficult to accurately predict the dynamic response of the MST under varying operating conditions. Inaccurate prediction can easily lead to overheating of the heating surfaces, accelerating the oxidation, creep, and fatigue processes of heat-resistant steels such as T91, and even causing serious accidents such as heating surface tube rupture. Although existing combustion-hydraulic coupled CFD models can achieve high-precision characterization of the thermodynamic field, their multiphysics iterative solution is computationally expensive and cannot meet the millisecond-level response requirements of online control. Therefore, how to utilize historical operating data to construct a high-precision data-driven model and achieve advanced and accurate perception of thermodynamic parameters has become an important way to overcome the key technical bottlenecks in the flexible operation of coal-fired power units.

[0004] To overcome the bottleneck of high-precision data-driven models in boiler modeling applications, convolutional neural networks (CNNs), with their weight-sharing mechanism and parallel computing advantages, have gradually gained attention in the field of thermal energy engineering. Research shows that using one-dimensional convolution (1D-CNN) and its improved variants can effectively improve the model's efficiency in capturing local features such as abrupt changes in air and coal parameters, and significantly shorten inference time. Subsequent research has further attempted to introduce Transformer networks, utilizing self-attention mechanisms to capture long-term temporal global dependencies, or employing temporal convolutional networks (TCNs) to expand the temporal perception range through dilated convolutions. However, existing temporal models still suffer from a mismatch between algorithms and physical mechanisms when applied to boiler systems with both high inertia and rapid adjustments. While conventional CNNs, TCNs, and DFC-CNNs possess local perception capabilities, their convolutional kernel scale is fixed before training, lacking a dynamic adjustment mechanism. This limitation of insufficient local perception and the solidification of static feature scales fundamentally conflict with the multi-timescale dynamic characteristics of boiler main steam temperature: the desuperheating water parameter has high sensitivity, corresponding to a small receptive field; while the long inertial process caused by coal feed rate requires a broad temporal perspective, i.e., a large receptive field. Due to the lack of an adaptive fusion mechanism for the two time scales, existing models struggle to simultaneously consider feature extraction across time scales, leading to prediction biases under deep peak-shaving and variable load conditions. Summary of the Invention

[0005] The problem this invention aims to solve is to provide a method for predicting and modeling the main steam temperature of a coal-fired boiler, comprising the following steps: Step 1: Extract historical data from the boiler monitoring system and preprocess it. Use the random forest algorithm to optimize candidate features and select core features. The preprocessing includes using the 3σ rule to process abnormal data. Step 2: Construct the selected core features into a two-dimensional temporal tensor, input it into the first standard convolutional layer, and extract the low-level local features; Step 3: Input the feature tensor output by the first standard convolutional layer into the selective kernel convolutional layer, and adaptively weight the features at different time scales to generate a multi-scale fused feature tensor. Step 4: Input the multi-scale fusion feature tensor output by the selective kernel convolutional layer into the second standard convolutional layer to extract high-dimensional deep features and obtain a high-dimensional feature tensor. Step 5: Decode the high-dimensional feature tensor output by the second standard convolutional layer through a fully connected network to output the predicted value of the main steam temperature.

[0006] Furthermore, a random forest algorithm is used for feature optimization to select the core feature set that plays a dominant role in the evolution of the main steam temperature. Specifically, the following calculations are performed: The weight of the j_cand-th candidate feature , represented as: ; in, Let j_cand be the arithmetic mean of the importance of the j-th candidate feature across all decision trees. For the j_cand-th candidate feature in the whole tree The importance score in the equation, where M represents the total number of candidate features from the boiler side and the turbine side; When the weight of a candidate feature is greater than a set threshold, the feature is considered a core feature, and all core features constitute the core feature set.

[0007] Furthermore, the first standard convolutional layer extracts local features from the lower layers, represented as:

[0008] In the formula, It is a ReLU nonlinear activation function. This represents the set of input feature maps used in the computation. Tensors input to the first standard convolutional layer The Each feature channel This represents the discrete convolution operation. For the first standard convolutional layer, corresponding to the th The first output feature map A learnable convolutional kernel weight matrix; For the first layer Learnable bias terms corresponding to each output feature map; Features The output feature tensor of the first standard convolutional layer is obtained by stacking and concatenating along the channel dimension. .

[0009] Furthermore, in step 3, the feature tensor output by the first standard convolutional layer is input into a selective kernel convolutional layer to adaptively weight features at different time scales, generating a multi-scale fused feature tensor. Specifically: Input feature tensor The stream is split into two independent convolutional paths. and Each feature branch representing a different receptive field is generated. and , represented as:

[0010] The first path Adopting standards Convolution, the second path For expansion rate of Convolution, dilation rate Greater than 1; Then a fusion operation is performed, represented as:

[0011] Fusion features Batch normalization is performed, and then the vector is compressed into a low-dimensional vector through a fully connected layer. , represented as:

[0012] in, This represents the learnable weight matrix corresponding to the fully connected layer. This represents the fused features after batch normalization. The global channel statistical feature vector is obtained by performing global average pooling. It is the ReLU activation function; Soft attention weights for different scale branches are generated using the Softmax function. and , represented as

[0013] in, and The linear transformation matrices corresponding to different scale branches, and the soft attention weights satisfy the normalization constraint. ; Perform feature fusion to obtain the first Channel characteristics , represented as:

[0014] Features of each channel By concatenating the data, we obtain the multi-scale fused feature tensor generated by the selective kernel convolutional layer. .

[0015] Furthermore, in step 4, the multi-scale fused feature tensor output by the selective kernel convolutional layer is input into the second standard convolutional layer to extract high-dimensional deep features, as shown below:

[0016] in, Output tensors for selective kernel convolutional layers The Each feature map For the second standard convolutional layer, corresponding to the first The first output feature map A learnable convolutional kernel weight matrix, This is a learnable bias term for the second standard convolutional layer; All output features Together they constitute a high-dimensional feature tensor .

[0017] Furthermore, in step 5, the high-dimensional feature tensor output by the second standard convolutional layer is decoded through a fully connected network to output the predicted value of the main steam temperature, specifically as follows: First, for high-dimensional feature tensors A flattening operation is performed to obtain a one-dimensional feature vector. This one-dimensional feature vector is then passed sequentially through stacked fully connected layers for non-linear mapping. The first fully connected layer... Layer output Represented as:

[0018] in, and These are the first fully connected layers. Layer weight matrix and bias terms, This is the ReLU activation function for a fully connected layer.

[0019] Furthermore, the ReLU activation function of the fully connected layer is specifically expressed as follows: .

[0020] Furthermore, mean squared error is used as the loss function, and the loss function is minimized through backpropagation algorithm to learn the mapping law from multivariate coupled state to main steam temperature.

[0021] Beneficial Effects: Compared with existing technologies, this invention focuses on the feature extraction stage. Based on the mechanism, it constructs a 70-dimensional candidate variable set from the wind, smoke, steam, and water systems, and uses the RF-Gini algorithm to select key features, aiming to reduce dimensionality while preserving physical information. Furthermore, the SK module in the convolutional layer utilizes an adaptive adjustment mechanism to effectively solve the problem of sluggish response of single-scale convolutional kernels to varying thermal inertial conditions, achieving precise decoupling of transient changes and steady-state trends in main steam temperature. This invention can meet the needs of actual industrial production, providing high-precision prediction information for refined operation under deep peak-shaving conditions.

[0022] A Selective Kernel (SK) mechanism is introduced into the CNN backbone network. Through multi-path feature extraction and adaptive weight allocation, the network can autonomously focus on the receptive field at different time scales according to the type of disturbance under the current operating conditions. This mechanism achieves dynamic matching with the multi-scale physical characteristics of the boiler from an algorithmic structure perspective for the first time, aiming to solve the trade-off problem between local and global features in traditional models. The effectiveness of the SK-CNN model is verified by simulation experiments on real data from a 1000 MW dual reheat ultra-supercritical unit. Attached Figure Description

[0023] Figure 1 This is a flowchart of the SK-CNN prediction model;

[0024] Figure 2 This is a diagram of the SK-CNN prediction model structure;

[0025] Figure 3 It is a diagram of the target boiler combustion system layout and three-dimensional furnace structure;

[0026] Figure 4 It is a 10-day trend chart of operational data covering low, medium, and high load areas;

[0027] Figure 5 This is a comparison curve of the prediction results of each model in different load areas (Region IV);

[0028] Figure 6 This is a scatter plot comparing the SK-CNN prediction model with the basic CNN model. Detailed Implementation

[0029] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0030] A method for predicting and modeling the main steam temperature of a coal-fired boiler based on a multi-branch convolutional neural network is proposed. In the feature extraction stage, a 70-dimensional candidate variable set is constructed from the wind, flue gas, steam, and water systems based on their mechanisms. The RF-Gini algorithm is then used to select key features, aiming to reduce dimensionality while preserving physical information. Furthermore, the selective kernel SK module in the convolutional layer utilizes an adaptive adjustment mechanism to effectively address the problem of sluggish response of single-scale convolutional kernels to variable thermal inertia conditions, achieving precise decoupling between transient changes and steady-state trends in the main steam temperature. Figure 1 As shown, the method of the present invention includes the following steps:

[0031] Step 1: Construct a hybrid modeling framework for time-series prediction of main steam temperature in a 1000 MW double reheat unit. For multi-timescale dynamic characteristics, this framework adopts a three-level cascaded structure of "feature optimization—multi-scale perception—nonlinear regression": First, random forest-Gini index is used to quantify variable importance and construct a low-redundancy physical feature space; then, a dynamic convolutional network based on the selective kernel SK mechanism is designed to achieve adaptive decoupling between high-frequency disturbances of the desuperheating water and low-frequency inertia of the combustion side; finally, a fully connected network completes the nonlinear mapping from high-dimensional features to temperature values.

[0032] Step Two: Extract historical data of the target boiler from the boiler monitoring system. For abnormal data caused by signal noise interference during boiler operation, process the data using the 3σ rule. Use the random forest algorithm for feature selection, and evaluate feature importance based on the purity gain before and after splitting decision tree nodes. Assume the input space contains... Candidate feature variables from the boiler side and the turbine side The main steam temperature (MST) prediction task is modeled as a regression problem. For any node in the regression tree... First, calculate the Gini index of the sample set. To measure the uncertainty of the data distribution at the current node:

[0033]

[0034]

[0035] in, This represents the number of categories or regression intervals after discretization. Represents a node The middle sample belongs to the first The probability estimate of the class. Indicates the first When candidate features are used as splitting attributes, the current node of the decision tree The decrease in the Gini index before and after the split, this value directly characterizes the first... The local contribution of each candidate feature to the prediction of the main steam temperature at the current node; This indicates that after splitting a node using the target feature, the number of samples in the generated left child node accounts for a proportion of the number of samples in the parent node. Weighted proportions of the total sample size; It represents the Gini index corresponding to the left child node after the node is split with the target feature, and is used to measure the uncertainty of the distribution of sample data within the left child node; This indicates that after splitting a node using the target feature, the number of samples generated in the right child node accounts for a proportion of the number of samples in the parent node. The weighted proportion of the total sample size, and satisfying the following conditions: Normalization constraints; The Gini index represents the right child node generated after node splitting based on the target feature. It is used to measure the uncertainty of the distribution of sample data within the right child node.

[0036] In order to obtain features In the whole tree Importance rating The contribution of each node to the split nodes in the tree needs to be summed:

[0037]

[0038] in, For the first Among the trees A set of nodes with splitting characteristics.

[0039] Step 3: To comprehensively evaluate the role of variables in the entire random forest model and eliminate single-tree bias, features global importance Defined as all decision trees Arithmetic mean of importance:

[0040]

[0041] To eliminate the influence of different physical dimensions and facilitate horizontal comparison, the importance scores of all features are further normalized to obtain the final weight of the j_cand-th candidate feature. :

[0042]

[0043] Based on this weighted ranking, a cumulative importance threshold is set, and redundant variables with low weights, such as some coal mill inlet air temperatures, are eliminated to select the core feature set that plays a dominant role in the evolution of the main steam temperature (MST).

[0044] Step 4: Arrange the core features selected in the previous step according to the time step, and construct a dimension-wise array. The two-dimensional temporal tensor is used as the initial input to this layer, denoted as . .in The number of core features This represents the length of the sliding time window. The first standard convolutional layer performs a sliding window operation on the input data using a set of learnable filters to extract local features from the lower layers. The first standard convolutional layer's... Output feature map It can be specifically defined by the following formula:

[0045]

[0046] In the formula, This represents the discrete convolution operation. It is a ReLU nonlinear activation function; For the initial input tensor The One feature channel; For the first standard convolutional layer, corresponding to the th The first output feature map A learnable convolutional kernel weight matrix; For the first standard convolutional layer Learnable bias terms corresponding to each output feature map; This represents the set of input feature maps used in the computation. The feature map set extracted by this layer is stacked and concatenated along the channel dimension to form the input feature tensor of the selective kernel unit in step five. This provides a solid underlying foundation for subsequent multi-scale feature decoupling.

[0047] Step 5: Convert the feature tensor output by the first standard convolutional layer The input is fed into a selective kernel convolutional layer. This layer contains three tightly coupled computational paths: splitting, merging, and selecting. Figure 2 As shown, the aim is to achieve adaptive weighting of features at different time scales through a dynamic gating mechanism.

[0048] First, perform a splitting operation on the input feature tensor. The flow is split into two independent convolutional paths, each generating a feature branch representing a different receptive field. and To achieve parallel decoupling of the multi-scale dynamic characteristics of the main steam temperature:

[0049]

[0050] The first path Adopting standards Convolution focuses on capturing high-frequency transient changes. For the second path, to capture long-period thermal inertia trends without increasing the number of parameters, dilated convolution is introduced. In this invention, Set as expansion rate of Convolution expands the equivalent receptive field to This allows for coverage of a wider time window.

[0051] Then a fusion operation is performed, which merges the information from two receptive field branches of different sizes by summing element by element. :

[0052]

[0053] To mitigate gradient vanishing and accelerate convergence during deep network training, batch normalization (BN) is introduced before feature compression. The feature vectors processed by BN are then compressed into low-dimensional vectors through a fully connected layer (FC). To extract the nonlinear correlation between channels:

[0054]

[0055] in This represents the learnable weight matrix corresponding to the fully connected layer; This represents the fused features after batch normalization. The global channel statistical feature vector obtained by performing global average pooling; This is the ReLU activation function.

[0056] Finally, a selection operation is performed, using compact feature descriptions of low-dimensional vectors. It uses two parallel fully connected layers to remap features back to the original dimensions, and generates soft attention weights for different scale branches using the Softmax function. and ,

[0057]

[0058] in, and These correspond to the linear transformation matrices for different scale branches. Clearly, the soft attention weights... and Satisfy normalization constraints The first selective kernel convolutional layer Channel characteristics It is obtained by weighting and fusing the features of each branch according to their attention weights:

[0059]

[0060] This mechanism allows the network to automatically increase its capacity when it detects drastic load fluctuations. Prioritize using small receptive fields to capture transient changes; while increasing the receptive field during steady-state operation. By relying on a large receptive field to track trends, dynamic matching between the physical mechanism and the algorithm structure is achieved. The output features of each channel are concatenated to form the multi-scale fusion feature tensor of the selective kernel convolutional layer, denoted as... .

[0061] Step 6: Convert the multi-scale fusion feature tensor output by the selective kernel convolutional layer The data is directly input into the second standard convolutional layer for further refinement and compression of high-dimensional deep features. The second standard convolutional layer... Output feature map The calculation process is as follows:

[0062]

[0063] In the formula, Output tensors for selective kernel convolutional layers The Each feature map For the second standard convolutional layer, corresponding to the first The first output feature map A learnable convolutional kernel weight matrix, This is the learnable bias term for the second standard convolutional layer. All output feature maps together constitute the final high-dimensional feature tensor. .

[0064] Step 7: High-dimensional feature tensor output by the second standard convolutional layer It contains multi-scale spatiotemporal information, which needs to be decoded into the predicted value of the main steam temperature through a fully connected layer. First, the high-dimensional feature tensor... Performing a flattening operation yields a one-dimensional feature vector. :

[0065]

[0066] One-dimensional feature vector Nonlinear mapping is performed sequentially through stacked fully connected layers. The first fully connected layer... Layer output The definition is as follows:

[0067]

[0068] in and These are the first fully connected layers. Layer weight matrix and bias terms, is the ReLU activation function for the fully connected layers. To enhance the network's ability to fit the nonlinear characteristics of the boiler, the ReLU activation function is uniformly adopted for all hidden layers in the fully connected layers. Its piecewise linearity helps alleviate the gradient vanishing problem, and the specific expression is:

[0069]

[0070] After mapping by the fully connected layer, the final output layer of the network directly outputs the predicted scalar of the main steam temperature. .

[0071] Based on the modeling method for predicting the main steam temperature of a coal-fired boiler according to the present invention, the present invention also proposes an SK-CNN model for predicting the main steam temperature of a coal-fired boiler. The SK-CNN model includes a first standard convolutional layer, a selective kernel convolutional layer, a first standard convolutional layer, and a fully connected network.

[0072] The two-dimensional temporal tensor constructed from the core features is input into the first standard convolutional layer, where convolution is performed and the layers are processed by an activation function to extract low-level local features, outputting a feature tensor. ; the feature tensor The input is fed into a selective kernel convolutional layer, which adaptively weights features at different time scales to generate a multi-scale fused feature tensor. The input is fed into the second standard convolutional layer to extract high-dimensional depth features, resulting in a high-dimensional feature tensor. Then, through a fully connected network, the predicted value of the main steam temperature is output.

[0073] Considering the high sensitivity of industrial sites to large-deviation faults such as overheating of heated surfaces, the mean square error is used as the loss function of the SK-CNN model. To increase the penalty for outlier prediction errors:

[0074]

[0075] In the formula, The set of learnable parameters for the SK-CNN model The loss function is the independent variable; The total number of samples involved in the loss calculation; The actual value of the main steam temperature; The main steam temperature prediction value output by the SK-CNN model; This is the set of all learnable parameters for the SK-CNN model, including the convolutional kernel weights and bias terms of convolutional layers, the weight matrices and bias terms of fully connected layers, and all other network parameters to be optimized. This loss function is minimized using the backpropagation algorithm. The gradient descent method is used to update the network parameters. The specific parameter update formula is as follows:

[0076]

[0077] In the formula, For the first The set of SK-CNN model parameters at the next iteration; For the first The set of SK-CNN model parameters at the next iteration; The learning rate of the network; This represents the gradient of the loss function with respect to the parameters of the SK-CNN model. The SK-CNN model can thus adaptively learn the mapping from multivariate coupled states to the main steam temperature.

[0078] This embodiment takes a 1000 MW ultra-supercritical double reheat once-through boiler as the research object. The target boiler is as follows: Figure 3 As shown, the boiler adopts a tower-type layout, single furnace, and staged air supply. The furnace consists of four vertical water-cooled walls, defined as the front wall, rear wall, left wall, and right wall. All heating surfaces are vertically stacked along the height direction, from bottom to top: spiral tube water-cooled walls, superheaters of various stages (including the final superheater, reheaters (including high-pressure low-temperature reheater, low-pressure high-temperature reheater, and high-pressure high-temperature reheater), and economizer. The boiler is equipped with 26 layers of pulverized coal nozzles and 6 coal mills, employing a tangential combustion method. Eight layers of separate burnout air (SOFA) are arranged above the main combustion zone to control the flame center height and NOx emissions. The entire furnace uses a structured network, with the grid direction in the burner area perpendicular to the jet direction. To meet the temperature control requirements of the secondary reheat unit, the boiler introduces a flue gas recirculation (FGR) system. The FGR reinjects the low-temperature flue gas after the economizer to the bottom of the furnace, utilizing the increased flue gas velocity to enhance upper convective heat transfer. The arrows in the figure indicate the flow direction. (direction). In this example, main steam temperature is used as the core indicator for evaluating the safety and economy of unit operation. Historical data from approximately 10 days of continuous operation were collected from the target boiler's distributed control system (DCS) database at a sampling resolution of one minute. For example... Figure 4 As shown, the collected data covers most operating conditions, including high load range, medium load range, and low load range of approximately 300MW, while excluding shutdown ranges, in order to fully address changes in main steam temperature.

[0079] All experiments in this embodiment were conducted using the PyTorch deep learning framework on a high-performance computing platform equipped with an Intel Xeon Gold 6226R processor and a 24 GB NVIDIA Tesla GPU. To verify the predictive performance of the SK-CNN model in real-world industrial scenarios, a systematic comparative experiment was carried out based on the 1000MW ultra-supercritical unit full-condition dataset constructed earlier. Building upon this, multiple sets of control experiments were further designed to explore the differences in model performance before and after feature selection, the performance advantages and disadvantages of different deep learning models, and the impact of introducing the selective kernel SK module on model accuracy.

[0080] In comparative experiments with traditional deep neural network (DNN) and gated recurrent unit (GRU) models, the SK-CNN model of this invention demonstrates its predictive accuracy. Figure 5 As shown, the predicted curve of SK-CNN consistently maintains a high degree of consistency with the true value, accurately capturing the fluctuation trend of the main steam temperature. Furthermore, thanks to its multi-scale feature extraction capabilities, it keenly detects the reverse temperature fluctuations caused by metal exothermic reactions, always closely following the trajectory of the true value. In addition, SK-CNN exhibits minimal fluctuations in various metrics during repeated experiments, demonstrating that it not only makes accurate predictions but also remains highly stable when switching between various operating conditions, fully meeting the signal reliability requirements of optimized control for coal-fired power units.

[0081] In comparative experiments with the baseline standard CNN model, the positive impact of the selective kernel SK module on model performance was verified. Figure 6 As shown, Figure 6 Figure (a) shows the scatter distribution of predicted and true values ​​and the fitting results obtained using the SK-CNN of this invention. Figure 6 Figure (b) shows the scatter distribution of predicted and true values ​​and the fitting results obtained using a standard CNN: the "ideal matching line" in the figure is the reference line where the predicted value and the true value match perfectly, and the pink area represents the 95% confidence interval of the model. It can be seen that SK-CNN converges tightly near the diagonal, and its 95% confidence band is significantly narrowed, indicating a very strong linear correlation between the predicted result and the true value, and a small variance in the residual distribution. This reveals that the standard CNN has a bottleneck in feature extraction under extreme conditions, while SK-CNN can effectively correct this systematic bias. It demonstrates that SK-CNN can accurately capture key features using an adaptive mechanism, reduce noise interference from redundant information, and significantly improve the model's dynamic adaptability and fitting accuracy to complex and variable conditions.

[0082] In actual operation, the pursuit is not solely focused on operational economy or nitrogen oxide emissions; rather, it involves a comprehensive optimization of both economic efficiency and pollutant generation. This invention presents a main steam temperature prediction model for coal-fired boilers based on a multi-branch convolutional neural network, offering advantages such as high accuracy, stability, and wide applicability. The improved model and its application effectively decouple transient changes in main steam temperature from steady-state trends, significantly reducing large prediction error samples under extreme operating conditions, thus enhancing the model's adaptability to different operating conditions and its reliability in engineering applications.

[0083] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for predicting and modeling the main steam temperature of a coal-fired boiler, characterized in that, Includes the following steps: Step 1: Extract historical data from the boiler monitoring system and preprocess it. Use the random forest algorithm to optimize candidate features and select core features. Step 2: Construct the selected core features into a two-dimensional temporal tensor, input it into the first standard convolutional layer, and extract the low-level local features; Step 3: Input the feature tensor output by the first standard convolutional layer into the selective kernel convolutional layer, and adaptively weight the features at different time scales to generate a multi-scale fused feature tensor. Step 4: Input the multi-scale fusion feature tensor output by the selective kernel convolutional layer into the second standard convolutional layer to extract high-dimensional deep features and obtain a high-dimensional feature tensor. Step 5: Decode the high-dimensional feature tensor output by the second standard convolutional layer through a fully connected network to output the predicted value of the main steam temperature.

2. The method for predicting and modeling the main steam temperature of a coal-fired boiler according to claim 1, characterized in that, The random forest algorithm was used for feature optimization to select the core feature set that plays a dominant role in the evolution of the main steam temperature. Specifically, the following calculations were performed: The weight of the j_cand-th candidate feature , represented as: ； in, Let j_cand be the arithmetic mean of the importance of the j-th candidate feature across all decision trees. For the j_cand-th candidate feature in the whole tree The importance score in the equation, where M represents the total number of candidate features from the boiler side and the turbine side; When the weight of a candidate feature is greater than a set threshold, the feature is considered a core feature, and all core features constitute the core feature set.

3. The method for predicting and modeling the main steam temperature of a coal-fired boiler according to claim 1, characterized in that, The first standard convolutional layer extracts local features from the lower layers, represented as: In the formula, It is a ReLU nonlinear activation function. This represents the set of input feature maps used in the computation. Tensors input to the first standard convolutional layer The Each feature channel This represents the discrete convolution operation. For the first standard convolutional layer, corresponding to the th The first output feature map A learnable convolutional kernel weight matrix; For the first layer Learnable bias terms corresponding to each output feature map; Features The output feature tensor of the first standard convolutional layer is obtained by stacking and concatenating along the channel dimension. .

4. The method for predicting and modeling the main steam temperature of a coal-fired boiler according to claim 1, characterized in that, Step 3, which involves inputting the feature tensor output from the first standard convolutional layer into a selective kernel convolutional layer to adaptively weight features at different time scales and generate a multi-scale fused feature tensor, specifically involves: Input feature tensor The stream is split into two independent convolutional paths. and Each feature branch representing a different receptive field is generated. and , represented as: The first path Adopting standards Convolution, the second path For expansion rate of Convolution, dilation rate Greater than 1; Then a fusion operation is performed, represented as: Fusion features Batch normalization is performed, and then the vector is compressed into a low-dimensional vector through a fully connected layer. , represented as: in, This represents the learnable weight matrix corresponding to the fully connected layer. This represents the fused features after batch normalization. The global channel statistical feature vector is obtained by performing global average pooling. It is the ReLU activation function; Soft attention weights for different scale branches are generated using the Softmax function. and , represented as in, and Linear transformation matrices corresponding to different scale branches; Perform feature fusion to obtain the first Channel characteristics , represented as: Features of each channel By concatenating the data, we obtain the multi-scale fused feature tensor generated by the selective kernel convolutional layer. .

5. The method for predicting and modeling the main steam temperature of a coal-fired boiler according to claim 4, characterized in that, The soft attention weights satisfy the normalization constraint. .

6. The method for predicting and modeling the main steam temperature of a coal-fired boiler according to claim 1, characterized in that, In step 4, the multi-scale fused feature tensor output from the selective kernel convolutional layer is input into the second standard convolutional layer to extract high-dimensional deep features, as shown below: in, Output tensors for selective kernel convolutional layers The Each feature map For the second standard convolutional layer, corresponding to the first The first output feature map A learnable convolutional kernel weight matrix, This is a learnable bias term for the second standard convolutional layer; All output features Together they constitute a high-dimensional feature tensor .

7. The method for predicting and modeling the main steam temperature of a coal-fired boiler according to claim 1, characterized in that, In step 5, the high-dimensional feature tensor output by the second standard convolutional layer is decoded through a fully connected network to output the predicted main steam temperature value, specifically: First, for high-dimensional feature tensors A flattening operation is performed to obtain a one-dimensional feature vector. This one-dimensional feature vector is then passed sequentially through stacked fully connected layers for non-linear mapping. The first fully connected layer... Layer output Represented as: in, and These are the first fully connected layers. Layer weight matrix and bias terms, This is the ReLU activation function for a fully connected layer.

8. The method for predicting and modeling the main steam temperature of a coal-fired boiler according to claim 7, characterized in that, The ReLU activation function of a fully connected layer is expressed as follows: 。 9. The method for predicting and modeling the main steam temperature of a coal-fired boiler according to claim 1, characterized in that, The mean squared error is used as the loss function, and the loss function is minimized through the backpropagation algorithm to learn the mapping law from the multivariate coupled state to the main steam temperature.

10. The method for predicting and modeling the main steam temperature of a coal-fired boiler according to claim 1, characterized in that, The preprocessing includes handling outlier data using the 3σ rule.

Images