A load self-adaptive control system for a combined cycle unit during a starting stage

Abstract

The application discloses a kind of load self-adaptive control systems of combined cycle unit starting stage, specifically relates to load self-adaptive control technical field, including state perception modeling module by collecting and marking historical start operation data, constructs thermodynamic characteristic sample library and formulates unit dynamic response model, load prediction decision module utilizes model to predict real-time start parameter, and filters out optimal load increasing rate as control strategy;Closed-loop adaptive control module compares control strategy with unit actual feedback data, automatically generates and tracks load adjustment instruction, forms closed-loop management system, realizes the change from fixed rate start to load self-adaptive control, effectively judges and avoids thermal stress overrun and combustion instability risk.

Description

Technical Field

[0001] This invention relates to the field of load adaptive control technology, and more specifically, to a load adaptive control system for the start-up phase of a combined cycle unit. Background Technology

[0002] Gas-fired combined cycle (Gas-Steam) units, due to their high efficiency, low emissions, and rapid start-up and shutdown capabilities, play an increasingly important role in peak shaving and emergency backup in modern power systems. To ensure that the unit can respond quickly and safely to grid load demands during startup, while avoiding damage to critical components due to excessive thermal stress, precise control of the startup process is crucial. However, traditional combined cycle unit startup control strategies have the following problems: First, traditional start-up control typically uses a fixed rate or segmented fixed load increase rate. The control logic is based on a preset static curve and cannot be dynamically adjusted according to the unit's current actual thermal state, environmental conditions, and equipment aging. This one-size-fits-all approach is either too conservative, leading to excessive start-up time and fuel waste, or it carries risks, potentially causing excessive thermal stress in the motor cylinder or unstable combustion under specific operating conditions.

[0003] Secondly, traditional control methods are open-loop or semi-open-loop, lacking online evaluation and closed-loop correction of control performance. When the actual unit response deviates from expectations (e.g., due to sudden changes in ambient temperature or component performance degradation), the control system cannot adaptively adjust the subsequent load increase rate, easily leading to start-up failures or unplanned shutdowns.

[0004] Furthermore, in existing technologies, the unit's startup strategy is disconnected from the management of component lifespan losses. Operators often find it difficult to quantify the low-cycle fatigue life losses caused by each rapid startup to gas turbine hot-channel components or turbine rotors, making it impossible to make optimal decisions between "rapid grid connection" and "extending lifespan." The vast amount of historical data accumulated during startup has not been effectively mined, and no data-driven dynamic predictive model has been established to guide future startup control.

[0005] Therefore, there is an urgent need for a combined cycle unit start-up phase load adaptive control system that can predict potential risks based on the real-time status of the unit, adaptively generate the optimal load increase rate, and track the control effect in a closed loop, so as to improve the flexibility, safety and economy of unit start-up. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a load adaptive control system for the start-up phase of a combined cycle unit. This system can predict thermal stress and combustion risks under different load increase strategies based on historical start-up data using artificial intelligence algorithms, automatically generate and execute optimal load commands, and track control effects to achieve closed-loop management. The purpose of the present invention is to solve the problems existing in traditional start-up control methods, such as fixed rate, lack of adaptive capability, inability to quantify life loss, and lack of closed-loop correction mechanism, thereby improving the intelligence level of the start-up process of combined cycle units. Through the present invention, predictive adaptive control of start-up load can be achieved, reducing the workload of human intervention, realizing the systematization of unit health management, and continuously improving the accuracy and economy of control strategies through continuous learning and optimization.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A load adaptive control system for the start-up phase of a combined cycle unit includes: a state-aware modeling module, used to classify the thermodynamic state of the combined cycle unit during the start-up phase, collect historical operating data and mark safe and risk states to build a training sample library, and build a dynamic response prediction model based on the properties of the unit's thermodynamic parameters. The load prediction decision module is connected to the state perception modeling module. It is used to receive the dynamic response prediction model, predict the unit state data during the real-time startup process, generate the thermal stress over-limit probability under different load increase rates, and sort and select the load increase rate within the preset safety threshold as the optimal control command according to the thermal stress over-limit probability. The closed-loop adaptive control module, connected to the load prediction and decision module, is used to receive the optimal control command, compare and analyze the real-time operating parameters fed back by the unit, determine the final safe load increase rate, and automatically trigger the control system setpoint update and tracking to realize closed-loop adaptive control of the unit load.

[0008] In a preferred embodiment, the state-aware modeling module includes: The data classification unit is used to classify the thermodynamic state of the combined cycle unit into several categories based on the actual operating conditions during the start-up phase, including cylinder thermal stress state, waste heat boiler temperature rise rate state, and combustion chamber pressure fluctuation state. The data acquisition unit is used to classify the parameters during the unit startup process according to the defined categories, and collect historical operating data including compressor outlet temperature, turbine exhaust temperature, turbine rotor temperature, and generator active power. The sample construction unit is used to label the collected historical running data with status tags and perform standardization processing to form a training sample library; The model building unit is used to analyze the attributes of training data and test data in the training sample library, select non-contradictory training data attributes and their corresponding test data, and use one of the following to build a dynamic response prediction model: support vector machine, random forest, deep neural network, or long short-term memory network.

[0009] In a preferred embodiment, the load forecasting decision module includes: The data acquisition unit is used to acquire real-time operating data during the startup process of the combined cycle unit, including current load, load increase rate, component temperature and pressure; The model application unit is used to input the real-time operating data into the dynamic response prediction model to generate the thermal stress over-limit probability under different load increase rates; The result sorting unit is used to sort the load increase rates according to the probability of thermal stress exceeding the limit, and to filter out the set of safe load increase rates within the preset risk threshold. The priority determination unit is used to select the load increase rates corresponding to the N load increase rates with the lowest probability of thermal stress exceeding the limit and the top M real-time operating data of the unit response speed as candidate load increase rates, calculate the comprehensive score of the candidate load increase rates, and select the candidate load increase rate with the highest comprehensive score as the current optimal control command, where N and M are positive integers.

[0010] In a preferred embodiment, the closed-loop adaptive control module includes: The comparative analysis unit is used to compare the prediction results with the real-time feedback data of thermal stress, vibration and exhaust temperature of the unit, and to analyze the confidence level of each load increase rate in the predictive control command. The instruction determination unit is used to generate the final safe load increase rate by combining the current forecast results with the actual operating boundary of the unit. The work order generation unit is used to automatically generate load adjustment instructions based on the final safe load increase rate and push the instructions to the unit's distributed control system (DCS). The closed-loop tracking unit is used to monitor the execution of load adjustment commands, automatically trigger control parameter corrections based on the actual response of the unit, and track the load adjustment process until the target load is stably completed.

[0011] In a preferred embodiment, it further includes a data cleaning and alignment module, which is connected to the state awareness modeling module and is specifically used to: clean the multi-source heterogeneous data during the historical startup process and remove sensor outliers and communication interruption segments. Timestamp alignment and resampling are performed on data sources from various sampling frequencies; The aligned and cleaned structured data is stored in the training sample library for use by the state-aware modeling module.

[0012] In a preferred embodiment, a lifespan loss management module is further included, connected to the closed-loop adaptive control module, specifically for: Record the cumulative number of starts and equivalent operating hours of unit components; Monitor the cumulative low-cycle fatigue life loss of each unit component and transmit the life loss rate information to the closed-loop adaptive control module. When the load increase rate determined by the closed-loop adaptive control module causes the life loss rate to exceed the preset life threshold, an optimization prompt for the load decrease rate is automatically generated.

[0013] In a preferred embodiment, a boundary condition evaluation module is further included, connected to the load forecasting decision module, specifically for: Determine sensitivity weights based on boundary conditions; Data mining analysis was performed on the 10 to 20 boundary conditions with the highest sensitivity weights. Based on historical actual startup data, the prediction accuracy of the dynamic response prediction model under various boundary conditions is calculated. Based on the prediction accuracy results and the actual operating requirements of the unit, the final inner loop control parameter correction coefficient is determined and fed back to the load prediction decision module.

[0014] In a preferred embodiment, the closed-loop tracking unit includes: The status monitoring subunit is used to monitor the tracking deviations of unit load, main steam temperature, and cylinder differential expansion operating parameters; The early warning triggering subunit sends a reminder to the operators when the tracking deviation of the operating parameters exceeds the preset deviation threshold; When the tracking deviation still cannot converge after three automatic corrections, the over-limit intervention subunit generates an over-limit warning and suggests switching to manual control mode. The closed-loop verification subunit is used to verify the load control effect of this startup process based on the operating data after the target load has stabilized, and to archive the startup strategy after confirming that the control target has been achieved.

[0015] In a preferred embodiment, a data integration interface module is further included, for: It exchanges data with the unit's distributed control system (DCS) to obtain real-time process data; Exchange data with the life loss management module to obtain component life consumption data; The acquired real-time process data and component life consumption data are integrated and transmitted to the load prediction decision module as prediction input data; Establish a data source database to store and update data from the unit control system, vibration monitoring system, and performance calculation system.

[0016] In a preferred embodiment, a self-learning optimization module is further included, connected to the state-aware modeling module, for: Collect comparative data between predicted control commands and actual unit responses during each startup process; Analyze the trend of accuracy changes in dynamic response prediction models; When the accuracy reaches the set accuracy threshold, the model is retrained. The training sample library is updated based on the latest startup data, and the state-aware modeling module is notified to rebuild the dynamic response prediction model to achieve continuous self-optimization of the system.

[0017] The technical effects and advantages of this invention are as follows: 1. This invention realizes the transformation from fixed-rate start-up to load adaptive control, enabling early prediction and avoidance of risks such as excessive thermal stress and combustion instability, achieving rapid and safe grid connection. Secondly, this invention establishes a complete closed-loop control management mechanism, ensuring that each load adjustment command is effectively executed and corrected, preventing unit parameters from exceeding limits. Employing multiple AI algorithm technologies, it achieves high accuracy and reliability in load command prediction, and designs a self-learning optimization mechanism, enabling the system to continuously adapt to unit aging and environmental changes, improving control performance. Furthermore, it seamlessly integrates with the unit's existing DCS system, realizing automatic data flow and sharing, reducing the workload of operators.

[0018] 2. This invention can be widely applied to various gas-steam combined cycle generator sets, providing strong support for improving the grid peak-shaving response speed and the unit operation safety. At the same time, the technical ideas and methods of this invention can also be extended to the start-up control of other types of complex thermal systems, and have good promotional value. Attached Figure Description

[0019] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is an overall architecture diagram of a load adaptive control system for the start-up phase of a combined cycle unit according to the present invention; Figure 2 This is a structural block diagram of the state-aware modeling module of the present invention; Figure 3 This is a structural block diagram of the load forecasting decision module of the present invention; Figure 4 This is a structural block diagram of the closed-loop adaptive control module of the present invention; Figure 5 This is a flowchart illustrating the workflow of the data cleaning and alignment module of the present invention. Figure 6 This is a structural block diagram of the lifespan loss management module of the present invention; Figure 7 This is a schematic diagram illustrating the working principle of the boundary condition evaluation module of the present invention. Figure 8 This is a flowchart illustrating the operation of the closed-loop tracking unit of the present invention. Figure 9 This is a data flow diagram of the data integration interface module of the present invention; Figure 10 This is a schematic diagram illustrating the working principle of the self-learning optimization module of this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0021] Example 1: As Figure 1 As shown, the present invention provides a load adaptive control system for the start-up phase of a combined cycle unit, including a state perception modeling module 1, a load prediction and decision module 2, and a closed-loop adaptive control module 3. The state-aware modeling module 1 is used to classify the thermodynamic state of the combined cycle unit during the start-up phase, collect historical operating data and mark the safe and risk states to build a training sample library, and build a dynamic response prediction model based on the attributes of the unit's thermodynamic parameters. The load prediction decision module 2 is connected to the state perception modeling module 1 and is used to receive the dynamic response prediction model, predict the unit state data during the real-time startup process, generate the thermal stress over-limit probability under different load increase rates, and sort and select the load increase rate within the preset safety threshold as the optimal control command according to the thermal stress over-limit probability. The closed-loop adaptive control module 3 is connected to the load prediction and decision module 2. It is used to receive the optimal control command, compare and analyze the real-time operating parameters fed back by the unit, determine the final safe load increase rate, and automatically trigger the control system set value update and tracking to realize the closed-loop adaptive control of the unit load.

[0022] In this embodiment, the three core modules form a complete intelligent control system for data modeling, prediction and decision-making, and closed-loop control. Preferably, the state perception modeling module 1 and the load prediction and decision-making module 2 transmit data through an OPC UA interface, and the load prediction and decision-making module 2 and the closed-loop adaptive control module 3 communicate asynchronously through a high-speed real-time Ethernet, ensuring efficient and stable data flow between the system modules.

[0023] In practical applications, this system can be deployed on the industrial control server or plant-level monitoring information system (SIS) on the unit side, providing control strategy visualization through a human-machine interface. During system operation, the state-aware modeling module 1 first extracts historical startup data from the DCS and historical server to build a dynamic response prediction model; then, the load prediction decision module 2 analyzes the real-time startup parameters, predicts potential thermal stress risks, and recommends the optimal load increase rate; finally, the closed-loop adaptive control module 3 sends control commands to the DCS for execution and tracks the effects, forming a complete closed-loop adaptive control.

[0024] Example 2: Implementation of the State-Aware Modeling Module like Figure 2 As shown, the state-aware modeling module 1 includes a data classification unit 11, a data acquisition unit 12, a sample construction unit 13, and a model construction unit 14.

[0025] The data classification unit 11 is used to classify the thermodynamic state of the combined cycle unit into several categories based on the actual operating conditions during the start-up phase. In this embodiment, the thermodynamic state categories may include, but are not limited to: cylinder thermal stress state, waste heat boiler temperature rise rate state, combustion chamber pressure fluctuation state, shaft vibration state, etc. Preferably, the system can expand or adjust the state categories according to the actual unit configuration to ensure the completeness and applicability of the classification.

[0026] The data acquisition unit 12 is used to classify parameters during the unit startup process according to the defined categories, and collect historical operating data including compressor outlet temperature, turbine exhaust temperature, turbine rotor temperature, and generator active power. In practical applications, the data acquisition unit 12 can acquire historical startup data through OPC interface, Modbus protocol, or file import. Preferably, the data acquisition cycle can be set to automatically trigger after each startup to ensure the timeliness and representativeness of the data.

[0027] The sample construction unit 13 is used to label the collected historical operating data with status tags and perform standardization processing to form a training sample library. Standardization processing includes, but is not limited to, handling missing values, handling outliers, and standardizing data formats. In this embodiment, the sample construction unit 13 uses Z-score standardization to process numerical features, making their mean 0 and standard deviation 1; it uses One-Hot encoding to process operational condition category features, converting them into binary form. Preferably, for features with a missing value ratio exceeding 30%, they can be considered for removal or filled using interpolation; the specific processing method can be flexibly selected according to the data characteristics.

[0028] The model building unit 14 is used to analyze the attributes of training data and test data in the training sample library, select non-contradictory training data attributes and their corresponding test data, and construct a dynamic response prediction model using one of the following: Support Vector Machine, Random Forest, Deep Neural Network, or Long Short-Term Memory Network. Since the startup process has obvious time-series characteristics, this embodiment preferentially selects Long Short-Term Memory Network (LSTM) as the prediction model.

[0029] In this embodiment, the LSTM algorithm is taken as an example. Its core idea is to capture long-term dependencies in time series data. The implementation steps of the LSTM unit are as follows: 1. The forget gate determines which information is discarded from the cell's previous state: in, The output vector of the forget gate at time t (with values ​​ranging from 0 to 1) determines the proportion of information discarded from the cell state at the previous time step. The Sigmoid activation function maps input values ​​to the (0,1) interval. Here is the weight matrix for the forget gate. Let be the hidden state vector from the previous time step. The input vector at time t (e.g., current load, rate of temperature rise). For the bias term of the forget gate; 2. The input gate determines which new information is stored in the cell state: in, The output vector of the input gate determines which information needs to be updated. , The weight matrices correspond to the input gate and the candidate cell state, respectively. , For the corresponding bias term, Let be the candidate cell state vector, representing the new information to be stored. This is the hyperbolic tangent activation function, with an output range of (-1, 1). 3. Update cell status: in, This is the cell state vector updated at time t (long-term memory). Let be the cell state vector at time t-1. This is element-wise multiplication (Hadamard product). 4. The output gate determines the output based on the cell state: in, The output vector of the output gate determines how much information is output from the cell state. This is the weight matrix of the output gate. This is the bias term for the output gate. Let t be the hidden state vector (short-term memory, which is also the predicted output at that time, such as the predicted value of thermal stress).

[0030] The mathematical model of an LSTM network can be represented as a mapping function from the input sequence to the output sequence.

[0031] In practical applications, LSTM networks are typically set to 2-3 layers, with 50-200 hidden neurons per layer, which can be adjusted according to the data scale and computational resources. Preferably, the model building unit 14 also verifies and optimizes the built model to ensure its accuracy and generalization ability. Verification methods include, but are not limited to, time series cross-validation and hold-out methods. Model evaluation metrics include, but are not limited to, root mean square error (RMSE) and mean absolute percentage error (MAPE). When the model evaluation metrics reach preset thresholds (e.g., RMSE < 0.05, MAPE < 3%), the model is considered usable.

[0032] Example 3: Implementation of the load forecasting decision module like Figure 3 As shown, the load forecasting decision module 2 includes a data acquisition unit 21, a model application unit 22, a result sorting unit 23, and a priority determination unit 24.

[0033] The data acquisition unit 21 is used to acquire real-time operating data during the startup process of the combined cycle unit, including current load, load increase rate, and temperature and pressure of key components. In this embodiment, real-time data sources include, but are not limited to, the distributed control system (DCS), the vibration monitoring system (TDM), and the performance calculation system. The data acquisition unit 21 acquires this data through the OPC UA interface or database query method and performs preprocessing to ensure that the data format is consistent with the training data. Preferably, the data acquisition frequency can be set to 1Hz or higher to meet the requirements of real-time control.

[0034] The model application unit 22 is used to input real-time operating data into the dynamic response prediction model to generate the probability of thermal stress exceeding limits under different load increase rates. In this embodiment, the model application unit 22 loads the LSTM model trained by the state-aware modeling module 1 to perform rolling predictions on the real-time data and calculate the probability of thermal stress exceeding limits within the next 5-10 minutes under each candidate load increase rate (e.g., 1MW / min, 3MW / min, 5MW / min, 7MW / min, 9MW / min). Preferably, to improve processing efficiency, the model application unit 22 can employ parallel computing technology to simultaneously evaluate multiple candidate load increase rates.

[0035] The result sorting unit 23 is used to sort the load increase rates according to the probability of thermal stress exceeding the limit, and filter out the set of safe load increase rates within a preset risk threshold. In this embodiment, the preset risk threshold is determined based on historical model performance and equipment safety requirements, and is typically set to a thermal stress exceeding the limit threshold of 10%. For example, when the risk threshold is set to 10%, only load increase rates with a thermal stress exceeding the limit probability of less than 10% will be filtered out. Preferably, the risk threshold can be set differently for different components; for expensive components such as turbine rotors, the threshold can be appropriately reduced to increase the safety margin.

[0036] The priority determination unit 24 is used to select the load increase rates corresponding to the N load increase rates with the lowest probability of thermal stress exceeding the limit and the top M real-time operating data of the unit response speed as candidate load increase rates, and calculate the comprehensive score of the candidate load increase rates. The candidate load increase rate with the highest comprehensive score is selected as the current optimal control command, where N and M are positive integers. In this embodiment, the values ​​of N and M are determined according to actual control requirements, typically N is set to 2-3 and M is set to 3-5. This dual screening mechanism considers both safety (lowest risk) and economy (faster response speed), and can more comprehensively balance the needs of rapid start-up and safe operation. Preferably, the values ​​of N and M can be dynamically adjusted based on historical start-up effects and operator feedback to gradually optimize the control strategy.

[0037] In this embodiment, the overall score is calculated using the following formula: in, The score represents the overall score for the i-th candidate load increase rate; a higher score indicates a better strategy. For the i-th candidate load increase rate, The predicted probability (excessive thermal stress, unstable combustion, etc.) given by the model when using this load increase rate is in the range of [0,1]. The expected response speed score of the unit when adopting this load increase rate is normalized, ranging from 0 to 1. , Let be the weighting coefficient, satisfying Typical values (Security weight) (Response speed weight).

[0038] Example 4: Implementation of the Closed-Loop Adaptive Control Module like Figure 4 As shown, the closed-loop adaptive control module 3 includes a comparison analysis unit 31, an instruction determination unit 32, a work order generation unit 33, and a closed-loop tracking unit 34.

[0039] The comparative analysis unit 31 is used to compare the prediction results with the real-time feedback data of thermal stress, vibration, and exhaust temperature from the unit, and to analyze the confidence level of each load increase rate in the predictive control command. In this embodiment, the comparative analysis unit 31 calculates the deviation between the predicted value and the actual feedback value, including but not limited to: thermal stress prediction deviation, exhaust temperature prediction deviation, etc. Preferably, the comparative analysis uses an exponentially weighted moving average method, assigning higher weight to the most recent deviation to ensure the timeliness of the analysis results.

[0040] The instruction determination unit 32 is used to generate the final safe load increase rate by combining the current prediction results with the actual operating boundary of the unit. In this embodiment, the instruction determination unit 32 comprehensively considers prediction risk, real-time feedback deviation, and unit hard constraints (such as the maximum allowable temperature rise rate), and uses a weighted method to determine the final load increase rate. The weighted calculation formula is as follows: in, Indicates the final load factor. This represents the optimal load increase rate recommended by the load forecasting decision module. This is a correction factor, ranging from 0.2 to 0.5, which can be adaptively adjusted according to the actual deviation trend. This represents the actual thermal stress value reported by the unit. The predicted thermal stress value output by the dynamic response prediction model.

[0041] In practical applications, the correction coefficient can be adaptively adjusted according to the deviation trend; for example, it can be increased when the deviation continues to widen. The value is increased to enhance the correction strength. Preferably, the final load increase rate must also meet the upper and lower limit constraints of the unit protection logic to ensure the safety of control commands.

[0042] The work order generation unit 33 is used to automatically generate load adjustment instructions based on the final safe load increase rate and push the instructions to the distributed control system (DCS) of the unit. In this embodiment, the load adjustment instructions include the following key information: target load increase rate, duration, expected final load, safety constraints, etc. Preferably, the instructions are generated using a standardized communication protocol (such as MODBUS or OPC) and directly written into the control setpoints of the DCS to achieve fully automatic control.

[0043] The closed-loop tracking unit 34 monitors the execution of load adjustment commands, automatically triggers control parameter corrections based on the actual unit response, and tracks the load adjustment process until the target load is stably achieved. In this embodiment, the closed-loop tracking employs a PID incremental algorithm, dynamically adjusting the load increase rate based on the deviation between the actual load and the target load. The closed-loop tracking unit 34 automatically sends control correction commands according to state transition rules and time nodes to ensure smooth progress during startup. Preferably, the closed-loop tracking also includes a control effect evaluation function, verifying the effectiveness of the control strategy through subsequent stable operating data to form a complete control optimization closed loop.

[0044] Example 5: Implementation of the data cleaning and alignment module like Figure 5 As shown, the system of the present invention also includes a data cleaning and alignment module 4, which is connected to the state perception modeling module 1. It is used to clean the multi-source heterogeneous data during the historical startup process, remove sensor anomalies and communication interruption segments; perform timestamp alignment and resampling on data sources from different sampling frequencies (such as vibration monitoring, temperature acquisition, and power transmitters); and store the aligned and cleaned structured data in the training sample library for use by the state perception modeling module 1.

[0045] In this embodiment, the workflow of the data cleaning and alignment module 4 includes three main steps: data cleaning, time alignment, and resampling. During the data cleaning stage, the system identifies and removes sensor saturation values, abrupt changes, and null values ​​caused by communication packet loss. During the time alignment stage, the system uses the highest sampling frequency (e.g., 100Hz vibration data) as a reference and performs linear interpolation or zero-order hold on other low-frequency data (e.g., 1Hz temperature data) to ensure that all data are aligned on the same time coordinate. During the resampling stage, the system downsamples the aligned data to 1Hz for easier model training.

[0046] Preferably, the data cleaning and alignment module 4 employs an adaptive threshold cleaning algorithm to address the non-stationary characteristics of the startup process. This algorithm can distinguish between genuine parameter fluctuations and sensor malfunctions, improving the accuracy of data cleaning. Data segments that cannot be automatically repaired are marked by the system for easy review and correction by engineers.

[0047] In practical applications, the data cleaning and alignment module 4 achieves a data processing accuracy of over 95%, significantly reducing the labor costs of data preprocessing while improving the completeness and accuracy of the training sample library.

[0048] Example 6: Implementation of the Lifetime Loss Management Module like Figure 6 As shown, the system of the present invention also includes a life loss management module 5, which is connected to the closed-loop adaptive control module 3. It is used to record the cumulative number of starts and equivalent operating hours of key components of the unit (including gas turbine hot channel components and steam turbine high- and medium-pressure rotors); monitor the cumulative low-cycle fatigue life loss of each component, and transmit the life loss rate information to the closed-loop adaptive control module 3; when the load increase rate determined by the closed-loop adaptive control module 3 causes the life loss rate to exceed the preset threshold, it automatically generates a load reduction rate optimization prompt.

[0049] In this embodiment, the lifespan loss management module 5 mainly manages the lifespan-critical components within the combined cycle unit, including but not limited to: the gas turbine first-stage rotor blades, the gas turbine combustion chamber bushings, and the steam turbine high-pressure rotor. The lifespan loss management module 5 maintains a component lifespan ledger, recording basic component information (component number, material, design life, accumulated consumption) and the current loss rate.

[0050] Preferably, the life loss management module 5 employs a low-cycle fatigue life assessment model based on the Manson-Coffin formula, calculating the percentage of life loss caused by each startup in real time based on the temperature change amplitude and allowable cycle number during the startup process. The calculation formula is as follows: in, For thermal strain amplitude, This refers to the temperature change amplitude of critical components (such as the rotor) during startup. The coefficient of thermal expansion of the material; in, This represents the number of cycles a material can withstand at a given strain amplitude. This is a material constant, related to ductility. Here are the material constants and fatigue strength. in, The percentage of lifespan loss due to a single startup; When the cumulative lifespan loss percentage exceeds the preset lifespan threshold (such as 80% of the design lifespan), the system will automatically generate an early warning and suggest adopting a more conservative load increase strategy in subsequent startups.

[0051] In practical applications, the linkage mechanism between the life loss management module 5 and the closed-loop adaptive control module 3 can effectively balance startup speed and component lifespan. For example, when a critical component is nearing the end of its lifespan, the system will automatically reduce the maximum allowable load increase rate to extend its service life.

[0052] Example 7: Implementation of the Boundary Condition Evaluation Module like Figure 7 As shown, the system of the present invention also includes a boundary condition evaluation module 6, which is connected to the load forecasting decision module 2. It is used to formulate sensitivity weights based on boundary conditions such as ambient temperature, atmospheric pressure, and fuel calorific value; perform data mining analysis on the 10 to 20 boundary conditions with the highest sensitivity weights; calculate the prediction accuracy of the dynamic response prediction model under different boundary conditions by combining historical actual startup data; and determine the final inner loop control parameter correction coefficient based on the prediction accuracy results and the actual operating requirements of the unit, and feed it back to the load forecasting decision module 2.

[0053] In this embodiment, the boundary condition evaluation module 6 uses sensitivity analysis to evaluate the boundary condition weights. First, through orthogonal experimental design, the influence of changes in each boundary condition on the peak thermal stress of the unit is analyzed; the greater the influence, the higher the weight. The weight calculation formula is as follows: in, This represents the sensitivity weight for the nth boundary condition; a larger value indicates a more significant impact of that boundary condition on the peak thermal stress. This represents the change in peak thermal stress. Let n be the change in the nth boundary condition, for example, a 1°C change in ambient temperature or a 0.1 MJ / Nm³ change in fuel calorific value. 3 .

[0054] Based on the calculated sensitivity weights, the boundary condition assessment module 6 selects the top 10-20 boundary conditions for focused attention and analysis. Preferably, the boundary condition assessment can also incorporate interaction effect analysis to consider the coupling effects when multiple boundary conditions change simultaneously.

[0055] Regarding control parameter correction, the boundary condition evaluation module 6 establishes a mapping relationship between boundary conditions and control correction coefficients through table lookup or neural network fitting. Specifically, under different boundary conditions, offline simulation is used to find the correction coefficients that optimize the control effect, forming a correction coefficient table. In actual operation, the system retrieves the correction coefficients from the table in real time based on the current boundary conditions.

[0056] In practical applications, to accommodate different operational needs, the boundary condition assessment module 6 also considers the balance between the urgency of power grid commands and lifetime protection. When the power grid urgently needs peak power, the correction coefficient can be appropriately relaxed to prioritize response speed.

[0057] Example 8: Implementation of the Closed-Loop Tracking Unit like Figure 8 As shown, the closed-loop tracking unit 34 includes a status monitoring subunit 341, an early warning triggering subunit 342, an over-limit intervention subunit 343, and a closed-loop verification subunit 344.

[0058] The status monitoring subunit 341 is used to monitor the tracking deviations of key operating parameters such as unit load, main steam temperature, and cylinder differential expansion. In this embodiment, the status monitoring subunit 341 checks the deviation between the actual load and the set load in real time (e.g., every second), and calculates the absolute value and rate of change of the deviation. Preferably, the control deviation visualization is achieved using trend curves and digital dashboards to intuitively display the tracking status and changing trends of each parameter.

[0059] The warning triggering subunit 342 is used to send a reminder to the operator when the tracking deviation of a key operating parameter exceeds a preset deviation threshold. In this embodiment, the criteria for judging the tracking deviation of key operating parameters include, but are not limited to: load deviation exceeding 5% of the set value, main steam temperature deviating from the sliding pressure curve by more than ±8℃, and cylinder differential expansion exceeding 80% of the alarm value. When an anomaly is detected, the warning triggering subunit 342 will notify the operator through system messages, indicator lights, or voice. Preferably, the reminder level and method can be adjusted according to the severity of the deviation; for parameters approaching the protection setpoint, the reminder frequency can be increased and intervention can be suggested.

[0060] The over-limit intervention subunit 343 is used to generate an over-limit warning and suggest switching to manual control mode when the tracking deviation of the operating parameters still cannot converge after three automatic corrections. In this embodiment, the over-limit intervention level is higher than a general reminder, and warning information will be sent simultaneously to the shift leader, specialist, and the person in charge of the operations department, and recorded in the unit operation analysis system. Preferably, the over-limit intervention will also trigger a control strategy rollback mechanism to reduce the load increase rate to the next higher safety level to ensure unit safety.

[0061] The closed-loop verification subunit 344 is used to verify the load control effect of this startup process based on the operating data after the target load has stabilized, and to archive the startup strategy after confirming that the control target has been achieved. In this embodiment, the closed-loop verification adopts a dual mechanism: first, it statistically analyzes key indicators such as the load increase rate, maximum tracking deviation, and adjustment time of the startup process; second, it scores the verification results (e.g., a maximum score of 100 points) to generate a startup quality report. Verification methods include, but are not limited to, comparison with the control target and comparison with the historical best startup. Preferably, for startup processes with low scores, the system automatically analyzes the reasons and stores them in the knowledge base for subsequent model optimization.

[0062] In practical applications, the closed-loop tracking unit 34 realizes full monitoring and management of the start-up control process, ensuring that each load adjustment is smooth and reliable, forming a true closed-loop adaptive control.

[0063] Example 9: Implementation of the Data Integration Interface Module like Figure 9 As shown, the system of the present invention also includes a data integration interface module 7, which is used to exchange data with the unit's distributed control system (DCS) to obtain real-time process data; exchange data with the life loss management module 5 to obtain component life consumption data; integrate the obtained data and transmit it to the load prediction decision module 2 as prediction input data; and establish a data source database to store and update data from the unit control system, vibration monitoring system, and performance calculation system.

[0064] In this embodiment, the data integration interface module 7 adopts an industrial internet platform architecture to realize the collection, transformation, and aggregation of multi-source heterogeneous industrial data. During the collection phase, the system acquires real-time data from various subsystems through industrial protocols such as OPC UA, Modbus TCP, and IEC 104. During the transformation phase, the system performs range conversion, engineering unit conversion, and quality judgment on the data to ensure data consistency and availability. During the aggregation phase, the system stores the processed data in a real-time database and a historical database for use by other modules.

[0065] Preferably, the data integration interface module 7 adopts a combination of real-time processing and batch processing. For critical control data (such as current load and main steam temperature), a millisecond-level real-time processing mode is used; for analytical data (such as historical startup curves), a batch processing mode is used to improve system efficiency. The data synchronization frequency is set according to the data type and business requirements; for example, vibration data is synchronized every 20 milliseconds, load data is synchronized every 100 milliseconds, and historical data is archived daily.

[0066] In terms of data storage, the data source database adopts an architecture that combines time-series databases (such as InfluxDB) and relational databases. The time-series database is used to store frequently changing operating parameters, such as temperature, pressure, and vibration; the relational database is used to store structured ledger information, such as equipment parameters and startup count statistics.

[0067] In terms of security mechanisms, the data integration interface module 7 implements industrial firewall isolation and communication encryption to ensure secure interaction between the control network and the information network. All data exchange processes are logged for auditing and traceability.

[0068] In practical applications, the data integration interface module 7 serves as a bridge connecting the control system and the intelligent optimization system, providing comprehensive, accurate, and real-time data support for AI predictive analysis. Example 10: Implementation of the self-learning optimization module like Figure 10 As shown, the system of the present invention also includes a self-learning optimization module 8, which is connected to the state-aware modeling module 1. It is used to collect comparison data between the predicted control commands and the actual unit response during each startup process; analyze the accuracy change trend of the dynamic response prediction model; trigger model retraining when the accuracy reaches the set accuracy threshold or the accuracy improvement stagnates; update the training sample library according to the latest startup data, and notify the state-aware modeling module 1 to reconstruct the dynamic response prediction model, so as to realize the continuous self-optimization of the system.

[0069] In this embodiment, the self-learning optimization module 8 employs a strategy combining incremental learning and periodic reconstruction to periodically evaluate and optimize the performance of the prediction model. Evaluation metrics include, but are not limited to, the root mean square error (RMSE) of thermal stress prediction and the accuracy of the proposed load increase rate. The model performance evaluation uses a sliding window method, considering only data from the most recent 10-20 startups to ensure timely evaluation.

[0070] Preferably, the self-learning optimization module 8 analyzes the prediction performance under different operating conditions (such as cold start, warm start, and hot start), identifies poorly performing conditions, and optimizes them accordingly. Optimization methods include, but are not limited to: increasing the training samples for relevant operating conditions, adjusting network hyperparameters, and optimizing feature engineering. For example, when a large prediction error is found in cold start, the prediction accuracy can be improved by increasing the weight of cold start samples or extending the time backtracking window of the LSTM network.

[0071] Regarding the triggering mechanism for model retraining, the self-learning optimization module 8 sets two conditions: first, when the model performance reaches a preset target (e.g., RMSE < 0.03), the model performance is good enough to enter a stable operation phase; second, when the model performance improvement stagnates (e.g., performance change < 1% after 5 consecutive startup evaluations), new data needs to be introduced or the model structure adjusted to overcome the performance bottleneck. Preferably, the frequency of model retraining should not be too high, generally set to once a month or triggered after 10 new startup data are added cumulatively, in order to balance the optimization effect and the consumption of computing resources.

[0072] Regarding the updating of the training sample library, the self-learning optimization module 8 incorporates the latest startup results and prediction bias data, especially cases with previously large prediction errors, to help the model learn and improve. Simultaneously, it removes outdated or abnormal startup data to maintain the timeliness and representativeness of the sample library. Preferably, the sample library update employs a rolling window strategy, retaining startup data from the most recent 1-2 years, ensuring the inheritance of historical experience while avoiding interference from old data on the model.

[0073] In practical applications, the self-learning optimization module 8 realizes the intelligence and adaptability of the control system, enabling the system to continuously adjust and optimize according to unit aging, environmental changes, and operational requirements, thereby improving control accuracy and economy. For example, after the unit undergoes a major overhaul or burner upgrade, the system can quickly learn new thermodynamic characteristics through several startups and adjust the prediction model; when the power grid imposes new requirements on peak-shaving rates, the system can also adapt to the new objectives and maximize response speed while ensuring safety.

[0074] Through the detailed description of the above embodiments, the present invention provides a load adaptive control system for the start-up phase of a combined cycle unit, which realizes intelligent prediction of start-up load, adaptive command generation, and closed-loop control, significantly improving the safety, speed, and economy of the start-up process of the combined cycle unit. The system constructs a dynamic response prediction model through a state-aware modeling module, analyzes and makes decisions on real-time parameters through a load prediction decision module, executes and tracks control commands through a closed-loop adaptive control module, and provides comprehensive functional support through auxiliary modules such as data cleaning and alignment, lifetime loss management, boundary condition assessment, data integration, and self-learning optimization.

[0075] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0076] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0077] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0078] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0079] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A load adaptive control system for the start-up phase of a combined cycle power unit, characterized in that: include: The state-aware modeling module is used to classify the thermodynamic state of the combined cycle unit during the start-up phase, collect historical operating data and mark the safe and risk states to build a training sample library, and build a dynamic response prediction model based on the attributes of the unit's thermodynamic parameters. The load prediction decision module is connected to the state perception modeling module. It is used to receive the dynamic response prediction model, predict the unit state data during the real-time startup process, generate the thermal stress over-limit probability under different load increase rates, and sort and select the load increase rate within the preset safety threshold as the optimal control command according to the thermal stress over-limit probability. The closed-loop adaptive control module, connected to the load prediction and decision module, is used to receive the optimal control command, compare and analyze the real-time operating parameters fed back by the unit, determine the final safe load increase rate, and automatically trigger the control system setpoint update and tracking to realize closed-loop adaptive control of the unit load.

2. The load adaptive control system for the start-up phase of a combined cycle unit according to claim 1, characterized in that: The state-aware modeling module includes: The data classification unit is used to classify the thermodynamic state of the combined cycle unit into several categories based on the actual operating conditions during the start-up phase, including cylinder thermal stress state, waste heat boiler temperature rise rate state, and combustion chamber pressure fluctuation state. The data acquisition unit is used to classify the parameters during the unit startup process according to the defined categories, and collect historical operating data including compressor outlet temperature, turbine exhaust temperature, turbine rotor temperature, and generator active power. The sample construction unit is used to label the collected historical running data with status tags and perform standardization processing to form a training sample library; The model building unit is used to analyze the attributes of training data and test data in the training sample library, select non-contradictory training data attributes and their corresponding test data, and use one of the following to build a dynamic response prediction model: support vector machine, random forest, deep neural network, or long short-term memory network.

3. The load adaptive control system for the start-up phase of a combined cycle unit according to claim 1, characterized in that: The load forecasting decision module includes: The data acquisition unit is used to acquire real-time operating data during the startup process of the combined cycle unit, including current load, load increase rate, component temperature and pressure; The model application unit is used to input the real-time operating data into the dynamic response prediction model to generate the thermal stress over-limit probability under different load increase rates; The result sorting unit is used to sort the load increase rates according to the probability of thermal stress exceeding the limit, and to filter out the set of safe load increase rates within the preset risk threshold. The priority determination unit is used to select the load increase rates corresponding to the N load increase rates with the lowest probability of thermal stress exceeding the limit and the top M real-time operating data of the unit response speed as candidate load increase rates, calculate the comprehensive score of the candidate load increase rates, and select the candidate load increase rate with the highest comprehensive score as the current optimal control command, where N and M are positive integers.

4. The load adaptive control system for the start-up phase of a combined cycle unit according to claim 1, characterized in that: The closed-loop adaptive control module includes: The comparative analysis unit is used to compare the prediction results with the real-time feedback data of thermal stress, vibration and exhaust temperature of the unit, and to analyze the confidence level of each load increase rate in the predictive control command. The instruction determination unit is used to generate the final safe load increase rate by combining the current forecast results with the actual operating boundary of the unit. The work order generation unit is used to automatically generate load adjustment instructions based on the final safe load increase rate and push the instructions to the unit's distributed control system (DCS). The closed-loop tracking unit is used to monitor the execution of load adjustment commands, automatically trigger control parameter corrections based on the actual response of the unit, and track the load adjustment process until the target load is stably completed.

5. The load adaptive control system for the start-up phase of a combined cycle unit according to claim 1, characterized in that: It also includes a data cleaning and alignment module, which is connected to the state awareness modeling module and is specifically used to clean multi-source heterogeneous data during the historical startup process and remove sensor outliers and communication interruption segments. Timestamp alignment and resampling are performed on data sources from various sampling frequencies; The aligned and cleaned structured data is stored in the training sample library for use by the state-aware modeling module.

6. The load adaptive control system for the start-up phase of a combined cycle unit according to claim 1, characterized in that: It also includes a lifespan loss management module, which is connected to the closed-loop adaptive control module and is specifically used for: Record the cumulative number of starts and equivalent operating hours of unit components; Monitor the cumulative low-cycle fatigue life loss of each unit component and transmit the life loss rate information to the closed-loop adaptive control module. When the load increase rate determined by the closed-loop adaptive control module causes the life loss rate to exceed the preset life threshold, an optimization prompt for the load decrease rate is automatically generated.

7. The load adaptive control system for the start-up phase of a combined cycle unit according to claim 1, characterized in that: It also includes a boundary condition assessment module, which is connected to the load forecasting decision module and is specifically used for: Determine sensitivity weights based on boundary conditions; Data mining analysis was performed on the 10 to 20 boundary conditions with the highest sensitivity weights. Based on historical actual startup data, the prediction accuracy of the dynamic response prediction model under various boundary conditions is calculated. Based on the prediction accuracy results and the actual operating requirements of the unit, the final inner loop control parameter correction coefficient is determined and fed back to the load prediction decision module.

8. The load adaptive control system for the start-up phase of a combined cycle unit according to claim 4, characterized in that: The closed-loop tracking unit includes: The status monitoring subunit is used to monitor the tracking deviations of unit load, main steam temperature, and cylinder differential expansion operating parameters; The early warning triggering subunit sends a reminder to the operators when the tracking deviation of the operating parameters exceeds the preset deviation threshold; When the tracking deviation still cannot converge after three automatic corrections, the over-limit intervention subunit generates an over-limit warning and suggests switching to manual control mode. The closed-loop verification subunit is used to verify the load control effect of this startup process based on the operating data after the target load has stabilized, and to archive the startup strategy after confirming that the control target has been achieved.

9. A load adaptive control system for the start-up phase of a combined cycle unit according to claim 6, characterized in that: It also includes a data integration interface module, used for: It exchanges data with the unit's distributed control system (DCS) to obtain real-time process data; Exchange data with the life loss management module to obtain component life consumption data; The acquired real-time process data and component life consumption data are integrated and transmitted to the load prediction decision module as prediction input data; Establish a data source database to store and update data from the unit control system, vibration monitoring system, and performance calculation system.

10. The load adaptive control system for the start-up phase of a combined cycle unit according to claim 1, characterized in that: It also includes a self-learning optimization module, connected to the state-aware modeling module, for: Collect comparative data between predicted control commands and actual unit responses during each startup process; Analyze the trend of accuracy changes in dynamic response prediction models; When the accuracy reaches the set accuracy threshold, the model is retrained. The training sample library is updated based on the latest startup data, and the state-aware modeling module is notified to rebuild the dynamic response prediction model to achieve continuous self-optimization of the system.

Images