Modelica-based parameter identification and hybrid modeling method for economizer

Abstract

This invention discloses a hybrid modeling method for economizer parameter identification based on Modelica. The method uses Modelica to model the main mechanisms of the economizer. Addressing the unpredictable water pressure drop issue in economizer mechanism prediction, a DNN neural network model is used for parameter identification and data-driven modeling of the feedwater flow resistance coefficient. The DNN network parameters are derived and integrated into the economizer mechanism model. Modelica is used to create an interpretable and general mechanism model for the economizer, while the data-driven model improves the prediction accuracy of the mechanism model. Finally, the data-driven model parameters are exported and written into the mechanism model, enabling efficient engineering applications.

Description

Technical Field

[0001] This application relates to the field of thermal power plant unit equipment modeling technology, and in particular to a hybrid modeling method for economizer parameter identification based on Modelica. Background Technology

[0002] Modelica is a component-oriented modeling language specifically designed for modeling physical systems across multiple domains. Its mechanistic models are based on physical laws, describing the dynamic behavior of systems through equations, making it well-suited for simulating the interactions of complex systems such as thermodynamics, fluid dynamics, and electromechanical systems. Modelica supports component-oriented modeling methods, allowing devices and systems to be modeled as independent modules (or components), each module representing a unit device (such as a pump, valve, or boiler). These modules can be reused across different systems, greatly improving model reusability and maintainability. However, mechanistic models may contain idealized elements, leading to mismatches between model predictions and actual conditions.

[0003] Hybrid mechanistic-data modeling combines the interpretability of physical mechanism models with the accuracy of data-driven models, effectively improving the predictive power and robustness of the model. Physical mechanism models provide the fundamental laws of the system, ensuring the model's universality, while data-driven models can capture complex nonlinear relationships, compensating for the shortcomings of physical models and thus improving overall performance. Current hybrid mechanistic-data models mainly use various interfaces to transfer information between the two models; however, mechanistic models and data-driven models on different platforms have significant limitations in practical applications, especially in system modeling of Modelica unit devices, which can lead to extremely cumbersome systems. Summary of the Invention

[0004] The purpose of this application is to provide a hybrid modeling method for economizer parameter identification based on Modelica, so as to solve the technical problem of the transferability of hybrid models in practical applications in related technologies.

[0005] According to a first aspect of the embodiments of this application, a hybrid modeling method for economizer parameter identification based on Modelica is provided, characterized in that it includes:

[0006] Modelica was used to model the heat transfer process between flue gas and water in the economizer and the energy-momentum balance of the two streams, resulting in a mechanistic model.

[0007] The ideal feedwater flow resistance coefficient is calculated using first field data, which includes the inlet and outlet flue gas pressure drop of the economizer, feedwater flow rate, feedwater temperature, outlet water temperature, and the height difference between the inlet and outlet of the economizer.

[0008] The second field data is standardized, and then the input features of the DNN model are calculated using both PCC and MI methods. The second field data includes furnace coal quantity, feedwater flow rate, temperature, and pressure, and inlet flue gas temperature, pressure, and flow rate.

[0009] The DNN model is trained using the input features of the DNN model and the ideal water supply flow resistance coefficient as output features, and Bayesian optimization is used to fine-tune the DNN model structure and hyperparameters.

[0010] Derive the weights, biases, and activation functions of neurons in each layer of the optimized DNN model;

[0011] Write the weights, biases, and activation functions into the mechanism model. Create a new function in the mechanism model and set multiple parameters to store the weight and bias matrices, the mean and variance of the 6-dimensional input features.

[0012] In the newly created function, use the Modelica language to reconstruct the mapping relationships between neurons in each layer of the DNN model;

[0013] By calling the newly created function through the aforementioned mechanism model, the parameter identification of the water supply flow resistance coefficient can be achieved using the DNN model within the mechanism model.

[0014] Optionally, equations for the heat transfer and momentum balance of the flue gas and water in the economizer are constructed using Modelica, including:

[0015] Two streams, flue gas and water, are obtained and denoted by the subscripts g and w, respectively. The three parameters of temperature, pressure, and flow rate for the two streams are denoted by T, p, and qm, respectively. The input and output streams are distinguished by the subscripts i and o. The heat transfer equations for flue gas and water are constructed as follows:

[0016] ;

[0017] ;

[0018] in, The average temperature difference between the two streams. For heat transfer power, coefficient The heat transfer coefficient;

[0019] Construct the momentum balance equation for flue gas:

[0020] ;

[0021] in, This refers to the fixed pressure drop in the flue gas section due to factors such as height. The flow resistance coefficient of the flue gas. The density of the flue gas;

[0022] The momentum balance equation of water:

[0023] ;

[0024] in, The height difference between the inlet and outlet of the economizer flow stream, These are the parameters that need to be identified using a DNN model;

[0025] Construct the energy balance equation:

[0026] ;

[0027] in, For the internal volume of the economizer, This refers to the specific heat of the flue gas.

[0028] Optionally, a DNN model is used to identify the parameters of the feedwater flow resistance coefficient in the mechanistic model. Standardized training data is input. The DNN model includes an input layer, hidden layers, and an output layer. The input feature dimension of the DNN model is 6, including the furnace coal quantity, economizer feedwater flow rate and pressure, and inlet flue gas temperature, pressure, and flow rate. The output feature dimension is 1, which is the feedwater flow resistance coefficient. The number of intermediate hidden layers and the number of neurons in each layer are obtained through Bayesian optimization. In the DNN model, the data is passed from the input layer to the hidden layer, and then through all the hidden layers to finally reach the output layer, completing one forward pass. Each neuron receives the output from the previous layer, and the weighted sum is used to obtain the output result through the ReLU activation function.

[0029] Optionally, the reconstructed mapping relationship is as follows:

[0030]

[0031]

[0032]

[0033] This represents the standardized result of the model input. This represents the mapping result of the nth layer neuron. This represents the weight matrix of the neurons in the nth layer; Let be the bias matrix of the neurons in the nth layer. This represents the final mapping result, i.e., the predicted result of the required feedwater flow resistance coefficient, where... ReLU means comparing each element in matrix A with 0 and taking the larger value, which is the calculation method of ReLU. The last output layer does not have ReLU, so there is only a linear mapping. ReLU is the activation function.

[0034] According to a second aspect of the embodiments of this application, a hybrid modeling apparatus for economizer parameter identification based on Modelica is provided, comprising:

[0035] The modeling module is used to model the heat transfer process of flue gas and water in the economizer and the energy-momentum balance of the two streams using Modelica, so as to obtain the mechanism model.

[0036] The first calculation module is used to calculate the ideal feedwater flow resistance coefficient using the first field data, which includes the inlet and outlet flue gas pressure drop of the economizer, feedwater flow rate, feedwater temperature, outlet water temperature, and the height difference between the inlet and outlet of the economizer.

[0037] The second calculation module is used to standardize the second field data and then use PCC and MI methods to weight and calculate the input features of the DNN model. The second field data includes furnace coal quantity, feedwater flow rate, temperature, and pressure, and inlet flue gas temperature, pressure, and flow rate.

[0038] The training optimization module is used to train the DNN model using the input features of the DNN model and the ideal water supply flow resistance coefficient as output features, and to fine-tune the DNN model structure and hyperparameters using Bayesian optimization.

[0039] The export module is used to export the weights, biases, and activation functions of neurons in each layer of the optimized DNN model.

[0040] The writing module is used to write the weights and biases into the mechanism model. A new function is created in the mechanism model, and multiple parameters are set to store the weight and bias matrix, the mean and variance of the 6-dimensional input features.

[0041] The reconstruction module is used to reconstruct the mapping relationships between neurons in each layer of a DNN model using the Modelica language in a newly created function;

[0042] The calling module is used to call the newly created function through the mechanism model to realize the parameter identification of the water supply flow resistance coefficient using the DNN model in the mechanism model.

[0043] According to a third aspect of the embodiments of this application, an electronic device is provided, characterized in that it includes:

[0044] One or more processors;

[0045] Memory, used to store one or more programs;

[0046] When the one or more programs are executed by the one or more processors, the one or more processors perform the method as described in the first aspect.

[0047] According to a third aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described in the first aspect.

[0048] The technical solutions provided by the embodiments of this application may include the following beneficial effects:

[0049] This application adopts the mechanistic modeling method of Modelica, and the resulting final unit device model can be transferred to various similar scenarios, with strong reusability and scalability.

[0050] A data-driven parameter identification method was adopted to overcome the problem of low prediction accuracy in mechanistic models.

[0051] By exporting and writing data-driven models into Modelica functions, a unified unit device model of mechanistic data hybrid models on the Modelica platform is realized, solving the problem that current hybrid models require interfaces for model reading in practical engineering applications.

[0052] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0053] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0054] Figure 1 This is a flowchart illustrating a hybrid modeling method for economizer parameter identification based on Modelica, according to an exemplary embodiment.

[0055] Figure 2 is a schematic diagram showing a comparison between the model prediction results after using the present invention and the mechanistic model prediction results without parameter identification, according to an exemplary embodiment.

[0056] Figure 3 This is a schematic diagram of a hybrid modeling device for economizer parameter identification based on Modelica, according to an exemplary embodiment. Detailed Implementation

[0057] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0058] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0059] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0060] Figure 1 This is a flowchart illustrating a hybrid modeling method for economizer parameter identification based on Modelica, according to an exemplary embodiment. Figure 1 As shown, this method, when applied to a terminal, may include the following steps:

[0061] Step S1: Use Modelica to model the heat transfer process of flue gas and water in the economizer and the energy-momentum balance of the two streams to obtain the mechanism model;

[0062] Specifically, the mechanism model mainly consists of the following heat transfer equation, flue gas momentum balance equation, water momentum balance equation, and energy balance equation. The construction of each equation will be explained in detail below.

[0063] Create a new model in Modelica, define the variables for the flue gas and water inlet / outlet streams: flow rate, temperature, and pressure, and write the following equations into the model's equations. First, the flue gas and water streams are represented by the subscripts g and w, respectively. The three parameters for the two streams—temperature, pressure, and flow rate—are represented by T, p, and qm, respectively. The input and output streams are distinguished by the subscripts i and o. The heat transfer equation is:

[0064]

[0065]

[0066] in The average temperature difference between the two streams. For heat transfer power, For heat exchange area, coefficient Fitting is performed using training data.

[0067] Flue gas momentum balance equation:

[0068]

[0069] in This refers to the fixed pressure drop in the flue gas section due to factors such as height. The flow resistance coefficient of the flue gas. This represents the density of the flue gas.

[0070] The momentum balance equation of water:

[0071]

[0072] in The height difference between the inlet and outlet of the economizer's flow stream.

[0073] Energy balance equation:

[0074]

[0075] For the internal volume of the economizer, This refers to the specific heat of the flue gas.

[0076] Modelica, as a non-causal modeling method, uses equations to describe the heat exchange process of the economizer. In the actual use of the model, as long as a sufficient number of model boundary conditions are known, the remaining boundary conditions can be calculated, rather than being limited to determining the outlet flow only through the inlet flow.

[0077] Step S2: Calculate the ideal flow resistance coefficient of the feedwater using the first field data. The first field data is collected by the thermal power plant sensors and includes flue gas pressure drop, feedwater flow rate, feedwater temperature, outlet water temperature, and economizer inlet and outlet height difference.

[0078] Specifically, the ideal water supply flow resistance coefficient for each group was calculated using first-hand field data. The specific calculation formula is as follows:

[0079]

[0080]

[0081] The density of water is calculated using average temperature and pressure. This allows for the calculation of the ideal heat transfer coefficient for the process, causing the DNN model's predictions to bias towards the desired ideal value.

[0082] Step S3: Standardize the second field data. The second field data is also collected by the factory sensors. Then, use PCC and MI methods to calculate the input features of the DNN model. The second field data includes the furnace coal quantity, feedwater flow rate, temperature, and pressure, and the inlet flue gas temperature, pressure, and flow rate.

[0083] Specifically, the training data for the data-driven model is standardized:

[0084]

[0085]

[0086]

[0087] For the standardized data, two data correlation calculation methods, PCC and MI, were used:

[0088]

[0089] This is the partial correlation coefficient of XY after the control variable Z, where This represents the Pearson correlation coefficient between X and Y. This represents the Pearson correlation coefficient between Y and Z. This represents the Pearson correlation coefficient between X and Z. The calculated result reflects the linear relationship between X and Y after excluding the influence of Z. The partial correlation coefficient is a number in the range [-1, 1], where -1 indicates a negative linear correlation, 1 indicates a linear correlation, and 0 indicates no linear correlation.

[0090] The formula for calculating MI is as follows:

[0091]

[0092] Mutual information It is a joint distribution of x and y in With marginal distribution The relative entropy. The result of mutual information calculation is a positive number greater than 0, and does not exceed the entropy of x and y themselves. .

[0093] The weighted sum of the two terms is used as the final evaluation metric. The weighting formula is as follows:

[0094]

[0095] in This represents the evaluation score of the nth input feature. This represents the partial correlation coefficient between the nth input feature and the output after controlling for all other input features. The last term represents the input feature. The ratio of mutual information with Y to the entropy of Y itself. Therefore, through... The relative magnitudes of the input features and the number of input features (n) are considered to select appropriate input features for subsequent neural network training and prediction. In this case, the input features, based on mechanistic analysis, may be related to the feedwater flow resistance coefficient. These seven items are: boiler coal quantity, economizer feedwater flow rate, feedwater temperature, feedwater pressure, flue gas flow rate, inlet flue gas temperature, and inlet flue gas pressure. Excluding feedwater temperature, the remaining six items are used as input features for the DNN network. The output feature to be predicted in this case is the economizer feedwater flow resistance coefficient.

[0096] In data-driven modeling, the selection of input features is crucial to model performance. Mechanistic analysis methods based on scientific models in physics, chemistry, or engineering provide a clear theoretical foundation for feature selection, incorporating domain knowledge. However, these models often oversimplify the system and may ignore important features or treat them as constants in practice, potentially hindering the effectiveness of data-driven models. Furthermore, some variables are far removed from the inputs in the system, making it difficult to quantify their impact on the output. MI and PCC measure the statistical correlation between two variables, indicating how much uncertainty in one variable is reduced by understanding the other. It can capture both linear and complex nonlinear relationships between input and target variables. It makes fewer assumptions about the data distribution and is applicable to various data types, including continuous and discrete variables. It can also quantitatively measure the importance of different input variables to the target variable.

[0097] Step S4: Train the DNN model using the input features of the DNN model and the ideal water supply flow resistance coefficient as output features, and use Bayesian optimization to fine-tune the DNN model structure and hyperparameters.

[0098] Specifically, a DNN model is used to identify the parameters of the feedwater flow resistance coefficient in the mechanistic model. Standardized training data is input. The DNN model includes an input layer, hidden layers, and an output layer. The input features in the DNN model have a dimension of 6 (S3), including furnace coal quantity, economizer feedwater flow rate and pressure, and inlet flue gas temperature, pressure, and flow rate. The output features have a dimension of 1, which is the feedwater flow resistance coefficient. The number of intermediate hidden layers and the number of neurons in each layer are obtained through Bayesian optimization. In the DNN model, data is passed from the input layer to the hidden layer, and then through all the hidden layers to finally reach the output layer, completing one forward pass. Each neuron receives the output from the previous layer, and the weighted sum is used to obtain the output result through the ReLU activation function.

[0099] Bayesian optimization is used for hyperparameter tuning of DNN networks. Bayesian hyperparameter tuning simulates the objective function, i.e., the relationship between the model's hyperparameters and the evaluation metric MSE, by constructing a surrogate model. This includes setting:

[0100] Learning rate at 10 -6 -10 -1 ;

[0101] The number of hidden layers is between 1 and 5.

[0102] The number of neurons in each layer ranges from 32 to 128.

[0103] This defined hyperparameter space provides the search range for optimization. The objective function is then defined. In this function, a set of hyperparameter combinations is input at each call, and a DNN model is built using these parameters. The Adam optimizer is used, and the MSE value of the validation set is used to measure performance during model training. Different hyperparameter combinations are evaluated through multiple calls to the objective function. The surrogate model is a Gaussian process, and in each iteration, a sampling function is used to balance exploration and utilization, gradually narrowing the range to find a better solution. Finally, the optimal hyperparameter combination is output. In this case, the final hyperparameters obtained are 3 hidden layers with 69, 126, and 126 neurons respectively, and a learning rate of 0.00906155284684755.

[0104] Bayesian parameter tuning yields a relatively good set of hyperparameters for the DNN model, and the results of several sets of hyperparameters obtained during the Bayesian iteration process can guide subsequent hyperparameter selection. The DNN model is used here for two reasons: firstly, DNN models generally have higher accuracy, consistently outperforming GPR and LSTM models in experiments; secondly, the DNN model has a simpler framework, with parameters primarily consisting of the weights and biases of neurons in each layer and the activation function. Unlike LSTM, it doesn't require storing the current cell state and passing it to the next time step, which is too difficult to implement in Modelica. In contrast, GPR doesn't change the number of parameters with varying training data, even with large datasets. GPR requires storing the covariance matrix of training parameters, and increasing the amount of training data significantly increases the number of parameters in the final encapsulated model. Therefore, the DNN model is easier to implement while maintaining high accuracy.

[0105] Step S5: Export the weights, biases, and activation functions of neurons in each layer of the optimized DNN model;

[0106] Specifically, the weights and bias matrices of each layer of the trained DNN are derived, resulting in four weight matrices and four bias matrices: W1[69,6], W2[126,69], W3[126,126], W4[1,126], B1[69,1], B2[126,1], B3[126,1], and B4[1,1]. The activation function type of the network is also derived.

[0107] Step S6: Write the weights and biases into the mechanism model. Create a new function in the mechanism model and set multiple parameters to store the weight and bias matrix, the mean and variance of the 6-dimensional input features.

[0108] Specifically, create a new function in the mo file, set six input Reals and one output Real, and set eight parameter matrices to store the weight matrix and bias matrix derived from the DNN model, namely W1[69,6], W2[126,69], W3[126,126], W4[1,126], B1[69,1], B2[126,1], B3[126,1], B4[1,1], and a parameter matrix to store the mean and variance of the input features.

[0109] Step S7: Use the Modelica language in the newly created function to reconstruct the mapping relationship between neurons in each layer of the DNN model;

[0110] Specifically, the reconstructed mapping relationship is as follows:

[0111]

[0112]

[0113]

[0114] This represents the standardized result of the model input. This represents the mapping result of the nth layer neuron. This represents the weight matrix of the neurons in the nth layer; Let be the bias matrix of the neurons in the nth layer. This represents the final mapping result, i.e., the predicted result of the required feedwater flow resistance coefficient, where... ReLU means comparing each element in matrix A with 0 and taking the larger value, which is the calculation method of ReLU. The last output layer does not have ReLU, so there is only a linear mapping. ReLU is the activation function.

[0115] Used for data standardization, a set of input features is standardized to x[6,1] in the function, and the output feature is y[1,1]. Then the following calculations are performed:

[0116]

[0117] The y value represents the water flow resistance coefficient that needs to be identified as a parameter.

[0118] By writing the DNN model as a function, it can be debugged independently in Modelica or called in different models, improving its reusability. Furthermore, by performing computations as a function, it eliminates the need to write all parameters into memory during model runtime, reducing the computer load and speeding up model execution.

[0119] Step S8: Call the newly created function through the mechanism model to identify the parameters of the water supply flow resistance coefficient using the DNN model in the mechanism model.

[0120] This solves the problem that current mainstream hybrid models use various interfaces to achieve information transfer between two models. However, the mechanism models and data-driven models of different platforms have great limitations in practical applications, especially in the system modeling of Modelica unit device models, which will make the system extremely cumbersome.

[0121] See final results Figure 2 The horizontal axis represents the number of data sets, and the vertical axis represents the pressure drop value in Pascals. The figure shows the actual pressure drop value, the mechanistic model prediction results when using a constant value as the feedwater flow resistance coefficient, and the results of the hybrid model after using DNN model parameter identification in predicting the economizer pressure drop. From these three figures, it can be seen that the prediction accuracy is greatly improved after using parameter identification.

[0122] As can be seen from the above embodiments, this application adopts the mechanistic modeling method of Modelica, so the final unit device model obtained can be transferred to various similar scenarios, exhibiting strong reusability and scalability. A data-driven parameter identification method is employed to overcome the problem of low model prediction accuracy in mechanistic models. By exporting the data-driven model and writing it into Modelica functions, an integrated unit device model of the mechanistic-data hybrid model is realized on the Modelica platform, solving the problem that current hybrid models require an interface for model reading in practical engineering applications.

[0123] Corresponding to the aforementioned embodiments of the hybrid modeling method for economizer parameter identification based on Modelica, this application also provides embodiments of a hybrid modeling apparatus for economizer parameter identification based on Modelica.

[0124] Figure 3 This is a block diagram of a hybrid modeling apparatus for economizer parameter identification based on Modelica, according to an exemplary embodiment. (Refer to...) Figure 3 The device includes:

[0125] Modeling module 1 is used to model the heat transfer process of flue gas and water in the economizer and the energy-momentum balance of the two streams using Modelica, so as to obtain the mechanism model;

[0126] The first calculation module 2 is used to calculate the ideal feedwater flow resistance coefficient using the first field data, which includes the inlet and outlet flue gas pressure drop of the economizer, feedwater flow rate, feedwater temperature, outlet water temperature, and the height difference between the inlet and outlet of the economizer.

[0127] The second calculation module 3 is used to standardize the second field data and then use PCC and MI methods to weight and calculate the input features of the DNN model; the second field data includes furnace coal quantity, feedwater flow rate, temperature, and pressure, and inlet flue gas temperature, pressure, and flow rate.

[0128] Training optimization module 4 is used to train the DNN model using the input features of the DNN model and the ideal water supply flow resistance coefficient as output features, and to fine-tune the DNN model structure and hyperparameters using Bayesian optimization.

[0129] Export module 5 is used to export the weights, biases, and activation functions of neurons in each layer of the optimized DNN model;

[0130] The writing module 6 is used to write the weights, biases, and activation functions into the mechanism model. A new function is created in the mechanism model, and multiple parameters are set to store the weight and bias matrix, the mean and variance of the 6-dimensional input features.

[0131] Reconstruction Module 7 is used to reconstruct the mapping relationships between neurons in each layer of the DNN model using the Modelica language in a newly created function;

[0132] Module 8 is used to call the newly created function through the mechanism model to identify the parameters of the water supply flow resistance coefficient using the DNN model in the mechanism model.

[0133] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0134] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0135] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the above-described hybrid modeling method for economizer parameter identification based on Modelica.

[0136] Accordingly, this application also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the above-described hybrid modeling method for economizer parameter identification based on Modelica.

[0137] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0138] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A hybrid modeling method for economizer parameter identification based on Modelica, characterized in that, include: Modelica was used to model the heat transfer process between flue gas and water in the economizer and the energy-momentum balance of the two streams, resulting in a mechanistic model. The ideal feedwater flow resistance coefficient is calculated using first field data, which includes the inlet and outlet flue gas pressure drop of the economizer, feedwater flow rate, feedwater temperature, outlet water temperature, and the height difference between the inlet and outlet of the economizer. The second field data is standardized, and then the input features of the DNN model are calculated using both PCC and MI methods. The second field data includes furnace coal quantity, feedwater flow rate, temperature, and pressure, and inlet flue gas temperature, pressure, and flow rate; the PCC is the Pearson correlation coefficient, and MI is the coefficient for calculating the joint distribution of x and y. With marginal distribution The relative entropy; the weighted calculation formula is as follows: , This represents the evaluation score of the nth input feature. This represents the partial correlation coefficient between the nth input feature and the output after controlling for all other input features. The last term represents the input feature. The ratio of mutual information with Y to the entropy of Y itself; The DNN model is trained using the input features of the DNN model and the ideal water supply flow resistance coefficient as output features, and Bayesian optimization is used to fine-tune the DNN model structure and hyperparameters. Derive the weights, biases, and activation functions of neurons in each layer of the optimized DNN model; Write the weights, biases, and activation functions into the mechanism model. Create a new function in the mechanism model and set multiple parameters to store the weight and bias matrices, the mean and variance of the 6-dimensional input features. In the newly created function, use the Modelica language to reconstruct the mapping relationships between neurons in each layer of the DNN model; By calling the newly created function through the aforementioned mechanism model, the parameter identification of the water supply flow resistance coefficient can be achieved using the DNN model within the mechanism model. The reconstructed mapping relationship is as follows: ； ； ； This represents the standardized result of the model input. This represents the mapping result of the nth layer neuron. This represents the weight matrix of the neurons in the nth layer; Let be the bias matrix of the neurons in the nth layer. This represents the final mapping result, i.e., the predicted result of the required feedwater flow resistance coefficient, where... ReLU means comparing each element in matrix A with 0 and taking the larger value, which is the calculation method of ReLU. The last output layer does not have ReLU, so there is only a linear mapping. ReLU is the activation function.

2. The method according to claim 1, characterized in that, Equations for heat transfer and momentum balance of flue gas and water in the economizer were constructed using Modelica, including: Two streams, flue gas and water, are obtained and denoted by the subscripts g and w, respectively. The three parameters of temperature, pressure, and flow rate for the two streams are denoted by T, p, and qm, respectively. The input and output streams are distinguished by the subscripts i and o. The heat transfer equations for flue gas and water are constructed as follows: ； ； in, The average temperature difference between the two streams. For heat transfer power, coefficient Fitting data using training data; Construct the momentum balance equation for flue gas: ； in, To maintain a constant voltage drop, The flow resistance coefficient of the flue gas. The density of the flue gas; The momentum balance equation of water: ； in, The height difference between the inlet and outlet of the economizer's flow stream; Construct the energy balance equation: ； in, For the internal volume of the economizer, This refers to the specific heat of the flue gas.

3. The method according to claim 1, characterized in that, A DNN model is used to identify the parameters of the feedwater flow resistance coefficient in the mechanistic model. Standardized training data is input. The DNN model consists of an input layer, hidden layers, and an output layer. The input feature dimension of the DNN model is 6, including the furnace coal quantity, economizer feedwater flow rate and pressure, and inlet flue gas temperature, pressure, and flow rate. The output feature dimension is 1, which is the feedwater flow resistance coefficient. The number of intermediate hidden layers and the number of neurons in each layer are obtained through Bayesian optimization. In the DNN model, the data is passed from the input layer to the hidden layer, and then through all the hidden layers to finally reach the output layer, completing one forward pass. Each neuron receives the output from the previous layer, and the weighted sum is then passed through the ReLU activation function to obtain the output result.

4. A hybrid modeling device for economizer parameter identification based on Modelica, characterized in that, include: The modeling module is used to model the heat transfer process of flue gas and water in the economizer and the energy-momentum balance of the two streams using Modelica, so as to obtain the mechanism model. The first calculation module is used to calculate the ideal feedwater flow resistance coefficient using the first field data, which includes the inlet and outlet flue gas pressure drop of the economizer, feedwater flow rate, feedwater temperature, outlet water temperature, and the height difference between the inlet and outlet of the economizer. The second calculation module is used to standardize the second field data and then use PCC and MI methods to weight and calculate the input features of the DNN model. The second field data includes furnace coal quantity, feedwater flow rate, temperature, and pressure, and inlet flue gas temperature, pressure, and flow rate; the PCC is the Pearson correlation coefficient, and MI is the coefficient for calculating the joint distribution of x and y. With marginal distribution The relative entropy; the weighted calculation formula is as follows: , This represents the evaluation score of the nth input feature. This represents the partial correlation coefficient between the nth input feature and the output after controlling for all other input features. The last term represents the input feature. The ratio of mutual information with Y to the entropy of Y itself; The training optimization module is used to train the DNN model using the input features of the DNN model and the ideal water supply flow resistance coefficient as output features, and to fine-tune the DNN model structure and hyperparameters using Bayesian optimization. The export module is used to export the weights, biases, and activation functions of neurons in each layer of the optimized DNN model. The writing module is used to write the weights, biases, and activation functions into the mechanism model. A new function is created in the mechanism model, and multiple parameters are set to store the weight and bias matrices, the mean and variance of the 6-dimensional input features. The reconstruction module is used to reconstruct the mapping relationships between neurons in each layer of a DNN model using the Modelica language in a newly created function; The calling module is used to call the newly created function through the mechanism model to realize the parameter identification of the water supply flow resistance coefficient using the DNN model in the mechanism model; The reconstructed mapping relationship is as follows: ； ； ； This represents the standardized result of the model input. This represents the mapping result of the nth layer neuron. This represents the weight matrix of the neurons in the nth layer; Let be the bias matrix of the neurons in the nth layer. This represents the final mapping result, i.e., the predicted result of the required feedwater flow resistance coefficient, where... ReLU means comparing each element in matrix A with 0 and taking the larger value, which is the calculation method of ReLU. The last output layer does not have ReLU, so there is only a linear mapping. ReLU is the activation function.

5. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-3.

6. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-3.

Images