Automatic optimization and anti-freezing control method for cold end system of thermal power generating unit

Abstract

The invention discloses an automatic optimization and anti-freezing control method for a cold end system of a thermal power generating unit. According to the invention, remote real-time monitoring and control by operators can be realized. And then based on the actual operation condition of the unit of the Qingdong aluminum industry power generation limited company, combining with the environmental parameters of the power plant, establishing a whole-plant thermodynamic system simulation model of the direct air cooling unit, and carrying out deeper numerical simulation research on the variable working condition operation characteristics of the direct air cooling system of the power plant so as to develop a cold end automatic optimization control platform. The optimal operation mode of the cold-end system of the unit under different thermoelectric load and environmental conditions is obtained through the neural network model, so that real-time control and adjustment of cold-end system equipment such as a fan and a plugging and anti-freezing device are completed, and the unit can be controlled and adjusted in real time under the background of deep adjustment in winter. The anti-freezing measures can be automatically executed, adjusted and optimized according to the environment and operation conditions, the reliability of the system is improved, the operation safety of the unit is guaranteed, and good economic benefits and social benefits are achieved.

Description

Technical Field

[0001] This invention belongs to the field of thermal power cold end technology, specifically a method for automatic optimization and antifreeze control of the cold end system of thermal power units. Background Technology

[0002] Direct air cooling systems are a cooling technology used in large thermal power plants or nuclear power plants. They utilize ambient air to directly cool the exhaust steam from the turbine, unlike traditional wet cooling towers that rely on water as the cooling medium. This system uses a large array of finned tube radiators to condense the exhaust steam inside the tubes, while high-powered fans force airflow outside to remove the heat. The entire process involves no water evaporation, achieving a water saving rate of over 90%. Direct air cooling systems offer significant advantages in water-scarce regions. Due to their simple structure, low investment, and significant water-saving effects, they have been widely adopted in recent years in China, especially in the coal-rich but water-scarce northern regions, and are expected to be one of the mainstream water-saving technologies for thermal power units in northern my country for some time to come.

[0003] However, because the air-cooled island is installed in the natural environment, the operation of direct air-cooled units is greatly affected by environmental conditions. In cold northern regions, the air intake temperature of the air-cooled radiators is very low in winter. The extreme minimum temperature in the Qingtongxia area can reach -25°C in winter. Therefore, when starting and stopping the unit in winter, the steam entering the air-cooled island may freeze prematurely before it fully reaches the air-cooled radiators, disrupting the steam-water circulation. During normal operation of the unit, the exhaust steam flow of the thermal power unit will also decrease due to the need for deep peak shaving or steam extraction for external heating, posing a risk of freezing to the air-cooled unit. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic optimization and antifreeze control method for the cold end system of thermal power units in order to solve the problems mentioned above.

[0005] The technical solution adopted in this invention is as follows: an automatic optimization and antifreeze control method for the cold end system of thermal power units, the method comprising the following steps: Step S1: Investigate the latest technologies of direct air-cooled generator sets that are newly built or upgraded with thermal infrared imaging temperature measurement systems and air-cooled island sealing and anti-freezing devices, as well as related equipment manufacturers, universities, and enterprises.

[0006] Step S2: Based on the principle of thermal infrared imaging, develop an online monitoring system for the temperature field of the air-cooled island, which includes a thermal infrared imager, a network system, and a back-end data processing server, and equip it with software to realize thermal image processing, recording, alarm, and structural transmission functions.

[0007] Step S3: Design an air-cooled island sealing and anti-freezing device including power, guide rail, traction and other devices. Install a device with an adjustable coverage area of ​​0% to 100% at the fan inlet. Use figure-eight steel wire wheel group for traction to realize remote / local fully automatic control and DCS screen control.

[0008] Step S4: Use EBSILON software to establish a simulation model of the entire plant's thermal system, simulate the operating characteristics of the main equipment, and obtain the analysis results of the cold end system under varying operating conditions by changing environmental parameters, loads, etc.

[0009] Step S5: Condense prior knowledge into a neural network surrogate model, optimize hyperparameters, train the model based on the analysis results of the cold end system under varying operating conditions, and learn the relationship between heat rate, back pressure, efficiency and the operating status of cold end equipment.

[0010] Step S6: Develop an automatic optimization control platform for the cold end based on a neural network model. Input the unit's operating and environmental parameters, optimize the cold end operation mode with the goals of economy and safety, and realize real-time closed-loop control and remote monitoring interface for the fan and sealing device.

[0011] Step S7: Implement engineering and technical modifications, and analyze and judge the modification effect through performance tests.

[0012] As a preferred embodiment of the present invention, in step S1, the research work mainly focuses on newly built or renovated direct air-cooled generator sets, with a focus on cases of units that have been equipped with thermal infrared imaging temperature measurement systems and air-cooled island sealing and anti-freezing devices. At the same time, visits are made to equipment manufacturers, relevant universities and enterprises that provide such renovation services to comprehensively collect the current application status of thermal infrared imaging temperature measurement technology in air-cooled island monitoring, the latest technical information on the structural design, operational reliability and control logic of sealing and anti-freezing devices, and to sort out the advantages and disadvantages and practical application effects of different technical solutions, so as to accumulate technical reserves and practical references for subsequent system development and device design.

[0013] As a preferred embodiment of the present invention, in step S2, the online monitoring system for the temperature field of the air-cooled island, developed based on the principle of thermal infrared imaging, includes a thermal infrared imager network system and a back-end data processing server as its core components. The network system includes transmission equipment such as network cable switches and optical transceivers. The thermal imager is installed in a suitable location on the air-cooled island to cover the entire monitoring area. After being connected through a network switch, real-time thermal image data is transmitted to a remote data server via optical fiber. The thermal infrared analysis software running in the background processes the collected thermal images in real time and records the temperature distribution. When a local temperature anomaly is detected, an alarm function is automatically triggered to ensure that operators can keep abreast of the dynamics of the temperature field of the air-cooled island and provide data support for anti-freezing control and optimized operation.

[0014] As a preferred embodiment of the present invention, in step S3, the designed air-cooled island sealing and antifreeze device consists of a power unit, a guide rail device, a traction device, a tensioning wheel device, a sealing device, and a control system. The sealing device is installed at the air inlet of the fan unit, and its coverage area can be flexibly adjusted between 0 and 100%. The traction device adopts an 8-shaped steel wire wheel design, which is more reliable and has better synchronization than the conventional roller traction method. The control system has a high level of automation and supports both remote and local control modes to achieve fully automatic operation. Operators can quickly complete the antifreeze operation through remote control, reducing operation time and improving real-time performance and practicality. At the same time, the system is connected to the DCS system to form an intuitive control screen, reducing the risk of misoperation and ensuring accurate and reliable fan start-stop control.

[0015] In a preferred embodiment of the present invention, in step S4, the unit type and equipment characteristic curves are converted into physical constraints for the neural network. Based on this, the optimization range for the neural network structure parameters is determined: for example, the number of network layers is set to 3-7, the neuron type uses a hybrid of fully connected and convolutional units, the activation function candidate set is ReLU / LeakyReLU / Sigmoid, the batch size is 32-128, and the learning rate is 0.001-0.01. The hyperparameters are then iteratively optimized using the Tree-structured Parzen Estimator (TPE) algorithm of the Optuna framework to finally determine a model structure that balances fitting accuracy and computational efficiency.

[0016] The cold-end system variable operating condition analysis results obtained from the EBSILON simulation in step S4 (covering more than 5,000 sets of data under boundary conditions such as ambient temperature -20℃~40℃, unit load 30%~100% THA, fan frequency 15~50Hz, and plugging device opening 0%~100%) were organized into a structured dataset. The dataset includes 23-dimensional parameters such as input features (ambient temperature, humidity, unit power supply load, condenser heat load, fan operating frequency, and plugging device opening) and output labels (heat rate, back pressure, and cold-end system efficiency). The dataset was preprocessed by Z-score standardization and outlier removal (based on the 3σ criterion). This dataset was used to train a neural network surrogate model. During training, overfitting was avoided through dynamic learning rate adjustment (cosine annealing strategy) and early stopping mechanism. A historical data playback mechanism was also introduced to achieve active autonomous learning (e.g., increasing sampling weights for key operating conditions such as high load and low ambient temperature). Ultimately, a surrogate model was constructed that can accurately map the nonlinear relationship between the operating status of cold-end equipment (fan speed, blockage position, etc.) and the unit's heat rate, back pressure, and efficiency. The model test set R... 2 The value reaches above 0.98, and the mean absolute error (MAE) is controlled within 2%.

[0017] As a preferred embodiment of the present invention, in step S4, to balance economy and security, a weighted loss function incorporating physical constraints is adopted during the model training phase, as shown in the following formula: ; In the formula: H pred H true This represents the model-predicted heat rate versus the actual heat rate (kJ / kWh). P b Predicted back pressure (kPa), Pb,lim represents the upper limit of safe back pressure (kPa). η pred η true Indicates the predicted efficiency versus the actual efficiency (%). α, β, γ represent weighting coefficients (dynamically adjusted according to the unit's operating priority, typically α=0.6, β=0.3, γ=0.1).

[0018] To address the nonlinear characteristics of the heat dissipation efficiency of air-cooled islands being affected by the coupling of ambient temperature and airflow, a heat transfer-based correction term is embedded in the hidden layer of the neural network: ; In the formula: η ac This indicates the corrected heat dissipation efficiency of the air-cooled island; η0 represents the baseline efficiency (%) under the design conditions; T amb Tdes represents the ambient temperature (°C), and Tdes represents the design ambient temperature (°C). Q air Q air,rated This indicates the difference between actual air volume and rated air volume (m³). 3 / h); c1 and c2 represent empirical coefficients (obtained by fitting variable working condition data, typically c1=0.12, c2=0.08).

[0019] When optimizing neural network hyperparameters using the Optuna framework, a comprehensive scoring function is defined with the goal of improving model prediction accuracy and generalization ability. The scoring formula is as follows: ; In the formula: R 2 The coefficient of determination (measures the degree of explanation for the variation in the output variable; the closer to 1, the better). RMSE val RMSE train This represents the root mean square error between the validation set and the training set. The ratio of the two reflects the degree of overfitting (the smaller the ratio, the stronger the generalization ability).

[0020] In a preferred embodiment of the present invention, step S5 first involves condensing key prior knowledge based on the type characteristics of the direct air-cooled unit, including the energy balance relationship in the first law of thermodynamics, the heat transfer characteristic curve of the air-cooled island radiator, and the power characteristic equation of the fan, etc., and transforming these physical laws into implicit constraints of the neural network. Next, the optimization range of the neural network hyperparameters is determined, with the number of network layers set to 3 to 7, the neuron type adopting a hybrid structure of fully connected and convolutional units, the activation function candidate set including ReLU, LeakyReLU, and Sigmoid, the batch size selected between 32 and 128, and the learning rate range determined to be 0.001 to 0.01. Subsequently, the hyperparameters are iteratively optimized using the Tree-structured Parzen Estimator algorithm of the Optuna framework. After multiple rounds of testing, the final model structure is determined to be a 5-layer network, with the number of neurons in each layer being 128, 64, 32, 16, and 3 respectively, the activation function selected being LeakyReLU, the batch size set to 64, and the learning rate adjusted to 0.005. The model training phase utilizes a cold-end system variable operating condition analysis dataset. This dataset encompasses over 5000 data points under various boundary conditions, including ambient temperatures ranging from -20℃ to 40℃, unit loads from 30% to 100% THA, fan frequencies from 15 to 50Hz, and plugging device opening from 0% to 100%. Input features include 23 dimensions such as ambient temperature, humidity, unit power load, condenser heat load, fan operating frequency, and plugging device opening. Output labels correspond to heat rate, back pressure, and cold-end system efficiency. In the data preprocessing phase, Z-score normalization is used to normalize the input features, and outliers are removed using the 3σ criterion. During training, a cosine annealing strategy is employed to dynamically adjust the learning rate, combined with an early stop mechanism to prevent overfitting. A historical data playback mechanism is also introduced to increase the sampling weight of key operating conditions such as high load and low ambient temperature. The resulting neural network surrogate model accurately maps the nonlinear relationship between the operating status of cold-end equipment and heat rate, back pressure, and efficiency. The coefficient of determination R0 on the model test set is [value missing]. 2 The accuracy reaches above 0.98, and the average absolute error is controlled within 2%.

[0021] In a preferred embodiment of the present invention, in step S6, the cold-end automatic optimization control platform uses the neural network proxy model trained in step S5 as its core algorithm. It takes real-time operating parameters of the unit (such as current power supply load, main steam pressure, condenser heat load, and inlet and outlet temperatures of each unit in the air-cooled island) and environmental parameters (ambient temperature, relative humidity, and wind speed) as input variables. Through its built-in multi-objective optimization logic, it comprehensively balances economic and safety objectives. The economic objective focuses on minimizing the unit's heat consumption rate, calculating heat consumption values ​​under different cold-end equipment combinations using the model. The safety objective focuses on preventing the risk of freezing in the air-cooled island, combining real-time temperature field data from the monitoring system in step S2 to ensure that adjustments to fan speed and sealing device opening do not cause local tube bundle temperatures to fall below freezing. The platform handles target conflicts through a dynamic weight allocation mechanism (such as prioritizing safety in extreme low temperature environments) and uses heuristic search algorithms (such as rapid iterative optimization based on historical best solutions) to solve the optimal operating mode of the cold end system under the current operating conditions in milliseconds. The specific outputs include the target operating frequency of each air-cooled fan, the real-time adjustment opening of the blocking device, and the backup control strategy under abnormal operating conditions (such as the load transfer scheme when a fan fails).

[0022] As a preferred embodiment of the present invention, in step S6, the platform has a dual mode of fully automated control and manual intervention. In automatic mode, the system converts the optimization results into control commands and sends them to the fan inverter and the blocking device actuator in real time via the industrial bus to achieve closed-loop regulation of the cold end equipment. At the same time, it continuously receives equipment feedback signals (such as the actual fan speed and the position feedback of the blocking mechanism) and verifies the control effect through a neural network model. If a deviation occurs (such as the actual back pressure being higher than the optimal value), a secondary optimization is immediately triggered. To support operator decision-making, the platform has developed a visual operation interface, integrating a real-time data dashboard (dynamically displaying key indicators such as heat rate, back pressure, and air-cooled island temperature field distribution), trend curve analysis (historical heat rate and ambient temperature correlation curves, and optimal back pressure curves under different loads), an abnormal alarm module (automatic pop-up prompts when the local temperature is lower than the antifreeze threshold or the heat rate deviates from the benchmark value by 5%), and a manual control panel (allowing operators to temporarily lock the fan frequency or adjust the blocking opening and record the impact of intervention operations on economic efficiency). The interface design follows the DCS operation habits of the power industry, adopts a layered menu structure, and uses red, yellow and green to distinguish the over-limit, warning and normal status of key parameters, so as to ensure that operators can quickly grasp the system status and remotely complete start-up, shutdown and adjustment operations.

[0023] As a preferred embodiment of the present invention, in step S7, during the engineering modification implementation phase, a thermal infrared imager temperature measurement system and an air-cooled island sealing and antifreeze device need to be installed according to the design scheme. The hardware and software system of the cold-end automatic optimization control platform needs to be deployed, and the equipment wiring and debugging and communication docking with the unit's DCS system need to be completed. After the modification is completed, the system function and operation effect are verified through performance tests. The test content includes testing the accuracy and response speed of the temperature field monitoring system, the adjustment accuracy and antifreeze effect of the sealing device, and the optimization efficiency and control stability of the cold-end automatic optimization control platform under different ambient temperatures and unit load conditions. By comparing the unit heat consumption rate, back pressure and antifreeze accident incidence rate before and after the modification, the improvement effect of the modification on the economy and safety of the cold-end system is comprehensively evaluated.

[0024] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In this invention, a refined simulation platform is first built based on the actual structural parameters and operating characteristics of thermal power units. The platform simulates the dynamic changes of key parameters such as ambient temperature, humidity, unit load, fan frequency, and sealing opening under extreme and normal operating conditions, effectively expanding the learning space of the subsequent neural network model. Simultaneously, during the simulation, a strict variable control method is used to isolate the influence of different factors on the performance of the cold-end system, ensuring that the dataset accurately reflects the nonlinear mapping relationship between various operating parameters and economic and safety indicators, avoiding the problem of fuzzy patterns caused by multi-variable coupling interference in actual operating data. This enables the trained neural network surrogate model to accurately capture the dynamic characteristics of the cold-end system under complex operating conditions, providing reliable predictive capabilities for the automatic optimization control platform. This ensures that while pursuing the minimization of heat rate, the risk of air-cooled island freezing can be identified and avoided in a timely manner, ultimately achieving safe and economical operation of the cold-end system across the entire operating range.

[0025] 2. In this invention, the online temperature field monitoring system for air-cooled islands, developed based on the principle of thermal infrared imaging, achieves comprehensive real-time sensing of the surface temperature of the air-cooled island through non-contact measurement. Compared with traditional contact temperature measurement methods, it effectively avoids problems such as limited measurement point layout and complex equipment installation and maintenance, and can completely capture the temperature distribution characteristics of each area of ​​the air-cooled island. The system's background service functions not only support real-time data monitoring and anomaly warning, but also immediately trigger alarms when local temperature anomalies occur (such as approaching freezing point or overheating), giving operators time for emergency response. Furthermore, through historical data storage and trend analysis, it provides data support for cold-end system performance evaluation. More importantly, the system directly participates in cold-end control and regulation, using temperature field data as a key input for control strategy optimization, shifting cold-end system regulation from "experience-based judgment" to "real-time status sensing," thus improving the accuracy and timeliness of control decisions.

[0026] 3. In this invention, addressing the deviation between operational and design data caused by errors in measurement point data, changes in system state, and equipment performance degradation in power plant production, this study integrates prior knowledge such as the principles of thermal systems and equipment characteristic curves into a neural network surrogate model for the cold end system. This effectively overcomes the excessive reliance on sample quality in purely data-driven models. By introducing physical laws such as energy balance relationships and heat transfer characteristics as model constraints, the neural network surrogate model can maintain an accurate mapping of the cold end system's operational laws even when anomalies such as missing jumps or dead numbers exist in the production data, avoiding model distortion caused by data quality issues. This modeling method, which integrates prior knowledge, enables the neural network model to possess both data fitting capabilities and physical interpretability, ensuring the stability and reliability of the cold end system control strategy under complex operating conditions and providing a reliable predictive basis for subsequent automatic optimization.

[0027] 4. In this invention, the original unit antifreeze measures relied on manual experience for adjustment, which suffered from problems such as response lag and insufficient accuracy. However, the cold-end automatic optimization control platform, developed based on a plant-wide thermal system simulation model and neural network model, fully integrates real-time data from the online temperature field monitoring system of the air-cooled island with the rapid adjustment capability of the air-cooled island sealing and antifreeze device. By dynamically analyzing multi-dimensional parameters such as heat and power load, ambient temperature, and humidity, the platform can automatically solve for the optimal operating mode of the cold-end system under different operating conditions, achieving precise matching between fan frequency and sealing device opening. This ensures that the temperature in each area of ​​the air-cooled island is maintained within a safe range (avoiding freezing or overheating) while minimizing the unit's heat consumption rate. This closed-loop control mode not only replaces the subjectivity and lag of traditional manual operation but also fully utilizes the adjustment potential of the cold-end equipment, enabling the cold-end system to achieve a dynamic balance between safety and economy, significantly improving the unit's operating efficiency across the entire operating range. Attached Figure Description

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0029] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0030] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0031] Figure 1 This is a schematic diagram of the online temperature field monitoring system software for the air-cooled island of the present invention; Figure 2 This is a schematic diagram of the online temperature field monitoring system for the hollow cold island in this invention; Figure 3 This is a schematic diagram of the control system of the hollow cooling island sealing device in this invention; Figure 4 This is a schematic diagram of the EBSILON software modeling and calculation in this invention; Figure 5 This is a schematic diagram of the process principle in this invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] Example: Reference Figure 1-5 , An automatic optimization and antifreeze control method for the cold-end system of thermal power units, the method includes the following steps: Step S1: Investigate the latest technologies of direct air-cooled generator sets that are newly built or upgraded with thermal infrared imaging temperature measurement systems and air-cooled island sealing and anti-freezing devices, as well as related equipment manufacturers, universities, and enterprises.

[0034] Step S2: Based on the principle of thermal infrared imaging, develop an online monitoring system for the temperature field of the air-cooled island, which includes a thermal infrared imager, a network system, and a back-end data processing server, and equip it with software to realize thermal image processing, recording, alarm, and structural transmission functions.

[0035] Step S3: Design an air-cooled island sealing and anti-freezing device including power, guide rail, traction and other devices. Install a device with an adjustable coverage area of ​​0% to 100% at the fan inlet. Use figure-eight steel wire wheel group for traction to realize remote / local fully automatic control and DCS screen control.

[0036] Step S4: Use EBSILON software to establish a simulation model of the entire plant's thermal system, simulate the operating characteristics of the main equipment, and obtain the analysis results of the cold end system under varying operating conditions by changing environmental parameters, loads, etc.

[0037] Step S5: Condense prior knowledge into a neural network surrogate model, optimize hyperparameters, train the model based on the analysis results of the cold end system under varying operating conditions, and learn the relationship between heat rate, back pressure, efficiency and the operating status of cold end equipment.

[0038] Step S6: Develop an automatic optimization control platform for the cold end based on a neural network model. Input the unit's operating and environmental parameters, optimize the cold end operation mode with the goals of economy and safety, and realize real-time closed-loop control and remote monitoring interface for the fan and sealing device.

[0039] Step S7: Implement engineering and technical modifications, and analyze and judge the modification effect through performance tests.

[0040] In step S1, the survey mainly focuses on newly built or renovated direct air-cooled generator sets, with a particular emphasis on cases where thermal infrared imaging temperature measurement systems and air-cooled island sealing and anti-freezing devices have been added. At the same time, visits are made to equipment manufacturers, relevant universities and enterprises that provide such renovation services to comprehensively collect the current application status of thermal infrared imaging temperature measurement technology in air-cooled island monitoring, the latest technical information on the structural design, operational reliability and control logic of sealing and anti-freezing devices, and to sort out the advantages and disadvantages and practical application effects of different technical solutions, so as to accumulate technical reserves and practical references for subsequent system development and device design.

[0041] In step S2, the online temperature field monitoring system for the air-cooled island, developed based on the principle of thermal infrared imaging, includes a thermal infrared imager network system and a back-end data processing server as its core components. The network system includes transmission equipment such as network cable switches and optical transceivers. The thermal imager is installed in a suitable location on the air-cooled island to cover the entire monitoring area. After being connected through a network switch, real-time thermal image data is transmitted to a remote data server via optical fiber. The thermal infrared analysis software running in the background processes the collected thermal images in real time and records the temperature distribution. When a local temperature anomaly is detected, an alarm function is automatically triggered to ensure that operators can keep abreast of the dynamics of the air-cooled island temperature field and provide data support for anti-freezing control and optimized operation.

[0042] In step S3, the designed air-cooled island sealing and antifreeze device consists of a power unit, a guide rail device, a traction device, a tensioning wheel device, a sealing device, and a control system. The sealing device is installed at the air inlet of the fan unit, and its coverage area can be flexibly adjusted between 0 and 100%. The traction device adopts an 8-shaped steel wire wheel design, which is more reliable and has better synchronization than the conventional roller traction method. The control system has a high level of automation and supports both remote and local control modes to achieve fully automatic operation. Operators can quickly complete the antifreeze operation through remote control, reducing operation time and improving real-time performance and practicality. At the same time, the system is connected to the DCS system to form an intuitive control screen, reducing the risk of misoperation and ensuring accurate and reliable fan start-up and shutdown control.

[0043] In step S4, the unit type and equipment characteristic curves are transformed into physical constraints for the neural network. Based on this, the optimization range of the neural network structure parameters is determined: for example, the number of network layers is set to 3-7, the neuron type uses a hybrid of fully connected and convolutional units, the candidate set of activation functions is ReLU / LeakyReLU / Sigmoid, the batch size is 32-128, and the learning rate is 0.001-0.01. The hyperparameters are then iteratively optimized using the Tree-structured Parzen Estimator algorithm of the Optuna framework to finally determine a model structure that balances fitting accuracy and computational efficiency.

[0044] The variable operating condition analysis results of the cold-end system obtained from the EBSILON simulation in step S4 were compiled into a structured dataset, including input features and output labels, and preprocessed by Z-score standardization and outlier removal. This dataset was used to train a neural network surrogate model. During training, overfitting was avoided through dynamic learning rate adjustment and an early stop mechanism. A historical data playback mechanism was also introduced to achieve active autonomous learning. Ultimately, a surrogate model capable of accurately mapping the nonlinear relationship between the operating status of cold-end equipment and the unit's heat rate, back pressure, and efficiency was constructed. The model test set R... 2 The value reaches 0.98 or higher, and the mean absolute error (MAE) is controlled within 2%. In step S4, to balance economy and safety, a weighted loss function incorporating physical constraints is used during the model training phase, as shown in the following formula: ; In the formula: H pred H true This represents the model-predicted heat rate versus the actual heat rate (kJ / kWh). P b Predicted back pressure (kPa), Pb,lim represents the upper limit of safe back pressure (kPa). η pred η true Indicates the predicted efficiency versus the actual efficiency (%). α, β, γ represent weighting coefficients (dynamically adjusted according to the unit's operating priority, typically α=0.6, β=0.3, γ=0.1).

[0045] To address the nonlinear characteristics of the heat dissipation efficiency of air-cooled islands being affected by the coupling of ambient temperature and airflow, a heat transfer-based correction term is embedded in the hidden layer of the neural network: ; In the formula: η ac This indicates the corrected heat dissipation efficiency of the air-cooled island; η0 represents the baseline efficiency (%) under the design conditions; T amb Tdes represents the ambient temperature (°C), and Tdes represents the design ambient temperature (°C). Q air Q air,rated This indicates the difference between actual air volume and rated air volume (m³). 3 / h); c1 and c2 represent empirical coefficients (obtained by fitting variable working condition data, typically c1=0.12, c2=0.08).

[0046] When optimizing neural network hyperparameters using the Optuna framework, a comprehensive scoring function is defined with the goal of improving model prediction accuracy and generalization ability. The scoring formula is as follows: ; In the formula: R 2 The coefficient of determination (measures the degree of explanation for the variation in the output variable; the closer to 1, the better). RMSE val RMSE train This represents the root mean square error between the validation set and the training set. The ratio of the two reflects the degree of overfitting (the smaller the ratio, the stronger the generalization ability).

[0047] In step S5, key prior knowledge is first extracted based on the type characteristics of the direct air-cooled unit, including the energy balance relationship in the first law of thermodynamics, the heat transfer characteristic curve of the air-cooled island radiator, and the power characteristic equation of the fan. These physical laws are then transformed into implicit constraints for the neural network. Next, the optimization range of the neural network hyperparameters is determined, with the number of network layers set to 3 to 7, the neuron type adopting a hybrid structure of fully connected and convolutional units, the activation function candidate set including ReLU, LeakyReLU, and Sigmoid, the batch size selected between 32 and 128, and the learning rate ranged from 0.001 to 0.01. Subsequently, the hyperparameters are iteratively optimized using the Tree-structured Parzen Estimator algorithm of the Optuna framework. After multiple rounds of testing, the final model structure is determined to be a 5-layer network, with the number of neurons in each layer being 128, 64, 32, 16, and 3 respectively, the activation function selected being LeakyReLU, the batch size set to 64, and the learning rate adjusted to 0.005. The model training phase uses a cold-end system variable operating condition analysis dataset, which covers more than 5,000 sets of data under various boundary conditions, such as ambient temperature from -20℃ to 40℃, unit load from 30% to 100% THA, fan frequency from 15 to 50Hz, and plugging device opening from 0 to 100%. The input features include 23-dimensional parameters such as ambient temperature, humidity, unit power supply load, condenser heat load, fan operating frequency, and plugging opening. The output labels correspond to heat rate, back pressure, and cold-end system efficiency.

[0048] In step S5, the data preprocessing stage uses Z-score normalization to normalize the input features and removes outliers using the 3σ criterion. During training, a cosine annealing strategy is used to dynamically adjust the learning rate, combined with an early stopping mechanism to prevent overfitting. A historical data replay mechanism is also introduced to increase the sampling weight of key operating conditions such as high load and low ambient temperature. The resulting neural network surrogate model accurately maps the nonlinear relationship between the operating status of cold-end equipment and heat dissipation rate, back pressure, and efficiency. The coefficient of determination R0 on the model test set is [value missing]. 2 The accuracy reaches above 0.98, and the average absolute error is controlled within 2%.

[0049] In step S6, the cold-end automatic optimization control platform uses the neural network proxy model trained in step S5 as its core algorithm. It takes the unit's real-time operating parameters and environmental parameters as input variables and comprehensively balances economic and safety objectives through its built-in multi-objective optimization logic. The economic objective focuses on minimizing the unit's heat rate, calculating heat consumption values ​​under different cold-end equipment combinations using the model. The safety objective prioritizes preventing the risk of air-cooled island freezing, ensuring that adjustments to fan speed and sealing device opening do not cause local tube bundle temperatures to drop below freezing point, based on real-time temperature field data from the monitoring system in step S2. The platform handles objective conflicts through a dynamic weight allocation mechanism and uses a heuristic search algorithm to solve for the optimal operating mode of the cold-end system under current conditions within milliseconds. Specific outputs include the target operating frequency of each air-cooled fan, the real-time adjustment opening of the sealing device, and backup control strategies for abnormal operating conditions.

[0050] In step S6, the platform features both fully automated control and manual intervention modes. In automatic mode, the system converts the optimization results into control commands, which are then sent in real-time to the wind turbine inverter and the actuator of the blocking device via the industrial bus, achieving closed-loop regulation of the cold-end equipment. Simultaneously, it continuously receives equipment feedback signals and uses a neural network model to reverse-check the control effect; if a deviation occurs, a secondary optimization is immediately triggered. To support operator decision-making, the platform has developed a visual operation interface, integrating real-time data dashboards, trend curve analysis, anomaly alarm modules, and a manual control panel. The interface design follows the DCS operation habits of the power industry, employing a layered menu structure. Key parameters are distinguished by red, yellow, and green to indicate over-limit, warning, and normal states, ensuring that operators can quickly grasp the system status and remotely complete start-up, shutdown, and adjustment operations.

[0051] In step S7, during the engineering renovation implementation phase, the thermal infrared imaging temperature measurement system and the air-cooled island sealing and antifreeze device need to be installed according to the design plan. The hardware and software system of the cold-end automatic optimization control platform should be deployed, and the equipment wiring and debugging and communication docking with the unit's DCS system should be completed. After the renovation is completed, the system functions and operating effects should be verified through performance tests. The test content includes testing the accuracy and response speed of the temperature field monitoring system, the adjustment accuracy and antifreeze effect of the sealing device, and the optimization efficiency and control stability of the cold-end automatic optimization control platform under different ambient temperatures and unit load conditions. By comparing the unit heat consumption rate, back pressure and antifreeze accident incidence before and after the renovation, the improvement effect of the renovation on the economy and safety of the cold-end system should be comprehensively evaluated.

[0052] From the above, we can conclude that: This invention first establishes a refined simulation platform based on the actual structural parameters and operating characteristics of thermal power units. The platform simulates the dynamic changes of key parameters such as ambient temperature, humidity, unit load, fan frequency, and sealing opening under extreme and normal operating conditions, effectively expanding the learning space of the subsequent neural network model. Simultaneously, the simulation process employs a strict variable control method to isolate the influence of different factors on the performance of the cold-end system, ensuring that the dataset accurately reflects the nonlinear mapping relationship between various operating parameters and economic and safety indicators. This avoids the problem of ambiguity caused by multi-variable coupling interference in actual operating data. This enables the trained neural network surrogate model to accurately capture the dynamic characteristics of the cold-end system under complex operating conditions, providing reliable predictive capabilities for the automatic optimization control platform. It ensures that while minimizing heat rate, it can promptly identify and avoid the risk of air-cooled island freezing, ultimately achieving safe and economical operation of the cold-end system across the entire operating range.

[0053] This invention presents an online temperature field monitoring system for air-cooled islands, developed based on the principle of thermal infrared imaging. This system achieves comprehensive real-time sensing of the surface temperature of the air-cooled island through non-contact measurement. Compared to traditional contact temperature measurement methods, it effectively avoids problems such as limited measurement point placement and complex equipment installation and maintenance, and can completely capture the temperature distribution characteristics of each area of ​​the air-cooled island. The system's backend service functions not only support real-time data monitoring and anomaly warnings, triggering alarms immediately when local temperature anomalies occur (such as near freezing point or overheating), giving operators time for emergency response, but also provide data support for cold-end system performance evaluation through historical data storage and trend analysis. More importantly, the system directly participates in cold-end control and regulation, using temperature field data as a key input for control strategy optimization. This shifts cold-end system regulation from "experience-based judgment" to "real-time status perception," improving the accuracy and timeliness of control decisions.

[0054] This invention addresses the deviation between operational and design data caused by errors in measurement data, system state changes, and equipment performance degradation in power plant production. It integrates prior knowledge such as thermodynamic system principles and equipment characteristic curves into a neural network surrogate model for the cold-end system, effectively overcoming the over-reliance on sample quality in purely data-driven models. By introducing physical laws such as energy balance relationships and heat transfer characteristics as model constraints, the neural network surrogate model can maintain an accurate mapping of the cold-end system's operational patterns even when production data exhibits anomalies such as missing jumps or dead numbers, avoiding model distortion caused by data quality issues. This modeling method, which integrates prior knowledge, endows the neural network model with both data fitting capabilities and physical interpretability, ensuring the stability and reliability of the cold-end system control strategy under complex operating conditions and providing a credible predictive basis for subsequent automatic optimization.

[0055] In this invention, the original unit antifreeze measures relied on manual experience for adjustment, resulting in problems such as response lag and insufficient accuracy. However, the cold-end automatic optimization control platform, developed based on a plant-wide thermal system simulation model and a neural network model, fully integrates real-time data from the online temperature field monitoring system of the air-cooled island with the rapid adjustment capability of the air-cooled island sealing and antifreeze device. By dynamically analyzing multi-dimensional parameters such as heat and power load, ambient temperature, and humidity, the platform can automatically solve for the optimal operating mode of the cold-end system under different operating conditions, achieving precise matching between fan frequency and sealing device opening. This ensures that the temperature in each area of ​​the air-cooled island is maintained within a safe range (avoiding freezing or overheating) while minimizing the unit's heat consumption rate. This closed-loop control mode not only replaces the subjectivity and lag of traditional manual operation but also fully utilizes the adjustment potential of the cold-end equipment, achieving a dynamic balance between safety and economy in the cold-end system, significantly improving the unit's operating efficiency across the entire operating range.

[0056] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An automatic optimization and antifreeze control method for the cold end system of a thermal power unit, characterized in that: The method includes the following steps: Step S1: Investigate the latest technologies of direct air-cooled generator sets that are newly built or upgraded with thermal infrared imaging temperature measurement systems and air-cooled island sealing and anti-freezing devices, as well as related equipment manufacturers, universities and enterprises; Step S2: Based on the principle of thermal infrared imaging, develop an online monitoring system for the temperature field of the air-cooled island, which includes a thermal infrared imager, a network system and a back-end data processing server, and equip it with software to realize thermal image processing, recording, alarm and structural transmission functions. Step S3: Design an air-cooled island sealing and anti-freezing device including power, guide rail and traction device. Install a device with an adjustable coverage area of ​​0% to 100% at the air inlet of the fan. Use figure-eight steel wire wheel group for traction to realize remote / local fully automatic control and DCS screen control. Step S4: Use EBSILON software to establish a simulation model of the entire plant's thermal system, simulate the operating characteristics of the equipment, and obtain the analysis results of the cold end system under varying operating conditions by changing environmental parameters and loads. Step S5: Refine prior knowledge and embed it into a neural network surrogate model, optimize hyperparameters, and train the model based on the analysis results of the cold end system under varying operating conditions to learn the relationship between heat rate, back pressure, efficiency and the operating status of cold end equipment. Step S6: Develop an automatic optimization control platform for the cold end based on a neural network model. Input the unit's operating and environmental parameters, optimize the cold end operation mode with the goals of economy and safety, and realize real-time closed-loop control and remote monitoring interface for the fan and sealing device. Step S7: Implement engineering and technical modifications, and analyze and judge the modification effect through performance tests.

2. The automatic optimization and antifreeze control method for the cold end system of a thermal power unit as described in claim 1, characterized in that: In step S1, the survey focuses on newly built or renovated direct air-cooled generator sets, with particular attention to cases where thermal infrared imaging temperature measurement systems and air-cooled island sealing and anti-freezing devices have been added. At the same time, visits are made to equipment manufacturers, relevant universities and enterprises that provide such renovation services. The survey comprehensively collects the current application status of thermal infrared imaging temperature measurement technology in air-cooled island monitoring, the latest technical information on the structural design, operational reliability and control logic of sealing and anti-freezing devices, and sorts out the advantages and disadvantages and practical application effects of different technical solutions, so as to accumulate technical reserves and practical references for subsequent system development and device design.

3. The automatic optimization and antifreeze control method for the cold end system of a thermal power unit as described in claim 1, characterized in that: In step S2, the online temperature field monitoring system for the air-cooled island, developed based on the principle of thermal infrared imaging, includes a thermal infrared imager network system and a back-end data processing server. The network system encompasses network cable switches, optical transceivers, and transmission equipment. The thermal imager is installed in a suitable location on the air-cooled island to cover the entire monitoring area. After being connected via a network switch, real-time thermal image data is transmitted to a remote data server via optical fiber. The thermal infrared analysis software running in the background processes the collected thermal images in real time and records the temperature distribution. When a local temperature anomaly is detected, an alarm function is automatically triggered to ensure that operators can promptly grasp the dynamics of the air-cooled island temperature field and provide data support for anti-freezing control and optimized operation.

4. The automatic optimization and antifreeze control method for the cold end system of a thermal power unit as described in claim 1, characterized in that: In step S3, the designed air-cooled island sealing and antifreeze device consists of a power unit, a guide rail device, a traction device, a tensioning wheel device, a sealing device, and a control system. The sealing device is installed at the air inlet of the fan unit, and its coverage area can be flexibly adjusted between 0 and 100%. The traction device adopts an 8-shaped steel wire wheel design, which is more reliable and has better synchronization than the conventional roller traction method. The control system has a high level of automation and supports both remote and local control modes to achieve fully automatic operation. Operators can quickly complete the antifreeze operation through remote control, reducing operation time and improving real-time performance and practicality. At the same time, the system is connected to the DCS system to form an intuitive control screen, reducing the risk of misoperation and ensuring accurate and reliable fan start-up and shutdown control.

5. The automatic optimization and antifreeze control method for the cold end system of a thermal power unit as described in claim 1, characterized in that: In step S4, the unit type and equipment characteristic curves are converted into physical constraints for the neural network; based on this, the optimization range of the neural network structure parameters is determined. The variable operating condition analysis results of the cold-end system obtained from the EBSILON simulation in step S4 were organized into a structured dataset, including input features and output labels, and preprocessed by Z-score standardization and outlier removal. This dataset was used to train a neural network surrogate model. During training, overfitting was avoided through dynamic learning rate adjustment and an early stop mechanism. A historical data playback mechanism was also introduced to achieve active autonomous learning. Finally, a surrogate model capable of accurately mapping the nonlinear relationship between the operating status of cold-end equipment and the unit's heat rate, back pressure, and efficiency was constructed. The model test set R... 2 The value reaches above 0.98, and the average absolute error is controlled within 2%.

6. The automatic optimization and antifreeze control method for the cold end system of a thermal power unit as described in claim 1, characterized in that: In step S4, to balance economy and safety, a weighted loss function incorporating physical constraints is used during the model training phase, as shown in the following formula: ； In the formula: H pred H true This represents the difference between the model-predicted heat rate and the actual heat rate. P b Predicted back pressure, Pb,lim represents the upper limit of safe back pressure; η pred η true This represents the difference between predicted efficiency and actual efficiency. α, β, γ represent weighting coefficients; To address the nonlinear characteristics of the heat dissipation efficiency of air-cooled islands being affected by the coupling of ambient temperature and airflow, a heat transfer-based correction term is embedded in the hidden layer of the neural network: ； In the formula: η ac This indicates the corrected heat dissipation efficiency of the air-cooled island; η0 represents the baseline efficiency under the design conditions; T amb Tdes represents the ambient temperature, while Tdes represents the design ambient temperature. Q air Q air,rated This indicates the difference between actual air volume and rated air volume. c1 and c2 represent empirical coefficients; When optimizing neural network hyperparameters using the Optuna framework, a comprehensive scoring function is defined with the goal of improving model prediction accuracy and generalization ability. The scoring formula is as follows: ； In the formula: R 2 Indicates the coefficient of determination of the model; RMSE val RMSE train This represents the root mean square error between the validation set and the training set; the ratio of the two reflects the degree of overfitting.

7. The automatic optimization and antifreeze control method for the cold end system of a thermal power unit as described in claim 1, characterized in that: In step S5, key prior knowledge is first extracted based on the type characteristics of the direct air-cooled unit, including the energy balance relationship in the first law of thermodynamics, the heat transfer characteristic curve of the air-cooled island radiator, and the power characteristic equation of the fan. These physical laws are then transformed into implicit constraints of the neural network. Next, the optimization range of the neural network hyperparameters is determined, with the number of network layers set to 3 to 7, the neuron type adopting a hybrid structure of fully connected and convolutional units, the activation function candidate set including ReLU, LeakyReLU, and Sigmoid, the batch size selected between 32 and 128, and the learning rate range determined to be 0.001 to 0.

01. Subsequently, the hyperparameters are iteratively optimized using the Tree-structured Parzen Estimator algorithm of the Optuna framework. After multiple rounds of testing, the model structure is finally determined to be a 5-layer network, with the number of neurons in each layer being 128, 64, 32, 16, and 3 respectively, the activation function selected being LeakyReLU, the batch size set to 64, and the learning rate adjusted to 0.

005.

8. The automatic optimization and antifreeze control method for the cold end system of a thermal power unit as described in claim 1, characterized in that: In step S6, the cold-end automatic optimization control platform uses the neural network proxy model trained in step S5 as the core algorithm basis, takes the real-time operating parameters of the unit and environmental parameters as input variables, and comprehensively balances the economic and safety objectives through the built-in multi-objective optimization logic; among them, the economic objective focuses on minimizing the unit's heat consumption rate, and calculates the heat consumption value under different cold-end equipment combination operation modes through the model; The safety objective focuses on preventing the risk of freezing of the air-cooled island. By combining the real-time temperature field data of the monitoring system in step S2, it ensures that the adjustment of the fan speed and the opening of the sealing device will not cause the local tube bundle temperature to drop below the freezing point. The platform handles target conflicts through a dynamic weight allocation mechanism and uses a heuristic search algorithm to solve the optimal operating mode of the cold end system under the current operating conditions in milliseconds. The specific outputs include the target operating frequency of each air-cooled fan, the real-time adjustment opening of the sealing device, and the backup control strategy under abnormal operating conditions.

9. The automatic optimization and antifreeze control method for the cold end system of a thermal power unit as described in claim 1, characterized in that: In step S6, the platform has both fully automated control and manual intervention modes. In automatic mode, the system converts the optimization results into control commands and sends them to the fan inverter and the sealing device actuator in real time via the industrial bus to achieve closed-loop regulation of the cold end equipment. At the same time, it continuously receives equipment feedback signals and uses a neural network model to reverse verify the control effect. If a deviation occurs, a second optimization is immediately triggered. To support operators' decision-making, the platform has developed a visual operation interface that integrates real-time data dashboards, trend curve analysis, anomaly alarm modules, and manual control panels. The interface design follows the DCS operation habits of the power industry, adopts a layered menu structure, and uses red, yellow and green to distinguish the over-limit, warning and normal status of key parameters, so as to ensure that operators can quickly grasp the system status and remotely complete start-up, shutdown and adjustment operations.

10. The automatic optimization and antifreeze control method for the cold end system of a thermal power unit as described in claim 1, characterized in that: In step S7, during the engineering renovation implementation phase, the thermal infrared imager temperature measurement system and the air-cooled island sealing and antifreeze device need to be installed according to the design plan. The hardware and software system of the cold end automatic optimization control platform should be deployed, and the equipment wiring and debugging and communication docking with the unit's DCS system should be completed. After the renovation is completed, the system function and operation effect should be verified through performance tests. The test content includes testing the accuracy and response speed of the temperature field monitoring system, the adjustment accuracy and antifreeze effect of the sealing device, and the optimization efficiency and control stability of the cold end automatic optimization control platform under different ambient temperatures and unit load conditions.

Images