A data-driven optimization auxiliary decision-making method for steam turbine cold end

Abstract

This invention provides a data-driven auxiliary decision-making method for cold-end optimization of steam turbines, relating to the field of steam turbine cold-end system optimization and control technology. The method includes: preprocessing historical data of thermal power units to obtain a quasi-steady-state database of unit operation; establishing a data-driven model between turbine back pressure and condenser-related variables using support vector regression; establishing an optimization model for maximizing the net power of the power plant based on the relationship between the total power of the turbine and the operating power of the cold-end system; and solving the optimization model using artificial neural networks and a hybrid improved particle swarm optimization algorithm to obtain the optimal cold-end operation scheme. The hybrid improved particle swarm optimization algorithm includes simulated annealing and particle swarm optimization algorithms. The decision-making method provided by this invention can realize the operation planning of the cold-end system under optimal conditions, thereby maximizing the safe and efficient operation of thermal power units and achieving energy conservation and emission reduction.

Description

Technical Field

[0001] This invention relates to the field of steam turbine cold-end system optimization control technology, and more specifically, to a data-driven steam turbine cold-end optimization auxiliary decision-making method. Background Technology

[0002] The cold-end system of a steam turbine is a crucial auxiliary system for thermal power units in power plants. Generally, lower back pressure during unit operation is more conducive to increasing power generation, but this is usually achieved at the cost of increased power consumption in the cold-end equipment (circulating water pumps). There exists an optimal operating mode for the cold-end equipment that maximizes the unit's net power (unit load minus plant auxiliary power). The operating status of the cold-end system affects the power plant's net power, but due to long-term extensive management practices, most thermal power plants have not properly adjusted their cold-end systems, resulting in low power generation performance. Furthermore, environmental parameters can influence the operating status of the cold-end system, thus affecting power generation. These problems have become more prominent in recent years due to poor management and abnormal global climate change. Therefore, research on cold-end optimization technology is of great significance in the context of energy efficiency and low-carbon production.

[0003] Traditional methods for optimizing the cold-end operation of steam turbines are primarily based on thermodynamic tests and condenser variable-condition calculation models. First, thermodynamic tests are used to obtain the relationship between the unit's slight power increase and back pressure, and the relationship between circulating water flow and circulating water pump power consumption under different operating modes of the cold-end equipment. Then, based on the condenser variable-condition calculation model, the net power gain generated by different cold-end equipment operating modes under the current boundary conditions (steam turbine inlet parameters, circulating water inlet temperature, etc.) is calculated. However, after a period of operation, due to condenser heat exchange surface contamination, equipment aging, and other reasons, the actual operating performance of the unit will change, deviating from the test results and calculation models. This cannot reflect the actual operating conditions of the unit in real time, and thermodynamic tests are time-consuming and labor-intensive, making it difficult to conduct further thermodynamic tests for correction. Currently, most newly built thermal power plants in China have constructed digital handover systems. Through digital handover during the power plant's infrastructure construction period, design data, construction data, and installation and commissioning information of the main and auxiliary equipment can be widely collected and linked with 3D models, laying a solid data foundation for the research and development of intelligent production and operation systems for power plants. The smart power system can use big data and AI technology to establish accurate models of each subsystem of the cold end system, and realize auxiliary decision-making for the optimization of the cold end of the steam turbine.

[0004] In the cold-end system of a power plant, all components participate in heat exchange, and this heat exchange process affects the power generation performance of the steam turbine. Many scholars focus on the heat transfer characteristics of core components such as condensers and cooling towers, and their research mainly concentrates on improving heat transfer models and studying the impact of environmental or structural parameters on their performance. The performance of the condenser has a direct impact on power production. Studies have found that factors such as ambient temperature, relative humidity, wind speed, wind direction, steam flow rate, and the mass of steam present in the turbine can all affect condenser performance.

[0005] Generally, there are two methods for analyzing the performance characteristics of condensers. The first method, the analytical method, uses empirical equations to calculate heat transfer or pressure drop and uses condenser design data to describe performance under non-design conditions. These empirical equations are derived from relevant experimental studies. However, the performance represented by empirical equations and design data may not be accurate because equipment performance degrades after a certain operating time, and this performance degradation is often difficult to express using empirical formulas. The second method involves detailed numerical simulations using computational fluid dynamics to study heat and mass transfer phenomena under various environmental and operating conditions. However, this method is typically computationally expensive and not suitable for real-time identification of optimal operation. Predicting the real-time performance of actual condensers using these two mathematical process modeling methods is difficult because it requires solving hundreds of nonlinear equations and simplifying assumptions, leading to unquantifiable uncertainties and irregularities. Furthermore, if the goal is to predict condenser performance hundreds of time points in the near-real-time scenario, the traditional methods are insufficient due to computational cost requirements.

[0006] Cold-end optimization aims to achieve different optimization objectives by using appropriate algorithms. Current research often focuses on high energy efficiency or low energy consumption. However, the industrial environments of different cooling water systems vary. While existing low-carbon concepts can significantly contribute to energy conservation without considering industrial output indicators, they may adversely affect other aspects of industrial productivity. For example, the power plant's generation performance may deteriorate compared to before optimization, indicating that optimizing the cooling water flow rate reduces power generation performance. Therefore, optimizing cold-end systems also requires considering real-time modeling and prediction within the system, as well as corresponding optimization algorithms.

[0007] In existing technologies, turbine cold-end optimization mainly falls into two categories: one is the correction curve method based on thermodynamic tests, which requires shutdown testing and cannot reflect performance degradation after long-term operation; the other is variable-condition calculation based on mechanistic models, which requires numerous assumptions and is prone to model inaccuracies due to factors such as condenser fouling and equipment aging. In recent years, although some scholars have attempted to use neural networks for backpressure prediction, these methods are mostly limited to single-component modeling, and the optimization algorithms often employ simple genetic algorithms or particle swarm optimization, which are prone to getting trapped in local optima, and the computational speed is insufficient for online applications. Therefore, there is an urgent need for a cold-end optimization decision-making method that can automatically adapt to equipment aging, is computationally fast, and has strong global convergence. Summary of the Invention

[0008] The present invention aims to at least address one of the technical problems in the prior art: traditional models of condensers in cold-end systems cannot reflect the aging process of equipment; complex calculation processes reduce the accuracy of model predictions; cold-end optimization algorithms do not consider the real-time model of the system; and intelligent algorithms are prone to getting trapped in local optima in multiple iterations, resulting in low computational efficiency.

[0009] Specifically, the technical problems to be solved by this invention include:

[0010] (1) Traditional mechanism models cannot reflect the impact of condenser fouling and equipment aging on heat exchange performance in real time, resulting in optimization results deviating from reality;

[0011] (2) Existing data-driven models are mostly static offline models, which cannot be corrected online as working conditions change;

[0012] (3) Conventional optimization algorithms (such as particle swarm optimization and genetic algorithm) are prone to getting stuck in local optima when dealing with multi-peak and nonlinear optimization problems in cold-end systems. Furthermore, due to the complex heat exchange iteration involved, the calculation is time-consuming and it is difficult to achieve real-time decision-making.

[0013] Therefore, this invention provides a data-driven auxiliary decision-making method for optimizing the cold end of a steam turbine.

[0014] This invention proposes a data-driven optimization auxiliary decision-making method for the cold end of a steam turbine, comprising:

[0015] Historical data of thermal power units are preprocessed to obtain a quasi-steady-state database of unit operation; wherein, the quasi-steady-state database records historical data of thermal power units under stable operating conditions;

[0016] Support vector regression (SVR) is used to establish a data-driven model between turbine back pressure and condenser-related variables, with condenser-related variables as input and turbine back pressure as output. The data-driven model is trained using historical data in a quasi-steady-state database and is corrected online as the database is updated to automatically adapt to equipment aging and environmental changes.

[0017] Based on the relationship between the total power of the steam turbine and the operating power of the cold end system, an optimization model for maximizing the net power of the power plant is established. The operating power of the cold end system is obtained by iteratively calculating the inlet and outlet temperatures of each component based on the data-driven model, referring to the temperature balance process of cooling water circulating in the cold end system.

[0018] By combining artificial neural networks and a hybrid improved particle swarm optimization algorithm to solve the optimization model, the optimal cold-end operation scheme is obtained. The hybrid improved particle swarm optimization algorithm integrates simulated annealing into the iterative process of particle swarm optimization. After each particle swarm update, inferior solutions are accepted with simulated annealing probability, thereby enhancing the global search capability and avoiding getting trapped in local optima.

[0019] According to the above-described technical solution of the present invention, a data-driven auxiliary decision-making method for optimizing the cold end of a steam turbine may further have the following additional technical features:

[0020] In the above technical solution, the preprocessing of historical data of thermal power units to obtain a quasi-steady-state database of unit operation includes:

[0021] Using the time window method, the normalized standard deviation s of the measured parameter x within the sliding window is calculated. If s is less than a preset threshold ε, the window is determined to be a steady-state condition, and the average data within the window is taken as a quasi-steady-state sample. The formula for calculating the normalized standard deviation s is:

[0022]

[0023] Where s represents the normalized standard deviation; N represents the number of data samples, i.e., the window length; t represents the current running data instruction number, i.e., the current time; and i represents the summation calculation number. This represents the average value of the data within the sliding window. The preset threshold represents the normalization standard value.

[0024] In the above technical solution, the step of preprocessing historical data of thermal power units to obtain a quasi-steady-state database of unit operation further includes:

[0025] Starting with the first calculated steady-state sample, the time step between any two adjacent samples in the final selected sample set should be greater than a given value L, where L is 10 to 30 minutes, to ensure the independence between samples.

[0026] In the above technical solution, the condenser-related variables include at least one of the following: exhaust mass, condensate flow rate, condensate enthalpy, ambient temperature, circulating water flow rate, and wind speed.

[0027] In the above technical solution, the data-driven model is obtained by training historical data using a support vector regression machine; the trained model takes the following form:

[0028]

[0029]

[0030]

[0031] in, Indicates the temperature of the condensate; This indicates the enthalpy of condensate; The surface velocity is represented by A, and the cross-sectional area is represented by A. This indicates the heat loss from exhaust gas; This indicates the flow loss of the exhaust gas; This indicates the density of ambient air under constant pressure. This indicates the specific heat capacity of ambient air under constant pressure. The ambient temperature is represented by K, the overall heat transfer coefficient by F, the total heat transfer area by NTU, and the number of heat transfer units by NTU. Indicates the back pressure of the steam turbine. This represents the regression function in support vector regression, which maps the input to a high-dimensional feature space through a kernel function.

[0032] Specifically, based on steady-state operation sample data in the quasi-steady-state database, a data-driven model is established from historical operation data to determine the relationship between turbine back pressure and condenser-related variables.

[0033] In the above technical solution, the step of solving the optimization model by combining artificial neural networks and a hybrid improved particle swarm optimization algorithm includes:

[0034] First, a rapid prediction model for the operating water temperature of the cold-end system is established using a backpropagation artificial neural network (BP-ANN). Steam load and circulating water flow rate are used as inputs, and condenser outlet water temperature and cooling tower outlet water temperature are used as outputs. The neural network is trained using a sample set generated by iterative heat transfer calculations, enabling it to predict water temperature instantaneously, replacing the complex iterative heat transfer process. The sample set is generated as follows: the steam load is discretized from 50% to 110% of the rated load into X intervals, and the circulating water flow rate is discretized from the minimum to the maximum adjustable range into Y intervals, resulting in X×Y initial schemes. Iterative heat transfer calculations are performed using a condenser variable operating condition model combined with cooling tower characteristics. The iteration convergence accuracy is set to a temperature difference of less than 0.01℃, obtaining the steady-state water temperature corresponding to each scheme, forming X×Y sets of "input-output" samples.

[0035] Then, the optimization variables are discretized, and with the goal of maximizing the net power of the power plant, a hybrid improved particle swarm optimization algorithm (SA-PSO: simulated annealing particle swarm optimization algorithm) is used for global optimization. The optimization variables include circulating water flow rate and pump combination. In each fitness calculation, the trained neural network is called to predict the water temperature, and then the turbine power and cold end power are calculated to obtain the net power.

[0036] The final output includes the optimal cooling water flow rate, stable water temperature, cooling water pump combination scheme, and the corresponding maximum net power.

[0037] In the above technical solution, the step of establishing an optimization model for maximizing the net power of the power plant based on the relationship between the total power of the steam turbine and the operating power of the cold-end system includes:

[0038] Based on the circulation process of cooling water in the cold end system, thermal calculations are performed on each cold end system component in sequence on the basis of the data-driven model to obtain the inlet and outlet temperatures of each cold end system component, and then the power of the cold end system is calculated based on the inlet and outlet temperatures.

[0039] Calculate the turbine power based on the power correction line provided by the manufacturer;

[0040] An optimization model for maximizing the net power of a power plant is established based on the power of the cold-end system and the power of the turbine. The constraints of the optimization model for maximizing the net power of the power plant include the terminal difference, subcooling, condensate outlet temperature, and pipe flow velocity.

[0041] In the above technical solution, the cold end system components include cooling water pipes, condensers, hot water pipes, cooling towers, and water supply components.

[0042] In the above technical solution, the step of calculating the turbine power based on the power correction line provided by the manufacturer includes:

[0043]

[0044] in, Indicates the turbine power. Indicates steam load. Indicates the rated power of the steam turbine. This represents the coefficients of the first power correction equation provided by the manufacturer. This represents the coefficients of the second power correction equation provided by the manufacturer. This indicates the back pressure of the steam turbine.

[0045] In the above technical solution, the step of establishing an optimization model to maximize the net power of the power plant based on the power of the cold-end system and the power of the steam turbine includes:

[0046]

[0047] in, This indicates the net power output of the power plant. Indicates the turbine power. This indicates the power of the cold end system.

[0048] In the above technical solution, the step of combining artificial neural networks and a hybrid improved particle swarm optimization algorithm to solve the optimization model and obtain the optimal cold-end operation scheme includes:

[0049] A black-box model of the system's operating water temperature was established using a backpropagation artificial neural network;

[0050] The steam load and circulating water flow rate of the condenser are discretized into several schemes, and the system operating water temperature corresponding to each scheme is obtained by heat transfer calculation using an iterative method.

[0051] All schemes and their corresponding calculation results are imported as samples into the backpropagation artificial neural network for data training; among them, the results of iterative heat transfer calculation are used as the target set.

[0052] The simulated annealing algorithm and the particle swarm optimization algorithm are combined to solve the optimization model for maximizing the net power of the power plant. The optimal cooling water flow rate of the condenser, the stable water temperature, the optimal combination of cooling water pumps, and the maximum net power of the power plant are obtained through a joint iterative solver, thereby providing the optimal cold end decision scheme.

[0053] In the above technical solution, the method proposed in this invention also includes an online update mechanism for the data-driven model: real-time monitoring of the backpressure prediction error of the data-driven model; if the prediction error exceeds 0.3 kPa for three consecutive points, model retraining is triggered; simultaneously, a periodic update cycle (e.g., every 7 days) is set, and the data-driven model is retrained using newly added quasi-steady-state data. The quasi-steady-state database adopts a "first-in, first-out" principle, maintaining a database capacity of 5000 latest samples; each new sample undergoes steady-state testing (normalized standard deviation threshold ε is set to 0.02) and the time interval between it and existing samples is greater than 15 minutes before it can be added to the database. Model retraining is performed in a background thread, and the old model is used until training is complete to ensure continuous operation of the optimization system.

[0054] In summary, due to the adoption of the above-mentioned technical features, the beneficial effects of the present invention are:

[0055] This invention proposes a data-driven optimization auxiliary decision-making method for the cold end of a steam turbine. It utilizes intelligent algorithms to establish a data-driven model of the subsystems within the cold end system. Based on this, an optimization model for maximizing the net power of the power plant is established, ultimately achieving optimal operation of the cold end system and maximizing the safe, efficient, energy-saving, and consumption-reducing operation of the thermal power unit.

[0056] Specifically, this invention eliminates erroneous measurements through data reconciliation and obtains the most reliable data values ​​by imposing physical constraints on a specific system, thereby improving the stability and computational speed of the data-driven model. Historical data accumulation during the aging process of the condenser is considered to determine the mathematical relationship between the turbine back pressure and related variables. This artificial intelligence algorithm largely ensures the high reliability of interpolation and improves the accuracy of model predictions. Furthermore, by integrating simulated annealing and particle swarm optimization, the global search capability of particles in early iterations is improved, and the local search capability is enhanced in later iterations, preventing getting trapped in local optima and shortening computation time.

[0057] Synergistic effects and significant advancements of this invention: This invention is not a simple superposition of existing technologies, but rather an organic combination of three core modules—an online-updated SVR data-driven model, a BP-ANN surrogate model, and an SA-PSO hybrid optimization algorithm—resulting in unexpected technical effects.

[0058] SVR online updates solve the model aging problem, ensuring that optimization is always based on the current device status;

[0059] The BP-ANN surrogate model reduces the time for a single fitness calculation from seconds to milliseconds, making real-time optimization possible;

[0060] The SA-PSO hybrid algorithm avoids getting trapped in local optima through simulated annealing, thus ensuring global optimization capabilities.

[0061] The three technologies work together to achieve model adaptability, real-time computation, and global convergence, which are impossible to achieve simultaneously with traditional methods. This results in significant economic benefits, such as a net power increase of 1.2MW and annual electricity savings of 8.4 million kWh, which is something that no single technological improvement can achieve.

[0062] Specifically, the beneficial effects can be summarized as follows:

[0063] (1) Dynamic adaptive capability: Through continuous updates of the quasi-steady-state database and online correction of the SVR model, the optimization scheme always fits the current actual condition of the equipment, overcoming the defect of traditional methods failing due to equipment aging.

[0064] (2) Balancing global optimization and real-time performance: The BP-ANN surrogate model is introduced to replace the complex heat exchange iteration, reducing the single fitness calculation time from seconds to milliseconds (single prediction <1ms); at the same time, the hybrid improved particle swarm algorithm is integrated with the simulated annealing mechanism, which greatly improves the convergence speed while ensuring global search capability. Compared with the standard PSO, the convergence time is shortened from 120 seconds to 15 seconds (8 times faster), and it completely avoids getting trapped in local optima, with a net power increase of 1.2MW, meeting the requirements of online real-time optimization.

[0065] (3) Systemic and Portable: This invention is not only for condensers, but also includes the entire cold end system (pipelines, condensers, cooling towers, etc.) in the optimization and adopts modular modeling, which makes it easy to be transferred to different units and has strong versatility.

[0066] Additional aspects and advantages of the invention will become apparent in the following description or may be learned by practice of the invention. Attached Figure Description

[0067] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0068] Figure 1 This is a flowchart of a data-driven auxiliary decision-making method for optimizing the cold end of a steam turbine, according to an embodiment of the present invention. Detailed Implementation

[0069] like Figure 1 As shown, in order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of this application can be combined with each other.

[0070] 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 therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0071] The following reference Figure 1 This describes a data-driven auxiliary decision-making method for optimizing the cold end of a steam turbine, provided by some embodiments of the present invention.

[0072] Some embodiments of this application provide a data-driven auxiliary decision-making method for optimizing the cold end of a steam turbine.

[0073] like Figure 1 As shown, some embodiments of the present invention propose a data-driven auxiliary decision-making method for optimizing the cold end of a steam turbine, comprising at least the following steps S1-S4. It should be noted that steps S1 to S4 are not strictly performed in sequence; those skilled in the art can adjust the order of the steps as needed, and different steps can also be performed simultaneously.

[0074] S1. Preprocess historical data of thermal power units to obtain a quasi-steady-state database of unit operation; wherein, the quasi-steady-state database records historical data of thermal power units under stable operating conditions.

[0075] Specifically, the boundary conditions during the operation of thermal power units are complex and variable. The units need to adjust their operating state in real time according to grid load commands and external environmental conditions. Some variable operating condition data cannot accurately reflect the regularity of equipment and system operation. In some embodiments, step S1 uses a time window method to obtain a quasi-steady-state database of unit operation, wherein the normalized standard deviation s of the measured parameter x represented by the selected state within the time window should be less than a given threshold; including:

[0076]

[0077] Where s represents the normalized standard deviation; N represents the number of data samples, i.e., the window length; t represents the current running data instruction number, i.e., the current time; and i represents the summation calculation number. This represents the average value of the data within the sliding window; This represents the preset threshold for the normalization standard value, typically set between 0.01 and 0.05. When... When the data meets the steady-state screening requirements, it can be added to the quasi-steady-state database.

[0078] In some embodiments, additional time step rules are applied to ensure the representativeness of the data samples in the quasi-steady-state database under specific operating conditions. Specifically, starting from the first calculated steady-state sample, the time step between any two adjacent samples in the final selected sample set should be greater than a given value L. This given value can be flexibly set according to actual conditions; in one specific embodiment, L is 10-30 minutes to ensure the independence between samples.

[0079] By filtering historical data of the unit through the above steps, a database under stable operating conditions was established.

[0080] S2. A data-driven model between turbine back pressure and condenser-related variables is established using support vector regression.

[0081] The relationship between turbine back pressure and condenser-related variables is of great significance. However, direct analytical formulas or numerical simulation techniques suffer from inaccuracies or excessively long computation times. This embodiment employs support vector regression to establish a data-driven model that expresses this non-explicit relationship from operational data. External variables (condenser-related variables) are selected as input vectors to represent the physical relationship in the analytical method. These condenser-related variables include exhaust gas mass, condensate flow rate, condensate enthalpy, ambient temperature, circulating water flow rate, and wind speed. It is understood that under certain operating conditions, only one or more of the aforementioned condenser-related variables may be selected to establish the data-driven model.

[0082] To ensure that the data-driven model can reflect the current state of the equipment in real time, this invention designs an online model update mechanism: every certain period (e.g., once a week) or when the model prediction error exceeds a threshold, the SVR model is retrained using the latest quasi-steady-state database, so that the model always approximates the input-output characteristics of the actual system. This dynamic update capability is unmatched by traditional fixed models and is also the key to this invention's adaptation to equipment aging.

[0083] In one specific embodiment, the data-driven model is trained on historical data using a support vector regression machine. Its input variables include exhaust gas mass, condensate flow rate, condensate enthalpy, ambient temperature, circulating water flow rate, and wind speed. The model output is the turbine back pressure. The trained model takes the following form:

[0084]

[0085]

[0086]

[0087] in, Indicates the temperature of the condensate; This indicates the enthalpy of condensate; The surface velocity is represented by A, and the cross-sectional area is represented by A. This indicates the heat loss from exhaust gas; This indicates the flow loss of the exhaust gas; This indicates the density of ambient air under constant pressure. This indicates the specific heat capacity of ambient air under constant pressure. The ambient temperature is represented by K, the overall heat transfer coefficient by F, the total heat transfer area by NTU, and the number of heat transfer units by NTU. Indicates the back pressure of the steam turbine. This represents the regression function in support vector regression, which maps the input to a high-dimensional feature space through a kernel function.

[0088] If a set of steady-state operation sample data with high accuracy and reliability is identified from the quasi-steady-state database established in step S1, a data-driven model between turbine back pressure and condenser-related variables can be established from historical operation data to express the characteristics of the condenser during actual operation.

[0089] S3. Based on the relationship between the total power of the steam turbine and the operating power of the cold end system, establish an optimization model for maximizing the net power of the power plant; wherein, the operating power of the cold end system is calculated and obtained according to the data-driven model with reference to the temperature balance process of cooling water circulating in the cold end system.

[0090] Specifically, a temperature equilibrium process occurs when the cooling water circulates within the cold-end system. Uncirculated water is first heated in the condenser, and then flows into the cooling tower through hot water pipes for cooling. The cooling water is then lifted by a circulating water pump and pumped back to the condenser through cooling water pipes. After several cycles, the water temperature at the inlet and outlet of each component tends to be constant. In this embodiment, the invention simulates the above process through programming, treating each iteration as a water circulation process. Based on the data-driven model of the condenser in step S2, thermal calculations are sequentially performed on the cooling water pipes, condenser, hot water pipes, cooling tower, and water supply components to obtain their inlet and outlet temperatures.

[0091] In one specific embodiment, step S3 includes:

[0092] Based on the circulation process of cooling water in the cold end system, thermal calculations are performed on each cold end system component in sequence on the basis of the data-driven model to obtain the inlet and outlet temperatures of each cold end system component, and then the power of the cold end system is calculated based on the inlet and outlet temperatures.

[0093] Calculate the turbine power based on the power correction line provided by the manufacturer; the calculation formula is:

[0094]

[0095] in, Indicates the turbine power. Indicates steam load. Indicates the rated power of the steam turbine. This represents the coefficients of the first power correction equation provided by the manufacturer. This represents the coefficients of the second power correction equation provided by the manufacturer. This indicates the back pressure of the steam turbine.

[0096] An optimization model for maximizing the net power of the power plant is established based on the power of the cold-end system and the power of the turbine. Without considering the energy consumption of the primary power system, the optimization model for maximizing the net power of the power plant can be expressed as follows:

[0097]

[0098] in, This indicates the net power output of the power plant. Indicates the turbine power. This represents the power of the cold-end system. Furthermore, the optimization model considers constraints such as terminal temperature difference, subcooling, condensate outlet temperature, and pipe flow velocity.

[0099] S4. Solve the optimization model by combining artificial neural networks and a hybrid improved particle swarm optimization algorithm to obtain the optimal cold-end operation scheme; the hybrid improved particle swarm optimization algorithm includes simulated annealing algorithm and particle swarm optimization algorithm.

[0100] In some embodiments, step S4 includes:

[0101] A black-box model of the system's operating water temperature was established using a backpropagation artificial neural network (BP-ANN).

[0102] The steam load and circulating water flow rate of the condenser in the model established in step S3 are discretized into several schemes. The system operating water temperature corresponding to each scheme is obtained by heat transfer calculation using an iterative method. The more schemes there are, the better the subsequent training effect and the more accurate the output results. In this embodiment, the number of schemes is 3000.

[0103] All schemes and their corresponding calculation results are imported as samples into a backpropagation artificial neural network for data training; the results of iterative heat transfer calculations are used as the target set for training these samples. Thus, a black-box model built using a neural network replaces complex heat transfer calculations, enabling rapid prediction of stable operating water temperature.

[0104] The optimization model for maximizing the net power of the power plant is solved by combining simulated annealing and particle swarm optimization (PSO). A joint iterative solver obtains the optimal cooling water flow rate of the condenser, the stable water temperature, the optimal combination of cooling water pumps, and the maximum net power of the power plant, thus providing the optimal cold-end decision scheme. Specifically, a hybrid improved PSO algorithm is used to iterate and optimize all combinations. PSO is an intelligent optimization method based on simulating bird foraging behavior, finding targets by updating population velocity and position. When solving complex functions using PSO, there is a possibility that the optimization result may get trapped in local optima. The optimization model in step S3 is solved by combining simulated annealing and PSO. The optimal cooling water flow rate of the condenser, the stable water temperature, the optimal combination of cooling water pumps, and the maximum net power of the power plant are obtained through a joint iterative solver, providing the optimal cold-end decision scheme.

[0105] Specifically, the method of solving the optimization model by combining artificial neural networks and a hybrid improved particle swarm optimization algorithm includes:

[0106] First, a rapid prediction model for the operating water temperature of the cold-end system is established using a backpropagation artificial neural network (BP-ANN). The model uses steam load and circulating water flow rate as inputs, and condenser outlet water temperature and cooling tower outlet water temperature as outputs. The neural network is trained using a sample set generated by iterative heat transfer calculations, enabling it to predict water temperature instantaneously, replacing the complex iterative heat transfer process. The sample set is generated as follows: the steam load is discretized from 50% to 110% of the rated load into 50 intervals, and the circulating water flow rate is discretized from the minimum to the maximum adjustable range into 60 intervals, resulting in 3000 initial schemes. A mature condenser variable operating condition model combined with cooling tower characteristics is used for iterative heat transfer calculations. The iteration convergence accuracy is set to a temperature difference of less than 0.01℃, obtaining the steady-state water temperature (condenser outlet water temperature and cooling tower outlet water temperature) corresponding to each scheme, forming 3000 sets of "input-output" samples. 80% is randomly selected as the training set, 10% as the validation set (to prevent overfitting), and 10% as the test set. In one specific embodiment, the BP-ANN structure is as follows: an input layer with 2 nodes (steam load, circulating water flow rate), two hidden layers with 12 neurons each, using the ReLU activation function, and an output layer with 2 nodes (two water temperatures). The training objective is a root mean square error of less than 0.05℃. The Adam optimizer is used with an initial learning rate of 0.001 and 500 iterations. Training stops when the validation set loss no longer decreases after 10 consecutive iterations. The deviation between the BP-ANN's predicted water temperature and the iterative calculation results is within 0.03℃, and the single prediction time is less than 1ms.

[0107] Then, the optimization variables are discretized, and a hybrid improved particle swarm optimization algorithm is used for global optimization with the goal of maximizing the net power of the power plant. The optimization variables include circulating water flow rate and pump combination. In each fitness calculation, a trained neural network is used to predict the water temperature, and then the turbine power and cold-end power are calculated to obtain the net power. As a specific implementation, the parameters of the hybrid improved particle swarm optimization algorithm are set as follows: number of particles: 30; inertia weight ω linearly decreases from 0.9 to 0.4; learning factor c1=c2=2.0; maximum number of iterations: 200. Simulated annealing: initial temperature T0=1000℃, annealing rate α=0.95, termination temperature Tend=1℃. After each particle updates its position, a new fitness is calculated. If it is better than the current optimal solution, it is directly accepted; otherwise, a suboptimal solution is accepted with probability P=exp(-ΔE / T), where ΔE is the fitness difference (net power difference, normalized), and T is the current temperature. The role of the inferior solution acceptance mechanism: In the early, high-temperature phase, even inferior solutions have a high probability of acceptance, maintaining population diversity; in the later, as the temperature decreases, the probability of accepting inferior solutions approaches zero, strengthening local search. Testing has shown that this parameter enables the algorithm to converge to the global optimum in about 15 generations, and the results are stable across multiple runs.

[0108] The final output includes the optimal cooling water flow rate, stable water temperature, cooling water pump combination scheme, and the corresponding maximum net power.

[0109] The specific process of the hybrid improved particle swarm optimization algorithm is as follows:

[0110] (1) Initialize the particle swarm, with each particle representing a set of decision variables (such as circulating water flow rate and number of pumps in operation).

[0111] (2) Calculate the fitness (i.e. net power) of each particle, where the required water temperature is quickly predicted by the trained BP-ANN;

[0112] (3) Update the individual optimal and global optimal;

[0113] (4) Update particle velocity and position according to the particle swarm algorithm;

[0114] (5) Introduce simulated annealing operation: calculate the fitness of the new position. If it is better than the original position, it is accepted. Otherwise, the inferior solution is accepted with probability exp(-ΔE / T), where T is the current temperature and ΔE is the fitness difference.

[0115] (6) Reduce the temperature T according to the annealing plan;

[0116] (7) Repeat steps (2)-(6) until the termination condition is met (such as the maximum number of iterations or the temperature drops to the threshold).

[0117] In the early stages, the hybrid algorithm is more likely to accept inferior solutions due to the higher temperature, thus maintaining population diversity and avoiding premature convergence. In the later stages, as the temperature decreases, it gradually degenerates into a standard particle swarm optimization algorithm, which strengthens local search and thus achieves a balance between global and local search.

[0118] In the early stages, the algorithm is more likely to accept inferior solutions due to the higher temperature, thus maintaining population diversity and avoiding premature convergence. In the later stages, as the temperature decreases, it gradually degenerates into a standard particle swarm optimization algorithm, which strengthens local search.

[0119] In this specification, the illustrative expressions of the terms used do not necessarily refer to the same embodiments or examples. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0120] Taking a 600MW thermal power unit as an example, one year of its operating data was collected, and 5000 quasi-steady-state samples were obtained after preprocessing. A backpressure model was established using SVR, and the model achieved an R² of 0.97 and a root mean square error of 0.15 kPa on the test set. The optimization method of this invention was compared with the conventional particle swarm optimization (PSO) algorithm. The parameters of the standard PSO were set as follows: 30 particles, inertia weight ω=0.8, c1=c2=2.0, maximum number of iterations 200, and no simulated annealing operation. The fitness calculation also used the original heat transfer iteration (non-BP-ANN) to ensure fairness. The results show that the average convergence time of the standard PSO is 120 seconds (using heat transfer iteration), and it gets trapped in local optima 9 times out of 30 independent runs (the optimal value deviates from the true optimum by >0.3MW); while the average convergence time of the proposed SA-PSO+BP-ANN is only 15 seconds, and it finds the global optimum every time, resulting in a net power increase of 1.2MW. Based on 7,000 hours of operation per year, it saves 8.4 million kWh of electricity annually, equivalent to saving approximately 2,700 tons of standard coal, demonstrating significant economic benefits and energy conservation and emission reduction effects.

[0121] Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention shall be included within the scope of protection of this invention.

Claims

1. A data-driven optimization auxiliary decision-making method for the cold end of a steam turbine, characterized in that, include: Historical data of thermal power units are preprocessed to obtain a quasi-steady-state database of unit operation; wherein, the quasi-steady-state database records historical data of thermal power units under stable operating conditions; Support vector regression is used to establish a data-driven model between turbine back pressure and condenser-related variables, with condenser-related variables as input and turbine back pressure as output. The data-driven model is trained using historical data in a quasi-steady-state database and is corrected online as the database is updated to automatically adapt to equipment aging and environmental changes. Based on the relationship between the total power of the steam turbine and the operating power of the cold end system, an optimization model for maximizing the net power of the power plant is established. The operating power of the cold end system is obtained by iteratively calculating the inlet and outlet temperatures of each component based on the data-driven model, referring to the temperature balance process of cooling water circulating in the cold end system. By combining artificial neural networks and a hybrid improved particle swarm optimization algorithm to solve the optimization model, the optimal cold-end operation scheme is obtained. The hybrid improved particle swarm optimization algorithm integrates simulated annealing into the iterative process of particle swarm optimization. After each particle swarm update, inferior solutions are accepted with simulated annealing probability, thereby enhancing the global search capability and avoiding getting trapped in local optima. The method for solving the optimization model by combining artificial neural networks and a hybrid improved particle swarm optimization algorithm includes: First, a rapid prediction model for the operating water temperature of the cold-end system is established using a backpropagation artificial neural network. The model uses steam load and circulating water flow rate as inputs, and condenser outlet water temperature and cooling tower outlet water temperature as outputs. The neural network is trained using a sample set generated by iterative heat transfer calculations, enabling it to predict water temperature instantaneously, replacing the complex iterative heat transfer process. The sample set is generated as follows: the steam load is discretized from 50% to 110% of the rated load into X intervals, and the circulating water flow rate is discretized from the minimum to the maximum adjustable range into Y intervals, resulting in X×Y initial schemes. Iterative heat transfer calculations are performed using a condenser variable operating condition model combined with cooling tower characteristics. The iteration convergence accuracy is set to a temperature difference of less than 0.01℃, obtaining the steady-state water temperature corresponding to each scheme, forming X×Y sets of "input-output" samples. Then, the optimization variables are discretized, and a hybrid improved particle swarm optimization algorithm is used to perform global optimization with the goal of maximizing the net power of the power plant. The optimization variables include circulating water flow rate and pump combination. In each fitness calculation, the trained neural network is called to predict the water temperature, and then the turbine power and cold end power are calculated to obtain the net power. The final output includes the optimal cooling water flow rate, stable water temperature, cooling water pump combination scheme, and corresponding maximum net power. The optimization model is solved by combining artificial neural networks and a hybrid improved particle swarm optimization algorithm to obtain the optimal cold-end operation scheme, including: A black-box model of the system's operating water temperature was established using a backpropagation artificial neural network; The steam load and circulating water flow rate of the condenser are discretized into several schemes, and the system operating water temperature corresponding to each scheme is obtained by heat transfer calculation using an iterative method. All schemes and their corresponding calculation results are imported as samples into the backpropagation artificial neural network for data training; among them, the results of iterative heat transfer calculation are used as the target set. The simulated annealing algorithm and the particle swarm optimization algorithm are combined to solve the optimization model for maximizing the net power of the power plant. The optimal cooling water flow rate of the condenser, the stable water temperature, the optimal combination of cooling water pumps, and the maximum net power of the power plant are obtained through a joint iterative solver, thereby providing the optimal cold end decision scheme.

2. The data-driven auxiliary decision-making method for cold-end optimization of steam turbines according to claim 1, characterized in that, The preprocessing of historical data from thermal power units to obtain a quasi-steady-state database of unit operation includes: Using the time window method, the normalized standard deviation s of the measured parameter x within the sliding window is calculated. If s is less than a preset threshold ε, the window is determined to be a steady-state condition, and the average data within the window is taken as a quasi-steady-state sample. The formula for calculating the normalized standard deviation s is: Where s represents the normalized standard deviation; N represents the number of data samples, i.e., the window length; t represents the current running data instruction number, i.e., the current time; and i represents the summation calculation number. This represents the average value of the data within the sliding window. The preset threshold represents the normalization standard value.

3. The data-driven auxiliary decision-making method for cold-end optimization of steam turbines according to claim 2, characterized in that, The preprocessing of historical data from thermal power units to obtain a quasi-steady-state database of unit operation also includes: Starting with the first calculated steady-state sample, the time step between any two adjacent samples in the final selected sample set should be greater than a given value L, where L is 10 to 30 minutes, to ensure the independence between samples.

4. The data-driven auxiliary decision-making method for cold-end optimization of steam turbines according to claim 1, characterized in that, The condenser-related variables include at least one of the following: exhaust mass, condensate flow rate, condensate enthalpy, ambient temperature, circulating water flow rate, and wind speed.

5. The data-driven auxiliary decision-making method for cold-end optimization of steam turbines according to claim 4, characterized in that, The data-driven model is trained on historical data using a support vector regression machine. The trained model takes the following form: in, Indicates the temperature of the condensate; This indicates the enthalpy of condensate; The surface velocity is represented by A, and the cross-sectional area is represented by A. This indicates the heat loss from exhaust gas; This indicates the flow loss of the exhaust gas; This indicates the density of ambient air under constant pressure. This indicates the specific heat capacity of ambient air under constant pressure. The ambient temperature is represented by K, the overall heat transfer coefficient by F, the total heat transfer area by NTU, and the number of heat transfer units by NTU. Indicates the back pressure of the steam turbine. This represents the regression function in support vector regression, which maps the input to a high-dimensional feature space through a kernel function. Specifically, based on steady-state operation sample data in the quasi-steady-state database, a data-driven model is established from historical operation data to determine the relationship between turbine back pressure and condenser-related variables.

6. The data-driven auxiliary decision-making method for cold-end optimization of steam turbines according to claim 1, characterized in that, The optimization model for maximizing the net power of the power plant is established based on the relationship between the total power of the steam turbine and the operating power of the cold-end system, including: Based on the circulation process of cooling water in the cold end system, thermal calculations are performed on each cold end system component in sequence on the basis of the data-driven model to obtain the inlet and outlet temperatures of each cold end system component, and then the power of the cold end system is calculated based on the inlet and outlet temperatures. Calculate the turbine power based on the power correction line provided by the manufacturer; An optimization model for maximizing the net power of a power plant is established based on the power of the cold-end system and the power of the turbine. The constraints of the optimization model for maximizing the net power of the power plant include the terminal difference, subcooling, condensate outlet temperature, and pipe flow velocity.

7. The data-driven auxiliary decision-making method for cold-end optimization of steam turbines according to claim 6, characterized in that, The cold end system components include cooling water pipes, condensers, hot water pipes, cooling towers, and water supply components.

8. The data-driven auxiliary decision-making method for cold-end optimization of steam turbines according to claim 6, characterized in that, The calculation of turbine power based on the power correction line provided by the manufacturer includes: in, Indicates the turbine power. Indicates steam load. Indicates the rated power of the steam turbine. This represents the coefficients of the first power correction equation provided by the manufacturer. This represents the coefficients of the second power correction equation provided by the manufacturer. This indicates the back pressure of the steam turbine.

9. A data-driven auxiliary decision-making method for optimizing the cold end of a steam turbine according to claim 8, characterized in that, The optimization model for maximizing the net power of the power plant, based on the power of the cold-end system and the power of the turbine, includes: in, This indicates the net power output of the power plant. Indicates the turbine power. This indicates the power of the cold end system.

10. A data-driven auxiliary decision-making method for cold-end optimization of steam turbines according to claim 1, characterized in that, It also includes an online update mechanism for the data-driven model: real-time monitoring of the backpressure prediction error of the data-driven model; if the prediction error exceeds a set threshold for three consecutive points, model retraining is triggered; at the same time, a regular update cycle is set to retrain the data-driven model using newly added quasi-steady-state data; the quasi-steady-state database adopts the first-in-first-out principle to maintain a fixed number of the latest samples; each new sample can be added to the database only after it has undergone steady-state testing and the time interval between it and existing samples is greater than a set time; model retraining is performed in a background thread, and the old model is still used until training is completed to ensure the continuous operation of the optimization system.

Images