Boiler thermal efficiency soft measurement method and system based on machine learning and storage medium

Abstract

The invention discloses a boiler thermal efficiency soft measurement method and system based on machine learning, and a storage medium. The method comprises the following steps: collecting historical operation data of a boiler; preprocessing the data; the method comprises the following steps: constructing a BiGRU-KAN-Attention model; training the model by using the training set, and determining an optimal network structure and parameter setting; inputting test set data into the optimized BiGRU-KAN-Attention model, obtaining a prediction result, performing reverse normalization processing, and then analyzing and evaluating the prediction result by using an evaluation function; and inputting boiler operation data acquired in real time into the trained BiGRU-KAN-Attention model, and outputting a dynamic thermal efficiency prediction value. The method can effectively overcome the defects of a single model, achieves the high-precision dynamic measurement of the thermal efficiency of the boiler, and remarkably improves the measurement precision and real-time performance of the thermal efficiency of the boiler.

Description

Technical Field

[0001] This invention belongs to the field of boiler thermal efficiency measurement technology, and specifically relates to a soft measurement method, system and storage medium for boiler thermal efficiency based on machine learning. Background Technology

[0002] Coal-fired power units are initially designed as a stable power source for the power system. However, under the conditions of deep peak shaving, these units are forced to operate at low loads deviating from their design conditions for extended periods, resulting in severe damage to boiler thermal efficiency. Improving boiler thermal efficiency under complex operating conditions such as deep peak shaving and transient conditions requires establishing an accurate thermal efficiency prediction model. Therefore, this paper conducts research on dynamic modeling of thermal efficiency.

[0003] Research on boiler combustion process modeling includes two branches: mechanistic methods and data-driven methods. Mechanistic methods are based on kinetic models and the law of energy conservation involved in combustion and related processes, and then establish mathematical models reflecting the process mechanisms. Previous researchers have established models describing boiler thermal efficiency and NOx based on the combustion mechanism of in-furnace coal distribution and over-fire air (OFA). x Mathematical models of emissions were established using a simplified combustion system, resulting in a discrete-time linear state-space model. Computational Fluid Dynamics (CFD) is also a common research method for establishing mechanistic models of boiler combustion processes. For example, a coupled CFD model of the boiler's flue gas and steam-water sides was established based on mechanistic methods, using a Dell 3600 workstation with a 3GHz main frequency for four-thread parallel computation. However, solving a single operating condition takes about a week, which is unacceptable for coal-fired power plants. Taking boiler thermal efficiency prediction as an example, coking, wear, and replacement of heating surfaces all affect the thermodynamic process characteristics. Therefore, with increasing operating time, the mechanistic model will gradually become mismatched with actual operating conditions. Moreover, the combustion process of coal-fired boilers is complex, involving aerodynamics, two-phase flow, heat and mass transfer processes in the furnace. The accuracy and real-time performance of online predictions of thermal efficiency calculated using mechanistic models are difficult to guarantee. In summary, the applicability of mechanistic modeling is limited.

[0004] Compared to mechanistic methods, data-driven methods leverage machine learning algorithms to directly extract mapping relationships between system parameters from data, resulting in faster and more accurate modeling. For industrial processes where the mechanisms are not yet fully understood, data-driven modeling methods, primarily based on neural networks, are particularly suitable.

[0005] The modeling method in the prior art does not consider the time scale in the input dimension, belongs to the category of static models, and is applicable to the stable state of the boiler. Specifically, when the boiler is started from a cold state or switched to different loads, the historical working conditions will have an impact on the current thermal efficiency. This is because when the load of the boiler changes, the heat storage working medium will be released or rebalanced. When the boiler is in a stable running state, the heat storage effect can generally improve the combustion efficiency and thermal efficiency. Therefore, it is necessary to establish a dynamic model for the time sequence characteristics of the thermal efficiency of the boiler, that is, to include the historical working conditions in the model input.

[0006] BiGRU can capture the context dependence of time sequence data through forward and backward GRU units, comprehensively extract dynamic features, and has strong bidirectional time sequence modeling capability. However, BiGRU relies on linear weight matrix combination features, and it is difficult to eliminate complex nonlinear coupling between variables. However, KAN can accurately map complex nonlinear relationships through a learnable spline function (such as B-spline), improve model generalization, and improve nonlinear fitting capability. The two can complement each other. The BiGRU-KAN hybrid model solves the precision and interpretability bottleneck of traditional soft measurement methods in dynamic nonlinear systems through bidirectional time sequence feature extraction and mathematically driven nonlinear decoupling, and provides an efficient and transparent algorithm core for intelligent monitoring of industrial processes. However, when dealing with large-scale inputs, the BiGRU-KAN hybrid model may ignore the key features of individual time points, resulting in a decrease in model learning ability and affecting the prediction performance. Therefore, the introduction of the Attention mechanism can highlight the contribution of different time step features to the prediction result, thereby improving the model accuracy. SUMMARY

[0007] To solve the above problems, the purpose of the present application is to provide a boiler thermal efficiency soft measurement method, system and storage medium based on machine learning, which combines BiGRU layers to capture time sequence features using bidirectional units and KAN nonlinear decoupling capability, and adds an Attention layer to weight the correlation of each time point feature to the prediction target, thereby constructing a hybrid model that can effectively overcome the shortcomings of a single model and achieve high-precision dynamic measurement of the boiler thermal efficiency.

[0008] To achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0009] A boiler thermal efficiency soft measurement method based on machine learning, comprising the following steps:

[0010] Step 1: Collecting historical operation data from the boiler DCS system;

[0011] Step 2: Preprocessing the data obtained in step 1, including missing data processing, data dimensionality reduction, data reconstruction, and then dividing the data set into a training set, a validation set and a test set;

[0012] Step 3, a BiGRU-KAN-Attention model is constructed, which is composed of an input layer, a BiGRU network, a KAN classifier, an Attention layer and an output layer connected in sequence, wherein: the input layer receives the boiler operation data, the BiGRU network captures time series features with a bidirectional unit, the KAN classifier performs nonlinear feature decoupling, the Attention layer weights the correlation degree of each time point feature to the prediction target, and the output layer outputs the predicted boiler thermal efficiency;

[0013] Step 4, the BiGRU-KAN-Attention model is trained using the training set, and the BiGRU-KAN-Attention model parameters are optimized by an optimization algorithm to determine the best network structure and parameter settings;

[0014] Step 5, input the test set data into the optimized BiGRU-KAN-Attention model, obtain the prediction results and perform inverse normalization processing, then analyze and evaluate the prediction results using an evaluation function, and verify the prediction performance of the model under different operating conditions;

[0015] Step 6, input the real-time collected boiler operation data into the trained BiGRU-KAN-Attention model to output the dynamic thermal efficiency prediction value.

[0016] Further, in step 1, the collected historical operation data includes thermal efficiency and auxiliary variables related thereto.

[0017] Further, in step 2, the missing data processing step is: abnormal data is detected using 3σ-rule, and abnormal data is replaced based on the following formula:

[0018] △x i (t)=x i (t)-x i (t-1)

[0019]

[0020] Wherein, △x i (t) represents the auxiliary variable x i The change rate of the adjacent two time points, represents the average change rate of △x i (t) before time t, σ i (t) represents the standard deviation, t=2,…,n is the sampling time.

[0021] Further, in the step 2, the data dimension reduction step is: using random forest algorithm to sort the importance of auxiliary variables, and screening the top 10 important variables as input variables.

[0022] Further, in the step 2, the data reconstruction step is: using sliding window method to convert time series data into two-dimensional tensor, wherein the time dimension contains 20 continuous sampling time points, and the feature dimension is the screened 10 variables.

[0023] Further, in the step 3, the BiGRU network includes at least two GRUs, and a channel equalization block (CE-Block) is connected after each GRU layer to suppress the distribution difference of feature values between channels.

[0024] Further, in the step 4, the hyperparameters of the BiGRU-KAN-Attention model are set as: batch size 32, epoch 200, optimizer Adam, learning rate 0.005, and activation function ReLU.

[0025] A system for implementing the machine learning-based boiler thermal efficiency soft measurement method, comprising:

[0026] A data acquisition module for real-time acquisition of boiler DCS system operation data;

[0027] A preprocessing module for performing missing data processing, data dimension reduction, and data reconstruction;

[0028] A model operation module for deploying the BiGRU-KAN-Attention model and performing prediction calculation;

[0029] A visualization terminal for displaying real-time boiler thermal efficiency prediction results.

[0030] A storage medium storing a computer program, wherein the computer program is executed by a processor to implement the machine learning-based boiler thermal efficiency soft measurement method.

[0031] Further, the storage medium is integrated into a coal-fired power plant DCS system or an edge computing terminal, which receives boiler operation data in real time and outputs thermal efficiency prediction results.

[0032] Beneficial effects: the application can effectively overcome the shortcomings of single model, realize high-precision dynamic measurement of boiler thermal efficiency, and significantly improve the precision and real-time performance of boiler thermal efficiency measurement by combining the time series feature capturing ability of BiGRU network bidirectional unit, the KAN nonlinear decoupling ability and the ability of Attention layer to emphasize the feature weight that contributes significantly to the prediction result, and constructing a dynamic comprehensive model. The application designs a dynamic soft measurement method suitable for boiler operation data, which can adapt to multivariate and nonlinear characteristics under complex working conditions, and has strong robustness and generalization ability. Specifically, the application uses BiGRU to capture time series features of boiler operation data in bidirectional unit, and captures the nonlinear relationship in the data; then, the extracted features are input into the KAN layer for fast nonlinear decoupling to improve the fitting ability; then, the Attention layer discriminates the specific time point feature weight that contributes significantly to the prediction result, and emphasizes and optimizes the weighted features and calculates the prediction result through the output layer, so as to realize real-time and dynamic measurement of boiler thermal efficiency; finally, the model is evaluated by using real data of 600MW coal-fired boiler under variable load conditions. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 A step diagram of the boiler thermal efficiency dynamic soft measurement method based on BiGRU-KAN-Attention in the embodiment;

[0034] Figure 2 A data reconstruction schematic diagram in the embodiment, wherein (a) is a time series variable, and (b) is a tensor diagram obtained by a sliding window;

[0035] Figure 3 A data set segmentation schematic diagram in the embodiment;

[0036] Figure 4 A structure schematic diagram of a gated network (GRU);

[0037] Figure 5 A structure schematic diagram of a bidirectional gated network (BiGRU);

[0038] Figure 6 A KAN structure schematic diagram;

[0039] Figure 7 An Attention structure schematic diagram;

[0040] Figure 8 A BiGRU-KAN-Attention model structure diagram in the embodiment;

[0041] Figure 9 A prediction result diagram of BiGRU-KAN-Attention on the test set in the embodiment;

[0042] Figure 10 Figure for prediction results of KAN on test set in the embodiment;

[0043] Figure 11 Figure for prediction results of BiGRU on test set in the embodiment. DETAILED DESCRIPTION

[0044] The application will be further explained in connection with the accompanying drawings and embodiments.

[0045] EMBODIMENT

[0046] As shown in the figure, the boiler thermal efficiency soft measurement method based on machine learning in the embodiment comprises the following steps: Figure 1

[0047] (1) Collect historical operation data from the boiler DCS system.

[0048] The research object of the embodiment is a 600MW supercritical thermal power unit W flame boiler. The boiler is a supercritical parameter, W flame combustion, variable pressure operation direct current boiler. The main design parameters of the boiler are shown in Table 1. The power plant is located in the hinterland of southwest China, the climate is humid, and the low-quality anthracite with less than 12% volatile matter is burned, which leads to a large amount of coking in the furnace. The coking phenomenon seriously affects the combustion atmosphere and heat transfer in the furnace, making it more difficult to predict the boiler thermal efficiency through the mechanism model, so the data-driven modeling method is more suitable for such industrial scenarios.

[0049] Table 1 Main design parameters of the boiler

[0050]

[0051] According to the combustion mechanism, expert experience and suggestions of operation engineers, the present application integrates 53 related variables such as wind, coal and water to form auxiliary variables, and the prediction target is the boiler efficiency, as shown in Table 2. The selection of auxiliary variables comprehensively considers the influence factors such as the air dynamic field in the furnace, the wind-coal two-phase flow field and the working medium heat absorption. Among them, the secondary air controls the main air required for combustion in the furnace, the mill inlet capacity wind flow controls the distribution of the wind-coal two-phase flow field in the furnace, and the main steam pressure, main steam temperature and other variables reflect the working medium heat absorption level in the furnace.

[0052] (2) Preprocess the obtained data.

[0053] The historical data of about 7 days of continuous operation are collected from the target boiler DCS, a total of 5019 samples, and the sampling period is 2 minutes. The original data is preprocessed to meet the model input requirements, and the data preprocessing process includes missing data processing, data dimensionality reduction, data reconstruction and data set division.

[0054] (2.1) Missing data processing​

[0055] Due to the complex environment of boiler production site, signal mixed with noise, leading to the existence of abnormal values in different degrees of the original data stored in the DCS. This embodiment adopts 3σ-rule to detect abnormal data, and replaces the abnormal data based on the following formula:

[0056] Δx i (t) = x i (t) - x i (t-1)

[0057]

[0058] Where, Δx i (t) represents the auxiliary variable x i The rate of change in adjacent two time, represents the average rate of change of Δx i (t) before time t, σ i (t) represents the standard deviation, t = 2, …, n is the sampling time.

[0059] Table 2 Auxiliary variable table

[0060]

[0061] (2.2) Data dimension reduction

[0062] This embodiment is aimed at the high-dimensional nonlinear and large inertia industrial properties of the boiler combustion system, and proposes to use feature engineering to process the auxiliary variables, and analyzes the rationality of supervised dimension reduction in the combustion process based on real historical data.

[0063] Random Forest (RF) is an extended variant of Bagging, which is particularly suitable for evaluating the importance between high-dimensional nonlinear variables. Because RF uses several independent decision trees for parallel prediction, random and with replacement sampling makes the out-of-bag data naturally have the function of evaluating the deterioration of model prediction error, and the degree of deterioration indirectly indicates the importance of the variable. For each decision-making decision tree, the bag-in data is used to train the tree, and the bag-out data (OOB) is used to evaluate the tree. The deterioration of prediction accuracy will be realized by adding noise interference to important variables. Before and after the interference, the importance measure of the input variable is given by the following formula:

[0064]

[0065] Where, The prediction errors of the out-of-bag data before and after the interference on the decision tree b, respectively, B is the number of decision trees. The sum of the importance of all input variables is 1, that is, represents the number of auxiliary variables.

[0066] The RF algorithm is used to evaluate the feature importance of 53 comprehensive auxiliary variables affecting thermal efficiency, and the decision tree is set to 500. The sum of the importance of the top 5 variables is more than 0.9. The author combines the prior knowledge of the boiler, and the F layer secondary air door which mainly affects the combustion in the furnace is also included in the input variable. In addition, the load has a greater impact on the overall unit, so the input variable is determined to be 10 by comprehensive consideration. The total weight of the top 10 variables is 0.933, and the reordering is shown in Table 3.

[0067] Table 3 Importance of variables related to thermal efficiency (mean ± standard value)

[0068]

[0069] The variable importance calculated by RF follows the change rule of boiler thermal efficiency. The model complexity and universality are compromised, and the top 10 important variables are selected as the input variables of the thermal efficiency model in this embodiment.

[0070] (2.3) Data reconstruction

[0071] The input format of two-dimensional BiGRU is a continuous two-dimensional tensor. In order to meet the input requirements, the input variables need to be reconstructed. As time series data, the sliding window method is used to divide the data according to the feature dimension and time dimension. The time dimension contains 20 sampling time points, i.e. H = 20; the feature dimension is 10 variable types, i.e. W = 10; and the window sliding step is 1. The label corresponding to each two-dimensional tensor is the thermal efficiency of the last time dimension of the tensor, Figure 2 The data reconstruction schematic diagram is shown.

[0072] (2.4) Data set segmentation

[0073] The original 5019 sampling data are divided into 5000 standard input matrices, and the original sampling serial number of the label (thermal efficiency) corresponding to each matrix is 20-5019. We reorder it to 1-5000 to make it consistent with the matrix serial number, that is, the serial number of the input matrix and the label is counted from 1, as shown in Figure 3 The data set segmentation schematic diagram is shown. The segmented data is divided into training set, validation set and test set according to the ratio of 6:2:2, and the data set maintains the original sampling order.

[0074] (3) constructing a BiGRU-KAN-Attention model, the BiGRU-KAN-Attention model is composed of an input layer, a BiGRU network, a KAN network, an Attention layer and an output layer connected in turn, wherein: the input layer receives the boiler operation data, the BiGRU network utilizes a bidirectional unit to capture time sequence features, the KAN classifier performs nonlinear feature decoupling, the Attention layer weights the correlation degree of each time point feature to the prediction target, and the output layer outputs the predicted boiler thermal efficiency.

[0075] The Gate Recurrent Unit (GRU) is a variant of the Recurrent Neural Network (RNN), and the core mechanism revolves around two key gating operations: update gate and reset gate. As shown in Figure 4 , the update gate determines the proportion of the past hidden state retained in the current state, affecting the continuity of information; the reset gate determines how the model integrates information based on the current input and the previous state, thereby adapting to the dynamic changes in the sequence. The specific process of the GRU forward transmission is as follows:

[0076] r t =σ(W r ·[X t ,h t-1 ]+b r )

[0077] z t =σ(W z ·[X t ,h t-1 ]+b z )

[0078]

[0079] In the formula: r t is the reset gate, X t is the current input, h t-1 is the hidden information at the last time, sigma is the sigmoid function, [] represents matrix stacking, W r is the weight of the reset gate, b r is the bias of the reset gate, z t is the update gate, W z is the weight of the update gate, b z is the bias of the update gate, h xt is the candidate hidden state content, tanh is the hyperbolic tangent function, W h is the weight of the candidate hidden state, b h is the bias of the candidate hidden state, and h t is the hidden state.

[0080] To effectively capture past and future information in the sequence, Bi-directional Gated Recurrent Unit (BiGRU) is widely used. As shown in Figure 5 BiGRU integrates two GRU units, one of which propagates information along the time sequence in the forward direction, and the other of which backtracks in the reverse direction, so that the model has the ability to review the past and look into the future. The output of each BiGRU unit at time t is jointly affected by the bias at that time, the forward GRU and the backward GRU output, as follows:

[0081]

[0082] where GRU(·) is the calculation process of GRU, is the output value of the GRU forward hidden layer at time t, is the output value of the GRU backward hidden layer at time t, m is the GRU forward hidden layer output weight, n is the GRU backward hidden layer output weight, c t is the bias of the corresponding hidden layer. t

[0083] KAN is a new type of neural network architecture based on Kolmogorov-Arnold representation theorem, Figure 6 is a schematic diagram of KAN neural network structure. In the traditional multi-layer perceptron (MLP) model, the weight parameters are usually set at the network edge, and the fixed activation function is placed on the neuron. Unlike MLP, KAN model places a learnable activation function at the network edge, which is adaptively adjusted with network training, so that the network weight parameters are replaced by univariate spline function, so that the network can fit the complex network structure with fewer parameters while maintaining flexibility. KAN model is composed of external function and internal function, and its mathematical expression is:

[0084]

[0085] where x is an n-dimensional input vector; is the internal function, the domain is [0, 1], and the value domain is the real number set R; ψ q is the external function, and the domain and value domain are both R.

[0086] The attention mechanism learns from the way humans focus on processing information, like the brain's process of screening information. The core of deep learning based on attention mechanism is to let the machine learn to perceive important and unimportant parts of the data, and its structure is shown in Figure 7 The calculation steps are as follows:​

[0087] (1) Encode the input sequence [y1, y2,..., y n ] to get a set of query vectors q; then score the similarity between them [s1, s2,..., s n ], the scoring function is as follows:

[0088] s i = V T tanh(Wy i + Uq)

[0089] Where V, W, U are all parameter matrices to be learned.

[0090] (2) Normalize using the Softmax function to get each input vector a i , the formula is as follows:

[0091]

[0092] (3) Calculate the weighted sum according to the weight and the corresponding value vector to get the final hidden layer state value output a, the formula is as follows:

[0093]

[0094] In view of the various excellent performances of bidirectional gated recurrent unit BiGRU, KAN network and Attention mechanism, the three are combined to construct BiGRU-KAN-Attention network, which fully utilizes their respective advantages and overcomes their shortcomings.

[0095] Figure 8 The structure of BiGRU-KAN-Attention network is that its BiGRU network is composed of two single-layer GRUs, KAN completes nonlinear decoupling, Attention network highlights weighting, and is the final output layer. The BiGRU-KAN-Attention model process is as follows: first, the network data is collected through the network data capture module, and the collected historical network data is preprocessed, including feature mapping digitization, normalization and dimension reconstruction, etc., to form a training data set; then, after receiving the feature data, the BiGRU layer captures the time series features by using the bidirectional unit; next, the Attention layer weights the correlation degree of each time point feature to the prediction target; finally, the weighted features are summarized to obtain the optimal BiGRU-KAN-Attention measurement model. In online measurement, first, the real-time network data generated in the previous few seconds is collected through the network data capture module, and the real-time network data is preprocessed at the same time, then the preprocessed test data set is sent into the BiGRU-KAN-Attention model to complete the measurement classification of the input network connection.

[0096] (4) The BiGRU-KAN-Attention model is trained using the training set, the BiGRU-KAN-Attention model parameters are optimized through an optimization algorithm, and the optimal network structure and parameter settings are determined;

[0097] Table 4 Structure of BiGRU-KAN-Attention when the feature dimension is 10 variables

[0098]

[0099] To ensure the lightweight of the model structure, the BiGRU-KAN-Attention is only equipped with 2 GRUs, KAN decoupling layers, KAN combination layers, and Attention layers. As shown in Table 4, the hyperparameters of the BiGRU-KAN-Attention are determined through multiple training and verification. In addition, the batch size is 32, the epoch is 200, the optimizer is Adam, and the learning rate is 0.005.

[0100] (5) The test set data is input into the optimized BiGRU-KAN-Attention model, the prediction results are obtained and are subjected to inverse normalization processing, and then the prediction results are analyzed and evaluated using an evaluation function, so as to verify the prediction performance of the model under different working conditions;

[0101] (6) The real-time collected boiler operation data are input into the trained BiGRU-KAN-Attention model, and the dynamic thermal efficiency prediction value is output.

[0102] To evaluate the performance of the BiGRU-KAN-Attention thermal efficiency dynamic prediction model, different modeling methods are compared with the BiGRU-KAN-Attention, and the mean and standard deviation of the model prediction error come from twenty independent experiments.

[0103] The BiGRU-KAN-Attention and the BiGRU model are compared on the test set. The BiGRU model adopts a single hidden layer structure, and since the input is a 20x10 tensor, the hidden layer nodes contain 20 GRUCells respectively. The output of the last Cell of the BiGRU is taken as the overall prediction result of the model. The mean and standard deviation of the prediction error of the BiGRU, KAN and BiGRU-KAN-Attention are shown in Table 5.

[0104] Table 5 Prediction performance of BiGRU, KAN and BiGRU-KAN-Attention on the test set

[0105]

[0106] The RMSE of BiGRU-KAN-Attention was 0.133±0.008%, the MAE was 0.065±0.007%, and the R... 2 The value was 0.978 ± 0.003, and all performance metrics were superior to the BiGRU and KAN models. The prediction curve for a randomly selected experiment from 20 experiments is shown below. Figure 9 As shown, the corresponding unit load curve is also as follows. Figure 9 As shown. Figure 10 and Figure 11 The images show the prediction results of BiGRU and KAN on the test set, respectively. In summary, the thermal efficiency prediction model based on BiGRU-KAN-Attention is capable of handling both transient and steady-state conditions.

[0107] The present invention also provides a system for implementing the aforementioned machine learning-based soft measurement method for boiler thermal efficiency, comprising:

[0108] The data acquisition module acquires real-time operating data from the boiler's DCS system.

[0109] The preprocessing module performs missing data processing, data dimensionality reduction, and data reconstruction.

[0110] The model computation module deploys the BiGRU-KAN-Attention model and performs prediction calculations.

[0111] The visualization terminal displays real-time boiler thermal efficiency prediction results.

[0112] The present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the machine learning-based soft measurement method for boiler thermal efficiency of the present invention. Furthermore, the storage medium is integrated into a DCS system or edge computing terminal of a coal-fired power plant to receive boiler operating data in real time and output boiler thermal efficiency prediction results.

[0113] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A soft measurement method for boiler thermal efficiency based on machine learning, characterized in that: Includes the following steps: Step 1: Collect historical operating data from the boiler DCS system; Step 2 involves preprocessing the data obtained in Step 1, including: handling missing data, data dimensionality reduction, and data reconstruction. Then, the dataset is divided into training set, validation set, and test set. Step 3: Construct a BiGRU-KAN-Attention model. The BiGRU-KAN-Attention model consists of an input layer, a BiGRU network, a KAN classifier, an Attention layer, and an output layer connected in sequence. The input layer receives boiler operation data, the BiGRU network uses bidirectional units to capture time series features, the KAN classifier performs nonlinear feature decoupling, the Attention layer weights the correlation of the prediction target according to the features at each time point, and the output layer outputs the predicted boiler thermal efficiency. Step 4: Train the BiGRU-KAN-Attention model using the training set, optimize the parameters of the BiGRU-KAN-Attention model using the optimization algorithm, and determine the optimal network structure and parameter settings. Step 5: Input the test set data into the optimized BiGRU-KAN-Attention model, obtain the prediction results and perform inverse normalization, and then use the evaluation function to analyze and evaluate the prediction results to verify the prediction performance of the model under different working conditions. Step 6: Input the real-time collected boiler operation data into the trained BiGRU-KAN-Attention model and output the dynamic thermal efficiency prediction value.

2. The method according to claim 1, characterized in that: In step 1, the collected historical operating data includes thermal efficiency and related auxiliary variables.

3. The method according to claim 1, characterized in that: In step 2, the missing data processing steps are as follows: 3σ-rule is used to detect outlier data, and outlier data is replaced based on the following formula: △x i (t)=x i (t)-x i (t-1) Where, △x i (t) represents the auxiliary variable x i The rate of change between two adjacent time points Indicate △x i (t) represents the average rate of change before time t, σ i (t) represents the standard deviation, and t = 2, ..., n represents the sampling time.

4. The method according to claim 1, characterized in that: In step 2, the data dimensionality reduction step is as follows: the auxiliary variables are ranked by importance using the random forest algorithm, and the top 10 most important variables are selected as input variables.

5. The method according to claim 1, characterized in that: In step 2, the data reconstruction steps are as follows: the time series data is converted into a two-dimensional tensor using the sliding window method, wherein the time dimension contains 20 consecutive sampling times and the feature dimension consists of 10 filtered variables.

6. The method according to claim 1, characterized in that: In step 3, the BiGRU network includes at least two GRU layers, each of which is followed by a channel equalization block (CE-Block) to suppress the distribution differences of feature values ​​between channels.

7. The method according to claim 1, characterized in that: In step 4, the hyperparameters of the BiGRU-KAN-Attention model are set as follows: batch size 32, epoch 200, optimizer Adam, learning rate 0.005, and activation function ReLU.

8. A system for implementing the method of any one of claims 1-7, characterized in that, include: The data acquisition module acquires real-time operating data from the boiler's DCS system. The preprocessing module performs missing data processing, data dimensionality reduction, and data reconstruction. The model computation module deploys the BiGRU-KAN-Attention model and performs prediction calculations. The visualization terminal displays real-time boiler thermal efficiency prediction results.

9. A storage medium, characterized in that: The storage medium stores a computer program that, when executed by a processor, implements the machine learning-based soft measurement method for boiler thermal efficiency as described in any one of claims 1-7.

10. The storage medium according to claim 9, characterized in that: The storage medium is integrated into the DCS system or edge computing terminal of a coal-fired power plant to receive boiler operation data in real time and output thermal efficiency prediction results.

Images