Flue gas waste heat gradient utilization system optimization method based on deep transfer learning

Abstract

The invention discloses a flue gas waste heat gradient utilization system optimization method based on deep transfer learning, and the method comprises the steps: firstly constructing a flue gas waste heat system simulation model under multiple working conditions based on an EBSILON simulation platform, and generating a simulation data set covering flue gas waste heat system key operation parameters of different loads, environment conditions and the like; then, establishing a deep neural network pre-training model based on simulation data, adding linear and nonlinear calibration networks on the basis of the deep neural network pre-training model, performing hyper-parameter updating on the calibration networks by using field operation data, and migrating the pre-training model to a feature space adaptive to the operation data; on the basis, flue gas bypass parameters are dynamically adjusted in combination with a genetic algorithm, and maximum utilization of waste heat resources is achieved. According to the method, the limitation of scarcity of field effective operation data samples on model training is overcome, and the overall energy efficiency of the system is improved through intelligent regulation and control.

Description

Technical Field

[0001] The present invention belongs to the field of energy utilization and intelligent control technology, and specifically relates to a flue gas waste heat cascade utilization system optimization method based on deep transfer learning, which is suitable for industrial flue gas waste heat recovery systems such as coal-fired power plants, steel mills, and chemical plants. Background Art

[0002] Flue gas waste heat recovery technology is of great significance for improving the energy utilization efficiency of industrial enterprises, reducing energy waste, and lowering production costs. The large amount of flue gas generated by industrial enterprises during the production process usually carries high-temperature waste heat. If it is not effectively utilized, it will lead to a large amount of energy waste and aggravate environmental pollution problems. However, due to the complex and changeable industrial production conditions, parameters such as flue gas flow, temperature, and pressure fluctuate violently. Traditional flue gas waste heat cascade utilization systems mostly use empirical design or simple control strategies. These methods lack real-time adaptability and cannot accurately respond to dynamic changes in industrial sites, resulting in a significant reduction in flue gas waste heat recovery efficiency.

[0003] Furthermore, cascaded flue gas heat utilization systems typically include complex structures such as multi-stage heat exchangers and flue gas bypass control devices, making coordinated optimization of all system components crucial. Currently, most companies still control these systems manually or semi-automatically, making it difficult to ensure optimal operation of each device under various load conditions. Furthermore, existing control strategies based on empirical formulas or simple models struggle to achieve a comprehensive balance between real-time performance and accuracy, particularly under conditions of frequent load fluctuations. In recent years, the development of artificial intelligence (AI), particularly deep learning, has provided new avenues for optimizing the control of industrial systems. However, training deep learning models typically requires a large amount of sample data across multiple operating conditions. In practice, units often operate under a single operating condition, and data-driven approaches face data scarcity. Therefore, relying solely on field data makes it difficult to quickly build accurate deep learning models. Furthermore, the operating conditions of flue gas heat utilization systems vary significantly across industrial sites, making it difficult to quickly and effectively migrate existing models to new application scenarios.

[0004] Therefore, in order to solve the above problems, it is urgent to propose an intelligent optimization method that can effectively utilize simulation data, realize rapid model migration and adapt to complex industrial site conditions, so as to significantly improve the operating efficiency of the flue gas waste heat cascade utilization system. Summary of the Invention

[0005] Purpose of the invention: The present invention provides a flue gas waste heat cascade utilization system optimization method based on deep transfer learning, which can make full use of high-quality simulation data and limited field data, quickly build an accurate and well-generalized intelligent optimization model, and realize efficient and precise optimization of the flue gas waste heat recovery system in a complex industrial environment.

[0006] Technical solution: The present invention provides a method for optimizing a flue gas waste heat cascade utilization system based on deep transfer learning, comprising the following steps:

[0007] S1: Establish a simulation model of the flue gas waste heat cascade utilization system based on the EBSILON platform, and obtain high-quality simulation sample data for multiple operating conditions after verifying the accuracy of the simulation model;

[0008] S1.1: Establish simulation models of key equipment based on the manufacturer's thermal balance diagram and the actual structure and operating parameters of the system;

[0009] S1.2: Set the flue gas flow rate, temperature, pressure, heat exchanger efficiency, bypass valve opening and environmental parameters;

[0010] S1.3: Simulate system operating conditions under different load conditions, including full load, part load, and low load conditions;

[0011] S1.4: Compare the output of the simulation model with the design value and calculate the relative error between the two;

[0012] S1.5: Determine whether the error index meets the preset accuracy requirements. If not, repeat S1.2 to S1.4 until the error value meets the requirements.

[0013] S1.6: Based on the verification results, optimize the key parameters in the simulation model to ensure that the output results of the optimized model are consistent with the actual operation data;

[0014] S1.7: Use the API provided by EBSILON to write MATLAB scripts to automatically batch read the simulation data files generated by EBSILON;

[0015] S1.8: Preprocess simulation data to generate standardized datasets suitable for deep learning.

[0016] S2: Deep transfer learning and parameter optimization between simulation data and field operation data;

[0017] S2.1: Pre-train the deep neural network model using the diverse operating condition simulation dataset generated in step 1;

[0018] S2.2: Extract the feature representation of the flue gas waste heat cascade utilization system through pre-training to obtain a preliminary pre-training model;

[0019] S2.3: Collect on-site historical operating data of the flue gas waste heat utilization system;

[0020] S2.4: Preprocess field operation data;

[0021] S2.5: Based on the pre-trained model, update the calibration network using the training set of field operation data;

[0022] S2.6: Optimize model parameters using the backpropagation algorithm to minimize the error between the predicted value and the actual value;

[0023] S2.7: Use the validation set of field data to evaluate the prediction accuracy of the fine-tuned model and calculate the model performance indicators, including mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R). 2 ;

[0024] S2.8: Based on the model after the network is calibrated and updated using operational data, the flue gas bypass parameters are optimized using an intelligent optimization algorithm. Through iterative optimization, the optimal flue gas bypass parameter configuration is obtained.

[0025] S2.9: Apply the optimized flue gas bypass parameter configuration to the real-time control of the flue gas waste heat utilization system, and verify the optimization effect through system feedback.

[0026] Furthermore, the key equipment in step S1.1 includes a steam turbine, a boiler, a flue gas bypass, a heat exchanger, and a fan.

[0027] Furthermore, the simulation data acquisition process described in step S1.8 is as follows: calling the unit thermal system model under variable operating conditions, obtaining simulation results under different operating conditions by modifying parameters and setting bypass key parameter constraints, and obtaining waste heat utilization system simulation data sets under different operating conditions including energy consumption indicators such as coal consumption rate and power generation heat consumption through thermal calculation.

[0028] Furthermore, the pre-training process of the deep neural network model described in step S2.1 is as follows:

[0029] The input of the deep neural network model is the key parameters of the flue gas bypass, and the output is the unit boiler efficiency η b , turbine heat rate q and coal consumption rate b:

[0030]

[0031] Among them, Q zq is the main steam heat; Q gs is the heat of feed water; Q zr is the heat of reheat steam; Q jws The heat of the desuperheating water in the two-stage superheater; Q gp is the heat of the air preheater bypass high pressure economizer; Q dp is the heat of the air preheater bypass low-pressure economizer; B is the coal consumption of the boiler; Pe is the generator output power; q is the turbine heat rate, η b is the boiler efficiency; η pis the pipeline efficiency; Q net It is the low calorific value of standard coal;

[0032] During the pre-training process, the learning rate lr1 is set to 0.1 and the number of training times Ne1 is set to 40, and the weights and biases in the hidden layer are iteratively updated; the calculation formula of the deep neural network is as follows:

[0033]

[0034] in, is the expected output of unit i in μ mode; w ij 、w jk is the weight of the corresponding unit; i, j, k are the corresponding input, hidden and output units; P is the number of input patterns; g is the activation function.

[0035] Furthermore, the preprocessing of the field operation data in step S2.4 includes data cleaning, missing value filling, normalization, and outlier detection and processing.

[0036] Furthermore, in step S2.5, when the amount of on-site operation data is limited during the calibration process, the learning rate lr2 is set to 0.001 and the number of training times Ne2 is set to 100, so as to reduce fluctuations while ensuring the training speed.

[0037] Furthermore, the implementation process of step S2.5 is as follows:

[0038] In the migration framework, the mapping relationship between the simulation data model and the actual data is expressed as follows:

[0039] y H =F(x,y L )

[0040] Among them, x is the model input parameter; y H and y L are the output values ​​of the field operation data and the corresponding simulation data respectively;

[0041] The mathematical expression of the calibration network is:

[0042] F=F1+F n1 =αF1(x,y L )+(1-α)F n1 (x,y L )

[0043] Among them, F1 and F n1 are the linear term and nonlinear term in the function F(τ) respectively; α is the weight factor of the linear term in the function F(τ).

[0044] Furthermore, the implementation process of step S2.6 is as follows:

[0045] Minimize the loss function to reduce the difference between the predicted value and the actual value:

[0046]

[0047] in, is the input sample; is the output sample; ||·||2 represents the L2 norm.

[0048] Furthermore, the flue gas bypass parameter configuration in step S2.8 includes the flue gas flow distribution ratio, the bypass valve opening, and the bypass economizer water inlet flow.

[0049] Beneficial effects: Compared with the existing technology, the beneficial effects of the present invention are: the present invention combines deep learning and transfer learning technology to quickly construct an accurate and highly generalized flue gas waste heat cascade utilization system optimization model, and realizes real-time adaptive optimization of the system under complex and dynamically changing industrial conditions; through the automated model optimization process, the present invention effectively reduces the complexity of manual control, significantly improves the waste heat recovery efficiency of the system, and reduces the energy consumption and operating costs of the enterprise; at the same time, the present invention has a high degree of versatility and portability, and can be widely used in flue gas waste heat recovery systems in different industrial fields, promoting the development of industrial energy conservation and emission reduction and green production technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is an overall overview diagram of the unit in the present invention;

[0051] Figure 2 This is a flow chart of the cascade utilization of waste heat from flue gas in the present invention;

[0052] Figure 3 This is a simulation model diagram of the EBSILON system for cascade utilization of flue gas waste heat in the present invention;

[0053] Figure 4 This is a comparison chart of the heat rate between the simulation model output value and the design value in the present invention;

[0054] Figure 5 A flow chart of the model migration method based on linear and nonlinear networks proposed in the present invention;

[0055] Figure 6 This is a diagram showing the training results of a deep neural network using simulated sample data in the present invention;

[0056] Figure 7 This is a comparison chart between the field operation data pre-training + calibration update method and direct training in the present invention. DETAILED DESCRIPTION

[0057] The present invention is further described in detail below with reference to the accompanying drawings:

[0058] The present invention proposes a method for optimizing a flue gas waste heat cascade utilization system based on deep transfer learning. The method involves obtaining simulation sample data, establishing a pre-training model based on the simulation data, fine-tuning the pre-training model based on operating data, and optimizing the unit energy consumption index using an intelligent optimization algorithm. The method specifically includes the following steps:

[0059] S1: Acquisition of high-quality simulation sample data based on the EBSILON simulation platform.

[0060] S1.1: Based on the manufacturer's heat balance diagram and actual structural parameters, build a detailed simulation model of key equipment including the turbine, boiler, flue gas bypass, heat exchanger, fan, etc., an overall overview of the unit and a detailed schematic diagram of the flue gas waste heat utilization system, such as Figure 1 and Figure 2 Then, they are connected according to fixed interface rules to construct the overall flue gas waste heat utilization system simulation model. The overall simulation model is as follows Figure 3 shown.

[0061] S1.2: Accurately set key operating parameters such as flue gas flow, temperature, pressure, heat exchanger efficiency, and bypass valve opening to simulate typical operating conditions such as full load, part load, low load, and deep peak load operation to ensure representative and diverse simulation data.

[0062] S1.3: Compare and analyze the simulation results with the design values, quantify the error w, evaluate the accuracy of the model, and determine whether the error index meets the preset accuracy requirements (w < 5%). If not, adjust the parameters and repeat the simulation. The heat rate comparison between the design value and the simulation value is as follows: Figure 4 shown.

[0063] S1.4: Utilize the built-in API of EBSILON to write MATLAB scripts to automatically export simulation data in batches. This script calls the unit's thermal system model under varying operating conditions. By modifying parameters and setting key bypass parameter constraints, simulation results are obtained under different operating conditions. Thermal calculations are then used to generate a simulation dataset for the waste heat utilization system under different operating conditions, including energy consumption indicators such as coal consumption rate and power generation heat consumption.

[0064] Use MATLAB's actxserver function to create an EBSILON COM object and establish a connection with EBSILON. Then, use the COM object to load the EBSILON model files for the flue gas waste heat utilization system under each sub-operating condition. Use the API interface to set the simulation parameter operating conditions, simulation time, and parameter limits. The optimization objectives and specific constraints are as follows:

[0065] minb=f(β,Q1,Q2,T1,T2)

[0066]

[0067] Among them, b is the standard coal consumption rate; β is the bypass flue gas ratio; Q1 is the high bypass economizer inlet water flow; Q2 is the low bypass economizer inlet water flow; T1 is the air preheater inlet flue gas temperature; T2 is the bypass midpoint flue gas temperature.

[0068] Next, call the EBSILON simulation method, run the simulation, and generate data. Then, read the simulation result data through the API interface. Finally, use the loop structure to batch read the simulation data under different working conditions and save it to a file or MATLAB variable.

[0069] S1.5: Perform data preprocessing on the exported simulation data to generate a standardized dataset suitable for deep learning.

[0070] Step S2: Optimize the parameters of the flue gas waste heat cascade utilization system based on the deep transfer learning method, such as Figure 5 As shown, specifically including:

[0071] S2.1: Pre-train the deep neural network model based on simulation data, extract the characteristic laws of the flue gas waste heat cascade utilization system, and build an initial model. The training results of the simulation sample data are as follows: Figure 6 shown.

[0072] The simulation conditions mainly include standard heat rate acceptance condition (Turbine Heat Acceptance, THA), 75% THA, 50% THA and deep peak load shaving 30% THA. The input of the deep neural network model is the key parameters of flue gas bypass, and the output is the unit boiler efficiency η b , power generation heat consumption q and coal consumption rate b.

[0073]

[0074] Among them, Q zq is the main steam heat; Q gs is the heat of feed water; Q zr is the heat of reheat steam; Q jws The heat of the desuperheating water in the two-stage superheater; Q gp is the heat of the air preheater bypass high pressure economizer; Q dp is the heat of the air preheater bypass low-pressure economizer; B is the coal consumption of the boiler; Pe is the generator output power; q is the turbine heat rate, η b is the boiler efficiency; η p is the pipeline efficiency, set to 0.99; Q netThe lower calorific value of standard coal is 29307.6 kJ·(kg) -1 .

[0075] The training process selects appropriate hyperparameters. Considering the large amount of simulation data, the weights and biases in the hidden layer are iteratively updated with a large learning rate lr1 of 0.1 and a training number Ne1 of 40 during the pre-training process. The calculation formula of the deep neural network is as follows:

[0076]

[0077] in, is the expected output of unit i in μ mode; w ij 、w jk is the weight of the corresponding unit; i, j, k are the corresponding input, hidden and output units; P is the number of input patterns; g is the activation function.

[0078] During the pre-training process, the weights and biases in the hidden layer are iteratively updated with a larger learning rate lr1 and the number of training times Ne1.

[0079] S2.2: Collect real-time operating data of the flue gas waste heat utilization system on site and perform data preprocessing, including data cleaning, missing value filling and normalization, as well as outlier detection and processing. The specific methods are K-means clustering algorithm and interpolation method.

[0080] S2.3: Use a composite neural network to simultaneously characterize the simulation sample data model and its correction process for field operation data. During the calibration process, freeze the pre-trained model NN L The network hyperparameters remain unchanged, and only the calibration network NN is optimized. H1 and NN H2 parameters to ensure that the model can adapt to the field operation data while maintaining the learning content of the simulation data.

[0081] In the migration framework, the mapping relationship between the simulation data model and the actual data can be expressed as follows:

[0082] y H =F(x,y L )

[0083] Among them, x is the model input parameter; y H and y L are the output values ​​of the field operation data and the corresponding simulation data, respectively.

[0084] The mathematical expression of the calibration network can be expressed as follows:

[0085] F=F1+F n1 =αF1(x,y L )+(1-α)Fn1 (x,y L )

[0086] Among them, F1 and F n1 are the linear term and nonlinear term in the function F(τ) respectively; α is the weight factor of the linear term in the function F(τ).

[0087] When the amount of on-site operation data is limited during the calibration process, a lower learning rate lr2 and number of training times Ne2 are used to reduce fluctuations while ensuring the training speed.

[0088] S2.4: Optimize and fine-tune the model parameters through the back propagation algorithm to minimize the error between the predicted results and the actual running data. The comparison between the training effect after fine-tuning and the result of direct training is shown in the figure below. Figure 7 shown.

[0089] The difference between the predicted value and the actual value is reduced by minimizing the loss function, which is defined as follows:

[0090]

[0091] in, is the input sample; is the output sample; ||·||2 represents the L2 norm.

[0092] S2.5: Combine intelligent optimization algorithms to optimize key control parameters of the system, including flue gas flow distribution ratio, bypass valve opening, flue gas bypass economizer water inlet flow, etc.

[0093] Genetic algorithms are used to simulate the natural selection process (selection, crossover, and mutation) to search for the optimal solution. The optimized parameter configuration is applied to the actual system to observe the changes in system performance and verify whether the system energy consumption is reduced through engineering examples.

[0094] In combination with actual data, the present invention provides optional specific implementation plans as follows:

[0095] The effectiveness of the proposed method is verified by the actual operation data of the flue gas waste heat cascade utilization system of the target thermal power plant. The overall overview of the target unit is as follows: Figure 1As shown. At present, in order to reduce the coal consumption of coal-fired units, the deep utilization of flue gas waste heat has become the key to improving their thermal performance and economic benefits. Traditional waste heat utilization routes such as low-temperature economizers and water-based air heaters are all arranged after the air preheater and can only recover low-grade waste heat. The air preheater bypass flue gas waste heat utilization system adopts a parallel bypass design, which performs cascade heating on the feed water system by diverting part of the high-temperature flue gas. The system mainly includes two heat exchange links: on the one hand, high-temperature flue gas is used to heat the boiler feed water, and on the other hand, medium and low-temperature condensate is preheated. Through this cascade utilization method, the high-energy extraction steam of the steam turbine can be effectively squeezed out, allowing it to continue to expand and do work in the steam turbine, thereby improving the overall power generation efficiency of the unit and increasing the power output. However, due to the diversity of the unit's feed coal composition and the complexity of the operating conditions, the flue gas temperature fluctuation phenomenon is more prominent, which to a certain extent increases the difficulty of designing the unit's waste heat utilization plan. Based on deep transfer learning, the present invention proposes an optimization method for the flue gas waste heat cascade utilization system driven by the fusion of simulation data and operation data.

[0096] The inputs to the deep neural network were the bypass flue gas ratio, high bypass economizer inlet water flow, low bypass economizer inlet water flow, high bypass economizer inlet water temperature, high bypass economizer outlet water temperature, low bypass economizer inlet water temperature, low bypass economizer outlet water temperature, bypass midpoint flue gas temperature, air preheater inlet flue gas temperature, air preheater outlet flue gas temperature, and air preheater bypass outlet flue gas temperature. Partial operational data for April 2024 was collected from the DCS system, totaling 6,000 samples, as the training set. The collected data was preprocessed, including data cleaning, normalization, and data segmentation, and divided into a training set (70%), a validation set (20%), and a test set (10%).

[0097] During the model training and fine-tuning phase, a deep neural network (DNN) model was used for pre-training. The model consisted of three hidden layers with 128 neurons in each layer, the activation function was ReLU, and the loss function was the mean square error (MSE). The model was pre-trained using simulated data, with 100 training rounds and a batch size of 32. After training, the weights and parameters of the pre-trained model were saved. Next, transfer learning technology was used to update the parameters of the calibration network of the model using the training set of field data, with 50 training rounds and a batch size of 32. After the update was completed, the prediction accuracy of the model was evaluated using the validation set to ensure that the mean square error (MSE) was less than 0.01 and the coefficient of determination (R 2 ) is greater than 0.95.

[0098] During the parameter optimization phase, a genetic algorithm (GA) was used to optimize parameters with the goal of minimizing system energy consumption. The population size was set to 50, and the number of iterations was set to 100. The optimized parameters included the flue gas flow distribution ratio, bypass valve opening, and economizer inlet water flow. The fine-tuned model was used to predict system performance under different parameter configurations. The genetic algorithm was then used to search for the optimal parameter configuration, outputting the optimal parameters and their corresponding system performance.

[0099] Finally, during the verification phase, the optimized parameter configuration was applied to the actual system, running it for a week while collecting system operational data. The optimization results were verified by comparing performance indicators before and after optimization. The bypass parameter configuration table is shown in Table 1.

[0100] Table 1 shows the performance comparison before and after optimization.

[0101]

[0102]

[0103] During the experiment, the air preheater bypass flue gas ratio was increased from 14.15% to 19.20%, and the opening of the feedwater valve and condensate valve damper was adjusted synchronously to adapt to the corresponding changes in the inlet water flow. The experimental results show that under this optimized operating condition, the average system coal consumption rate within the unit operating range increased from 319.53g·(kW·h) -1 Reduced to 318.91g·(kW·h) -1 , the decrease reached 0.62g·(kW·h) -1 The boiler efficiency dropped by only 0.06%, which is almost negligible; the power generation heat consumption was reduced by 81.5 kJ·(kW·h) -1 The above operating data clearly demonstrates the significant impact of bypass parameter changes on the unit's thermal performance. By increasing the air preheater bypass flue gas ratio and optimizing the valve opening, the coal consumption rate and power generation heat rate were effectively reduced while maintaining stable boiler efficiency. The experimental results are consistent with the simulation results in this paper, verifying the accuracy of the simulation model and the effectiveness of the optimization method.

[0104] The above content merely illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of this technical solution belong to the technical idea proposed by the present invention and fall within the protection scope of the claims of the present invention.

Claims

1. A flue gas waste heat cascade utilization system optimization method based on deep transfer learning, characterized in that: The following steps are involved: S1: Establish a simulation model of the flue gas waste heat cascade utilization system based on the EBSILON platform, and obtain high-quality simulation sample data for multiple operating conditions after verifying the accuracy of the simulation model; S1.1: Establish simulation models of key equipment based on the manufacturer's thermal balance diagram and the actual structure and operating parameters of the system; S1.2: Set the flue gas flow rate, temperature, pressure, heat exchanger efficiency, bypass valve opening and environmental parameters; S1.3: Simulate system operating conditions under different load conditions, including full load, part load, and low load conditions; S1.4: Compare the output of the simulation model with the design value and calculate the relative error between the two; S1.5: Determine whether the error index meets the preset accuracy requirements. If not, repeat S1.2 to S1.4 until the error value meets the requirements. S1.6: Based on the verification results, optimize the key parameters in the simulation model to ensure that the output results of the optimized model are consistent with the actual operation data; S1.7: Use the API provided by EBSILON to write MATLAB scripts to automatically batch read the simulation data files generated by EBSILON; S1.8: Preprocess simulation data to generate standardized datasets suitable for deep learning. S2: Deep transfer learning and parameter optimization between simulation data and field operation data; S2.1: Pre-train the deep neural network model using the diverse operating condition simulation dataset generated in step 1; S2.2: Extract the feature representation of the flue gas waste heat cascade utilization system through pre-training to obtain a preliminary pre-training model; S2.3: Collect on-site historical operating data of the flue gas waste heat utilization system; S2.4: Preprocess field operation data; S2.5: Based on the pre-trained model, update the calibration network using the training set of field operation data; S2.6: Optimize model parameters using the backpropagation algorithm to minimize the error between the predicted value and the actual value; S2.7: Use the validation set of field data to evaluate the prediction accuracy of the fine-tuned model and calculate the model performance indicators, including mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R). 2 ; S2.8: Based on the model after the network is calibrated and updated using operational data, the flue gas bypass parameters are optimized using an intelligent optimization algorithm. Through iterative optimization, the optimal flue gas bypass parameter configuration is obtained. S2.9: Apply the optimized flue gas bypass parameter configuration to the real-time control of the flue gas waste heat utilization system, and verify the optimization effect through system feedback.

2. The method for optimizing a flue gas waste heat cascade utilization system based on deep transfer learning according to claim 1, characterized in that: The key equipment in step S1.1 includes a steam turbine, a boiler, a flue gas bypass, a heat exchanger, and a fan.

3. The method for optimizing a flue gas waste heat cascade utilization system based on deep transfer learning according to claim 1, characterized in that: The simulation data acquisition process described in step S1.8 is as follows: calling the unit thermal system model under variable operating conditions, obtaining simulation results under different operating conditions by modifying parameters and setting bypass key parameter constraints, and obtaining the waste heat utilization system simulation data set under different operating conditions including energy consumption indicators such as coal consumption rate and power generation heat consumption through thermal calculation.

4. The method for optimizing a flue gas waste heat cascade utilization system based on deep transfer learning according to claim 1, characterized in that: The implementation process of pre-training the deep neural network model described in step S2.1 is as follows: The input of the deep neural network model is the key parameters of the flue gas bypass, and the output is the unit boiler efficiency η b , turbine heat rate q and coal consumption rate b: Among them, Q zq is the main steam heat; Q gs is the heat of feed water; Q zr is the heat of reheat steam; Q jws The heat of the desuperheating water in the two-stage superheater; Q gp is the heat of the air preheater bypass high pressure economizer; Q dp is the heat of the air preheater bypass low-pressure economizer; B is the coal consumption of the boiler; Pe is the generator output power; q is the turbine heat rate, η b is the boiler efficiency; η p is the pipeline efficiency; Q net It is the low calorific value of standard coal; During the pre-training process, the learning rate lr1 is set to 0.1 and the number of training times Ne1 is set to 40, and the weights and biases in the hidden layer are iteratively updated; the calculation formula of the deep neural network is as follows: in, is the expected output of unit i in μ mode; w ij 、w jk is the weight of the corresponding unit; i, j, k are the corresponding input, hidden and output units; P is the number of input patterns; g is the activation function.

5. The method for optimizing a flue gas waste heat cascade utilization system based on deep transfer learning according to claim 1, characterized in that: The preprocessing of the field operation data described in step S2.4 includes data cleaning, missing value filling, normalization, and outlier detection and processing.

6. The method for optimizing a flue gas waste heat cascade utilization system based on deep transfer learning according to claim 1, characterized in that: In step S2.5, when the amount of on-site running data is limited during the calibration process, the learning rate lr2 is set to 0.001 and the number of training times Ne2 is set to 100, so as to reduce fluctuations while ensuring the training speed.

7. The method for optimizing a flue gas waste heat cascade utilization system based on deep transfer learning according to claim 1, characterized in that: The implementation process of step S2.5 is as follows: In the migration framework, the mapping relationship between the simulation data model and the actual data is expressed as follows: y H =F(x,y L ) Among them, x is the model input parameter; y H and y L are the output values ​​of the field operation data and the corresponding simulation data respectively; The mathematical expression of the calibration network is: F=F1+F n1 =αF1(x,y L )+(1-α)F n1 (x,y L ) Among them, F1 and F n1 are the linear term and nonlinear term in the function F(τ) respectively; α is the weight factor of the linear term in the function F(τ).

8. The method for optimizing a flue gas waste heat cascade utilization system based on deep transfer learning according to claim 1, characterized in that: The implementation process of step S2.6 is as follows: Minimize the loss function to reduce the difference between the predicted value and the actual value: in, is the input sample; is the output sample; ||·||2 represents the L2 norm.

9. The method for optimizing a flue gas waste heat cascade utilization system based on deep transfer learning according to claim 1, characterized in that: The flue gas bypass parameter configuration in step S2.8 includes the flue gas flow distribution ratio, the bypass valve opening, and the bypass economizer water inlet flow.

Images