A Method and System for Quantitative Characterization of Ash and Fouling on Heating Surfaces of Power Plant Boilers Based on Transfer Learning

Abstract

This invention belongs to the field of ash and fouling monitoring technology, and discloses a quantitative characterization method for ash and fouling on boiler heating surfaces based on transfer learning. This invention utilizes simulation software to build a boiler-furnace coupling model of the boiler under study, and employs a method combining simulation modeling and actual verification to study the quantitative characterization method for ash and fouling on boiler heating surfaces based on transfer learning. By utilizing a model and ash and fouling labels established and validated in simulation data, the modeling is transferred to the same actual heating surface for quantitative characterization modeling, reducing modeling difficulty, increasing modeling speed, requiring less data, and effectively fitting the ash accumulation trend of the actual heating surface. This effectively solves the problem that current boiler unit actual operating data lacks characterization labels for ash and fouling, which cannot provide good guidance for subsequent soot blowing optimization, and has certain application prospects.

Description

Technical Field

[0001] This invention belongs to the field of ash and fouling monitoring technology for heated surfaces, and particularly relates to a quantitative characterization and modeling method and system for ash and fouling on boiler heated surfaces based on transfer learning. Background Technology

[0002] Under the "dual carbon" context, energy conservation and pollution reduction are urgent priorities for all countries. Over the past decade, the proportion of installed capacity from new energy sources has continued to increase, but their intermittent, random, and fluctuating characteristics mean that the power support capacity of new energy sources is still insufficient. Therefore, coal-fired power generation, as the "ballast" for ensuring my country's power security, will continue to bear the heavy burden of basic security for a long time to come. Currently, coal-fired power generation is gradually shifting from a basic power source supporting basic loads to a regulating power source supporting the smoothing of new energy fluctuations and ensuring power supply security. Flexible operation of coal-fired power plants is imperative. Because coal contains a large amount of ash, the flue gas produced during combustion carries ash and comes into contact with the outer surfaces of different heating surfaces of the boiler. Over time, this ash gradually adheres and deposits on these surfaces, making ash accumulation and slagging unavoidable during the operation of coal-fired power plant boilers. This is mainly related to boiler structure, load, and coal quality. For a specific boiler, the coal feed rate, internal boiler temperature level, and flue gas velocity vary under different loads, thus affecting ash and slagging on the heating surface. Different coal types have different ash content and composition, resulting in significant differences in their ash accumulation and slagging characteristics. The flexible nature of coal-fired power plant boilers, with their flexible coal quality and load, means that traditional timed and quantitative soot blowing strategies do not match the actual soot blowing requirements of the boiler's heating surfaces, leading to either over-blowing or under-blowing. Therefore, there is an urgent need to develop quantitative characterization technology for ash and fouling on heating surfaces to ensure the safety and economy of boiler operation. Based on existing publicly available information, traditional mechanism-based or data-driven methods for ash and fouling monitoring modeling suffer from problems such as complex mechanisms that are difficult to simplify and a lack of labeled data. Furthermore, most of these methods are only applicable to scenarios with basic loads and fixed coal types, exhibiting significant errors in ash and fouling monitoring under flexible operating conditions, leading to the blindness and randomness of soot blowing on heating surfaces. Therefore, it is of great significance to overcome the limitations of traditional mechanism-based and data-driven modeling and establish a high-precision ash and fouling monitoring model.

[0003] Deep learning and transfer learning, as important branches of machine learning, are widely used in anomaly detection and prediction in industrial equipment and monitoring systems. Their powerful learning capabilities can capture the dependencies between time series data and the nonlinear expressions between feature parameters. However, achieving good results with deep learning algorithms often requires strong, high-quality data. For industrial equipment operation and monitoring systems, data acquisition is time-consuming and costly. For tasks requiring fixed parameters like anomaly monitoring, anomalies and malfunctions in real-world scenarios are often few-sample events with limited data labels, thus limiting the application of deep learning algorithms. While unsupervised learning, as a branch of deep learning, has been widely applied in scenarios lacking labeled data, learning hidden representations through dimensionality reduction and feature extraction has shown good results, it requires remodeling and parameter optimization to adapt to different learning tasks when dealing with different research objects. Furthermore, the massive amounts of data and complex, variable operating conditions limit the applicability and generalization of deep learning modeling methods. Transfer learning, on the other hand, leverages the concept of knowledge transfer to apply feature relationships learned in one domain to similar domains, further addressing the challenges of small sample sizes and dataset bias in real-world industrial scenarios. Analysis of publicly available data reveals that current applications of anomaly detection and diagnosis in power plant boilers are primarily concentrated in equipment such as coal mills and pulverizing systems, with limited application in other equipment. In actual power plant operation, the slow deposition of ash-laden flue gas on different heating surfaces after pulverized coal combustion in the furnace can lead to decreased heat transfer efficiency and safety hazards, exhibiting similarities to the deterioration and malfunctions of industrial equipment. Therefore, leveraging the advantages of transfer learning methods in addressing practical problems such as lack of labels and dataset bias, we can explore new approaches for high-precision quantitative characterization and monitoring modeling of ash contamination on boiler heating surfaces. Summary of the Invention

[0004] To address the challenges of complex simplification and lack of labeled data in traditional mechanism-based or data-driven methods for ash pollution monitoring modeling, this invention provides a transfer learning-based method for quantitative characterization of ash pollution on boiler heating surfaces. The method outputs dynamic operating data of the boiler heating surface and labeled data representing the degree of ash pollution through a dynamic simulation model. Following a transfer learning process of "model pre-training - model transfer - fine-tuning strategy - model verification," a pre-model for quantitative characterization of ash pollution on the heating surface based on simulation data is built and verified. The model is then transferred to actual boiler heating surfaces, demonstrating that the transfer learning model can effectively fit the ash pollution trend within the soot blowing cycle of the boiler heating surface in actual operating data.

[0005] This invention is implemented as follows: a method for quantitative characterization and modeling of ash and fouling on boiler heating surfaces based on transfer learning, comprising:

[0006] S1: Based on the boiler dynamic simulation model, obtain full-load, full-condition simulation operation data when the ash and dirt coefficients of each heating surface change;

[0007] S2: Select the operating dataset S1 of the heating surface in the simulation model as the source domain, and select the operating dataset of the corresponding heating surface in the actual coal-fired power plant boiler as the target domain.

[0008] S3: Based on machine learning algorithms such as autoencoder and support vector regression, a quantitative characterization model of boiler heating surface ash and pollution in the source domain is built and the parameters are adjusted, optimized and trained to obtain the pre-trained model with the best fitting result of heating surface ash and pollution.

[0009] S4: Transfer the pre-trained model from S3 to the target domain dataset, and use the method of fine-tuning the model parameters to make the model adapt to the ash and pollution quantification characterization task under the same boiler heating surface actual operating data.

[0010] S5: Utilize actual soot blowing actions to determine the ash accumulation trend and degree of ash contamination on the heated surface during different soot blowing cycles, and complete the verification and application of the transfer learning model.

[0011] Furthermore, the acquisition of full-load, full-condition operating data under the ash and dirt coefficients of each heating surface in S1 is conducted using a boiler dynamic simulation model. In the module attribute settings, the change in the ash and dirt coefficient of a specific heating surface is used as an indicator feature. The range is set according to the specific contamination situation of the object, generally between 0.6 and 1. A coefficient of 1 (100%) represents a clean heating surface. As the coefficient decreases, it indicates that the cleanliness of the heating surface begins to decline due to ash accumulation. When the coefficient rises back to 1, it indicates that the heating surface has recovered its cleanliness due to soot blowing. The simulation data sampling interval is 10 seconds to 1 minute. For each heating surface, original simulation datasets are generated using multiple parameters, including load, coal feed rate, working fluid flow rate, working fluid temperature, working fluid pressure, and flue gas temperature and pressure. The selection of feature data needs to be combined with the actual boiler measurement points for subsequent quantitative characterization of ash and dirt on each heating surface of the boiler.

[0012] Furthermore, the dataset partitioning in S2 is based on the fundamental concepts of transfer learning. Transfer learning utilizes the idea of ​​knowledge transfer to apply feature relationships learned in one domain to similar domains, dividing the dataset into source domains D. s (Source domain) and target domain D t(Target domain) The source domain data is generally labeled and of high quality. Valuable knowledge, such as feature representations, is then obtained from the source domain dataset through relevant algorithms. The target domain data, on the other hand, is generally unlabeled and its distribution or features are similar to those of the source domain. Algorithms can be used to transfer similar feature representations from the source domain to the target domain, helping the target domain to perform data prediction or classification and improving the performance of the target task. In this case, the simulation data of the heated surface and the ash pollution coefficient labels are selected as the source domain, and the actual operating data of the corresponding heated surface is selected as the target domain, with relatively consistent input parameters for the model.

[0013] Furthermore, the pre-trained model with the optimal fitting result for the ash contamination of the heated surface built in S3 is specifically constructed as follows: For the source domain data, machine learning algorithms such as unsupervised learning algorithms (Autoencoder, AE) or traditional Support Vector Regression (SVR) are combined. The encoder part of the AE model is used as a feature extractor. Data samples are input into the encoder part for data dimensionality reduction and deep feature extraction. Then, the feature vectors and data labels are input into the SVR model to perform regression. The obtained predicted value of the ash contamination degree of the heated surface is compared with the ash contamination coefficient label of the source domain. Through parameter tuning and optimization, the optimal model framework and parameter settings are obtained, which is represented as the pre-trained model AE-SVR for transfer learning, and subsequently used for model transfer learning for the target domain.

[0014] The autoencoder AE described in S3 includes only the encoding portion:

[0015] Encoder E will input feature vector Transform into latent feature vectors , .

[0016] In the formula, As input features, These are the encoder parameters. The encoder part performs feature extraction and dimensionality reduction on the input data.

[0017] The Support Vector Regression (SVR) described in S3 is based on the idea of ​​mapping sample data to a high-dimensional space through a nonlinear mapping function, and then performing linear regression in that high-dimensional space. By constructing a loss function that minimizes the training set data labels and the model fit values, a target function f(x) is ultimately determined, expressed as:

[0018] .

[0019] In the formula, w is the weight coefficient vector; is the high-dimensional feature space obtained by mapping the input space x; b is the bias.

[0020] Furthermore, in S4, the pre-model from S3 is transferred to the target domain. Since the distribution differences between different datasets during transfer can lead to a deterioration in model performance, deep transfer learning requires different fine-tuning strategies to adapt the parameters or model to the new dataset and improve model performance. For the AE-SVR model, a fine-tuning strategy is adopted that modifies the parameters of the AE encoder to adapt to the target domain data.

[0021] Furthermore, in S5, during the verification of the transfer results of the ash and dirt quantitative characterization model for the heating surface based on actual operating data, the actual boiler unit lacks label values ​​that characterize the degree of ash and dirt on the heating surface. That is, the simulation dataset of the source domain has labels, while the dataset of the target domain does not. Therefore, the AE-SVR model fitting results of each level of heating surface in different soot blowing cycles are judged by the actual soot blowing execution actions to determine whether they reflect the correct ash accumulation trend and degree of ash and dirt, thereby verifying the effectiveness and feasibility of the AE-SVR transfer learning method.

[0022] Another objective of this invention is to provide a quantitative characterization system for boiler heating surface ash fouling based on transfer learning, comprising:

[0023] The simulation data acquisition module, based on a dynamic simulation model with detailed mechanisms, acquires full-load, full-condition simulation operation data when the ash and dirt coefficients of each heating surface change.

[0024] Run the data acquisition module for modeling, verification, and application of quantitative characterization of contamination on heated surfaces;

[0025] The selection module is used to select the operating dataset of the heating surface in the simulation model as the source domain and the operating dataset of the corresponding heating surface in the actual coal-fired power plant boiler as the target domain.

[0026] The training module is used to build a quantitative characterization model of boiler heating surface ash and pollution based on autoencoder and support vector regression algorithms, and to perform parameter tuning, optimization and training to obtain the pre-trained model with the best fitting result of heating surface ash and pollution.

[0027] The transfer module is used to transfer the pre-trained model to the target domain dataset and to fine-tune the model parameters so that the model can adapt to the ash and pollution quantification characterization task under the same boiler heating surface actual operating data.

[0028] The verification module is used to determine the ash accumulation trend and degree of ash contamination on the heated surface during different soot blowing cycles by using actual soot blowing actions, and to complete the verification and application of the transfer learning model.

[0029] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the method for quantitative characterization of ash and fouling on boiler heating surfaces based on transfer learning.

[0030] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for quantitative characterization of ash and fouling on boiler heating surfaces based on transfer learning.

[0031] Another objective of this invention is to provide an information data processing terminal for implementing the boiler heating surface ash and dirt quantitative characterization system based on transfer learning.

[0032] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0033] First, while unsupervised learning methods have been widely applied in scenarios lacking labeled data, they require remodeling when facing different research objects. Large amounts of data and varying operating conditions limit their applicability and generalization. Transfer learning, on the other hand, transfers models trained and validated in the source domain to similar target domains. Under varying data distributions and feature dimensions, it better utilizes source domain knowledge to assist in modeling the target domain, aligning with the similarity of datasets across different heating surfaces in this paper. Therefore, this invention utilizes simulation software to build a dynamic simulation model of the boiler, employing a method combining simulation modeling and actual verification to study a transfer learning-based method for quantitative characterizing ash and dirt accumulation on boiler heating surfaces. By using a model and ash and dirt labels established and validated in simulation data, and transferring them to the same actual heating surfaces for quantitative characterization, the modeling difficulty is reduced, the modeling speed is increased, the required data volume is small, and the ash accumulation trend of the actual heating surfaces can be effectively fitted. This effectively solves the problem of the lack of ash and dirt labels in current boiler unit operating data, which hinders guidance for subsequent soot blowing optimization and has promising application prospects.

[0034] Secondly, traditional mechanism-based and data-driven modeling suffers from problems such as the difficulty in simplifying complex mechanisms, the lack of labeled data, and applicability primarily to scenarios with basic loads and fixed coal types. For example, the commonly used cleanliness factor is affected by rapid load fluctuations, failing to accurately reflect the actual degree of ash pollution and thus unable to provide guidance for subsequent soot blowing optimization. Data-driven modeling, on the other hand, builds predictive models based on mechanistic results, and therefore cannot effectively address the issue of significant biases in monitoring results under flexible operating conditions. Unsupervised learning-based methods are suitable for modeling scenarios lacking labeled data, but they require remodeling for different research objects, and the large amount of data and variable operating conditions limit their applicability and generalization.

[0035] This invention follows a transfer learning process of "model pre-training, model transfer, fine-tuning strategy, and model validation." It transfers a model trained and validated on a source domain (simulation dataset) to a similar target domain (the same actual heated surface). When the data distribution and feature dimensions differ between the two domains, source domain knowledge is used to assist in modeling the target domain. This invention reduces the difficulty of modeling on real-world operational data lacking characterizing ash accumulation labels. By utilizing ash accumulation label and feature representations from the simulation dataset, it helps to effectively fit the ash accumulation trend of the actual operational data and validate the effectiveness of actual ash blowing actions. Furthermore, only a small number of heated surface objects in the simulation data need to be modeled and their relevant feature representations obtained. The heated surface objects in the actual data are modeled using transfer learning, and local or global parameter fine-tuning is used to adapt to the actual objects, further reducing the time and economic costs of modeling and training, and improving the model's applicability and generalization.

[0036] Third, does the technical solution of this invention solve a technical problem that people have long desired to solve but have never been able to successfully solve?

[0037] Under the dual-carbon context, my country's energy system is accelerating its transformation towards "clean, intelligent, flexible, and efficient" power generation. Currently, coal-fired power generation is gradually shifting from a primary power source supporting base loads to a regulating power source. Furthermore, due to fluctuations in the coal market, many power plants adopt coal blending to reduce energy costs, such as using high-alkali coal, which is abundant in both price and reserves. However, this coal is highly susceptible to fouling and slagging. Therefore, with the current situation of flexible unit load and coal quality, the ash and slagging characteristics inside the boiler differ from the past. This often leads to a mismatch between the traditional timed and quantitative soot blowing strategy and the actual soot blowing needs of the heating surfaces, resulting in localized or even global "underblowing" or "overblowing" of the heating surfaces. Therefore, it is necessary to develop high-precision quantitative characterization technology for heating surface ash and fouling to ensure the safety and economy of boiler operation. Traditional mechanism-based and data-driven modeling suffers from several drawbacks, including complex mechanisms that are difficult to simplify, a lack of labeled data, and applicability primarily to scenarios with basic loads and fixed coal types. For example, the commonly used cleanliness factor is affected by rapid load fluctuations, failing to accurately reflect the actual degree of ash pollution and thus unable to provide guidance for subsequent soot blowing optimization. Data-driven modeling, on the other hand, builds predictive models based on mechanistic results, thus failing to adequately address the significant biases in monitoring results under flexible operation. This invention first analyzes the advantages of transfer learning in solving practical problems such as lack of labels and dataset bias in industrial system monitoring. Existing literature shows that the application of transfer learning in power plant boilers is currently concentrated on equipment such as coal mills and pulverizing systems, with limited application in other equipment. In actual power plant operation, due to the large number of boiler heating surfaces, remodeling is required for different research objects. The large amount of data and variable operating conditions limit the applicability and generalization of the models. Furthermore, actual operational data is affected by frequent load fluctuations, environmental noise, and other factors, resulting in low data quality. This leads to low accuracy in ash pollution monitoring models built based on actual operational data, and the time cost of parameter tuning and optimization is high, lacking accurate guidance for soot blowing. Transfer learning, on the other hand, transfers models trained and validated in the source domain to a similar target domain. Under conditions of changing data distribution and feature dimensions, it better utilizes knowledge from the source domain to assist in modeling the target domain. Therefore, transfer learning can be used to explore new approaches for ash pollution quantitative representation. Ash pollution monitoring models built on simulation datasets can be transferred to actual operational datasets to help quantitatively represent ash pollution on the corresponding boiler heating surfaces. This approach offers fast modeling speed, requires less data, and can effectively fit ash pollution trends.

[0038] Fourth, existing technical issues:

[0039] 1. Insufficient accuracy of traditional ash accumulation detection methods: Currently, most ash accumulation monitoring of boiler heating surfaces relies on ash pollution monitoring systems based on simplified convective heat transfer mechanisms. These methods have limited accuracy and cannot accurately quantify the degree of ash pollution. Furthermore, the monitoring results deviate significantly from the actual situation when operating conditions change (variable load, variable coal type), leading to reduced equipment operating efficiency, increased fuel consumption, and high maintenance costs.

[0040] 2. Lack of Labeled Data in Data-Driven Models: The most challenging aspect of characterizing boiler heating surfaces is obtaining labeled data, i.e., the actual degree of fouling on the heating surfaces under all operating conditions. Existing machine learning models often use simplified convective heat transfer mechanisms to generate labeled data from ash monitoring. These models themselves have inherent flaws, and modeling with the data generated by them suffers from inherent deficiencies and low accuracy. Aside from slightly faster computation speed, they have no practical application value.

[0041] 3. Difficulty in synchronizing ash accumulation with soot blowing optimization: Traditional soot blowing methods are difficult to adjust flexibly according to the ash accumulation on the boiler's heating surfaces. They often involve cleaning according to a fixed cycle, leading to over-cleaning or cleaning only when there is excessive ash accumulation. This model cannot make real-time adjustments based on the actual ash accumulation during boiler operation, affecting the overall efficiency and economy of the boiler.

[0042] The technical solution and significant technological advancements of this invention:

[0043] 1. Achieve broad adaptability of the model by combining transfer learning:

[0044] This invention pre-trains a quantitative characterization model for boiler heating surface ash and fouling by combining autoencoders (AE) and support vector regression (SVR), and then performs transfer learning on target domain data, enabling the model to adapt to different boiler operating conditions. Compared to traditional single-model training, this invention's transfer learning model, through data interaction between the source and target domains, achieves efficient application of the model in different power plants or boilers, greatly improving the model's adaptability and generalization ability. This innovation solves the problem of insufficient generalization in existing models, allowing the model to be used across equipment and environments.

[0045] 2. Precise gray contamination characterization based on deep feature extraction:

[0046] This invention proposes using an autoencoder (AE) to extract deep features from boiler operating data, thereby achieving accurate characterization of ash and pollution levels. Compared to traditional machine learning models based on shallow feature extraction, this invention can uncover deep features in boiler operating data and, combined with the SVR algorithm for nonlinear mapping and regression calculations, significantly improves the accuracy of ash and pollution quantification. This hybrid modeling approach combining deep learning and machine learning results in a 15%-20% improvement in accuracy compared to traditional methods.

[0047] 3. Simultaneous dynamic characterization of ash and dirt and optimization of dust removal:

[0048] This invention can analyze the ash and fouling accumulation on boiler heating surfaces in real time and dynamically adjust the soot blowing strategy through a transfer learning model. By accurately predicting ash accumulation trends and quantifying ash fouling, this invention provides a flexible soot cleaning optimization scheme, solving the problems of traditional soot cleaning being too fixed and lacking flexibility. Soot cleaning optimization not only avoids over-cleaning but also effectively reduces the decline in boiler operating efficiency caused by ash accumulation. It is estimated that after optimizing the soot cleaning cycle, boiler operating efficiency can be improved by approximately 5%-10%, and fuel costs can be reduced by about 5%.

[0049] 4. Improve boiler operating efficiency and reduce maintenance costs:

[0050] By employing precise ash quantification and dynamic ash removal optimization strategies, this invention significantly improves boiler operating efficiency and reduces energy consumption during boiler operation. Furthermore, equipment maintenance costs are also reduced, as boiler maintenance personnel can use models to predict ash accumulation, avoiding frequent shutdowns for inspection and maintenance. The expected annual reduction in combined boiler fuel and maintenance costs is approximately 10%-15%.

[0051] Significant technological advancements:

[0052] 1. Innovative application of dynamic simulation models in the quantification of boiler ash and pollutants:

[0053] Two major challenges exist in the quantification of boiler ash and fouling: a lack of abundant and effective labeling data, and the difficulty of reverse engineering through detailed mechanism modeling. This invention, for the first time, applies a detailed mechanism dynamic simulation model to generate simulation labeling data. Because the simulation model considers the detailed mechanisms of the entire boiler process, it achieves high simulation accuracy and solves the problem of a lack of abundant and effective labeling data.

[0054] 1. Innovative application of transfer learning in the quantification of boiler ash pollution:

[0055] This invention is the first to introduce transfer learning into the characterization of ash and fouling on boiler heating surfaces, achieving hybrid modeling based on simulation and real data. This solves the problems of difficulty in reverse engineering detailed mechanism modeling and poor model transferability in existing technologies. Through fine-tuning strategies, the model can adapt to the actual operating conditions of different boilers, enabling cross-platform model application and filling the gap in the quantitative characterization technology of ash and fouling on boiler heating surfaces.

[0056] 2. High-precision quantification and real-time prediction capabilities for ash and dirt contamination:

[0057] By extracting deep features from boiler operating data using an autoencoder, this invention can capture the dynamics of ash accumulation on boiler heating surfaces and achieve high-precision ash and fouling characterization by combining it with SVR (Surface Dynamics Reflection). Compared to existing machine learning methods based on shallow feature extraction, this invention can more accurately predict the ash and fouling accumulation on boiler heating surfaces under different operating conditions, solving the problem of low ash accumulation prediction accuracy in existing technologies.

[0058] 3. Optimization of intelligent dust removal strategy:

[0059] The intelligent ash removal strategy of this invention, based on the accurate prediction of ash accumulation trends using a transfer learning model, can dynamically adjust the ash removal cycle and intensity. This avoids both energy waste caused by excessive ash removal and the negative impact of excessive ash accumulation on boiler heat transfer efficiency. This intelligent ash removal optimization strategy achieves automated management of boiler ash removal, making boiler operation more efficient and economical.

[0060] In summary, this invention not only solves the technical problems of accuracy, model transfer, and ash removal optimization in the quantitative characterization of boiler ash and pollution, but also significantly improves boiler operating efficiency and reduces fuel and maintenance costs, thus possessing extremely high industrial application value. Attached Figure Description

[0061] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a flowchart of the boiler heating surface ash and fouling quantitative characterization method based on transfer learning provided in the embodiments of the present invention;

[0063] Figure 2 This is a schematic diagram of the ash and dirt fitting results of the AE-SVR model for the final stage superheater in the source domain simulation dataset provided in this embodiment of the invention;

[0064] Figure 3 This is a schematic diagram of the ash and dirt fitting results of the AE-SVR model for the final stage superheater in the actual operating dataset of the target domain provided in this embodiment of the invention;

[0065] Figure 4 This is a schematic diagram comparing the ash and dirt fitting results of the AE-SVR model of the final stage superheater provided in this embodiment of the invention with the calculation results of the cleanliness factor method;

[0066] Figure 5 This is a block diagram of the boiler heating surface ash and dirt quantitative characterization system based on transfer learning provided in an embodiment of the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0068] Industrial Application Example 1: Ash Monitoring and Ash Removal Optimization of Heating Surface in Coal-fired Power Plant Boilers

[0069] In large coal-fired power plants, ash accumulation on boiler heating surfaces is a significant issue affecting boiler efficiency and safety. Ash buildup reduces heat transfer efficiency, increases fuel consumption, and can also cause overheating of the heating surfaces and equipment damage. Therefore, accurately quantifying the ash condition of boiler heating surfaces in real time and optimizing it through effective ash removal strategies is crucial for power plants to improve their economic efficiency and operational effectiveness.

[0070] The ash pollution quantitative characterization and modeling method based on transfer learning provided by this invention can accurately characterize and predict the ash situation on the heated surface under the actual operating conditions of coal-fired power plant boilers.

[0071] 1. Data Acquisition: First, boiler operating data under different loads and conditions, especially the variation data of the ash and pollution coefficient, is obtained through boiler simulation models. Actual power plant boiler operating data is acquired through sensor monitoring, including relevant parameters such as temperature, pressure, and flue gas composition.

[0072] 2. Modeling Process: An autoencoder and support vector regression model are used to model the ash and pollution quantification of boiler simulation data (source domain) to obtain the optimal pre-trained model. Then, the pre-trained model is transferred to the actual boiler operating data of a power plant (target domain), and the model parameters are fine-tuned to adapt the model to the ash and pollution quantification task under actual boiler operating conditions.

[0073] 3. Soot cleaning optimization: By analyzing the ash accumulation trend and ash pollution under different soot blowing cycles, we can help power plants optimize their soot blowing strategies and avoid reduced heat transfer efficiency or equipment damage caused by excessive soot blowing or untimely soot cleaning.

[0074] Effects and benefits:

[0075] Improving boiler efficiency: This method can quantify ash accumulation in real time, guide power plants to formulate reasonable ash removal plans, optimize boiler operating efficiency, and is expected to improve boiler thermal efficiency by 2% to 3%.

[0076] Reduce maintenance costs: By avoiding equipment damage caused by excessive dust cleaning or excessive dust accumulation, the frequency of maintenance and downtime can be reduced, and annual maintenance costs can be reduced by 10% to 15%.

[0077] Fuel cost savings: The optimized ash removal scheme with real-time monitoring of ash accumulation effectively reduces fuel consumption caused by decreased boiler heat transfer efficiency, saving more than 5% in fuel costs.

[0078] Industrial Application Example 2: Ash Monitoring and Prediction of Boiler Heating Surface in Urban Central Heating Systems

[0079] In urban centralized heating systems, boilers, especially during high-load operation in winter, are prone to ash accumulation on their heating surfaces. This can reduce boiler heating efficiency and even affect the heating effect for residents. Traditional ash accumulation detection relies on regular manual inspections, which lacks real-time accuracy and can easily lead to unstable heating or energy waste.

[0080] Applying the transfer learning-based ash and dirt quantitative characterization modeling method of the present invention to urban centralized heating systems can accurately predict the ash accumulation on boiler heating surfaces under complex operating conditions, optimize the ash removal frequency, and ensure the stable operation of the heating system.

[0081] 1. Source and target domain data selection: The source domain ash simulation data is obtained through the boiler simulation model of the centralized heating system, and the real-time operating data of the boiler is obtained through field sensors as the target domain data, including parameters such as temperature, pressure, steam output, and flue gas composition.

[0082] 2. Model Building and Transfer Learning: An autoencoder and support vector regression algorithm were used to model the ash accumulation characteristics of the source domain data. Transfer learning was then applied to actual boiler operation data. After fine-tuning, the model was able to accurately predict the ash accumulation trend on the heating surface under different operating conditions.

[0083] 3. Ash removal management: Based on the predicted ash accumulation trend, the system can intelligently generate ash removal plans to avoid the decline in heating efficiency caused by excessive ash accumulation, and also avoid energy waste caused by frequent ash removal.

[0084] Effects and benefits:

[0085] Improve heating efficiency: By monitoring ash accumulation in real time, the boiler can be operated efficiently, and the heating efficiency is expected to increase by more than 5%.

[0086] Stable heating service: Accurately predict boiler ash accumulation to effectively avoid heating interruptions or instability caused by excessive ash accumulation, ensuring the quality of heating services and improving resident satisfaction.

[0087] Reduce dust removal frequency: Through accurate dust accumulation characterization and prediction, unnecessary dust removal operations can be reduced, with an estimated 20% reduction in annual dust removal frequency, while also lowering operation and maintenance costs.

[0088] These two embodiments demonstrate the industrial application of the present invention in different types of boilers, especially how to improve boiler operating efficiency, reduce costs, and achieve intelligent system management through transfer learning and ash and sludge quantitative characterization technology.

[0089] To address the problems existing in the prior art, this invention provides a method for quantitative characterization and modeling of ash and fouling on boiler heating surfaces based on transfer learning. The invention will be described in detail below with reference to the accompanying drawings.

[0090] like Figure 1 As shown in the embodiment of the present invention, the method for quantitative characterization and modeling of boiler heating surface ash and fouling based on transfer learning includes:

[0091] S1: Based on the boiler dynamic simulation model, obtain full-load, full-condition simulation operation data when the ash and dirt coefficients of each heating surface change;

[0092] S2: Select the operating dataset S1 of the heating surface in the simulation model as the source domain, and select the operating dataset of the corresponding heating surface in the actual coal-fired power plant boiler as the target domain.

[0093] S3: Based on autoencoder and support vector regression algorithms, a quantitative characterization model of boiler heating surface ash and fouling under the source domain is built and the parameters are adjusted, optimized and trained to obtain the pre-trained model with the best fitting result of heating surface ash and fouling.

[0094] S4: Transfer the pre-trained model from S3 to the target domain dataset, and use the method of fine-tuning the model parameters to make the model adapt to the ash and pollution quantification characterization task under the same boiler heating surface actual operating data.

[0095] S5: Utilize actual soot blowing actions to determine the ash accumulation trend and degree of ash contamination on the heated surface during different soot blowing cycles, and complete the verification and application of the transfer learning model.

[0096] The acquisition of full-load, full-condition operating data for each level of heating surface in S1, under varying ash and fouling coefficients, was conducted using a boiler-furnace coupled dynamic simulation model established with Apros. In the module attribute settings, the change in the ash and fouling coefficient of specific boiler heating surfaces was used as an indicator feature, ranging from 0.8 to 1. A coefficient of 1 (100%) represented a clean heating surface; as the coefficient decreased, it indicated that the cleanliness of the heating surface began to decline due to ash accumulation. When the coefficient rose back to 1, it indicated that the heating surface had recovered its cleanliness due to soot blowing. The simulation data sampling interval was 1 minute. Original simulation datasets, consisting of multiple parameters including load, coal feed rate, working fluid flow rate, working fluid temperature, working fluid pressure, and flue gas temperature and pressure, were used for subsequent quantitative characterization of ash and fouling on each level of boiler heating surfaces.

[0097] The dataset partitioning in S2 is based on the fundamental concepts of transfer learning. Transfer learning utilizes the idea of ​​knowledge transfer to apply feature relationships learned in one domain to similar domains, dividing the dataset into source domains D. s (Sourcedomain) and Targetdomain D t (Target domain) The source domain data is generally labeled and of high quality. Valuable knowledge, such as feature representations, is then obtained from the source domain dataset through relevant algorithms. The target domain data, on the other hand, is generally unlabeled and its distribution or features are similar to those of the source domain. Algorithms can be used to transfer similar feature representations from the source domain to the target domain, helping the target domain to perform data prediction or classification and improving the performance of the target task. In this case, the simulation data of the heated surface and the ash pollution coefficient labels are selected as the source domain, and the actual operating data of the corresponding heated surface is selected as the target domain, with relatively consistent input parameters for the model.

[0098] The pre-trained model with the optimal fitting result for the ash contamination of the heated surface built in S3 is as follows: For the source domain data, the unsupervised learning algorithm Autoencoder (AE) and the traditional machine learning algorithm Support Vector Regression (SVR) are combined. The encoder part of the AE model is used as a feature extractor. Data samples are input into the encoder part for data dimensionality reduction and deep feature extraction. Then, the feature vectors and data labels are input into the SVR model to perform regression. The obtained predicted value of the ash contamination degree of the heated surface is compared with the ash contamination coefficient label of the source domain. Through parameter tuning and optimization, the model framework and parameter settings with the optimal fitting result are obtained. This is represented as the pre-trained model AE-SVR for transfer learning, which is subsequently used for model transfer learning for the target domain.

[0099] The autoencoder AE described in S3 includes only the encoding portion:

[0100] Encoder E will input feature vector Transform into latent feature vectors , .

[0101] In the formula, As input features, These are the encoder parameters. The encoder part performs feature extraction and dimensionality reduction on the input data.

[0102] The Support Vector Regression (SVR) described in S3 is based on the idea of ​​mapping sample data to a high-dimensional space through a nonlinear mapping function, and then performing linear regression in that high-dimensional space. By constructing a loss function that minimizes the training set data labels and the model fit values, a target function f(x) is ultimately determined, expressed as:

[0103] .

[0104] In the formula, w is the weight coefficient vector; is the high-dimensional feature space obtained by mapping the input space x; b is the bias.

[0105] In S4, the pre-model from S3 is transferred to the target domain. Since the distribution differences between different datasets during transfer can lead to a deterioration in model performance, deep transfer learning requires different fine-tuning strategies to adapt the parameters or model to the new dataset and improve model performance. For the AE-SVR model, a fine-tuning strategy is adopted that modifies the parameters of the AE encoder to adapt to the target domain data.

[0106] In S5, during the verification of the transfer results of the ash and dirt quantitative characterization model for the heating surface based on actual operating data, the actual boiler unit lacks label values ​​that characterize the degree of ash and dirt on the heating surface. That is, the simulation dataset of the source domain has labels, while the dataset of the target domain does not. Therefore, the actual soot blowing action is used to judge whether the fitting results of the AE-SVR model of each level of heating surface in different soot blowing cycles reflect the correct ash accumulation trend and degree of ash and dirt, thereby verifying the effectiveness and feasibility of the AE-SVR transfer learning method.

[0107] like Figure 5 As shown in the figure, an embodiment of the present invention provides a quantitative characterization system for boiler heating surface ash and fouling based on transfer learning, comprising:

[0108] The operation data acquisition module is used to acquire full-load, full-condition simulation operation data based on the boiler dynamic simulation model when the ash and dirt coefficients of each heating surface change.

[0109] The selection module is used to select the operating dataset of the heating surface in the simulation model as the source domain and the operating dataset of the corresponding heating surface in the actual coal-fired power plant boiler as the target domain.

[0110] The training module is used to build a quantitative characterization model of boiler heating surface ash and pollution based on autoencoder and support vector regression algorithms, and to perform parameter tuning, optimization and training to obtain the pre-trained model with the best fitting result of heating surface ash and pollution.

[0111] The transfer module is used to transfer the pre-trained model to the target domain dataset and to fine-tune the model parameters so that the model can adapt to the ash and pollution quantification characterization task under the same boiler heating surface actual operating data.

[0112] The verification module is used to determine the ash accumulation trend and degree of ash contamination on the heated surface during different soot blowing cycles by using actual soot blowing actions, and to complete the verification and application of the transfer learning model.

[0113] An embodiment of the present invention provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method for quantitative characterization of ash and dirt on boiler heating surfaces based on transfer learning.

[0114] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for quantitative characterization of ash and dirt on boiler heating surfaces based on transfer learning.

[0115] An embodiment of the present invention provides an information data processing terminal, which is used to implement the boiler heating surface ash and dirt quantitative characterization system based on transfer learning.

[0116] This study focuses on a 1000MW ultra-supercritical double reheat unit. The boiler operates with double intermediate reheat and ultra-supercritical pressure variation. A simulation model of the unit was built using Apros dynamic simulation software for subsequent data acquisition. In the module attribute settings, the change in the ash pollution coefficient of the boiler's specific heating surfaces was used as the indicator feature. A coefficient of 1 (100%) represents a clean heating surface. As the coefficient decreases, the cleanliness of the heating surface begins to decline due to ash accumulation. When the coefficient rises back to 1, it indicates that the heating surface has recovered its cleanliness due to soot blowing. The simulation data sampling interval was 1 minute. For each heating surface level, the original simulation dataset, consisting of multiple parameters including load, coal feed rate, working fluid flow rate, working fluid temperature, working fluid pressure, and flue gas temperature and pressure, was used as the source domain. Meanwhile, a total of 380,965 sets of actual operating data from the power plant's DCS system from March 9, 2021 to May 10, 2022 were collected as the target domain. The sampling interval and related parameters were consistent with the simulation data. The field data had already undergone the preprocessing step of mean denoising before modeling.

[0117] Leveraging the concept of transfer learning, this invention aims to reduce the modeling difficulty by transferring pre-built models and ash-smudge label representations from simulated data (i.e., the source domain) to the same actual heated surface (i.e., the target domain). Therefore, for the source domain data, this invention combines the unsupervised learning algorithm Autoencoder (AE) with the traditional machine learning algorithm Support Vector Regression (SVR). The encoder part of the AE model is used as a feature extractor. Data samples are input into the encoder part for dimensionality reduction and deep feature extraction. Then, the feature vectors and data labels are input into the SVR model to perform regression. The obtained predicted values ​​of ash-smudge degree on the heated surface are compared with the ash-smudge coefficient labels of the source domain. Through parameter tuning and optimization, it is verified that when the number of neurons in the input hidden layer is 32 and 16, the training batch size is 16, and the training iterations are 200, the model can be sufficiently pre-trained, achieving the highest label fitting accuracy. Finally, the optimal model framework and parameter settings are obtained, which are represented as the pre-trained model of transfer learning, as shown below. Figure 2 As shown, the quantitative characterization curve of the final-stage superheater based on the AE-SVR pre-trained model in the source domain has the highest degree of fit with the trend of the original ash pollution coefficient label. It is believed that it can be used as a pre-trained model for subsequent model transfer and model fine-tuning in the target domain.

[0118] Because actual boiler units lack labeled values ​​to characterize the degree of ash fouling on the heating surfaces—that is, the source domain simulation dataset has labels while the target domain dataset does not—the AE-SVR model fitting results for each level of heating surface in different soot blowing cycles are judged by actually performing soot blowing actions to determine whether they reflect the correct ash accumulation trend and degree of ash fouling. For example... Figure 3 As shown, the fitting results of the AE-SVR model closely approximate the actual ash and dirt levels, and the trend is correct. Within the soot blowing action range, the fitting results exhibit a rapid upward phase, indicating that the actual soot blowing action effectively removes ash accumulation on the heated surface. Subsequently, the cleanliness of the heated surface gradually decreases as ash and dirt deposits further accumulate. Therefore, it can be concluded that even in the absence of quantitative ash and dirt characterization labels, transferring the simulation data modeling results can effectively fit the actual ash accumulation trend on the heated surface.

[0119] In summary, the verification results of this embodiment demonstrate that the transfer learning-based method for quantitative characterization of boiler heating surface ash and fouling can reduce the modeling difficulty for actual operating data lacking ash and fouling labels. By utilizing the ash and fouling label representations and feature representations in the simulation dataset, it helps to effectively fit the ash accumulation trend of actual operating data and verify the effectiveness of actual soot blowing actions. Furthermore, only a small number of heating surface objects in the simulation data need to be modeled to obtain relevant feature representations, while the heating surface objects in the actual data are modeled based on transfer learning. Through local or global parameter fine-tuning, the model adapts to the actual objects, further reducing the time and economic costs of modeling and training, and improving the applicability and generalization of the model.

[0120] To verify the effectiveness of the transfer learning-based modeling for monitoring ash fouling on boiler heating surfaces, this study selects current mainstream methods and compares the model results calculated based on the heat transfer mechanism to determine the degree of ash fouling / cleanliness factor, using the final-stage superheater as an example. Figure 4 As shown, it was found that using the model and ash label representation established in the final stage superheater based on simulation data, the ash fouling degree fitting results when transferred to the same actual heating surface are close to the actual ash fouling degree range and the trend is correct. However, the cleanliness factor calculation results fluctuate greatly, and the results remain basically unchanged for some time periods, indicating that it cannot reflect the correct ash accumulation trend, and its accuracy and stability are poor. In summary, the ash fouling quantitative characterization method based on transfer learning for boiler heating surfaces provides a new modeling approach for boiler ash fouling monitoring. It uses the data output by the simulation model and the labels representing ash fouling to help model ash fouling monitoring of the corresponding boiler heating surfaces in actual operation data, solving the problem of lacking ash fouling labels in practice. Moreover, the modeling speed is fast, the amount of data required is small, and it effectively fits the ash fouling trend, showing high feasibility.

[0121] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for quantitatively representing the amount of soot deposition on a heating surface of a boiler based on transfer learning, characterized by, Comprise the following steps: S1: based on three transmission and one reflection to establish a dynamic simulation model of the boiler, and verify with actual operation data, develop variable working condition and variable ash fouling coefficient dynamic simulation, obtain the full load full working condition simulation operation data when the ash fouling coefficient of each stage heating surface changes; S2: select the heating surface operation data set S1 in the simulation model as the source domain, select the operation data set of the corresponding same heating surface in the actual coal-fired power plant boiler as the target domain; S3: based on the autoencoder and support vector regression algorithm, build the ash fouling quantification characterization model of the boiler heating surface under the source domain and carry out parameter optimization and training, obtain the pre-training model with optimal fitting result of the ash fouling of the heating surface; S4: migrate the S3 pre-training model to the target domain data set, and adopt the method of fine-tuning model parameters to make the model adapt to the ash fouling quantification characterization task under the actual operation data of the same boiler heating surface; S5: use the actual blowing action to judge the ash deposition trend and ash fouling degree of the heating surface in different blowing periods, and complete the verification of the transfer learning model; S6: solidify the verified model in the industrial computer, connect with the DCS system through Modbus and OPC communication, obtain real-time operation data, obtain the fouling condition of each stage heating surface after model calculation, and the calculation result is further used for intelligent blowing decision or returned to the DCS for reference of the operation personnel; The full load full working condition operation data under the ash fouling coefficient of each stage heating surface in S1 is obtained by establishing a boiler coupling dynamic simulation model by Apros, and the change of the ash fouling coefficient of the specific heating surface of the boiler is taken as the index feature in the module attribute setting, the range is 0.6-1, the coefficient 1 represents the clean heating surface, and as the coefficient decreases, the clean degree of the heating surface surface begins to decrease due to ash deposition; When the coefficient rises to 1 again, it means that the heating surface is restored to a clean state due to the blowing operation; The simulation data sampling interval can be 10 seconds to 1 minute, and the original simulation data set composed of load, coal supply amount, working medium side flow, working medium side temperature, working medium side pressure, flue gas side temperature and pressure of each stage heating surface is used for subsequent ash fouling quantification characterization of each stage heating surface of the boiler; The data set in S2 is divided based on the basic concept of transfer learning; The specific process of the pre-training model with optimal fitting result of the ash fouling of the heating surface built in S3 is as follows: facing the source domain data, combining the unsupervised learning algorithm autoencoder AE and the traditional machine learning algorithm support vector regression SVR, taking the encoder part of the AE model as a feature extractor, inputting the data sample into the encoder part for data dimension reduction and deep feature extraction, then inputting the feature vector and data label into the SVR model for regression, comparing the predicted value of the ash fouling degree of the heating surface with the ash fouling coefficient label of the source domain, and through parameter optimization, obtaining the model framework and parameter setting with optimal fitting result, which is represented as the pre-training model AE-SVR of transfer learning, which is used for subsequent model transfer learning of the target domain.

2. The method of claim 1, wherein the method is characterized by, The autoencoder AE in S3 comprises only an encoding part: The encoder E converts the input feature vector x into a latent feature vector z, zi = E (xi | θE) ; In the formula, x is an input feature, and θE is a parameter of the encoder; the input data is feature extracted and reduced dimension by the encoder part; The basic idea of the support vector regression SVR in S3 is to map the sample data to a high-dimensional space through a nonlinear mapping function, and perform linear regression in the high-dimensional space; by constructing a loss function that minimizes the label of the training set data and the fitting value of the model, a target function f(x) is finally determined, which is expressed as: ； In the formula, w is a weight coefficient vector; is a high-dimensional feature space mapped from the input space x; b is a bias.

3. The method of claim 1, wherein the method is characterized by, In S4, the pre-trained model in S3 is migrated to the target domain, and different fine-tuning strategies are needed in deep transfer learning to adapt the parameters or model to the new data set and improve the model effect; for the AE-SVR model, a fine-tuning strategy of changing the parameters of the AE encoder part to adapt to the target domain data is adopted.

4. The method of claim 1, wherein the method is characterized by, In S5, in the migration result verification of the heated surface ash fouling quantification characterization model facing the actual operation data, there is a lack of label value representing the degree of ash fouling of the heated surface in the actual boiler unit, that is, the simulation data set of the source domain has a label, and the data set of the target domain does not have a label; whether the AE-SVR model fitting result of each level of the heated surface in different blowing periods reflects the correct ash deposition trend and ash fouling degree is judged by the actual blowing execution action, to verify the effectiveness and feasibility of the AE-SVR transfer learning method.

5. A system for quantitatively characterizing the fouling of a boiler heating surface based on transfer learning, which implements the method for quantitatively characterizing the fouling of a boiler heating surface based on transfer learning according to any one of claims 1-4, characterized in that, The boiler heated surface ash fouling quantification characterization system based on transfer learning comprises: An operation data acquisition module configured to acquire full-load full-condition simulation operation data of a change in a heated surface ash fouling coefficient based on a boiler dynamic simulation model; A selection module configured to select a heated surface operation data set in a simulation model as a source domain, and select an operation data set of a corresponding same heated surface in an actual coal-fired power plant boiler as a target domain; A training module configured to build a boiler heated surface ash fouling quantification characterization model under the source domain based on an autoencoder and a support vector regression algorithm, and to perform parameter tuning optimization and training to obtain a pre-trained model with an optimal heated surface ash fouling fitting result; A migration module configured to migrate the pre-trained model to a target domain data set, and to adopt a method of fine-tuning model parameters to adapt the model to the ash fouling quantification characterization task under actual operation data of the same boiler heated surface; A verification module configured to judge the ash deposition trend and ash fouling degree of the heated surface in different blowing periods by using actual blowing execution actions, and to complete migration learning model verification and application.

6. A computer device, comprising: The computer device comprises a memory and a processor, and the memory stores a computer program, which, when executed by the processor, causes the processor to execute the steps of the boiler heated surface ash fouling quantification characterization method based on transfer learning according to any one of claims 1-4.

7. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the boiler heated surface ash fouling quantification characterization method based on transfer learning according to any one of claims 1-4.

8. An information data processing terminal, characterized by The information data processing terminal is used for realizing the boiler heating surface soot fouling quantification characterization system based on migration learning. The information data processing terminal is used for realizing the boiler heating surface soot fouling quantification characterization system based on migration learning.

Images