Boiler water wall temperature prediction method, overtemperature early warning method, system and medium

Abstract

The present invention discloses a boiler water-cooled wall temperature prediction method, an overtemperature warning method, a system and a medium, and belongs to the technical field of boiler safe operation. The boiler water-cooled wall temperature prediction method of the present invention comprises the following steps: 1, evaluating the characteristic importance of auxiliary variables to the water-cooled wall temperature through the Gain mode of the LightGBM algorithm, and screening out the top D key variables; 2, optimizing the delay time between each auxiliary variable and the target variable by using the average mutual information AMI and differential evolution DE algorithm, and reconstructing the input variable data; 3, dynamically optimizing the hidden layer node parameters of the random configuration network SCN based on the hiking algorithm HOA, constructing an HOA-SCN model, and introducing L2 regularization to calculate the output weight; and 4, predicting the water-cooled wall temperature by using the trained HOA-SCN model. The prediction method of the present invention is used to significantly improve the prediction accuracy and robustness, thereby providing effective support for the safe operation of the boiler.

Description

Technical Field

[0001] The present invention belongs to the technical field of safe operation of boilers, and in particular relates to a method for predicting the temperature of a boiler water-cooled wall, an over-temperature early warning method, a system and a medium. Background Art

[0002] Ultra-supercritical (UFC) units, due to their high efficiency, have become the workhorse of the power grid. However, serious tube bursts in the boiler's "four tubes" have significantly impacted unit safety. Thermal power units frequently participate in deep peak-shaving operations, which increases boiler combustion instability, severely impacts heat transfer within the furnace, and compromises the safety of the water-cooled walls. This can even lead to heating surface leaks and even explosions, limiting the unit's ability to participate in peak-shaving operations. Accurately predicting furnace wall temperatures is key to resolving overheating issues.

[0003] In recent years, the use of data-driven methods to establish boiler combustion models and predict wall temperature has become a research hotspot in this field.

[0004] As we all know, traditional neural networks lack interpretability, and the iterative search for a set of appropriate weights and biases is time-consuming and computationally intensive, which makes it difficult to efficiently process large-scale industrial data, especially power production process data.

[0005] Stochastic Configuration Networks (SCNs) are an incremental learning method with universal approximation properties. They employ a supervised mechanism to randomly configure the parameters of hidden layer nodes under a set of inequality constraints, adaptively selecting their value range. This method offers the significant advantages of eliminating manual intervention in the network structure and adaptively selecting the number of hidden layer nodes. After determining the hidden layer parameters using a stochastic configuration algorithm, the output weights are calculated using pseudo-inverse theory. This mechanism makes SCN training faster than traditional gradient-based iterative algorithms. Currently, stochastic configuration networks have been applied to model various industrial scenarios. This present invention marks the first application of stochastic configuration networks to model the metal wall temperature of boiler heating surfaces. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to realize the application of random configuration network (SCN) in the modeling of boiler water-cooled wall temperature. In order to improve the efficiency of traditional random configuration network (SCN) search weights and deviations and improve the generalization and prediction performance of the model, the present invention introduces the Hiking Optimization Algorithm (HOA) into the random configuration network and establishes a boiler water-cooled wall temperature prediction model based on HOA-SCN. This model enables boiler operators to predict the metal temperature of the heating surface in advance and provide guidance for adjusting the tube wall overtemperature.

[0007] The present invention proposes a method for predicting boiler water wall temperature, comprising the following steps:

[0008] Step 1: Select auxiliary variables that have an important impact on the water wall temperature, and the auxiliary variables are D in total;

[0009] Step 2: Build a water wall temperature prediction model. The steps are as follows:

[0010] Step 2.1, given the objective function , assuming that the L-1 neural nodes of the hidden layer of the random configuration network SCN have been constructed; among them, represents the dimensional space of the input, Represents the output dimensional space, D is the input sample dimension, that is, the number of auxiliary variables, and m is the output sample dimension;

[0011] Step 2.2, calculate the input weight of the jth node in the hidden layer and the bias of the jth node in the hidden layer , and then get the optimal output weights of all nodes in SCN ; Where, 1≤j≤L-1;

[0012] Step 2.3, water wall temperature prediction model prediction value for:

[0013] ;

[0014] in, represents the weight matrix of the input layer and hidden layer in SCN, represents the weight matrix of the input layer in SCN, represents the weight matrix of the L-1th hidden layer in SCN;

[0015] Step 3: training the water wall temperature prediction model to obtain a trained water wall temperature prediction model for predicting the water wall temperature.

[0016] Furthermore, in step 1, the auxiliary variables having a significant impact on the water wall temperature are selected. The specific steps are as follows:

[0017] Step 1.1, collect historical data of auxiliary variables;

[0018] Step 1.2, performing data preprocessing on the collected auxiliary variable historical data, wherein the preprocessing includes abnormal data processing and normalization processing;

[0019] In step 1.3, the Gain mode in the LightGBM algorithm is used to evaluate the characteristic importance of the auxiliary variables on the wall temperature, and the auxiliary variables with the top D characteristic importance are selected as the auxiliary variables that have a significant impact on the water-cooled wall temperature.

[0020] Furthermore, the characteristic importance of the auxiliary variables to the wall temperature is evaluated using the following formula:

[0021] (1)

[0022] (2)

[0023] in, It is the reduction in the loss of water-wall tube metal temperature caused by node splitting, measured by the mean square error; is the mean square error of the parent node, and is the mean squared error between the left and right child nodes.

[0024] Furthermore, the auxiliary variables selected in step 1 include main steam pressure, fuel quantity, position of stationary and moving blades of the induced draft fan, load, superheat, secondary air of the F damper of the B3 burner, secondary air of the F damper of the B1 burner, secondary air of the F damper of the A3 burner, secondary air of the F damper of the F1 burner, and flue gas temperature at the air preheater outlet; the total number of auxiliary variables is D=10.

[0025] Furthermore, the auxiliary variable in step 1 is a reconstructed auxiliary variable after taking into account the delay time between the auxiliary variable and the water-cooled wall temperature;

[0026] The first The variables input to the water wall temperature prediction model are defined as , is the number of samples; x i ( t ) = [ x i , 1 ( t ), x i , 2 ( t ),..., x i , d ( t ),..., x i , D ( t )] , the target variable is expressed as ;

[0027] Since there is a delay between the target variable and the input variable, Indicates the d-th input variable With the target variable The delay time between

[0028] The input variables are reconstructed as follows:

[0029] X i ( t ) = [ x ^ i , 1 ( t ), x ^ i , 2 ( t ),..., x ^ i , d ( t ),..., x ^ i , D ( t )] (3)

[0030] (4)

[0031] in, Represents the reconstructed value of the d-th input variable.

[0032] Delay time The calculation method is as follows:

[0033] The average mutual information (AMI) method combined with the differential evolution (DE) algorithm is used to calculate the delay time between each auxiliary variable and the water-cooled wall temperature.

[0034] Furthermore, the water wall temperature prediction model described in step 2 is constructed as follows:

[0035] Step 2.1, given the objective function , assuming that L-1 neural nodes have been constructed as the hidden layer of SCN, then the network output expression of the current hidden layer is:

[0036] (5)

[0037] in β j = [ β j , 1 ,..., β j , m ] T represents the output weight of hidden layer node j; represents the activation function; and are the input weight and bias of the jth node in the hidden layer, Represents the input of the current hidden layer;

[0038] Step 2.2, calculate the input weight of the jth node in the hidden layer through the hiking algorithm HOA and the bias of the jth node in the hidden layer , and then get the optimal output weight , the specific steps are as follows:

[0039] Step 2.2.1, initialize the traveler's position;

[0040] Initialize the traveler z's position , as shown in the following equation:

[0041] (6)

[0042] in is a uniformly distributed number in the range [0,1]; represents the current location of traveler z at time t; and Represents the upper and lower bounds of the j-th dimension of the decision variables of the optimization problem;

[0043] Step 2.2.2, compare q with , q represents the qth traveler, , is the population size, i.e. the number of travelers;

[0044] If q≤ When , calculate the fitness function , and go to step 2.2.3;

[0045] If q> , determine the optimal position of the traveler, that is, the minimum fitness function , the last element in the minimum fitness function is the bias , and the transpose of the other elements is the input weight , and when It>Maxlt, It is the current number of iterations, and Maxlt is the maximum number of iterations, so the optimal output weights of all nodes are determined by formula (11) , go to step 2.2.6;

[0046] If q> , and when It≤Maxlt, go to step 2.2.3;

[0047] (1) If q≤ Then calculate the fitness function for:

[0048] (7)

[0049] (8)

[0050] h L ( X ) = [ g L ( ω L T x 1 + b L ), g L ( ω L T x 2 + b L ),..., g L ( ω L T x N + b L )] T (9)

[0051] In the formula is the network residual of the Lth hidden layer node under the nth sample, is a function whose regularization parameter has a positive real-valued domain; is the output of the Lth hidden layer node, is the activation function of the Lth hidden layer node, , Represent the input weight and bias of the Lth hidden layer node respectively; It is an introduced value and a measurement standard, and its formula is as follows:

[0052] (10)

[0053] in is the regularization parameter;

[0054] (2) If q> , determine the optimal position of the traveler, that is, the minimum fitness function ;

[0055] It and Maxlt are judged. If It>Maxlt, the best position is stored in each iteration, the input weight and bias are obtained, and the optimal output weight of all nodes is determined. :

[0056] (11)

[0057] in, express The Moore-Penrose generalized inverse matrix of H L = [ h 1 , h 2 ,..., h L ] represents the weight matrix of the input layer and the hidden layer, represents the optimal output weight of all nodes, = [ β 1 * , β 2 * ,..., β L * ] T , represents the output weights of all nodes, is the Frobenius norm, O represents the target value of each input sample, K represents the identity matrix, represents the regularization strength.

[0058] Step 2.2.3, the number of travelers z and the maximum number of travelers Make comparisons;

[0059] If z> , then assign the minimum fitness function, i.e. the best position of the traveler, to the z+1th traveler, and return to step 2.2.2 to judge It again;

[0060] If z≤ , get the best position of traveler z at the Itth iteration , go to step 2.2.4;

[0061] Step 2.2.4, calculate , is the speed of traveler z at time t, as shown in the following formula:

[0062] (12)

[0063] is the slope of the path or terrain; and Expressed by formula (13):

[0064] (13)

[0065] in, and represent the elevation difference and the distance covered by the hiker, respectively; is the slope angle of the path or terrain; and θ z , t ∈ [ 0 , 50 ] ; and the speed of traveler z at time t+1 is:

[0066] (14)

[0067] in, is a uniformly distributed number in the range [0,1]; and denote the speed of the hiker at time t+1 and time t respectively; is the location of the lead hiker, is the scan factor of hiker z at time t; then, update the traveler's position And go to step 2.2.5, as shown in the following formula:

[0068] (15)

[0069] Step 2.2.5, set is the new fitness function, Is the current fitness function, if ≤ , then store the best position , and return to step 2.2.3; if > , then directly return to step 2.2.3;

[0070] Step 2.2.6, add new nodes until the or node ,in is the allowable error; is the maximum number of hidden neurons;

[0071] Step 2.3, the predicted value y of the water-cooled wall temperature prediction model is:

[0072] (16)

[0073] The present invention also proposes a boiler water-cooled wall temperature over-temperature warning method, which predicts the water-cooled wall temperature based on the water-cooled wall temperature prediction model constructed by the present invention to obtain the water-cooled wall temperature; compares the water-cooled wall temperature with the maximum allowable temperature of the water-cooled wall material; if the water-cooled wall temperature is higher than the maximum allowable temperature of the water-cooled wall material, an over-temperature warning signal is issued.

[0074] The present invention also proposes a boiler water wall temperature prediction system, comprising:

[0075] The data acquisition module is used to obtain auxiliary variable historical data from the distributed control system DCS.

[0076] The preprocessing module is used to perform abnormal data smoothing replacement and normalization processing on the auxiliary variable historical data to obtain preprocessed auxiliary variable historical data.

[0077] The delay time optimization module is used to perform delay processing on the pre-processed auxiliary variable historical data to obtain the delayed auxiliary variable historical data.

[0078] The prediction module uses a water-cooled wall temperature prediction model to predict the water-cooled wall temperature based on the auxiliary variable historical data after the delay.

[0079] The present invention also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the boiler water-cooled wall temperature prediction method of the present invention is implemented.

[0080] Beneficial effect: In the prior art, feature selection is mainly based on linear analysis methods, ignoring other highly nonlinear feature variables, which limits the prediction model's ability to express predicted temperature. The present invention adopts the LightGBM dimensionality reduction method to screen out feature variables with high correlation.

[0081] There is a time delay in the coal grinding system, and the control variables are not synchronized with the wall temperature changes. The present invention uses a differential evolution DE algorithm combined with the average mutual information method AIM to deal with the time delay problem and ensure high-quality feature input.

[0082] HOA-SCN can select the optimal hyperparameters for each newly added node in the hidden layer. At the same time, the introduction of L2 regularization in the output layer not only improves the model's predictive performance but also enhances its robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 It is a data processing flow chart of the HOA-SCN model of the present invention;

[0084] Figure 2 It is a graph of collected historical data;

[0085] Figure 3 It is a trend chart of historical data after abnormal data processing and normalization;

[0086] Figure 4 is the importance map of the first ten input variables obtained using LightGBM;

[0087] Figure 5 is the delay time between the input variable and the wall temperature Estimation graph;

[0088] Figure 6 It is a random configuration network SCN network structure diagram;

[0089] Figure 7 This is a comparison chart of the iterative loss values of HOA-SCN under different activation functions.

[0090] Figure 8is a graph of input variables extracted from auxiliary variables using MIC;

[0091] Figure 9 This is a graph of the evaluation results of the first 10 input variables using RF;

[0092] Figure 10 It is a prediction performance diagram of the HOA-SCN of the present invention under four variable processing methods;

[0093] Figure 11 This is a comparison chart of the prediction results of HOA-SCN with and without delay time;

[0094] Figure 12 This is a comparison chart of the prediction results of HOA-SCN, LSTM, and BP on the test set;

[0095] Figure 13 This is a comparison chart of the prediction errors of HOA-SCN, LSTM and BP;

[0096] Figure 14 This is a scatter plot comparison of the prediction results of HOA-SCN, GSCN, and SCN on the test set;

[0097] Figure 15 This is a comparison chart of the prediction errors of HOA-SCN, SCN, and GSCN. DETAILED DESCRIPTION

[0098] Traditional neural networks lack interpretability, and the process of iteratively searching for a set of appropriate weights and biases is time-consuming and computationally intensive, which is particularly difficult for processing large-scale industrial data. The present invention establishes a water-cooled wall temperature prediction model based on a method of constructing a neural network structure based on a random configuration network strategy, where the water-cooled wall temperature refers to the metal wall temperature of the boiler heating surface. The random configuration network (SCN) is an incremental learning method with universal approximation properties. It uses a supervision mechanism to randomly configure the parameters of the hidden layer nodes under a set of inequality constraints to adaptively select their value ranges. This method has the significant advantages of manual intervention in the network structure and adaptive selection of the number of hidden layer nodes. After the hidden layer parameters are determined by the random configuration algorithm, the output weights are calculated based on the pseudo-inverse theory. This mechanism makes the training speed of SCN faster than the traditional gradient-based iterative algorithm. At present, the random configuration network (SCN) has been used to establish models for various industrial scenarios. The present invention is the first time that the random configuration network (SCN) has been applied to modeling the metal wall temperature of the boiler heating surface.

[0099] To improve the efficiency of searching for weights and biases in traditional random configuration networks (SCNs), and enhance the model's generalization and predictive performance, this paper introduces the hiking HOA algorithm into random configuration networks to establish a boiler water wall temperature prediction model based on the HOA-SCN. The HOA-SCN can select optimal hyperparameters for each newly added node in the hidden layer. Furthermore, the introduction of L2 regularization in the output layer not only improves the predictive performance of the boiler water wall temperature prediction model, but also enhances its robustness.

[0100] In this specific implementation, a W-flame boiler with a rated load of 660 MW was selected as the target boiler. The boiler is equipped with six double-inlet and double-outlet pulverizers, each equipped with four double-cyclone pulverized coal burners. The 24 pulverized coal burners are arranged in a row on the front and rear wall arches of the lower furnace. Twenty-six overfire air (OFA) nozzles are also located on the upper portions of the front and rear water-cooled walls, with air volume controlled by four air registers. The secondary air for each burner is individually controlled, with the air distribution system consisting of upper and lower air boxes. The secondary air dampers for each burner are adjusted individually; see Table 1 for their specific functions.

[0101] Table 1 Functions and operating procedures of secondary air damper

[0102]

[0103] Boiler water-wall temperature is also affected by coal quality. Because the target boiler is not equipped with an online coal quality monitoring system, the coal quality data from the fuel section cannot be synchronized with the sampling period of the auxiliary variables. Therefore, the auxiliary variables do not consider coal characteristics and assume that the coal quality is stable during the simulation.

[0104] The present invention provides a method for predicting the water wall temperature of a pulverized coal boiler, which specifically includes the following steps:

[0105] Step 1: Select auxiliary variables that can affect the water wall temperature, pre-process the auxiliary variables, evaluate the characteristic importance of the auxiliary variables to the wall temperature, and select important auxiliary variables as input features of the boiler water wall temperature prediction model;

[0106] Step 1.1, select auxiliary variables that can affect the water wall temperature;

[0107] Selecting multiple media parameters, such as air, smoke, water, and steam, as auxiliary variables for water-wall temperature prediction can fully reflect the complex heat exchange processes within the boiler and the impact of various factors on wall temperature. These parameters provide information from different perspectives, helping to capture the dynamic response of the boiler's thermal system. This improves the accuracy of water-wall temperature prediction, optimizes boiler operation and control, and enhances energy efficiency and boiler safety.

[0108] Adjustments to induced and forced air affect furnace temperature distribution, especially during the combustion process. When air volume is too low or airflow is unevenly distributed, heat is unevenly distributed within the boiler, leading to localized overheating or underheating of the wall temperature. Therefore, changes in air distribution parameters such as air volume and pressure have a significant impact on water-wall temperature. Flue gas temperature and velocity directly influence heat distribution, which in turn affects water-wall temperature. Secondary air and pulverizer volumetric airflow are crucial control parameters in the boiler, directly impacting boiler combustion, heat distribution, furnace temperature, and the stability and uniformity of water-wall temperature. Using secondary air and volumetric airflow as auxiliary variables for water-wall temperature can better reflect wall temperature fluctuations and provide valuable information for boiler operation optimization. Based on the aforementioned in-furnace influences on water-wall temperature, combustion control principles, and the recommendations of the operation engineer, a list of auxiliary variables for dynamic wall temperature prediction based on the HOA-SCN model was established. Table 2 shows a total of 60 auxiliary variables. These auxiliary variables were selected to capture the fundamental characteristics of the combustion process and consider the impact of key operating factors on wall temperature.

[0109] Table 2 Auxiliary variable list

[0110]

[0111] Step 1.2, collect historical data of auxiliary variables;

[0112] This specific implementation method collects 7 days of historical data from the Distributed Control System (DCS) database of the target boiler at a time granularity of 1 minute. The data collected over these 7 days covers most situations, such as load increase intervals, load decrease intervals, load change intervals, and steady-state intervals, to fully reflect the changes in water-cooled wall temperature, and the shutdown intervals are excluded. The original historical data is as follows: Figure 2 As shown in Figure 1, there are approximately 10,080 sampling points, the first 80% of which are used as the training set for the boiler water-cooled wall temperature prediction model, and the last 20% are used as the test set.

[0113] Step 1.3, preprocessing the collected auxiliary variable historical data;

[0114] Step 1.3.1, abnormal data processing;

[0115] Due to the complex on-site operating environment of the boiler, the signal is interfered by noise, resulting in abnormal values in the original data stored in the DCS. This specific implementation adopts a moving smoothing replacement method, as shown in formula (1):

[0116] (1)

[0117] Where N is the window size, which is set to 10 in this specific implementation. The original historical data data points. It is the smoothed data value, representing the smoothed value at time t; the smoothed value at time t replaces the original historical data at time t .

[0118] Step 1.3.2, normalization processing;

[0119] In order to improve the training efficiency of wall temperature dynamic prediction and avoid error fluctuations in the output layer, the minimum-maximum scaling method is used to normalize the auxiliary variables after abnormal data processing:

[0120] (2)

[0121] in , Represent the normalized value and the value after abnormal data processing. data points, and , They represent the maximum and minimum values of the data after abnormal data processing. Figure 3 The following is a partial trend chart showing the normalized original data.

[0122] Step 1.4, reduce the dimension of auxiliary variables;

[0123] Light Gradient Boosting Machine (LightGBM) is a gradient-boosted decision tree (GBDT) algorithm proposed and released by Microsoft in 2017. It is widely used in classification and regression tasks, as well as in data mining. This implementation uses LightGBM to evaluate the feature importance of auxiliary variables.

[0124] LightGBM has made several major optimizations on the traditional GDNT algorithm: Histogram-based decision tree algorithm, one-sided gradient sampling (GOSS), mutually exclusive feature bundling (EFB), and Leaf-wise leaf growth strategy with depth restriction. LightGBM has the ability to train in parallel and can quickly process large amounts of industrial data. Therefore, it is suitable for processing nonlinear and large-data-volume industrial objects such as boilers. In addition, when calculating the importance of features, the LightGBM algorithm has two models, one is Gain and the other is Split. Split measures the number of times each feature is used in all trees, focusing on the frequency of feature occurrence. Gain measures the contribution of each feature to the model performance, which is usually obtained by calculating the information gain of the feature at the splitting point in the tree model. In LightGBM, the greater the gain, the greater the contribution of the feature to the model. This specific implementation method chooses to use the Gain mode to evaluate the feature importance of auxiliary variables to wall temperature, as shown in the following formula:

[0125] (3)

[0126] (4)

[0127] in, It is the reduction in the loss (error) of feature f (i.e., the metal temperature of the water-cooled wall tube) when the node is split, measured by the mean square error (MSE). is the mean square error of the parent node, and is the mean squared error between the left and right child nodes.

[0128] The LightGBM algorithm is used to evaluate the feature importance of the 60 auxiliary variables in Table 2. At the same time, in order to find the optimal hyperparameters, a grid search is used for the LightGBM algorithm. The results show that the best effect is achieved when the number of decision trees is 500, the tree iteration depth is 7, and the learning rate is 0.05. Figure 4 The importance of the top 10 characteristic variables in the Gain mode. The total weight ratio of the top 10 characteristic variables (auxiliary variables) reaches 0.9696.

[0129] Based on the importance assessment results, the necessity of variables in waterwall temperature modeling was analyzed. Main steam pressure is directly related to steam flow, heat exchange efficiency, and unit load, with an importance of 0.6190. Excessively high or low main steam pressure can lead to reduced thermal efficiency or boiler damage, and therefore has a significant impact on waterwall temperature. Furthermore, the fuel level directly determines the boiler's combustion intensity, affecting heat output and flue gas temperature. Increasing the fuel level increases the heat generated by combustion, raising furnace temperature and, in turn, waterwall temperature. Regarding load, an increase in load typically requires more heat output, increasing furnace temperature and waterwall temperature. The position of the induced draft fan's stators and rotors controls air flow and distribution, significantly impacting the combustion process. Excessive or insufficient air supply can affect combustion efficiency and boiler temperature distribution. Adequate air flow ensures sufficient combustion, preventing localized overheating and uneven cooling. Superheat affects boiler steam temperature control, which in turn influences the heat transfer characteristics of the waterwall. The burner's secondary air volume directly affects combustion efficiency and temperature distribution. Insufficient secondary air volume may lead to incomplete combustion, thus affecting the wall temperature.

[0130] Analysis of these characteristic variables reveals that main steam pressure, total fuel flow, and load are the most direct factors affecting boiler water wall temperature, as they determine the boiler's heat output and heat load. The burner's secondary air (such as the induced draft fan's stator and rotor blades, and the burner's F-damping secondary air) optimizes the combustion process by regulating air flow, thereby controlling the boiler's local temperature distribution. Superheat and air preheater outlet flue gas temperature reflect the boiler's heat exchange efficiency and indirectly affect water wall temperature. Based on the above analysis, the importance of the variables calculated by LightGBM follows the variation pattern of boiler water wall temperature. By balancing model complexity with generalizability, this specific implementation selects the top 10 most important auxiliary variables as input variables for the water wall temperature prediction model: main steam pressure, fuel flow, induced draft fan's stator and rotor blade positions, load, superheat, B3 burner F-damping secondary air, B1 burner F-damping secondary air, A3 burner F-damping secondary air, F1 burner F-damping secondary air, and air preheater outlet flue gas temperature.

[0131] Step 2, estimate the variable delay time;

[0132] During coal combustion in a furnace, chemical energy is converted into thermal energy from water. This results in varying degrees of delay between each auxiliary variable and the wall temperature. Therefore, calculating the delay between the input variables and the target variable helps establish more accurate dynamic predictions of water-cooled wall temperature. However, estimating time delays based solely on mechanistic analysis is uncertain. Therefore, this paper uses the average mutual information (AMI) method to calculate the delay between the input variables and the wall temperature. Due to the interactions and coupling between the input variables, calculating the MI value individually can result in significant errors, necessitating the consideration of all input variables as a whole.

[0133] Through LightGBM dimensionality reduction, 10 input variables are obtained, and the variables input to the water-cooled wall temperature prediction model are defined as: , represents the dth input variable of the i-th input sample, i= is the number of samples. x i , d ( t ) = [ x i , 1 ( t ), x i , 2 ( t ),..., x i , d ( t ),..., x i , 10 ( t )] , and the target variable is expressed as For this embodiment, the target variable is the water wall temperature.

[0134] Since there is a delay between the target variable and the input variable, time The previous value of the input variable is the same as Related, that is and .in, is the input variable With the target variable The delay time between

[0135] The input variable matrix is reconstructed as shown below:

[0136] X i ( t ) = [ x ^ i , 1 ( t ), x ^ i , 2 ( t ),..., x ^ i , d ( t ),..., x ^ i , 10 ( t )] (5)

[0137] (6)

[0138] in, Represents the reconstructed value of the d-th input variable.

[0139] To estimate the optimal delay time, this embodiment uses a differential evolution (DE) algorithm to optimize the delay time so that it satisfies the maximum value of the average mutual information (AMI). This method can perform a global search under the constraints to find the optimal delay time for each auxiliary variable. The formula for the average mutual information (AMI) is:

[0140] (7)

[0141] in is the mutual information MI in and The value between , N represents the number of samples.

[0142] Figure 5 The delay between the 10 input variables and the target variable is shown in Figure 1. Changes in main steam pressure typically affect the boiler's heat exchange efficiency relatively quickly. When main steam pressure increases, the boiler's heat output increases, which in turn increases the water-wall temperature. However, due to the physical delays in steam flow and heat transfer, the water-wall temperature typically lags slightly behind changes in main steam pressure. After a 2-minute delay, fluctuations in main steam pressure are transmitted to the boiler wall, affecting the water-wall temperature for a period of time. Fuel level and load directly determine the boiler's combustion intensity and heat output. Increasing fuel level or load immediately increases the heat within the furnace, thereby raising the water-wall temperature. However, there is a significant time delay between coal delivery to the pulverizer and heat supply. The transfer of combustion heat from the furnace to the water-wall requires time, resulting in significant lag in the propagation of temperature changes within the boiler furnace and between various system components. The 9-minute delay is related to factors such as coal delivery, combustion reaction time, boiler heat capacity, and steam flow regulation. Changes in secondary air affect the rate and efficiency of the combustion reaction, requiring some time to be fully reflected in the boiler's heat load and water-wall temperature. Heat release and temperature distribution during combustion gradually stabilize, requiring several minutes to be transferred to the water-walls. Furthermore, temperature changes within the boiler typically exhibit a certain degree of inertia. This is particularly true in large boilers, where heat transfer from the furnace to the water-walls and other components takes time. Therefore, the effects of secondary air adjustments on water-wall temperature are not immediately apparent. Therefore, a delay of 1 to 4 minutes for secondary air is entirely reasonable.

[0143] Step 3: construct the HOA-SCN water wall temperature prediction model to perform temperature prediction, and input historical data into the water wall temperature prediction model to train the model;

[0144] The Hiking Optimization Algorithm (HOA) was inspired by hiking, identifying similarities between the search landscape of an optimization problem and the mountainous terrain that hikers traverse. By simulating the idiosyncrasies involved in a hiking expedition, the ultimate goal is to reach the summit of a peak.

[0145] Step 3.1, construct auxiliary variables and input them into the water-cooled wall temperature model of the present invention, such as Figure 1 As shown, the following steps are included:

[0146] Step 3.1.1, given a set of training data sets ,in is the input sample of HOA-SCN, where represents the i-th sample input to SCN, In this specific implementation, D=10, indicating ten auxiliary variables; the output sample is , Where N is the number of samples, D is the input sample dimension, and m is the output sample dimension. For this specific embodiment, the output sample of this application is the predicted water-cooled wall temperature, which is one-dimensional data, so m=1.

[0147] Step 3.1.2, given the objective function , assuming that L-1 neural nodes have been constructed as the hidden layer of SCN, then the network output expression of the current hidden layer is:

[0148] (8)

[0149] in β j = [ β j , 1 ,..., β j , m ] T represents the output weight of hidden layer node j; represents the activation function; and are the input weight and bias of the j-th node in the hidden layer, respectively.

[0150] Step 3.1.3, as Figure 6 As shown, the input weight of the jth node in the hidden layer is calculated by the hiking algorithm HOA and the bias of the jth node in the hidden layer Then determine the optimal output weight , the specific steps are as follows:

[0151] Step 3.1.3.1, initialize the position of traveler z , as shown in the following equation:

[0152] (9)

[0153] in is a uniformly distributed number in the range [0,1]; represents the current location of traveler z at time t; and Represents the upper and lower bounds of the j-th dimension of the decision variables of the optimization problem.

[0154] Step 3.1.3.2, if q≤ When , calculate the fitness function , and proceed to step 3.1.3.3;

[0155] q represents the qth traveler, , is the population size, i.e. the number of travelers, which is set to 60 in this paper;

[0156] If q> , determine the optimal position of the traveler, that is, the minimum fitness function , the last element in the minimum fitness function is the bias , and the transpose of other elements is the input weight , and when It>Maxlt, It is the current number of iterations, Maxlt is the maximum number of iterations, and then the optimal output weights of all nodes are determined by formula (14) , go to step 3.1.3.6;

[0157] If q> , and It≤Maxlt, It is the current number of iterations, Maxlt is the maximum number of iterations, and go to step 3.1.3.3;

[0158] (1) If q≤ Then calculate the fitness function for:

[0159] (10)

[0160] (11)

[0161] h L ( X ) = [ g L ( ω L T x 1 + b L ), g L ( ω L T x 2 + b L ),..., g L ( ω L T x N + b L )] T (12)

[0162] In the formula is the network residual of the Lth hidden layer node under the nth sample, is a function whose regularization parameter has a positive real-valued domain. is the output of the Lth hidden layer node, is the activation function of the Lth hidden layer node, , denote the input weight and bias of the Lth hidden layer nodes respectively. It is a specially introduced value and a measurement standard used to determine whether each candidate node meets specific conditions. Its formula is as follows:

[0163] (13)

[0164] in is the regularization parameter.

[0165] (2) If q> , determine the optimal position of the traveler, that is, the minimum fitness function .

[0166] It and Maxlt are judged. If It>Maxlt, It is the current number of iterations and Maxlt is the maximum number of iterations. The best position is stored in each iteration and the output weights of all nodes are determined. :

[0167] (14)

[0168] in, express The Moore-Penrose generalized inverse matrix of H L = [ h 1 , h 2 ,..., h L ] represents the weight matrix of the input layer and the hidden layer, represents the optimal output weight of all nodes, = [ β 1 * , β 2 * ,..., β L * ] T , represents the output weights of all nodes, is the Frobenius norm, O represents the target value of each input sample, K represents the identity matrix, represents the regularization strength.

[0169] Step 3.1.3.3, determine z and . z is the number of travelers, Maximum number of travelers.

[0170] If z> , then the minimum fitness function, that is, the best position of the traveler, is assigned to the z+1th traveler, and return to step 3.1.3.2 to judge It again.

[0171] If z≤ , get the best position of traveler z at the Itth iteration , proceed to step 3.1.3.4

[0172] Step 3.1.3.4, calculate , is the speed of traveler z at time t, as shown in the following formula:

[0173] (15)

[0174] is the slope of the path or terrain. Expressed by formula (16):

[0175] (16)

[0176] in, and represent the elevation difference and the distance covered by the hiker, respectively; is the slope angle of the path or terrain. And θ z , t ∈ [ 0 , 50 ] . The speed of traveler z at time t+1 is:

[0177] (17)

[0178] in, is a uniformly distributed number in the range [0,1]; and denote the speed of the hiker at time t+1 and time t respectively; is the location of the lead hiker, is the scan factor (SF) of hiker z at time t. Then, update the traveler's position And proceed to step 3.1.3.5, as shown in the following formula:

[0179] (18)

[0180] Step 3.1.3.5, set is the new fitness function, Is the current fitness function, if ≤ , then store the best position , and return to step 3.1.3.3. If > , then directly return to step 3.1.3.3.

[0181] Step 3.1.3.6, add new nodes until the or node ,in is the allowable error; is the maximum number of hidden neurons.

[0182] Finally, the predicted value y of the water-cooled wall temperature prediction model is:

[0183] (19)

[0184] Step 3.2, on the ill-posed problem of random configuration network SCN

[0185] During the construction process of the random configuration network SCN, the input weights and biases of the neural network are randomly assigned under the inequality supervision mechanism. The output weights of the random configuration network SCN are calculated using the least squares method. The output weights of the random configuration network SCN are:

[0186] (20)

[0187] There is a defect in the random configuration network (SCN). If the dimension of the input data is high and the data volume is large, the number of hidden layer nodes in the random configuration network (SCN) model will increase, which will easily lead to collinearity between the output matrix variables of the hidden layer nodes, resulting in pathological problems. This will cause the generalization ability of the random configuration network (SCN) model to drop sharply, resulting in overfitting. Therefore, this specific embodiment uses the L2 regularization method to improve the output weights of the random configuration network (SCN) to alleviate the pathological problems of the random configuration network (SCN). The optimal output weights of all nodes in the random configuration network (SCN) are: to:

[0188] (twenty one)

[0189] Step 3.3 Parameter optimization of water wall temperature prediction model

[0190] After multiple simulation experiments, the maximum number of iterations of hiking HOA in HOA-SCN is set to 20 times, and the population size is set to 60 times. Set to 0.9999999, and the regularization coefficient is set to 0.05. The activation function determines the activation state of the neuron and allows the model to map complex nonlinear relationships. In this specific implementation, three activation functions, Tanh, Sigmoid, and ReLU, are used to compare the model loss function values on the validation set and the training set. Figure 7 As shown, Figure 7 (a) is the model loss function value curve on the validation set and training set obtained when the Sigmoid activation function is used. Figure 7 (b) is the model loss function value curve on the validation set and training set obtained when the ReLU activation function is used. Figure 7 (c) in the figure is a curve of the model loss function value on the validation set and the training set obtained when the Tanh activation function is used. The loss curve shows that the Sigmoid activation function is more suitable for HOA-SCN.

[0191] Step 4: Add auxiliary parameters Input into the HOA-SCN water-cooled wall temperature prediction model for temperature prediction.

[0192] Considering the time delay associated with changes in furnace combustion conditions affecting wall temperature, this invention calculates the temperature rise δt for the next two minutes based on the average rate of temperature rise (°C / min) at the water-wall metal temperature measurement point over the past minute. This δt is then added to the water-wall metal temperature obtained from the prediction model to determine the overtemperature warning temperature. The resulting overtemperature warning temperature is then compared with the maximum allowable temperature of the water-wall material. When the overtemperature warning temperature exceeds or equals the maximum allowable temperature threshold of the material, an overtemperature warning signal is issued to alert boiler operators to take timely preventative measures (such as reducing fuel input or increasing the oxygen setpoint). This involves adjusting the air-to-coal ratio from the combustion side, thereby adjusting combustion conditions and reducing the frequency of overtemperatures in the boiler water-wall metal wall. The secondary air dampers are adjusted by adjusting the opening of the F-layer secondary air dampers, located above the furnace, and dynamically increasing the feedwater flow setpoint offset (for a short period before recovering). This influences the fuel-to-water ratio, thereby reducing water-wall temperature. This patented invention is beneficial for promoting safe operation of power plant boilers and extending equipment life.

[0193] Accordingly, the present invention further provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the boiler water wall temperature prediction method of the present invention.

[0194] To evaluate the performance of the HOA-SCN water wall temperature prediction model, different modeling methods were used for comparison. The mean and standard deviation of the model prediction errors were obtained from 20 independent simulation experiments.

[0195] (1) Comparison of dimensionality reduction methods

[0196] Based on HOA-SCN, we compared the impact of LightGBM feature dimensionality reduction on prediction performance with and without LightGBM, as well as other dimensionality reduction methods. Without dimensionality reduction, the input variables were 59, while the other dimensionality reduction methods used were the mainstream MIC and RF dimensionality reduction methods. For ease of comparison, we also selected 10 features as input variables. Figure 8 and Figure 9 They are the top ten variables under MIC and RF feature dimensionality reduction.

[0197] The mean predictive performance of HOA-SCN under the four variable processing methods is as follows: Figure 10 Table 3 shows the mean and standard deviation of the performance of the 20 trials. The results show that the LightGBM dimensionality reduction method outperforms both MIC and RF, demonstrating that it is superior to the other two methods in handling nonlinear relationships and feature interactions. Furthermore, the training time with dimensionality reduction is significantly faster than without dimensionality reduction.

[0198] Table 3 Evaluation indicators of HOA-SCN under four variable processing methods

[0199]

[0200] (2) Delay time estimation

[0201] In order to verify the impact of the delay time of input variables on the prediction performance of the model, a comparative experiment with and without delay time was conducted based on HOA-SCN. Figure 11 is the prediction curve of the variable with or without delay time, and Table 4 shows the corresponding evaluation indicators. Figure 11 (a) is the comparison curve between the predicted value and the true value of the variable with delayed time. Figure 11 (c) is the comparison curve between the predicted value and the true value of the variable without time delay. The prediction effect considering time delay is significantly better than that without considering time delay. Figure 11 (b) is the absolute error between the predicted value and the true value when the variable has a delay time. Figure 11 (d) in the figure is the absolute error between the predicted value and the true value when there is no time delay. When time delay is considered, the absolute error between the predicted value and the true value of half of the sampling points is less than 1.10439°C, while this data is 1.43922°C when time delay is not considered. Therefore, considering the delay time of the input variable can effectively improve the prediction performance of the model.

[0202] Table 4 Prediction performance of HOA-SCN with and without input variable delay

[0203]

[0204] (3) Different modeling methods

[0205] To evaluate the generalization performance of HOA-SCN, we compared the traditional BP neural network and LSTM neural network models with HOA-SCN. The hyperparameters of the BP network and LSTM neural network are shown in Table 5.

[0206] Table 5 Hyperparameter settings of BP and LSTM neural networks

[0207]

[0208] Figure 12 (a) is the water-cooled wall temperature prediction curve of the three models HOA-SCN, LSTM, and BP. Figure 12 (b) in the figure is the corresponding unit load curve. Figure 12 (c) is the water-cooled wall temperature prediction curve in the load reduction range. Figure 12 (d) is the water-cooled wall temperature prediction curve in the load increase range. Figure 12 (e) is the water-cooled wall temperature prediction curve under the stable load state. Figure 12It can be found in (c), (d) and (e) that when the unit load decreases, the wall temperature also decreases. When the load increases, the temperature of the boiler water-cooled wall also increases. When the load is stable, the wall temperature tends to fluctuate and change steadily. From the boiler operation mechanism, the decrease in unit load means that the combustion intensity in the furnace is weakened and the heat generated by combustion is reduced. Moreover, the load reduction is usually accompanied by the adjustment of the induced draft fan and the forced draft fan, and the air flow and flue gas flow of the boiler also change. These factors cause the wall temperature to drop. The reasons for the increase and decrease in load are similar. When the load is stable, the heat load and combustion process in the boiler tend to be stable, the fluctuation of the wall temperature will decrease, and it will tend to be stable for a period of time. The thermal inertia and regulation mechanism of the boiler system keep the temperature of the water-cooled wall within a relatively stable range, but small fluctuations may still occur. From Figure 12 The error statistics of the three models in Figures (f), (g), and (h) clearly show that the HOA-SCN model has the highest and most concentrated error values near zero, followed by the LSTM model and finally the BP model. The traditional BP model has a simple structure and limited ability to extract variable features, resulting in poor prediction performance. The LSTM model's inherent memory module causes prediction errors to gradually accumulate, thus affecting prediction performance.

[0209] Figure 13 is the prediction error of the three models, such as Figure 13 As shown in (b), the RMSE of HOA-SCN is 2.06043±0.02241℃, MAE is 1.62609±0.03779℃, MAPE is 0.42372±0.01128%, R 2 is 0.97553±0.00053. Figure 13 As shown in (a), the RMSE of BP is 4.87107±1.26054℃, MAE is 3.90565±1.07193℃, MAPE is 0.97775±0.35495%, and R2 is 0.85407±0.07467. Figure 13 As shown in (c), the RMSE of LSTM is 2.68964±0.47341℃, MAE is 2.18015±0.46092℃, MAPE is 0.56336±0.11967%, R 2 The accuracy of HOA-SCN is 0.95701±0.0159. Compared with BP, the prediction indicators of HOA-SCN are improved by 57.70%, 58.36%, 56.67%, and 12.45%, respectively. Compared with LSTM, the prediction indicators of HOA-SCN are improved by 23.39%, 25.41%, 24.79%, and 1.90%, respectively.

[0210] (4) Comparison under the SCN architecture

[0211] To verify the impact of the hiking HOA algorithm and L2 regularization on the model prediction performance, this specific implementation method conducted a comparative experiment on the baseline model SCN, the optimized model GSCN and the HOA-SCN. To ensure the accuracy of the comparison results, the hyperparameters of the three models were kept consistent, among which the shrinkage factor in HOA-SCN and GSCN was Both are 0.999999, and the Set to 10 -6 (Note that GSCN here uses GSCN-II, not GSCN-I). As for SCN, From the collection Select from .

[0212] Figure 14 Scatter plot of the prediction results of HOA-SCN, GSCN, and SCN on the test set. Figure 14 As can be seen from (c), the experimental data are close to both sides of the Perfect line, and the 95% confidence interval is very narrow, indicating that the predicted value of HOA-SCN is closer to the true value. Figure 14 In Figures (a) and (b), we can clearly see that the confidence bands for SCN and GSCN are relatively wide, and the predicted points of SCN deviate significantly from the true values, indicating that SCN's prediction performance is poor. The slopes of the fitted curves also show that the slopes for HOA-SCN, GSCN, and SCN are 1.03, 0.93, and 0.80, respectively, indicating that HOA-SCN has the strongest generalization ability and SCN has the weakest.

[0213] Figure 15 The prediction results of HOA-SCN are better than those of SCN and GSCN. Compared with SCN, the prediction results of HOA-SCN (RMSE, MAE, MAPE, R 2 ) were improved by 54.05%, 49.52%, 56.66%, and 10.23% respectively. It can be seen that the hiking HOA algorithm effectively improves the prediction performance of the SCN model.

[0214] To accurately predict boiler water-wall temperature under variable operating conditions, this specific implementation proposes a new method based on the LightGBM dimensionality reduction method, delay time estimation, and the HOA-SCN model. This method has a simple yet effective network structure and significantly improves prediction performance. To evaluate the model's generalization capabilities, simulation experiments were conducted on an in-service 660MW supercritical, opposed-fired boiler. The following conclusions are drawn:

[0215] (1) Through comparative experiments with and without LightGBM dimensionality reduction, it was found that the LightGBM dimensionality reduction method not only improves training speed but also improves model accuracy. Comparative experiments with different dimensionality reduction methods such as MIC and RF show that the LightGBM dimensionality reduction method is more suitable for data of nonlinear, high-dimensional industrial objects such as boilers.

[0216] (2) Comparative experiments with and without delay time show that delay time estimation of input variables can improve the prediction accuracy of the model.

[0217] (3) Although the traditional BP neural network modeling method has high performance when processing steady-state loads, under variable loads, the wall temperature changes with the load. The traditional modeling method has limited ability to capture variable characteristics, resulting in low prediction accuracy under transient loads. The LSTM memory module will cause the accumulation of prediction errors, thereby reducing the prediction performance. HOA-SCN shows good prediction performance in both steady and transient conditions.

[0218] (4) The HOA optimization algorithm is introduced into SCN, which significantly improves the prediction accuracy without significantly increasing the training complexity. Compared with SCN and GSCN, the RMSE of HOA-SCN is improved by 54.05% and 43.76% respectively. 2 They increased by 10.23% and 5.39% respectively.

Claims

1. A method for predicting boiler water wall temperature, characterized in that: The steps include: Step 1: Select auxiliary variables that have an important impact on the water wall temperature, and the auxiliary variables are D in total; Step 2: Build a water wall temperature prediction model. The steps are as follows: Step 2.1, given the objective function , assuming that the L-1 neural nodes of the hidden layer of the random configuration network SCN have been constructed; among them, represents the dimensional space of the input, Represents the output dimensional space, D is the input sample dimension, that is, the number of auxiliary variables, and m is the output sample dimension; Then the network output expression of the current hidden layer is: （5） in represents the output weight of hidden layer node j; represents the activation function; and are the input weight and bias of the jth node in the hidden layer, Represents the input of the current hidden layer; Step 2.2, calculate the input weight of the jth node in the hidden layer through the hiking algorithm HOA and the bias of the jth node in the hidden layer , and then get the optimal output weights of all nodes in SCN , where 1≤j≤L-1; the specific steps are as follows: Step 2.2.1, Initialize the traveler's position Initialize the traveler z's position , as shown in the following equation: （6） in is a uniformly distributed number in the range [0,1]; represents the current location of traveler z at time t; and Represents the upper and lower bounds of the j-th dimension of the decision variables of the optimization problem; Step 2.2.2, compare q with , q represents the qth traveler, , is the population size, i.e. the number of travelers; If q≤ When , calculate the fitness function , and go to step 2.2.3; If q> , determine the optimal position of the traveler, that is, the minimum fitness function , the last element in the minimum fitness function is the bias , and the transpose of the other elements is the input weight , and when It>Maxlt, It is the current number of iterations, and Maxlt is the maximum number of iterations, so the optimal output weights of all nodes are determined by formula (11) , go to step 2.2.6; If q> , and when It≤Maxlt, go to step 2.2.3; (1) If q≤ Then calculate the fitness function for: （7） （8） （9） In the formula is the network residual of the Lth hidden layer node under the nth sample, is a function whose regularization parameter has a positive real-valued domain; is the output of the Lth hidden layer node, is the activation function of the Lth hidden layer node, , Represent the input weight and bias of the Lth hidden layer node respectively; It is an introduced value and a measurement standard, and its formula is as follows: （10） in is the regularization parameter; (2) If q> , determine the optimal position of the traveler, that is, the minimum fitness function ; The number of iterations It and the maximum number of iterations Maxlt are judged. If It>Maxlt, the best position is stored in each iteration, the input weight and bias are obtained, and the optimal output weight of all nodes is determined. : （11） in, express The Moore-Penrose generalized inverse matrix of represents the weight matrix of the input layer and the hidden layer, represents the optimal output weight of all nodes, = , represents the output weights of all nodes, is the Frobenius norm, O represents the target value of each input sample, K represents the identity matrix, represents the regularization strength; Step 2.2.3, the number of travelers z and the maximum number of travelers Make comparisons; If z> , then assign the minimum fitness function, i.e. the best position of the traveler, to the z+1th traveler, and return to step 2.2.2 to determine the number of iterations It again; If z≤ , get the best position of traveler z at the Itth iteration , proceed to step 2.2.4 Step 2.2.4, calculate , is the velocity of traveler z at time t, as shown in the following formula: （12） is the slope of the path or terrain; and Expressed by formula (13): （13） in, and represent the elevation difference and the distance covered by the hiker, respectively; is the slope angle of the path or terrain; and ; and the speed of traveler z at time t+1 is: （14） in, is a uniformly distributed number in the range [0,1]; and denote the speed of the hiker at time t+1 and time t respectively; is the location of the lead hiker, is the scan factor of hiker z at time t; then, update the traveler's position And go to step 2.2.5, as shown in the following formula: （15） Step 2.2.5, set is the new fitness function, Is the current fitness function, if ≤ , then store the best position , and return to step 2.2.3; if > , then directly return to step 2.2.3; Step 2.2.6, add new nodes until the or node ,in is the allowable error; is the maximum number of hidden neurons; Step 2.3, water wall temperature prediction model prediction value for: （16）； in, represents the weight matrix of the input layer and hidden layer in SCN, represents the weight matrix of the input layer in SCN, represents the weight matrix of the L-1th hidden layer in SCN; Step 3: training the water wall temperature prediction model to obtain a trained water wall temperature prediction model for predicting the water wall temperature.

2. The method for predicting boiler water wall temperature according to claim 1, characterized in that: In step 1, select auxiliary variables that have a significant impact on the water-cooled wall temperature. The specific steps are as follows: Step 1.1, collect historical data of auxiliary variables; Step 1.2, performing data preprocessing on the collected auxiliary variable historical data, wherein the preprocessing includes abnormal data processing and normalization processing; In step 1.3, the Gain mode in the LightGBM algorithm is used to evaluate the characteristic importance of the auxiliary variables on the wall temperature, and the auxiliary variables with the top D characteristic importance are selected as the auxiliary variables that have a significant impact on the water-cooled wall temperature.

3. The method for predicting boiler water wall temperature according to claim 2, characterized in that: The characteristic importance of the auxiliary variables to the wall temperature is evaluated using the following formula: （1） （2） in, It is the reduction in the loss of water-wall tube metal temperature caused by node splitting, measured by the mean square error; is the mean square error of the parent node, and is the mean squared error between the left and right child nodes.

4. The method for predicting boiler water wall temperature according to claim 1, characterized in that: The auxiliary variables selected in step 1 include main steam pressure, fuel quantity, position of stationary and moving blades of the induced draft fan, load, superheat, secondary air of the F damper of the B3 burner, secondary air of the F damper of the B1 burner, secondary air of the F damper of the A3 burner, secondary air of the F damper of the F1 burner, and the flue gas temperature at the air preheater outlet; the total number of auxiliary variables is D = 10.

5. The method for predicting boiler water wall temperature according to claim 1, characterized in that: The auxiliary variable in step 1 is the auxiliary variable reconstructed after taking into account the delay time between the auxiliary variable and the water-cooled wall temperature; The i-th variable input to the water-wall temperature prediction model is defined as ,i= is the number of samples; , the target variable is expressed as ; Since there is a delay between the target variable and the input variable, Indicates the dth input variable of the input With the target variable The delay time between The input variables are reconstructed as follows: （3） （4） in, Represents the reconstructed value of the d-th input variable.

6. The method for predicting boiler water wall temperature according to claim 5, characterized in that: Delay time The calculation method is as follows: The average mutual information (AMI) method combined with the differential evolution (DE) algorithm is used to calculate the delay time between each auxiliary variable and the water-cooled wall temperature.

7. A boiler water wall overtemperature warning method, characterized in that: Based on the water-cooled wall temperature prediction model constructed in the prediction method according to claim 1, the water-cooled wall temperature is predicted to obtain the water-cooled wall temperature; The water wall temperature is compared with the maximum allowable temperature of the water wall material. If the water wall temperature is higher than the maximum allowable temperature of the water wall material, an over-temperature warning signal is issued.

8. Boiler water wall temperature prediction system, characterized by: Based on the water-wall temperature prediction model constructed in the water-wall temperature prediction method according to claim 1, the water-wall temperature prediction system includes: Data acquisition module, used to obtain auxiliary variable historical data from the distributed control system DCS; a preprocessing module, configured to perform abnormal data smoothing replacement and normalization processing on the auxiliary variable historical data to obtain preprocessed auxiliary variable historical data; The delay time optimization module is used to perform delay processing on the pre-processed auxiliary variable historical data to obtain the delayed auxiliary variable historical data; The prediction module uses the water-cooled wall temperature prediction model to predict the water-cooled wall temperature based on the delayed auxiliary variable historical data.

9. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the boiler water wall temperature prediction method according to any one of claims 1 to 6.

Images