Thermal power unit main parameter regulating system and method adaptive to power grid load fluctuation

Abstract

The present application relates to the field of intelligent control technology of thermal power plant, and provides a main parameter adjusting system of thermal power unit adapting to power grid load fluctuation, comprising a data acquisition and feature fusion module, an intelligent decision module and a control execution module; the intelligent decision module is embedded into a distributed control system (DCS) of the thermal power plant as an upper optimizer, and the control instructions generated by the intelligent decision module are issued to the bottom control loop for execution through the control execution module; the intelligent decision module comprises a steam temperature control intelligent agent, a reference quantity prediction intelligent agent and a feedwater and dry-wet state conversion intelligent agent; the steam temperature control intelligent agent outputs dynamic optimized steam temperature dynamic optimization bias vectors of all levels; the reference quantity prediction intelligent agent outputs fuel quantity and collaborative feedwater quantity vectors under different coal quality mixing in future time; and the feedwater and dry-wet state conversion intelligent agent outputs coordinated action instruction sequences of the feedwater pump, the feedwater regulating valve and the 361 valve equipment. Through the system, autonomous and flexible adjustment of the thermal power system to the power grid is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for thermal power plants, specifically to a system and method for adjusting the main parameters of thermal power units to adapt to grid load fluctuations. Background Technology

[0002] Currently, the installed capacity of new energy sources, represented by wind power and photovoltaics, is rapidly increasing, placing unprecedented demands on the flexible adjustment capabilities of thermal power units. This is driving a profound transformation of coal-fired power into a "next-generation coal-fired power," with its core positioning shifting from a traditional primary power source to a more flexible power source emphasizing regulation and support. Supercritical units, as the ballast of power supply and the main force in achieving this transformation, are facing the routine test of deep peak shaving and reliable flexible operation. Simultaneously, the industry is shifting from ensuring operational safety to pursuing an intelligent operation mode that optimizes safety, economy, and environmental protection. However, current operators still require cumbersome manual intervention and parameter adjustments, leading to performance fluctuations and assessment risks due to differences in human experience. Therefore, the traditional control and operation mode, heavily reliant on operator experience, has become a bottleneck restricting further improvement in the efficiency of thermal power units and hindering the achievement of modern power plant construction goals. One of the key paths to developing "next-generation coal-fired power" is to fundamentally transform thermal power plants from "experience-based decision-making" to "intelligent decision-making" through highly automated and intelligent means. Therefore, it is evident that integrating new-generation information technologies such as artificial intelligence and big data to develop intelligent control technologies with self-sensing, self-decision-making, and self-optimization capabilities is an inevitable choice for the thermal power industry to meet challenges and move towards the transformation and upgrading towards "new-generation coal-fired power". Summary of the Invention

[0003] The technical problem to be solved by this invention is how to build an intelligent decision-making system for key parameters of thermal power plants, so as to improve the ability of thermal power plants to adjust autonomously and flexibly in response to changes in the power grid.

[0004] The present invention solves the above-mentioned technical problems through the following technical means:

[0005] This invention provides a main parameter adjustment system for thermal power units that adapts to grid load fluctuations, including a data acquisition and feature fusion module, an intelligent decision-making module, and a control execution module. The intelligent decision-making module is embedded as an upper-level optimizer in the distributed control system (DCS) of the thermal power plant, and the control commands it generates are sent down to the lower-level control loop for execution through the control execution module. The data acquisition and feature fusion module is used to acquire multi-source time-series data related to steam temperature, fuel, and feedwater in the distributed control system of thermal power units in real time, and to perform preprocessing and feature engineering. The intelligent decision-making module includes a steam temperature control intelligent agent, a baseline quantity prediction intelligent agent, and a feedwater and dry-wet state conversion intelligent agent; the steam temperature control intelligent agent outputs dynamically optimized bias vectors for each level of steam temperature; the baseline quantity prediction intelligent agent outputs vectors of fuel quantity and coordinated feedwater quantity under different coal quality mixtures at several future time points; the feedwater and dry-wet state conversion intelligent agent outputs a sequence of coordinated action instructions for feedwater pumps, feedwater regulating valves, and 361 valve equipment.

[0006] Furthermore, the intelligent decision-making system for key parameters of the thermal power plant also includes an online learning and iterative optimization module, which is used to fine-tune and incrementally learn the intelligent agent model online based on real-time operating performance feedback.

[0007] Furthermore, the steam temperature control agent generates advanced feedforward compensation by sensing the upstream temperature change gradient; by analyzing the continuous deviation pattern of the downstream temperature of the boiler, it inversely infers the feedback correction amount for the fuel-water ratio source, realizing closed-loop optimization from the desuperheating water terminal execution to the global feedwater-coal ratio.

[0008] Furthermore, the reference quantity prediction agent learns the temporal correlation in historical data to achieve rolling prediction of fuel quantity and water supply quantity references for several future sampling periods, generating dynamic feedforward signals; and corrects the feedforward signals through a coal quality adaptive correction mechanism.

[0009] Furthermore, the water supply and dry / wet state transition decision-making agent generates a coordinated action instruction sequence by constructing a multi-classification state diagnostic model and using a combination of data-driven and expert rule methods.

[0010] Furthermore, the specific operation of the steam temperature control intelligent agent is as follows: S10. Based on the data acquisition and feature fusion module, obtain time-series feature data characterizing the dynamics and coupling relationship of the steam temperature control system. ; S11. Construct a deep neural network based on temporal feature data. Output optimized setpoint partial vector As shown in the following formula:

[0011] in, Representative parameters are Deep neural networks, which extract data from time series data The complex relationship between combustion disturbance and steam temperature response is automatically learned. These are specific loads and their operating parameters; ,in, This is the set value for the secondary desuperheating water. This is the set value for the first-stage desuperheating water. Set the superheat value; S12. Train the deep neural network using historical operating data, with the composite loss function being the minimization of steam temperature deviation at each stage and the amount of desuperheating water used. Initialize the parameters through supervised learning or offline reinforcement learning. S13. After the deployment of the steam temperature control intelligent agent, the system enters the online adaptive mode. The intelligent agent integrates real-time data and environmental conditions to generate control commands. These commands are applied to the boiler object to generate new states. The reward signal is calculated based on performance indicators, and the model parameters are fine-tuned using online learning algorithms to continuously adapt to equipment aging and changes in fuel characteristics.

[0012] Furthermore, the specific operations of the benchmark quantity prediction agent are as follows: S20. Based on the data acquisition and feature fusion module, obtain feature vectors characterizing fuel demand and water supply coordination. ; S21. Construct a multi-step dynamic prediction model with Long Short-Term Memory (LSTM) network as its core, using feature vectors. Using the time series as input and the fuel quantity, which has been filtered and validated, as the output label, a supervised learning model is constructed to generate a dynamic feedforward signal, as shown in the following equation:

[0013] in, Based on fuel quantity, As the water supply benchmark, These are the weight parameters obtained during network training. S22. Construct an online correction loop based on the equivalent calorific value method, calculate in real time the deviation between the theoretical output corresponding to the actual fuel quantity and the actual output of the unit, dynamically derive and update a coal quality correction coefficient, and use the coal quality correction coefficient as a multiplication factor to correct the dynamic feedforward signal. S23. Establish an online update mechanism to regularly use recent running data to incrementally train or fine-tune the parameters of the LSTM model.

[0014] Furthermore, the specific operation of the water supply and dry / wet state conversion intelligent agent is as follows: S30. Based on the data acquisition and feature fusion module, obtain the feature vector representing the overall operating status of the boiler. As shown in the following formula:

[0015] in, The rate of change of the feedwater-fuel ratio is... Reflects the energy balance state of the boiler core; and These are the inlet and outlet differential pressures and vibration amplitude of the boiler water circulation pump, respectively. The temperature difference between the walls of the steam-water separator is used to directly determine whether there is water inside. For valve opening degree 361, For valve 361 flow rate, This is the valve action sequence; S31. Construct and train a three-class classification model using the random forest algorithm, as shown below:

[0016] Based on the aforementioned input feature vector This is used to identify whether the boiler is currently in a wet, transitional, or dry state. S32. The system adopts a combination of data-driven and expert rule methods to generate control strategies. Specifically, based on historical best operating data and a simplified boiler hydrodynamic model, an optimization decision engine is constructed. When making a decision, the engine starts from the current state of the system and, under the condition of strictly following multiple safety constraints of wall temperature, water level, and pressure, calculates the operating trajectory that makes the conversion process smooth and fast through rolling optimization iteration, namely the coordinated action command sequence of the feedwater pump, feedwater regulating valve, and valve 361.

[0017] Furthermore, during the decision-making process of the optimization decision engine, if any key parameter deviates from the safe range, the system automatically interrupts the current operation sequence and switches to a safe mode.

[0018] This invention also provides a method for adjusting the main parameters of thermal power units to adapt to grid load fluctuations, based on the above system, including the following steps: Collect and integrate multi-source data related to steam temperature, fuel, and feedwater during the operation of thermal power units and perform feature engineering; Intelligent decision-making is carried out on the partial vector of steam temperature setpoints at various levels, the fuel baseline quantity and coordinated water supply baseline quantity at several future moments, and the water supply and dry-wet state conversion strategy through deep reinforcement learning model, LSTM prediction model, and random forest classification model, respectively. The decision-making instructions are issued to the implementing agencies after security verification. Based on the performance evaluation results, the model is trained online and its parameters are fine-tuned.

[0019] The advantages of this invention are: This invention focuses on the three key control objects of thermal power plants: steam temperature, fuel, and feedwater. It constructs intelligent agents with self-sensing, self-learning, self-decision-making, and self-execution capabilities, realizing autonomous optimization of thermal power unit control. This frees operators from tedious manual intervention and parameter adjustment, avoids performance fluctuations and assessment risks caused by differences in human experience, and makes thermal power plants more in line with the current role of flexible power sources that emphasize regulation capabilities and support functions. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an intelligent decision-making method for key parameters of a thermal power plant, according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1 This embodiment provides a main parameter regulation system for thermal power units that adapts to grid load fluctuations. It aims to overcome key technologies such as intelligent decoupling of multivariable strongly coupled systems, adaptive control under all operating conditions, and autonomous decision-making in complex process sequences. This will drive the leap from automation to intelligence in thermal power control technology, forming an independent and controllable core solution and providing key technical support for the intelligent core of "next-generation coal-fired power." Economically, by optimizing the control quality and automation level of the unit under complex operating conditions such as dry-wet transitions, it can directly improve the unit's grid connection performance and operational economy, significantly reducing assessment costs. At the industry level, successful implementation of this solution will form a replicable and scalable standard system and practical examples for intelligent optimization control of supercritical units. This will provide strong technical support for the entire thermal power industry to improve operational efficiency, achieve cost reduction and efficiency improvement, and intelligent upgrading, ultimately moving towards "next-generation coal-fired power." It has profound significance for ensuring the safe and stable operation of the new power system and promoting the clean and low-carbon transformation of energy.

[0023] The system includes a data acquisition and feature fusion module, an intelligent decision-making module, and a control execution module. The intelligent decision-making module is embedded as an upper-level optimizer in the distributed control system (DCS) of the thermal power plant, and the control commands it generates are sent down to the lower-level control loop for execution through the control execution module. The data acquisition and feature fusion module is used to acquire multi-source time-series data related to steam temperature, fuel, and feedwater in the distributed control system of thermal power units in real time, and to perform preprocessing and feature engineering. The intelligent decision-making module includes a steam temperature control agent, a baseline quantity prediction agent, and a feedwater and dry / wet state conversion agent. The steam temperature control agent outputs dynamically optimized bias vectors for each stage of steam temperature. By sensing the upstream temperature change gradient, it generates advanced feedforward compensation. By analyzing the continuous deviation pattern of the downstream boiler temperature, it inversely infers the feedback correction amount for the fuel-to-water ratio source, realizing closed-loop optimization from the desuperheating water end-point execution to the global feedwater-to-coal ratio. This module replaces the traditional cascade PID controller, specifically operating as follows: S10. Based on the data acquisition and feature fusion module, obtain time-series feature data characterizing the dynamics and coupling relationship of the steam temperature control system. Data such as temperature measurement sequence of superheaters and reheaters at various levels in the DCS, desuperheating water flow rate and valve opening, separator superheat, load command and fuel quantity, etc., are collected. Through preprocessing such as time alignment and anomaly filtering, a high-quality dataset is formed. On this basis, deep feature engineering is performed to extract key features such as "desuperheater inlet temperature change acceleration", "historical trend of the temperature difference between the desuperheater and superheater outlet", and "dynamic margin of steam temperature setpoint based on load change rate" to accurately characterize the dynamics and coupling relationship of the steam temperature system.

[0024] S11. Instead of relying on traditional rule bases with fixed thresholds, it constructs a dynamic mapping model with both "look-ahead" and "backward" capabilities. Its core is a deep neural network that receives real-time and historical sequences of superheater temperatures at each stage, superheat, load commands, and other multi-dimensional signals, and directly outputs a dynamically optimized setpoint bias vector based on time-series feature data. Output optimized setpoint partial vector As shown in the following formula:

[0025] in, Representative parameters are Deep neural networks, which extract data from time series data The complex relationship between combustion disturbance and steam temperature response is automatically learned. These are specific loads and their operating parameters; ,in, This is the set value for the secondary desuperheating water. This is the set value for the first-stage desuperheating water. Set the superheat value; S12. Train the deep neural network using historical operating data, with the composite loss function being the minimization of steam temperature deviation at each stage and the amount of desuperheating water used. Initialize the parameters through supervised learning or offline reinforcement learning. S13. After the deployment of the steam temperature control intelligent agent, the system enters the online adaptive mode. The intelligent agent integrates real-time data and environmental conditions to generate control commands. These commands are applied to the boiler object to generate new states. The reward signal is calculated based on performance indicators, and the model parameters are fine-tuned using online learning algorithms to continuously adapt to equipment aging and changes in fuel characteristics.

[0026] The steam temperature control agent, acting as an intelligent decision-maker at the upper layer of the DCS, outputs optimized commands that, after safety verification, are sent down to the lower-level actuators to form a closed-loop control. The system is also equipped with an online performance evaluation module, which can fine-tune the agent's strategy based on real-time operational results, enabling continuous iterative optimization.

[0027] The benchmark prediction agent outputs fuel quantity and coordinated water supply vectors under different coal quality mixtures at several future time points. By learning the temporal correlation in historical data, it achieves rolling prediction of fuel quantity and water supply benchmarks for several future sampling periods, generating a dynamic feedforward signal. The feedforward signal is corrected through a coal quality adaptive correction mechanism. This module aims to construct a dynamic feedforward setting model for fuel quantity that can adapt to changes in coal quality and operating conditions to improve variable load performance. Specifically, the operation is as follows: S20. Based on the data acquisition and feature fusion module, obtain feature vectors characterizing fuel demand and water supply coordination. ; such as load commands, total coal feed, primary air volume, total air volume, oxygen content, main steam pressure, etc., multi-dimensional historical and real-time data, key features are extracted, such as constructing "load commands and their first and second derivatives", "historical average fuel quantity under the same working conditions based on sliding time window", "main steam pressure deviation integral reflecting boiler heat storage status", etc., to form an input feature set that can comprehensively reflect the dynamics of fuel demand. S21. Construct a multi-step dynamic prediction model with Long Short-Term Memory (LSTM) network as its core, using feature vectors. Using the time series as input and the fuel quantity, which has been filtered and validated, as the output label, a supervised learning model is constructed to generate a dynamic feedforward signal, as shown in the following equation:

[0028] in, Based on fuel quantity, As the water supply benchmark, These are the weight parameters obtained from network training. This LSTM model learns complex temporal correlations in historical data to achieve rolling predictions of fuel baseline quantities and coordinated water supply baseline quantities for several future sampling periods, thereby generating dynamic feedforward signals that directly replace the fixed function generator in traditional coordinated control systems. The LSTM network structure endows the model with powerful temporal dependency learning capabilities, enabling it to automatically extract dynamic patterns of complex nonlinear disturbances such as coal quality changes and equipment performance degradation from historical data.

[0029] S22. Construct an online correction loop based on the equivalent calorific value method, calculate in real time the deviation between the theoretical output and the actual output of the unit corresponding to the actual fuel quantity, dynamically derive and update a coal quality correction coefficient, and use the coal quality correction coefficient as a multiplication factor to correct the dynamic feedforward signal; thereby quickly responding to abrupt changes in coal quality. S23. Establish an online update mechanism to regularly use recent running data to incrementally train or fine-tune the parameters of the LSTM model.

[0030] The intelligent agent for water supply and dry / wet state conversion outputs a coordinated action command sequence for the water supply pump, water supply regulating valve, and 361 valve equipment. By constructing a multi-classification state diagnostic model and using a combination of data-driven and expert rule-based methods to generate the coordinated action command sequence, this module achieves the goal of "one-click start / stop" and seamless dry / wet state conversion throughout the water supply system. Specifically, the operation is as follows: S30. Based on the data acquisition and feature fusion module, obtain the feature vector representing the overall operating status of the boiler. As shown in the following formula:

[0031] in, The rate of change of the feedwater-fuel ratio is... Reflects the energy balance state of the boiler core; and These are the inlet and outlet differential pressures and vibration amplitude of the boiler water circulation pump, respectively. The temperature difference between the walls of the steam-water separator is used to directly determine whether there is water inside. For valve opening degree 361, For valve 361 flow rate, This is the valve action sequence.

[0032] S31. Construct and train a three-class classification model using the random forest algorithm, as shown below:

[0033] Based on the aforementioned input feature vector It is used to identify whether the boiler is currently in a wet, transitional, or dry state. The model can not only accurately determine whether the boiler is currently in a wet, dry, or transitional state, but also predict the critical point of state transition 5-10 minutes in advance by analyzing subtle precursors such as changes in wall temperature gradient, providing a forward-looking window for intelligent decision-making.

[0034] S32. The system adopts a combination of data-driven and expert rule methods to generate control strategies. Specifically, based on historical best operating data and a simplified boiler hydrodynamic model, an optimization decision engine is constructed. In each decision, the engine takes the current state of the system as the starting point and, under the condition of strictly following multiple safety constraints such as wall temperature, water level, and pressure, calculates the operating trajectory that makes the conversion process smooth and fast through rolling optimization iteration. That is, the coordinated action command sequence of the feedwater pump, feedwater regulating valve, and valve 361, thereby replacing the traditional fixed program manual operation.

[0035] The detailed operating mechanism of the optimization decision engine is as follows: (1) Based on the simplified boiler hydrodynamic model, a discrete prediction relationship is established, specifically including: The next moment's steam-water separator wall temperature difference = current wall temperature difference + a1 × (feed water-fuel ratio change rate) + b1 × (361 valve opening change) The water level at the next moment = the current water level + a² × (change in water pump speed) - b² × (change in water supply regulating valve opening) The boiler water circulation pump outlet pressure at the next moment = current pressure - c × (361 valve opening change) The coefficients a1, a2, b1, b2, and c were obtained by fitting historical optimal operation data.

[0036] (2) Constructing a rolling optimization objective function through a simple weighted sum of squares, specifically including: At each decision-making moment, the engine solves the following optimization problem: finding a sequence of control instructions for several steps (e.g., the next 5 steps) from the current moment that minimizes the following total cost: weight_wall × (wall temperature difference) (Target wall temperature difference)² Weight_level × (water level) Target water level)² Weight_pressure × (pressure) Target pressure)² Weight_pump × (Change in pump speed)² Weight_valve × (Change in water supply valve opening)² Weight_361 × (361 valve opening change)² The target wall temperature difference, target water level, and target pressure are determined by a three-class classification model based on the current state (wet, transition, or dry) and the historical best transition trajectory. Each weight coefficient is a preset positive value to balance stability and speed.

[0037] (3) Optimize the solution process by introducing safety constraints, specifically including: The absolute value of the wall temperature difference must not exceed the safety limit (e.g., ±10℃). The water level must be maintained between the minimum and maximum safety lines; The differential pressure fluctuation at the inlet and outlet of the boiler water circulation pump must not exceed the allowable range; The opening degree and rate of change of each valve must not exceed the physical limit.

[0038] (4) Embedding expert rules into the optimization decision engine, specifically including: The weights are dynamically adjusted. For example, when the system is in a transition state and the wall temperature gradient rises too quickly, the expert rules automatically increase the weight of the "wall temperature difference tracking term" in the objective function and additionally increase the weight of the "361 valve opening change" to force the valve action to slow down and prevent excessive thermal stress.

[0039] An initial solution is provided, and the expert rule base stores several benchmark operation sequences for typical operating conditions (such as a timing template for "when transitioning from wet to dry state, first close valve 361, then increase the feedwater pump speed"). The optimization solver uses this benchmark sequence as the starting point for its search, rather than starting from zero, thereby significantly improving computational efficiency and avoiding unreasonable instructions.

[0040] (5) Optimize the rolling execution of decisions and safety interruption, specifically including: In each control cycle (e.g., every 10 seconds), the engine solves the above optimization problem to obtain the optimal sequence of control commands for the next few steps. The system only executes the command for the current step (i.e., the actions of the three devices), and then remeasures the state and optimizes again in the next cycle.

[0041] If, during the optimization process or execution, the measured value of any key parameter (wall temperature difference, water level, pressure) exceeds the preset warning threshold, the system will immediately stop the output of automatic optimization commands, switch to safety mode (e.g., gradually closing the water supply regulating valve at a fixed rate, slowly opening valve 361 to relieve pressure), and issue an alarm signal. The optimization decision engine will only be allowed to restart after the parameters have returned to a safe range and been confirmed.

[0042] Through the aforementioned rolling optimization mechanism that combines data-driven (model coefficients from historical data and reference trajectories from optimal transformation records) and expert rules (dynamic weighting and initial solution templates), this engine replaces the traditional fixed-program manual operation, realizing "one-click start / stop" and seamless switching between dry and wet states throughout the water supply system.

[0043] The intelligent decision-making system for key parameters of thermal power plants also includes an online learning and iterative optimization module, which is used to fine-tune and incrementally learn the agent model online based on real-time operational performance feedback.

[0044] Example 2 It should be further explained that, based on the same inventive concept, this embodiment provides a method for adjusting the main parameters of thermal power units to adapt to grid load fluctuations. Based on the system described in Embodiment 1, the specific implementation process is as follows: Figure 1 As shown, it includes the following steps: Collect and integrate multi-source data related to steam temperature, fuel, and feedwater during the operation of thermal power units and perform feature engineering; Intelligent decision-making is carried out on the partial vector of steam temperature setpoints at various levels, the fuel baseline quantity and coordinated water supply baseline quantity at several future moments, and the water supply and dry-wet state conversion strategy through deep reinforcement learning model, LSTM prediction model, and random forest classification model, respectively. The decision-making instructions are issued to the implementing agencies after security verification. Based on the performance evaluation results, the model is trained online and its parameters are fine-tuned.

[0045] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A main parameter regulation system for thermal power units adapted to grid load fluctuations, characterized in that, It includes a data acquisition and feature fusion module, an intelligent decision-making module, and a control execution module; the intelligent decision-making module is embedded as an upper-level optimizer in the distributed control system (DCS) of the thermal power plant, and the control commands it generates are sent down to the lower-level control loop for execution through the control execution module; The data acquisition and feature fusion module is used to acquire multi-source time-series data related to steam temperature, fuel, and feedwater in the distributed control system of thermal power units in real time, and to perform preprocessing and feature engineering. The intelligent decision-making module includes a steam temperature control intelligent agent, a baseline quantity prediction intelligent agent, and a feedwater and dry-wet state conversion intelligent agent; the steam temperature control intelligent agent outputs dynamically optimized bias vectors for each level of steam temperature; the baseline quantity prediction intelligent agent outputs vectors of fuel quantity and coordinated feedwater quantity under different coal quality mixtures at several future time points; the feedwater and dry-wet state conversion intelligent agent outputs a sequence of coordinated action instructions for feedwater pumps, feedwater regulating valves, and 361 valve equipment.

2. The thermal power unit main parameter adjustment system for adapting to grid load fluctuations according to claim 1, characterized in that, The system also includes an online learning and iterative optimization module, which is used to fine-tune and incrementally learn the agent model online based on real-time performance feedback.

3. The thermal power unit main parameter adjustment system for adapting to grid load fluctuations according to claim 1, characterized in that, The steam temperature control agent generates advanced feedforward compensation by sensing the upstream temperature change gradient; by analyzing the continuous deviation pattern of the downstream temperature of the boiler, it inversely infers the feedback correction amount for the fuel-water ratio source, realizing closed-loop optimization from the desuperheating water end execution to the global feedwater-coal ratio.

4. The thermal power unit main parameter adjustment system for adapting to grid load fluctuations according to claim 1, characterized in that, The benchmark quantity prediction agent learns the temporal correlation in historical data to achieve rolling prediction of fuel quantity benchmarks and water supply benchmarks for several future sampling periods, generating dynamic feedforward signals; and corrects the feedforward signals through a coal quality adaptive correction mechanism.

5. The thermal power unit main parameter adjustment system for adapting to grid load fluctuations according to claim 1, characterized in that, The water supply and dry / wet state transition decision-making agent generates a coordinated action instruction sequence by constructing a multi-classification state diagnostic model and using a combination of data-driven and expert rule methods.

6. The thermal power unit main parameter adjustment system for adapting to grid load fluctuations according to claim 3, characterized in that, The specific operation of the steam temperature control intelligent agent is as follows: S10. Based on the data acquisition and feature fusion module, obtain time-series feature data characterizing the dynamics and coupling relationship of the steam temperature control system. ; S11. Construct a deep neural network based on temporal feature data. Output optimized setpoint partial vector As shown in the following formula: in, Representative parameters are Deep neural networks, which extract data from time series data The complex relationship between combustion disturbance and steam temperature response is automatically learned. These are specific loads and their operating parameters; ,in, This is the set value for the secondary desuperheating water. This is the set value for the first-stage desuperheating water. Set the superheat value; S12. Train the deep neural network using historical operating data, with the composite loss function being the minimization of steam temperature deviation at each stage and the amount of desuperheating water used. Initialize the parameters through supervised learning or offline reinforcement learning. S13. After the deployment of the steam temperature control intelligent agent, the system enters the online adaptive mode. The intelligent agent integrates real-time data and environmental conditions to generate control commands. These commands are applied to the boiler object to generate new states. The reward signal is calculated based on performance indicators, and the model parameters are fine-tuned using online learning algorithms to continuously adapt to equipment aging and changes in fuel characteristics.

7. The thermal power unit main parameter adjustment system for adapting to grid load fluctuations according to claim 4, characterized in that, The specific operations of the benchmark quantity prediction agent are as follows: S20. Based on the data acquisition and feature fusion module, obtain feature vectors characterizing fuel demand and water supply coordination. ; S21. Construct a multi-step dynamic prediction model with Long Short-Term Memory (LSTM) network as its core, using feature vectors. Using the time series as input and the fuel quantity, which has been filtered and validated, as the output label, a supervised learning model is constructed to generate a dynamic feedforward signal, as shown in the following equation: in, Based on fuel quantity, As the water supply benchmark, These are the weight parameters obtained during network training. S22. Construct an online correction loop based on the equivalent calorific value method, calculate in real time the deviation between the theoretical output corresponding to the actual fuel quantity and the actual output of the unit, dynamically derive and update a coal quality correction coefficient, and use the coal quality correction coefficient as a multiplication factor to correct the dynamic feedforward signal. S23. Establish an online update mechanism to periodically use recent running data to incrementally train or fine-tune the parameters of the LSTM model.

8. The thermal power unit main parameter adjustment system for adapting to grid load fluctuations according to claim 5, characterized in that, The specific operations of the intelligent agent for water supply and dry / wet state conversion are as follows: S30. Based on the data acquisition and feature fusion module, obtain the feature vector representing the overall operating status of the boiler. As shown in the following formula: in, The rate of change of the feedwater-fuel ratio is... Reflects the energy balance state of the boiler core; and These are the inlet and outlet differential pressures and vibration amplitude of the boiler water circulation pump, respectively. The temperature difference between the walls of the steam-water separator is used to directly determine whether there is water inside. For valve opening degree 361, For valve 361 flow rate, This is the valve action sequence; S31. Construct and train a three-class classification model using the random forest algorithm, as shown below: Based on the aforementioned input feature vector This is used to identify whether the boiler is currently in a wet, transitional, or dry state. S32. The system adopts a combination of data-driven and expert rule methods to generate control strategies. Specifically, based on historical best operating data and a simplified boiler hydrodynamic model, an optimization decision engine is constructed. When making a decision, the engine starts from the current state of the system and, under the condition of strictly following multiple safety constraints of wall temperature, water level, and pressure, calculates the operating trajectory that makes the conversion process smooth and fast through rolling optimization iteration, namely the coordinated action command sequence of the feedwater pump, feedwater regulating valve, and valve 361.

9. The thermal power unit main parameter adjustment system for adapting to grid load fluctuations according to claim 8, characterized in that, During the decision-making process of the optimization decision engine, if any key parameter deviates from the safe range, the system automatically interrupts the current operation sequence and switches to safe mode.

10. A method for adjusting the main parameters of thermal power units to adapt to grid load fluctuations, characterized in that, Includes the following steps: Collect and integrate multi-source data related to steam temperature, fuel, and feedwater during the operation of thermal power units and perform feature engineering; Intelligent decision-making is carried out on the partial vector of each steam temperature setpoint, the fuel baseline quantity and the coordinated water supply baseline quantity at several future moments, and the water supply and dry-wet state conversion strategy through deep reinforcement learning model, LSTM prediction model, and random forest classification model, respectively. The decision-making instructions are issued to the implementing agencies after security verification. Based on the performance evaluation results, the model is trained online and its parameters are fine-tuned.

Images