Denoising diffusion probability model-based coal-fired unit digital twin modeling method and device and storage medium

Abstract

The invention provides a coal-fired unit digital twin modeling method and device based on a denoising diffusion probability model and a storage medium, and belongs to the field of digital twin modeling. The problems that an existing method is large in modeling difficulty and high in model complexity, and mechanism simplification and data quality dependence cannot be avoided are solved. Comprising the following steps: data preprocessing: selecting equipment parameters highly related to operation of a coal-fired unit as diffusion characteristics, selecting diffusion indexes according to the diffusion characteristics, and carrying out normalization processing on the diffusion characteristics and the diffusion indexes; on the basis of the probability denoising diffusion model, a lightweight MobileNet is adopted to replace a residual block in a UNet network, and a lighter and faster industrial digital twinning denoising diffusion probability model DT-DDPM is obtained; guiding a sampling process by adopting diffusion characteristics; carrying out high-fidelity digital twinborn modeling on the coal-fired unit by adopting an industrial digital twinborn denoising diffusion probability model and historical data of operation of the coal-fired unit; the method is applied to coal-fired unit digital twin modeling.

Description

Technical Field

[0001] The present application relates to the technical field of digital twin modeling, and in particular to a method, device and storage medium for digital twin modeling of a coal-fired unit based on a denoised diffusion probability model. Background Art

[0002] The power energy structure is gradually shifting from fossil fuels to new energy sources, meeting the new requirements of low carbon, energy conservation, and high efficiency. However, due to the uncertainty of renewable energy power supply, thermal power generation, represented by coal-fired units, still plays a vital role in the energy system, placing higher demands on deep peak-shaving and rapid load response capabilities. Intelligentization and digitalization have become the future trend in the low-carbon transformation of coal-fired power plants. However, coal-fired units suffer from high inertia, large delays, nonlinearity, and strong coupling, leading to poor modeling accuracy, slow computational speed, and insufficient robustness. This makes it difficult to achieve effective engineering applications for intelligent services such as fault diagnosis, predictive maintenance, and coordinated optimization. Therefore, high-fidelity modeling of coal-fired units has become a key technical challenge that urgently needs to be addressed. In recent years, digital twin technology, as a key technology in the digital transformation of coal-fired units, has attracted widespread attention from both academics and engineers, providing a new solution for high-fidelity modeling of coal-fired units.

[0003] At present, there are three main methods for digital twin modeling of coal-fired units at home and abroad: mechanism modeling, data-driven modeling, and hybrid-driven modeling. (1) Mechanism modeling has good interpretability and effectively reflects the physical characteristics of coal-fired units. However, due to the complexity of coal-fired units, it is inevitable to simplify the mechanism, resulting in low modeling accuracy and difficulty in meeting the high-fidelity requirements of digital twin models. (2) Data-driven modeling uses historical data and adopts artificial neural networks to fit the complex nonlinear relationships of coal-fired units. It has achieved good application results in energy consumption optimization, fault diagnosis, performance monitoring, intelligent decision-making and other fields. However, the data-driven modeling method relies on data quality and lacks interpretability. In the case of insufficient data or noisy data, the modeling accuracy will be reduced, affecting the generalization ability and robustness of the model. (3) Hybrid-driven modeling makes up for the shortcomings of the above two methods to a certain extent and integrates the advantages of mechanism and data. However, this method is difficult to model and has high model complexity. It still cannot avoid problems such as mechanism simplification and data quality dependence.

[0004] As the most advanced technology currently, generative artificial intelligence (AIGC), represented by GAN, CLIP, Transformer, and Diffusion, has achieved remarkable results in areas such as text image generation and cross-modal learning. It provides a new solution for digital twin modeling in the industrial field and effectively addresses the shortcomings of the above three types of digital twin modeling. Summary of the Invention

[0005] In order to solve the above technical problems, this application proposes a digital twin modeling method, device and storage medium for coal-fired units based on a denoising diffusion probability model.

[0006] The technical solution adopted in this application is: a coal-fired unit digital twin modeling method based on a denoised diffusion probability model, comprising the following steps:

[0007] S1: Data preprocessing: Equipment parameters that are highly correlated with the operation of coal-fired units are selected as diffusion features, diffusion indicators are selected based on the diffusion features, and the diffusion features and diffusion indicators are normalized;

[0008] S2: Basic model construction: Build a lighter and faster industrial digital twin denoising diffusion probability model DT-DDPM, specifically:

[0009] Constructing the MobileNet kernel: Based on the probabilistic denoising diffusion model, a lightweight MobileNet is used to replace the residual block in the UNet network to obtain an improved UNet and named it DT-UNet;

[0010] S3: using diffusion features to guide the sampling process;

[0011] S4: Use the industrial digital twin denoising diffusion probability model and historical data of coal-fired unit operation to build high-fidelity digital twin modeling for coal-fired units.

[0012] Furthermore, the equipment parameters highly relevant to the operation of the coal-fired unit selected in step S1 include unit load, coal supply, water supply, air supply and nitrogen oxide emissions. The above five groups of characteristic data are separated from the coal-fired unit operation data set to obtain a diffusion characteristic data set.

[0013] Furthermore, the Spearman rank correlation coefficient method was used to calculate the Spearman rank correlation coefficient between the operating variables and diffusion characteristics of the coal-fired units. The operating variables with the absolute value of the Spearman rank correlation coefficient greater than or equal to 0.5 and less than or equal to 1 were selected to construct the diffusion index data set.

[0014] Furthermore, the data in the diffusion feature dataset and the diffusion index dataset are processed using a maximum and minimum value normalization method based on a threshold interval to obtain a training set and a test set corresponding to the diffusion feature and the diffusion index.

[0015] Furthermore, the activation function in the DT-UNet network adopts the Mish function.

[0016] Furthermore, step S3 includes:

[0017] S31: Design a diffusion feature embedding layer in DT-UNet to map the diffusion features to a dimension that matches the feature dimension of the upsampling stage;

[0018] S32: The diffusion feature processed by the diffusion feature embedding layer at time i and its corresponding diffusion index are subjected to t-times noise addition, and the diffusion feature processed by the diffusion feature embedding layer is added to the reverse diffusion process to guide the stepwise denoising;

[0019] S33: Training the DT-DDPM model. During the generation process, the trained DT-DDPM model samples diffusion indicators from the standard normal distribution and performs iterative updates under the guidance of the diffusion characteristics, thereby generating index data of the coal-fired unit under the diffusion characteristics.

[0020] Furthermore, step S4 includes:

[0021] S41: Model training process: The data format of the diffusion index training dataset is converted into a one-dimensional matrix form, and then input into the DT-DDPM model for training, and the diffusion feature training dataset is used to perform the diffusion feature guided sampling process;

[0022] S42: Model testing process: Assume that the diffusion feature test data set has n diffusion features and randomly generate n Gaussian noise matrix sets;

[0023] The Gaussian noise matrix set is input into the model sampling process, and the model output result is obtained by gradually denoising under the guidance of the diffusion feature test data set;

[0024] S43: Accuracy evaluation of sampling results: Different indicators are used to evaluate the fit of the DT-DDPM model;

[0025] S44: Sampling generation speed evaluation: Multiple indicators are used to evaluate the response speed of the DT-DDPM model.

[0026] Furthermore, MobileNet contains a depthwise convolution and a pointwise convolution, where the depthwise convolution uses a 3×3 two-dimensional convolution layer and the pointwise convolution uses a two-dimensional convolution layer with a convolution kernel size of 1.

[0027] A computer device comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0028] A computer-readable storage medium stores a computer program / instruction thereon, which implements the steps of the method when executed by a processor.

[0029] The beneficial effects of this application compared to the existing technology are as follows: This application designs a faster and lighter industrial digital twin denoising diffusion probability model, which reduces the complexity of the digital twin model while ensuring modeling accuracy, improves the response speed, and enhances the reliability and robustness of the digital twin model of the coal-fired unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The present application will be further described below with reference to the accompanying drawings:

[0031] Figure 1 An overall flow chart of the modeling method provided in the embodiments of the present application;

[0032] Figure 2 A flowchart of data preprocessing provided in an embodiment of the present application;

[0033] Figure 3 The DT-DDPM model and DT-UNet network structure diagram provided in the embodiments of this application;

[0034] Figure 4 A flowchart of the training and testing of a digital twin model for a coal-fired unit provided in an embodiment of the present application;

[0035] Figure 5 This is a graph showing the accuracy evaluation results of the digital twin model of a coal-fired unit provided in an embodiment of the present application;

[0036] Figure 6 The evaluation results of the generation speed of the digital twin model of the coal-fired unit provided in the embodiment of this application are as follows: DETAILED DESCRIPTION

[0037] like Figures 1 to 6 As shown, based on the operating characteristics of coal-fired units, this application proposes a digital twin modeling method for coal-fired units based on a denoising diffusion probability model, which reduces the complexity of the digital twin model while ensuring the modeling accuracy, and enhances the robustness and generalization ability of the digital twin.

[0038] like Figure 1 As shown, the modeling method of this application includes the following steps:

[0039] S1: Data preprocessing: Data preprocessing is performed using diffusion feature and index extraction, and threshold maximum and minimum value normalization methods to scale the data to the same dimension and provide data support for digital twin modeling of coal-fired units.

[0040] The specific steps of data preprocessing include:

[0041] S11: Extracting diffusion features : After expert consultation, the unit load was selected ( ), coal supply ( ), water supply ( ), air volume ( )as well as Emissions are characterized as diffusion. Its mathematical description is as follows:

[0042] Assume that the coal-fired boiler operation data set is :

[0043] ;

[0044] At the same time, 、 、 、 and for The 5 columns in the table have the following indexes: , , , and .but . Diffusion feature dataset For each eigenvector A collection of . Among them, It can be expressed as ,in is the coal supply, is the water supply, To provide air volume, is the unit load, is the nitrogen oxide emission, then It can be expressed as:

[0045] .

[0046] S12: Extracting diffusion indicators :The Spearman rank correlation coefficient method is used to calculate the Spearman rank correlation coefficient between the variables and the diffusion characteristics to construct a set of diffusion indicators. The mathematical description is as follows:

[0047] Assume that the coal-fired boiler operation data set is , the diffusion feature dataset is , then the remaining data sets yes In the dataset The complement of . Definition The column index is .Will and Do the Spearman calculation:

[0048] ;

[0049] ;

[0050] in, express The i-th column and The Spearman correlation coefficient, , is the diffusion indicator dataset, for In the dataset The complement of .when When it is between 0.5 and 1, the data in this column has a strong positive / negative correlation with the diffusion characteristics and is suitable as a diffusion indicator for the digital twin of a coal-fired unit.

[0051] S13: Normalization: A threshold-based maximum and minimum normalization method is used to process the diffusion index and diffusion characteristics to construct the data set required for modeling. The mathematical description is as follows:

[0052] Assume data set , The column index is . yes The threshold of column i, is the lower threshold, is the upper threshold. Through the following formula, After normalization, we get :

[0053] ;

[0054] Finally, the training set and , test set and .

[0055] S2: Basic model construction: Construct a lighter and faster industrial digital twin denoising diffusion probability model (called: DT-DDPM), and design a faster and lighter UNet network (DT-UNet) suitable for digital twin modeling, which improves the algorithm operation efficiency.

[0056] The construction steps of DT-DDPM include:

[0057] S21: Constructing the MobileNet Kernel: Based on the probabilistic denoising diffusion model, a lightweight MobileNet was designed to replace the convolutional module in the UNet network. This model consists of one depthwise convolution and one pointwise convolution. The depthwise convolution uses a 3×3 two-dimensional convolution layer, while the pointwise convolution uses a two-dimensional convolution layer with a kernel size of 1. The newly designed UNet is named DT-UNet, and the probabilistic denoising diffusion model using DT-UNet as the kernel is called DT-DDPM.

[0058] S22: Select the Mish function as the activation function: Replace the activation function Relu in DDPM with the Mish function to make the model more nonlinear and improve the interpretability and gradient flow of the network.

[0059] S3: Diffusion feature-guided sampling process: A diffusion feature-guided sampling process mechanism is designed to enhance modeling accuracy and robustness.

[0060] The steps of diffusion feature guided sampling include:

[0061] S31: Design diffusion feature embedding layer: a linear fully connected layer is used to embed the diffusion features To process, use the formula Maps input features to output features of a specific shape.

[0062] in, is a linear fully connected layer, is the output feature, with shape and size , is the input feature, the shape and size are ; is the number of channels of DT-UNet; is the batch size; is the number of diffusion features; is the output feature size.

[0063] S32: Diffusion feature embedding: Assume diffusion feature dataset and Diffusion feature after diffusion feature embedding layer processing at time i , and its corresponding diffusion index After t times of noise addition, The status from Became During the reverse diffusion process, adding To guide arrive The mathematical description of the diffusion feature embedding mechanism in the model training process is as follows:

[0064] ;

[0065] ;

[0066] in, is the conditional probability distribution, is a normal distribution, is the mean function, is the covariance function, is the variance scheduling ratio, is the cumulative variance scheduling ratio, For noise prediction network.

[0067] DT-UNet is used to minimize the difference between predicted noise and true noise. and use As a loss function to train the digital twin model;

[0068] in For expectations, is the noise vector, is a multidimensional Gaussian distribution with a mean of 0 and a variance of 1. is the identity matrix.

[0069] The trained digital twin model is sampled from the standard normal distribution during the generation process. , in a specific Under the guidance of The following coal-fired unit index data:

[0070]

[0071] in, For the model at time step The denoised data, is the current time step of noisy data.

[0072] S4: Use the industrial digital twin denoising diffusion probability model and historical data of coal-fired units to build a high-fidelity digital twin model of the coal-fired unit. The specific steps include:

[0073] S41: Model training process: The data format is converted into a one-dimensional matrix, and then input into the DT-DDPM model for training, and used Perform a diffusion signature-guided sampling process.

[0074] S42: Model testing process: There are n diffusion features, and n Gaussian noise matrix sets are randomly generated .

[0075] Will Input model sampling process, in Under the guidance of , the model output is obtained by step-by-step denoising. .

[0076] S43: Accuracy evaluation of sampling results: The convergence loss value (CL), mean square error (MSE), mean absolute error (MAE) and R2 determination coefficient (R-Square) indicators are used to evaluate the degree of fit of the digital twin model.

[0077] S44: Sampling generation speed evaluation: Four indicators, namely average training iteration time (ATIT), batch generation speed (BGS), single feature sampling speed (SSS), and number of model parameters (NoP), are used to evaluate the response speed of the digital twin model.

[0078] The digital twin model constructed in this application can be used in the Pytorch GPU environment. The server hardware configuration requirements are no less than: i9-13900HX (2.2GHZ) processor, NVIDA RTX4080 (12G) GPU, and 32G memory.

[0079] In this example, a 600MW coal-fired unit was used as the research object. The operating data set recorded key equipment status parameters, including turbine power, economizer, steam drum, low-temperature superheater, platen superheater, high-temperature superheater, condenser, low-pressure heater, high-pressure heater, high-pressure cylinder, intermediate-pressure cylinder, and low-pressure cylinder, during the unit's operation with a load reduction from 100% to 40%. This application is further explained below with reference to the aforementioned 600MW coal-fired unit.

[0080] S1: Data preprocessing: Data preprocessing is performed using diffusion feature and index extraction, and threshold maximum and minimum value normalization methods to scale the data to the same dimension and provide data support for digital twin modeling of coal-fired units.

[0081] The specific implementation steps of S1 are as follows:

[0082] S11: Extracting diffusion features : After analyzing the operating data of the 600MW coal-fired unit and combining expert consultation, "turbine power", "steam drum feed water flow", "pulverized coal flow of pulverizer", "total air supply volume of secondary fan" and "nitrogen oxide emission" were selected as diffusion characteristics.

[0083] S12: Spearman correlation analysis was performed using the five diffusion features extracted from S11 and other measurement point data of the coal-fired unit. 64 data measurement point values ​​greater than 0.5 (such as high reabsorption heat, high reabsorption flue gas outlet temperature, high reabsorption flue gas average temperature, feedwater pump outlet water pressure, feedwater pump valve opening, feedwater pump power, etc.) were selected as diffusion indicators.

[0084] S13: Normalization: A maximum and minimum normalization method based on threshold interval is used to normalize the five diffusion features of S11 and the 64 diffusion indicators of S12.

[0085] The normalized data is randomly divided into 7:3 to obtain the training set and test set Among them, 5 diffusion features constitute the diffusion features of the training set and test set diffusion characteristics ; 64 diffusion indicators constitute the training set diffusion indicators and test set diffusion index

[0086] S2: Basic model construction: Construct a lighter and faster industrial digital twin denoising diffusion probability model (called: DT-DDPM), and design a faster and lighter UNet network (DT-UNet) suitable for digital twin modeling, which improves the algorithm operation efficiency.

[0087] The specific implementation steps of S2 are as follows:

[0088] The model proposed by S21 is used: the model parameters are set as follows: (1) Batch size =256; (2) learning rate lr = 1e-4; (3) number of channels =32; (4) Optimizer: Adam optimizer; (5) Beta range: 1e-4 to 0.02.

[0089] S3: Diffusion feature-guided sampling process: A diffusion feature-guided sampling process mechanism is designed to enhance modeling accuracy and robustness.

[0090] The specific implementation steps of S3 are as follows:

[0091] Diffusion feature number It is set to 5, and the five diffusion features "turbine power", "steam drum feed water flow", "pulverized coal flow of pulverizer", "total air supply of secondary fan" and "nitrogen oxide emission" are used as guiding inputs to the diffusion feature embedding layer.

[0092] S4: Use the industrial digital twin denoising diffusion probability model and historical data of unit operation to build high-assurance digital twin modeling for coal-fired units.

[0093] S41: Model training process: The data format is converted from [1,64] to a one-dimensional matrix form [1,8,8], and then input into the DT-DDPM model for training, and used Perform a diffusion signature-guided sampling process.

[0094] S42: Model testing process: There are n diffusion features, and n [1,8,8] Gaussian noise matrix sets are randomly generated .Will Input model sampling process, in Under the guidance of , the model output is obtained by step-by-step denoising. .

[0095] S43: Use convergence loss (CL), mean square error (MSE), mean absolute error (MAE) and R2 coefficient of determination (R-Square) to generate and Comparison is made to evaluate the fitting degree of the digital twin model of the coal-fired unit.

[0096] S44: The response speed of the digital twin model of the coal-fired unit is evaluated by four indicators: Average Training Iteration Time (ATIT), Batch Generation Speed ​​(BGS), Single Sampling Speed ​​(SSS), and Number of Model Parameters (NoP).

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A coal-fired unit digital twin modeling method based on a denoising diffusion probability model, characterized by: The following steps are involved: S1: Data preprocessing: Select equipment parameters that are highly relevant to the operation of coal-fired units as diffusion features, select diffusion indicators based on the diffusion features, and normalize the diffusion features and diffusion indicators; S2: Basic model construction: Build a lighter and faster industrial digital twin denoising diffusion probability model DT-DDPM, specifically: Constructing the MobileNet kernel: Based on the probabilistic denoising diffusion model, a lightweight MobileNet is used to replace the residual block in the UNet network to obtain an improved UNet and named it DT-UNet; S3: using diffusion features to guide the sampling process; S4: Use the industrial digital twin denoising diffusion probability model and historical data of coal-fired unit operation to build high-fidelity digital twin modeling for coal-fired units.

2. The method for digital twin modeling of a coal-fired unit based on a denoised diffusion probability model according to claim 1, characterized in that: The equipment parameters highly relevant to the operation of the coal-fired unit selected in step S1 include unit load, coal feed rate, water feed rate, air feed rate and nitrogen oxide emissions. The above five groups of characteristic data are separated from the coal-fired unit operation data set to obtain a diffusion characteristic data set.

3. The method for digital twin modeling of a coal-fired unit based on a denoised diffusion probability model according to claim 2, characterized in that: The Spearman rank correlation coefficient method was used to calculate the Spearman rank correlation coefficient between the operating variables and diffusion characteristics of the coal-fired units. The operating variables with the absolute value of the Spearman rank correlation coefficient greater than or equal to 0.5 and less than or equal to 1 were selected to construct the diffusion index data set.

4. The method for digital twin modeling of a coal-fired unit based on a denoised diffusion probability model according to claim 3 is characterized by: The data in the diffusion feature dataset and the diffusion index dataset are processed using the maximum and minimum value normalization method based on the threshold interval to obtain the training set and test set corresponding to the diffusion feature and diffusion index.

5. The method for digital twin modeling of a coal-fired unit based on a denoised diffusion probability model according to claim 1, characterized in that: The activation function in the DT-UNet network uses the Mish function.

6. The method for digital twin modeling of a coal-fired unit based on a denoised diffusion probability model according to claim 1, characterized in that: Step S3 includes: S31: Design a diffusion feature embedding layer in DT-UNet to map the diffusion features to a dimension that matches the feature dimension of the upsampling stage; S32: The diffusion feature processed by the diffusion feature embedding layer at time i and its corresponding diffusion index are subjected to t-times noise addition, and the diffusion feature processed by the diffusion feature embedding layer is added to the reverse diffusion process to guide the stepwise denoising; S33: Training the DT-DDPM model. During the generation process, the trained DT-DDPM model samples diffusion indicators from the standard normal distribution and performs iterative updates under the guidance of the diffusion characteristics, thereby generating index data of the coal-fired unit under the diffusion characteristics.

7. The method for digital twin modeling of a coal-fired unit based on a denoised diffusion probability model according to claim 1, characterized in that: Step S4 includes: S41: Model training process: The data format of the diffusion index training dataset is converted into a one-dimensional matrix form, and then input into the DT-DDPM model for training, and the diffusion feature training dataset is used to perform the diffusion feature guided sampling process; S42: Model testing process: Assume that the diffusion feature test data set has n diffusion features and randomly generate n Gaussian noise matrix sets; The Gaussian noise matrix set is input into the model sampling process, and the model output result is obtained by gradually denoising under the guidance of the diffusion feature test data set; S43: Accuracy evaluation of sampling results: Different indicators are used to evaluate the fit of the DT-DDPM model; S44: Sampling generation speed evaluation: Multiple indicators are used to evaluate the response speed of the DT-DDPM model.

8. The method for digital twin modeling of a coal-fired unit based on a denoised diffusion probability model according to claim 1, characterized in that: MobileNet contains a depthwise convolution and a pointwise convolution, where the depthwise convolution uses a 3×3 two-dimensional convolution layer and the pointwise convolution uses a two-dimensional convolution layer with a convolution kernel size of 1.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images