Method, system and equipment for predicting wall temperature of boiler superheater and medium

Abstract

The invention discloses a boiler superheater wall temperature prediction method, system and equipment and a medium, and belongs to the technical field of safe operation of boiler heating surfaces. Key variables with the cumulative contribution rate larger than or equal to 95% are screened through a random forest; a BCNN-BAN fusion model is constructed, the BCNN adopts a RoSwsh activation function to extract multi-scale features, and the BAN focuses key features through bilinear attention; adopting an improved aurora optimization algorithm to optimize hyper-parameters; and inputting real-time data and outputting a predicted value. According to the method, the problems of weak nonlinear fitting capability and poor generalization of a traditional model are solved, overtemperature early warning can be effectively realized, technical support is provided for safe operation of the thermal power generating unit under flexible peak regulation, and the method has remarkable engineering application value.

Description

Technical Field

[0001] This invention belongs to the technical field of safe operation of boiler heating surfaces, and particularly relates to a method, system, equipment and medium for predicting the wall temperature of a boiler superheater. Background Technology

[0002] Currently, flexible peak shaving by thermal power units has become a crucial support for the operation of the power system. Frequent low-load operation and start-up / shutdown processes subject high-temperature components such as superheaters to drastically fluctuating conditions for extended periods, significantly increasing the risk of localized overheating and tube wall damage. Abnormal superheater wall temperature is one of the main causes of boiler tube rupture accidents, and its monitoring and prediction are directly related to the safe and stable operation of the unit. Especially in ultra-supercritical boilers, a continuous rise in superheater wall temperature can not only cause creep and oxide scale peeling of metal materials, but may also lead to systemic safety hazards such as SCR catalyst poisoning. Therefore, constructing an efficient, accurate, and deployable superheater wall temperature prediction model is of great significance for achieving overheating early warning, optimizing operation control, and extending equipment life.

[0003] Currently, research on superheater wall temperature prediction mainly focuses on two directions: one is numerical simulation methods based on physical mechanisms, which typically employ computational fluid dynamics (CFD) models to jointly model the heat transfer processes on the flue gas and steam sides inside the boiler, thereby achieving high-precision spatial prediction of wall temperature distribution. The other is soft sensing methods based on machine learning, which establish a nonlinear mapping relationship between input variables and wall temperature through historical data, enabling rapid prediction and dynamic monitoring of wall temperature. The first method has good physical interpretability and adaptability to operating conditions, and can reveal the coupling relationship between local heat transfer processes and structural responses. However, it usually relies on a large number of boundary conditions and structural parameter settings, resulting in complex modeling and high computational resource consumption, making it difficult to meet the needs of rapid response and real-time analysis in actual operation.

[0004] In recent years, with the rapid development of artificial intelligence technology, data-driven modeling methods have gradually become a research hotspot in the field of superheater wall temperature prediction. These methods have shown significant advantages in improving prediction efficiency and reducing computational costs. Existing research has introduced various neural network structures, including Convolutional Neural Networks (CNNs), Long Short-Term Memory Networks (LSTMs), Gated Recurrent Units (GRUs), and Nonlinear Autoregressive Models (NARX), to enhance the ability to model temperature time-series features, thereby improving prediction accuracy and model generalization ability. Among them, CNNs and Temporal Convolutional Networks (TCNs) are widely used in temperature sequence modeling tasks due to their outstanding performance in extracting local patterns and short-term dependency features. For example, existing technologies include multi-step wall temperature prediction models based on encoder-decoder architectures, which significantly improve the model's trend capture ability and prediction accuracy; and a multi-scale temporal modeling framework integrating TCNs, CNNs, and Transformers, which exhibits excellent stability and accuracy under complex fluctuating conditions. To further improve the feature selection capability and generalization performance of the model, the existing technology introduces grey relational analysis to screen input variables and combines it with LSTM to construct a wall temperature time series prediction model. Existing technologies also include a deep neural network (GA-DNN) model optimized by a genetic algorithm, which identifies and warns of the outlet temperature and overheating zone of the screen-type superheater, and can accurately locate the overheating risk tube screen 5 minutes in advance.

[0005] While convolutional neural networks (CNNs) have demonstrated significant advantages in modeling capabilities and nonlinear feature learning, the layered structure design employed by mainstream methods often results in limited effective receptive fields and weak cross-layer information fusion mechanisms. Particularly in industrial scenarios with strong nonlinear coupling and dynamic time delays, such as modeling the wall temperature of high-temperature superheaters, existing methods still exhibit significant shortcomings in areas such as feature space redundancy suppression, robustness of multi-scale feature representation, and generalization performance. Summary of the Invention

[0006] To address the aforementioned technical challenges, this invention proposes a method, system, equipment, and medium for predicting boiler superheater wall temperature. By constructing a fusion model based on Broad Convolutional Neural Networks (BCNN) and Bilinear Attention Networks (BAN), and combining the RoSwish activation function with improved PLO hyperparameter optimization, multi-dimensional technological breakthroughs are achieved in boiler superheater wall temperature prediction. This enables accurate prediction of boiler superheater wall temperature, providing technical support for the safe operation of thermal power units. Experimental verification demonstrates excellent performance in prediction accuracy, generalization, and engineering adaptability, providing an innovative reference for research on boiler heating surface wall temperature prediction and strongly supporting the safe operation of thermal power units.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for predicting boiler superheater wall temperature includes the following steps:

[0009] Step 1: Collect historical operating data of the boiler from the distributed control system of the unit, preprocess the historical operating data, and construct a standardized dataset;

[0010] Step 2: Use the random forest algorithm to reduce the dimensionality of the original variables affecting the boiler superheater wall temperature, select the first m variables as model input, and divide the standardized dataset into training set and test set;

[0011] Step 3: Construct a boiler superheater wall temperature prediction model based on BCNN-BAN, which is composed of a wide convolutional neural network connected in series with a bilinear attention network. The activation function of BCNN is the rotating Swish function.

[0012] Step 4: Train a boiler superheater wall temperature prediction model based on BCNN-BAN using the training set data, and optimize the model hyperparameters using the improved Aurora Optimization Algorithm to obtain the optimized boiler superheater wall temperature prediction model based on BCNN-BAN.

[0013] Step 5: Input the test set data into the optimized BCNN-BAN-based boiler superheater wall temperature prediction model, obtain the prediction results and perform inverse normalization, and then use the evaluation function to analyze and evaluate the model prediction performance and verify its prediction performance under different operating conditions.

[0014] Step 6: Input the real-time collected operating parameters of the coal-fired power generation unit into the optimized BCNN-BAN-based boiler superheater wall temperature prediction model, and output the predicted value of the boiler superheater wall temperature to realize early warning of boiler superheater overheating risk.

[0015] Furthermore, in step 1, the collected historical operating data includes boiler superheater wall temperature and auxiliary variables; the preprocessing of the historical operating data includes: filling missing values ​​and outliers with linear interpolation according to time labels, and then performing normalization processing.

[0016] Furthermore, in step 2, the top m variables with a cumulative contribution rate greater than or equal to 95% are selected as model inputs, where m is the number of feature variables corresponding to a cumulative contribution rate ≥ 95%, which is determined based on the actual unit's operating data characteristics.

[0017] Furthermore, in step 3, the constructed boiler superheater wall temperature prediction model based on BCNN-BAN includes an input layer, a BCNN layer, a BAN layer, and an output layer.

[0018] Furthermore, in step 3, the input layer is used to receive the m-dimensional model input after dimensionality reduction by random forest; the BCNN layer adopts a multi-branch parallel convolutional structure, and extracts local features from the input data through K parallel convolutional branches; the activation function used by the BCNN layer is the Swish function, which is used to dynamically adjust the rotation angle and offset of the function to adapt to the nonlinear feature distribution of the boiler superheater operation data; the BCNN layer concatenates the feature maps of the K branches in the channel dimension to obtain the final features of the BCNN layer.

[0019] Furthermore, in step 3, the BAN layer receives the final features from the BCNN layer, obtains two sets of features through two linear transformations, and then performs bilinear pooling fusion; weights are assigned to each dimension of the fused features through an attention mechanism; the output layer maps the attention features to the predicted value of the boiler superheater wall temperature through a fully connected layer.

[0020] Furthermore, in step 4, the model hyperparameters include the number of neurons and the learning rate. The optimization range for the number of neurons is [32, 64, 128, 256], and the optimization range for the learning rate is [0.001, 0.0001, 0.00001, 0.000001].

[0021] The present invention also provides a boiler superheater wall temperature prediction system for implementing the above method, comprising:

[0022] The data acquisition module is used to acquire historical operating data of the unit's distributed control system;

[0023] The preprocessing module is used to perform missing data imputation, outlier handling, data dimensionality reduction, and normalization on the historical running data;

[0024] The model calculation module is used to deploy a boiler superheater wall temperature prediction model based on BCNN-BAN, and perform prediction calculations to obtain the prediction results of boiler superheater wall temperature.

[0025] A visualization terminal is used to display the predicted wall temperature of the boiler superheater online.

[0026] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the program to implement the steps of a method for predicting the wall temperature of a boiler superheater.

[0027] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described boiler superheater wall temperature prediction method.

[0028] Beneficial effects:

[0029] (1) This invention adopts a BCNN-BAN fusion architecture, which overcomes the limitations of traditional CNNs, such as limited receptive field and insufficient feature extraction breadth due to single attention mechanism. BCNN achieves broad extraction of global multi-scale features through multi-branch parallel convolution, while BAN mines high-order feature correlations through bilinear pooling and accurately focuses on key features by combining attention mechanism. The two work together to achieve feature extraction effects of broad coverage and accurate selection. Using only BCNN will cause the core influencing factors to be submerged due to feature redundancy; using only BAN is limited by the receptive field and cannot cover global features. When the two work together, BCNN is responsible for extensively extracting features of all dimensions of working conditions, while BAN accurately focuses on key features of wall temperature correlation through bilinear attention. This not only preserves the integrity of global features but also strengthens the weight of core features, solving the problem of incomplete or redundant feature extraction under complex working conditions in traditional models.

[0030] (2) This invention proposes a learnable parameter , The RoSwish activation function, compared to ReLU and traditional Swish, can adaptively adapt to the strong nonlinear characteristics of boiler operating data, improving the model's ability to fit complex operating conditions. The dynamic width of BCNN... The breadth of feature extraction (number of network branches / channels) can be adaptively adjusted according to changes in boiler operating conditions; while the learnable parameters of the RoSwish activation function... , It can dynamically adjust the amplitude of the feature response. The two form an amplitude-channel dual-dimensional adaptive mechanism, which enables the model to maintain an accurate fit to wall temperature changes under dynamic operating conditions such as unit peak shaving and fuel switching. Compared with ReLU and traditional Swish activation functions, the nonlinear expression capability is improved by more than 20%.

[0031] (3) This invention addresses the problems of premature convergence in traditional genetic algorithms and low accuracy in high-dimensional Bayesian optimization by employing improved strategies such as aurora elliptical walk value updates, particle collision mechanisms, and adaptive weights. Specifically, for the collaborative optimization scenario involving multiple hyperparameters (number of neurons, learning rate, kernel size, number of branches, etc.) in the BCNN-BAN fusion model, the convergence speed is improved by over 30%, significantly enhancing the model's global optimality. This improves both global convergence capability and convergence efficiency. Ultimately, the optimal hyperparameter combination selected for the BCNN-BAN model minimizes the prediction error (MAE), far surpassing the performance of traditional optimization algorithms.

[0032] (4) This invention integrates the random forest feature dimensionality reduction and the BCNN-BAN fusion model for accurate prediction. First, the key variables with a cumulative contribution rate of ≥95% are screened through random forest, and then input into the BCNN-BAN fusion model to achieve wall temperature prediction, which not only ensures the effectiveness of features, but also improves the practicality of engineering.

[0033] In summary, through the synergistic innovation of multiple technologies, this invention has achieved breakthroughs in prediction accuracy, operating condition adaptability, and engineering practicality, providing a new technical solution for predicting boiler heating surface wall temperature and strongly supporting the safe and efficient operation of thermal power units in flexible peak-shaving scenarios. Attached Figure Description

[0034] Figure 1 This is a flowchart of the boiler superheater wall temperature prediction method of the present invention;

[0035] Figure 2 This example shows the changes in boiler load and superheater wall temperature in the selected operating data.

[0036] Figure 3 The contribution of random forest dimensionality reduction features in the example;

[0037] Figure 4 This is a flowchart of the hyperparameter optimization process of the improved aurora optimization algorithm used in the embodiment.

[0038] Figure 5 This is a graph showing the prediction results of the validation set before hyperparameter optimization in the example.

[0039] Figure 6 This is a graph showing the prediction results of the validation set after hyperparameter optimization in the example.

[0040] Figure 7 The above are the prediction results of BCNN-BAN with hyperparameter optimization in the example.

[0041] Figure 8The above shows the prediction results of the boiler superheater wall temperature prediction model based on BCNN-BAN in the embodiment on the validation set.

[0042] Figure 9 The example shows the prediction results of the boiler superheater wall temperature prediction model based on BCNN on the validation set.

[0043] Figure 10 The example shows the prediction results of the BAN-based boiler superheater wall temperature prediction model on the validation set.

[0044] Figure 11 The graph shows how the model training loss function changes with the number of iterations in the example. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0046] like Figure 1 As shown, the present invention provides a method for predicting the wall temperature of a boiler superheater, comprising the following steps:

[0047] Step 1: Collect historical operating data of the boiler from the distributed control system of the unit, preprocess the historical operating data, and construct a standardized dataset;

[0048] Step 2: Use the random forest algorithm to reduce the dimensionality of the original variables affecting the boiler superheater wall temperature, determine the feature values, select the top m variables as model input, and finally divide the standardized dataset into training and test sets; Step 3: Construct a boiler superheater wall temperature prediction model based on BCNN-BAN (four-layer network: input layer, BCNN layer, BAN layer, and output layer connected in sequence); it is composed of a wide convolutional neural network connected in series with a bilinear attention network, where the activation function of BCNN is the rotation Swish function; Step 4: Train the boiler superheater wall temperature prediction model based on BCNN-BAN using the data in the training set, and optimize the model hyperparameters through the improved aurora optimization algorithm to obtain the optimized boiler superheater wall temperature prediction model based on BCNN-BAN;

[0049] Step 5: Input the test set data into the optimized BCNN-BAN-based boiler superheater wall temperature prediction model, obtain the prediction results and perform inverse normalization, and then use the evaluation function to analyze and evaluate the model prediction performance and verify its prediction performance under different operating conditions.

[0050] Step 6: Input the real-time collected operating parameters of the coal-fired power generation unit into the optimized BCNN-BAN-based boiler superheater wall temperature prediction model, and output the predicted value of the boiler superheater wall temperature to realize early warning of boiler superheater overheating risk.

[0051] Preferably, in step 1, the preprocessing of historical operating data of the distributed control system of the unit includes: first, using linear interpolation to fill in missing and outlier values ​​in the historical operating data according to time labels, and then performing normalization processing.

[0052] Preferably, in step 2, the top m variables with a cumulative contribution rate greater than or equal to 95% are selected as model inputs (m is the number of feature variables corresponding to a cumulative contribution rate ≥ 95%, which is determined based on the actual unit's operating data characteristics).

[0053] Preferably, in step 3, the constructed boiler superheater wall temperature prediction model based on BCNN-BAN includes an input layer, a BCNN layer, a BAN layer, and an output layer, and the activation function is the Rotation Swish function (RoSwish).

[0054] Preferably, in step 4, the model hyperparameters include the number of neurons and the learning rate, and the hyperparameter optimization of the neural network uses an intelligent optimization algorithm (improved Aurora Optimization Algorithm PLO).

[0055] Preferably, in step 4, the optimization range of the number of neurons is [32, 64, 128, 256], and the optimization range of the learning rate is [0.001, 0.0001, 0.00001, 0.000001]. The final optimized number of neurons is determined to be 128 and the learning rate is 0.0001.

[0056] Example

[0057] This invention implements a method for predicting boiler superheater wall temperature based on BCNN-BAN, including the following steps:

[0058] Step 1: Collect historical operating data from the unit's distributed control system (DCS) and preprocess it:

[0059] (1) Data Collection. Taking the high-temperature superheater of a 1000MW ultra-supercritical once-through reheat generator unit as the research object, historical data from 21:17:00 on May 22, 2025 to 19:57:00 on May 29, 2025 were extracted from the unit's distributed control system (DCS), totaling 10,000 samples with a sampling time interval of 60 seconds. Through in-depth analysis of the historical operating data, it was found that the wall temperature at measuring point 2 of the high-temperature superheater was consistently high regardless of the load stage. Therefore, the wall temperature at this measuring point was selected for analysis.

[0060] The selected data includes load and superheater wall temperature variations, as follows: Figure 2 As shown, the load variation range is [295MW, 1002MW], and the wall temperature variation range is [581℃, 628℃]. It can be seen that the selected data covers various complex operating conditions, and the superheater wall temperature may exceed the limit.

[0061] (2) Data Preprocessing. Given the complex environment of thermal power unit production sites and the susceptibility of acquired signals to noise interference, missing and outlier values ​​are inevitable in the large amount of raw data obtained. This embodiment uses a 3σ-rule to detect outlier data in boiler operation data and replaces outlier data based on the following formula:

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] in, Representing auxiliary variables The rate of change at adjacent times t and t-1, express The average rate of change before time t, Let t represent the standard deviation at time t, where t=2,…,n are the sampling times, and n represents the total sampling time. This represents the time index value.

[0067] The data after replacing outlier data is linearly normalized using the following formula to standardize the data size:

[0068] ;

[0069] In the formula, , Let represent the minimum and maximum values ​​of the i-th variable, respectively. Represents the original data. This is the result of normalization.

[0070] Step 2: Dimensionality reduction of the data based on random forest, and partitioning the dataset:

[0071] Based on historical data, the Random Forest (RF) algorithm was used to calculate the feature importance scores for 106 key parameters affecting superheater wall temperature, including unit load, total fuel consumption, total air volume, feedwater flow rate, secondary air flow rate, main steam pressure, main steam flow rate, and main steam temperature. This score was then used to select key feature variables and reduce input dimensionality. The results of the Random Forest analysis of candidate variables are shown below. Figure 3 As shown. Figure 3 The vertical axis represents the importance (or contribution rate) of each feature, while the curve on the bottom plane shows the cumulative importance of different numbers of features. According to the contribution rate curve, the cumulative importance of the first 67 variables is 95%, which basically covers the main information of the original data. Based on the mechanistic analysis, five input variables were removed: "Reheat steam temperature B side temperature ℃", "Reheater outlet temperature ℃", "Reheater pressure MPa", "Reheat steam temperature A side temperature ℃", and "Coal feed rate E (t / h)". Finally, the first m = 62 variables were selected as the input variables of the model, as shown in Table 1.

[0072] Table 1. 62 selected variables and their contributions

[0073]

[0074] The preprocessed subsequence dataset was divided into two sets: 7976 sets were selected as the training set and 1994 sets were selected as the test set, with a ratio of 8:2.

[0075] Step 3: Construct a boiler superheater wall temperature prediction model based on BCNN-BAN.

[0076] The boiler superheater wall temperature prediction model based on BCNN-BAN includes an input layer, a BCNN layer, a BAN layer, and an output layer. The activation function is the Rotation Swish function (RoSwish).

[0077] (1) Input layer: Receives the 62-dimensional input feature values ​​after dimensionality reduction in step 2. This data is represented in matrix form as follows:

[0078] ;

[0079] In the formula, H represents the input raw monitoring data matrix; H is the height of the input data, corresponding to the time series length or spatial row number of the monitoring data; W is the width of the input data, corresponding to the feature dimension or spatial column number of the monitoring data; C is the number of channels of the input data, corresponding to different types of monitoring parameters.

[0080] (2) BCNN layer:

[0081] To address the shortcomings of Wide Learning Systems (BLS) and One-Dimensional Convolutional Neural Networks (1D-CNN), this invention proposes Wide Convolutional Neural Networks (BCNN). Unlike traditional stacked convolutional structures, this algorithm borrows the "width" concept from BLS to improve the network structure. It employs a mechanism of parallel extraction of feature nodes and enhancement nodes, followed by fusion, thereby enhancing the model's feature learning capability. Specifically, feature extraction utilizes one-dimensional convolution operations with adjustable kernel size. This flexible feature extraction approach enhances the model's adaptability to multi-scale features.

[0082] (2.1) Multi-branch convolutional layer. BCNN adopts a multi-branch parallel convolutional structure, which extracts local features from the input data through K parallel convolutional branches. The convolutional operation of each branch is as follows:

[0083] ;

[0084] In the formula, This represents the output feature map of the k-th convolutional branch. K represents the total number of convolutional branches. The convolution operation includes local window sliding and weighted summation. The convolution kernel parameters for the k-th branch are... Represents the set of real numbers. and Here, represents the height and width of the k-th convolutional kernel, and C is the number of input channels. This represents the number of output channels for the k-th branch. This is the bias term for the k-th branch; Residual connections are used to alleviate the vanishing gradient problem in deep networks and enhance feature propagation.

[0085] (2.2) In the multi-branch convolutional layers of BCNN, the activation function is the Rotation Swish function (RoSwish), whose expression is:

[0086] ;

[0087] in, These are the feature values ​​after the convolution operation; For the Sigmoid function, and These are learnable parameters used to dynamically adjust the function's rotation angle and offset to adapt to the nonlinear characteristic distribution of boiler superheater operating data. This activation function incorporates learnable parameters. and Compared to the traditional Swish function, it has stronger adaptability and can better capture the nonlinear variation law of superheater wall temperature related features, thus improving the feature extraction capability of BCNN for monitoring data under complex operating conditions.

[0088] (2.3) Feature Fusion and Output. The feature maps of the K branches are concatenated along the channel dimension to obtain the final feature output of BCNN:

[0089] ;

[0090] In the formula, The fused features output by BCNN; This is a channel-level splicing operation; and These are the height and width of the fused feature map, respectively, determined by parameters such as convolution stride and padding.

[0091] (3) BAN layer:

[0092] BAN is a deep learning architecture that integrates feature interaction relationships through bilinear pooling and attention mechanisms and focuses on key influencing factors.

[0093] (3.1) Bilinear Feature Fusion. BAN receives features output from BCNN. Two sets of features are obtained through two linear transformations, and then bilinear pooling is performed to fuse them:

[0094] ;

[0095] in, Indicates bilinear fusion. This represents the first set of linear transformation characteristics; This represents the characteristics of the second group of linear transformations; and This is the weight matrix for the linear transformation; and This is the bias term for the linear transformation; For bilinear pooling operations, element-wise multiplication or outer product can be used to enhance the interaction between features; This is the feature vector after bilinear fusion.

[0096] (3.2) Calculation of attention weights

[0097] The attention mechanism is used to fuse the bilinear feature vector. Each dimension is weighted to highlight features that play a key role in predicting boiler superheater wall temperature:

[0098] ;

[0099] ;

[0100] In the formula, The feature vector after bilinear fusion The i-th dimension, D is the feature vector after bilinear fusion. The dimension; The feature vector after bilinear fusion The j-th dimension, j ; Let be the scoring function, which measures the importance of features through a linear transformation, where The weight vector of the scoring function. This is the bias term of the scoring function. For scoring functions; The attention weights for the i-th feature dimension are normalized using the softmax function to ensure that the sum of the weights is 1. The weighted attention features focus on key features, and exp(·) represents the exponential function.

[0101] (4) Output layer:

[0102] Attention features are mapped to predicted values ​​of boiler superheater wall temperature using a fully connected layer:

[0103] ;

[0104] In the formula, This is the predicted value for the boiler superheater wall temperature; This is the weight matrix of the output layer; This is the bias term for the output layer.

[0105] Step 4: Train a boiler superheater wall temperature prediction model based on BCNN-BAN using the training set data, and optimize the model hyperparameters by improving the Aurora Optimization Algorithm.

[0106] All experiments were conducted using the PyTorch deep learning framework. The hardware consisted of an AMD Ryzen 98945HS CPU with 32GB of RAM and 8GB of GPU memory, and Windows 11 as the operating system. Step 2 above selected 7976 datasets as the training set. To better evaluate the model's performance, this invention used Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²). 2 Evaluate model performance:

[0107]

[0108] In the formula, , , These represent the actual measured value, the predicted value, and the mean of the actual measured value, respectively. n represents the total number of samples, and t represents the index value.

[0109] To further improve the prediction accuracy and fitting effect of the network model, the hyperparameters of the model are optimized. This invention employs an improved aurora optimization algorithm to optimize the model's hyperparameters. The aurora optimization algorithm (PLO) is a metaheuristic optimization algorithm inspired by the natural aurora phenomenon, which achieves global optimization by simulating the motion characteristics, energy conversion, and spatial distribution patterns of aurora particles. Figure 4 The flowchart for hyperparameter optimization of the improved aurora optimization algorithm used in the embodiment is as follows: ① Initialize time, maximum number of iterations (set to 200), particle swarm (particles correspond to hyperparameter combinations, particle number set to 30), calculate the fitness of each particle (based on the model MAE+MAPE weighted value) and sort them; ② Update the cyclotron speed, aurora elliptical walk value and adaptive weight of each particle (preferably weight range of 0.4~0.9); ③ Determine whether the particle collision mechanism is triggered (triggered when the particle fitness has not improved for 5 consecutive generations), if so, regenerate some particles to escape the local optimum; ④ Iterate the particles using a greedy selection mechanism, retaining the 50% of particles with the best fitness; ⑤ Calculate the fitness of the new particle swarm and update the globally optimal particle; ⑥ Determine whether the termination condition is met (reaching the maximum number of iterations or the optimal fitness ≤ 1.5℃), if so, output the optimal particle (corresponding to the optimal hyperparameter), otherwise return to ② to continue iterating.

[0110] Based on the improved aurora optimization algorithm parameter optimization model prediction results, for example... Figures 5-6 As shown (using test set data). Before hyperparameter optimization ( Figure 5 The model exhibits significant bias in the sample intervals corresponding to the two subplots, with a tracking delay of approximately 2-3 minutes for the dynamic trend of wall temperature. Its ability to capture peak values ​​in areas of drastic local wall temperature changes is poor, with predicted values ​​differing from measured values ​​by 1%-2%. After hyperparameter optimization using the improved aurora optimization algorithm (…), Figure 6 The model's predictive performance on the validation set was significantly improved, with a dynamic trend tracking delay of less than 1 minute; the ability to capture peak wall temperature changes was greatly enhanced, with the deviation between predicted and measured values ​​being less than 1%.

[0111] The iterative process of hyperparameter optimization is shown in Table 2. One optimal record is selected and displayed every five iterations, for a total of 20 data points. After selecting the optimal hyperparameters, the final prediction performance of the model on the training and validation sets is as follows: Figure 7 As shown.

[0112] As shown in Table 2, after optimization using the improved aurora optimization algorithm, the optimal hyperparameters and their results are: lr=0.0522, N1=189, N2=N 1= 189, Weight_Decay=1e-05, MAE= 1.691, MAPE=0.277, RMSE=2.117, R 2 =0.883, training time is 25.11 s. Where Lr (Learning Rate) is the model learning rate, which controls the step size of parameter updates; Weight_Decay is the weight decay coefficient, used to prevent the model from overfitting; N1 is the number of neurons in the first stage of the model; N2 is the number of neurons in the second stage of the model.

[0113] Table 2 Hyperparameters of the improved aurora optimization algorithm during the iterative process

[0114]

[0115] For other parameters, the prediction performance was best when the number of neurons was 189. The Adam optimization algorithm was used to optimize the learning rate, and the prediction performance was best when the learning rate was 0.0001. Figure 7 When the superheater wall temperature fluctuates, the boiler superheater wall temperature prediction model based on BCNN-BAN proposed in this invention can accurately predict its changing trend. The model prediction results can predict the inflection point of the measured value change of the superheater wall temperature, remain stable during the relatively stable temperature period, and match the actual measured value well.

[0116] Step 5: Input the test set data into the optimized BCNN-BAN-based boiler superheater wall temperature prediction model, obtain the prediction results and perform inverse normalization processing, and use the evaluation function to analyze and evaluate the model prediction performance.

[0117] To verify the optimality of the proposed model, a comparative experiment was conducted between the boiler superheater wall temperature prediction model based on BCNN-BAN and the BCNN and BAN models. To reduce the influence of random factors on the prediction results of the various models, each neural network model underwent multiple repeated experiments using the same historical dataset. The prediction results of different models are compared below. Figures 8-10 As shown ( Figure 8 The results of the BCNN-BAN-based boiler superheater wall temperature prediction model on the validation set are shown. Figure 9 The example shows the prediction results of the boiler superheater wall temperature prediction model based on BCNN on the validation set. Figure 10 (This is the prediction result of the BAN-based boiler superheater wall temperature prediction model on the validation set in the example). Figures 8-10The blue curve represents the measured superheater wall temperature, the red curve represents the predicted superheater wall temperature, and the green bars indicate the error between the measured and predicted values ​​at each sample point. A comparison of quantitative evaluation indicators is shown in Table 3. Table 3 shows that the boiler superheater wall temperature prediction model based on BCNN-BAN has the lowest error rates (RMSE=2.117℃, MAE=1.691℃, MAPE=0.277%) and the lowest coefficient of determination (R²). 2 =0.883) is closest to 1, indicating that its prediction performance is optimal.

[0118] Depend on Figures 8-10 The comparison results with Table 3 show that the proposed BCNN-BAN model performs best in terms of prediction accuracy. BCNN, through its wide convolutional structure, can simultaneously capture multi-dimensional and multi-scale original features from the input data, overcoming the limitation of a single network in terms of feature coverage breadth. Meanwhile, BAN's bilinear attention mechanism, through bilinear pooling and weight allocation of features, accurately locates the key features most significantly affecting wall temperature and suppresses interference from irrelevant or noisy features, compensating for the shortcomings of simple wide convolution in distinguishing feature importance. The combination of the two is not a simple structural superposition, but rather leverages the "wide coverage" of BCNN to provide a rich feature base, and the "precise selection" of BAN to strengthen the contribution of key information. This allows the model to comprehensively capture potential influencing factors and focus on core driving variables when dealing with complex scenarios involving strong coupling of multiple variables and dynamic changes in boiler superheaters. Therefore, it outperforms single BCNN or BAN models in both prediction accuracy and adaptability to changing operating conditions.

[0119] Table 3. Performance metrics comparison between BCNN-BAN and baseline models

[0120]

[0121] also, Figure 11 The graph shows the variation of the model training loss function with the number of iterations in the embodiment (normalized result). The boiler superheater wall temperature prediction model based on BCNN-BAN gradually stabilizes and converges after approximately 20 iterations. Therefore, compared with the baseline prediction model, the boiler superheater wall temperature prediction model based on BCNN-BAN proposed in this invention not only has a more accurate prediction effect, but also exhibits superior dynamic performance in predicting the actual boiler heating surface wall temperature.

[0122] Step 6: Input the real-time collected generator set operating parameters into the trained BCNN-BAN-based boiler superheater wall temperature prediction model, and output the predicted superheater wall temperature value for future time periods.

[0123] This invention also provides a boiler superheater wall temperature prediction system for implementing the above method, comprising the following modules:

[0124] The data acquisition module is used to acquire historical operating data of the unit's distributed control system;

[0125] The preprocessing module is used to perform missing data imputation, outlier handling, data dimensionality reduction, and normalization on the historical running data;

[0126] The model calculation module is used to deploy a boiler superheater wall temperature prediction model based on BCNN-BAN, and perform prediction calculations to obtain the prediction results of boiler superheater wall temperature.

[0127] A visualization terminal is used to display the predicted wall temperature of the boiler superheater online.

[0128] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the program to implement the steps of the above-described method for predicting the wall temperature of a boiler superheater.

[0129] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for predicting boiler superheater wall temperature.

[0130] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0131] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0132] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0133] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0134] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0135] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting the wall temperature of a boiler superheater, characterized by, The method comprises the following steps: Step 1, collecting historical operation data of the boiler from a unit decentralized control system, and preprocessing the historical operation data to construct a standardized data set; Step 2, using a random forest algorithm to reduce the dimension of original variables affecting the wall temperature of the boiler superheater, selecting the first m variables as the model input, and dividing the standardized data set into a training set and a test set; Step 3, constructing a boiler superheater wall temperature prediction model based on BCNN-BAN, which is composed of a wide convolutional neural network in series with a bilinear attention network, and the activation function of the BCNN is a rotating Swish function; Step 4, training the boiler superheater wall temperature prediction model based on BCNN-BAN using the data of the training set, and optimizing the model hyperparameters through an improved aurora optimization algorithm to obtain an optimized boiler superheater wall temperature prediction model based on BCNN-BAN; Step 5, inputting the data of the test set into the optimized boiler superheater wall temperature prediction model based on BCNN-BAN, obtaining the prediction results and performing reverse normalization processing, then analyzing and evaluating the prediction performance of the model using an evaluation function, and verifying the prediction performance of the model under different working conditions; Step 6, inputting the real-time collected operation parameters of the coal-fired power generating unit into the optimized boiler superheater wall temperature prediction model based on BCNN-BAN, outputting the predicted value of the boiler superheater wall temperature, and realizing early warning of the over-temperature risk of the boiler superheater.

2. The boiler superheater wall temperature prediction method of claim 1, wherein, In step 1, the historical operation data includes the wall temperature of the boiler superheater and auxiliary variables; the preprocessing of the historical operation data includes: filling the missing values and abnormal values using linear interpolation according to the time label, and then performing normalization processing.

3. The method of claim 1, wherein, In step 2, the first m variables with a cumulative contribution rate greater than or equal to 95% are selected as the model input.

4. The method of claim 1, wherein, In step 3, the boiler superheater wall temperature prediction model based on BCNN-BAN comprises an input layer, a BCNN layer, a BAN layer and an output layer.

5. The method of claim 4, wherein, The input layer is used to receive the m-dimensional model input after random forest dimension reduction; the BCNN layer adopts a multi-branch parallel convolution structure, and locally extracts features of the input data through K parallel convolution branches; the rotating Swish function adopted by the BCNN layer is used to dynamically adjust the rotation angle and offset of the function to adapt to the nonlinear feature distribution of the boiler superheater operation data; the BCNN layer splices the feature maps of the K branches in the channel dimension to obtain the final features of the BCNN layer.

6. The boiler superheater wall temperature prediction method of claim 5, wherein, The BAN layer receives the final features of the BCNN layer, obtains two groups of features through two linear transformations, and then performs bilinear pooling fusion; the attention mechanism is used to assign weights to each dimension of the fused features; the output layer maps the attention features to the predicted value of the boiler superheater wall temperature through a fully connected layer.

7. The method of claim 1, wherein, In step 4, the model hyperparameters include the number of neurons and the learning rate, the optimization range of the number of neurons is [32, 64, 128, 256], and the optimization range of the learning rate is [0.001, 0.0001, 0.00001, 0.000001].

8. System for implementing the method for predicting the wall temperature of a boiler superheater according to any one of claims 1-7, characterized in that, It comprises: A data collection module is configured to acquire historical operation data of the unit decentralized control system. A preprocessing module is configured to perform missing data filling, outlier processing, data dimension reduction and normalization processing on the historical operation data. A model operation module is configured to deploy a boiler superheater wall temperature prediction model based on BCNN-BAN, and perform prediction calculation to obtain a boiler superheater wall temperature prediction result. A visual terminal is configured to display the boiler superheater wall temperature prediction result online. 9.An electronic device comprising a processor and a memory, the electronic device characterized by: The memory stores a computer program, and the processor executes the program to implement the steps of the boiler superheater wall temperature prediction method according to any one of claims 1-7. 10.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the boiler superheater wall temperature prediction method according to any one of claims 1-7.

Images