Dynamic self-adaptive control method and system for gas-steam combined cycle unit

Abstract

This invention relates to the field of energy technology, specifically to a dynamic adaptive control method and system for a gas-steam combined cycle unit. The method includes: constructing a multimodal model library containing multiple operating modes; real-time acquisition of unit operating parameters and environmental parameters, performing multi-source data fusion and system state assessment; multi-timescale prediction of unit load demand, fuel characteristics, and environmental parameters based on a deep learning model; dynamically selecting model accuracy based on the prediction results, implementing multimodal collaborative optimization decision-making to generate control commands; executing control commands and monitoring the execution effect, and adaptively adjusting model parameters and strategies based on feedback. This invention enables the unit to maintain efficient and stable operation under scenarios such as grid peak shaving, rapid load fluctuations, and complex environmental changes, while reducing pollutant emissions and extending the service life of key equipment.

Description

Technical Field

[0001] This invention relates to the field of energy technology, specifically to a dynamic adaptive control method and system for a gas-steam combined cycle unit. Background Technology

[0002] A combined cycle power plant (CCPP) is a high-efficiency power generation system that uses a gas turbine (Bretton cycle) and a steam turbine (Rankine cycle) to work together. The high-temperature flue gas emitted by the gas turbine is introduced into a waste heat boiler to generate steam, which drives the steam turbine to generate electricity in a secondary manner, significantly improving energy utilization efficiency. Due to its high efficiency, flexibility and low carbon emission characteristics, this type of unit is widely used in grid peak shaving, baseload power supply and renewable energy consumption scenarios.

[0003] However, the actual operational performance of CCPP (Combined Cycle Power Plant) is highly dependent on its control system. Existing control methods are mostly based on classical control theory (such as PID control) and steady-state model design. Although they perform well under steady-state or fixed operating conditions, they face significant limitations in actual operation: the system relies on static models and single control strategies, resulting in sluggish dynamic response when load changes rapidly, fuel composition fluctuates, or environmental conditions change drastically, making it difficult to keep up with the grid's peak-shaving needs in a timely manner; existing methods mostly perform local optimization for single subsystems or operating modes, lacking dynamic collaborative optimization for multiple operating modes (such as pure gas turbine mode, combined cycle mode, and different steam parameter modes), making it difficult to achieve a comprehensive optimal balance between efficiency, emissions, and equipment lifespan; existing control models have fixed accuracy and cannot be dynamically adjusted according to operating conditions, leading to reduced optimization effects under complex operating conditions due to model mismatch or computational resource limitations; at the same time, existing prediction methods are difficult to handle multivariable and nonlinear time-series data, and cannot provide high-precision predictions of load, fuel, and environmental parameters for proactive control.

[0004] These problems further lead to decreased system efficiency, excessive pollutant emissions, increased equipment wear and tear, and insufficient fuel adaptability under varying operating conditions. Although existing improvement technologies involve optimization of a single link or model predictive control based on a single mode, they still lack a global perspective on dynamic collaborative optimization of multiple modes across the entire system and cannot effectively handle uncertainties such as system nonlinearity, prediction errors, and model mismatch. Ultimately, this limits the flexibility and economy of combined cycle units in scenarios such as grid peak shaving and renewable energy integration.

[0005] Therefore, there is a need in the art for a dynamic adaptive control method and system for gas-steam combined cycle units to solve the above problems. Summary of the Invention

[0006] To address the aforementioned technical problems, namely the issues of decreased efficiency, excessive pollutant emissions, increased equipment wear, and insufficient fuel adaptability in existing gas-steam combined cycle units under varying operating conditions, this invention provides a dynamic adaptive control method and system for gas-steam combined cycle units. The aim is to enable the unit to maintain efficient and stable operation under scenarios such as grid peak shaving, rapid load fluctuations, and complex environmental changes, while simultaneously reducing pollutant emissions and extending the service life of key equipment.

[0007] In a first aspect, the present invention provides a dynamic adaptive control method for a gas-steam combined cycle unit, the method comprising the following steps:

[0008] S1: Construct a multimodal model library containing multiple operating modes, where each operating mode corresponds to a control model with different precision;

[0009] S2: Real-time acquisition of unit operating parameters and environmental parameters, and multi-source data fusion and system status assessment;

[0010] S3: Based on deep learning models, multi-timescale predictions are made of unit load demand, fuel characteristics and environmental parameters;

[0011] S4: Dynamically select the model accuracy based on the prediction results, implement multimodal collaborative optimization decision-making to generate control commands;

[0012] S5: Execute the control command and monitor the execution effect, and make adaptive adjustments to the model parameters and strategy based on the feedback.

[0013] In some preferred embodiments, step S1 specifically includes:

[0014] S11: The unit's operating modes are divided into steady-state operating mode, transient operating mode, and special environment mode;

[0015] S12: Establish a high-precision mechanism model, a medium-precision hybrid model, and a low-precision simplified model for each type of operating mode.

[0016] In some preferred embodiments, step S2 specifically includes:

[0017] S21: Collect combustion parameters of the gas turbine unit, thermodynamic parameters of the steam cycle system, and environmental sensor data;

[0018] S22: Perform time synchronization, outlier removal, and noise filtering on the collected data;

[0019] S23: Calculate real-time thermal efficiency, pollutant emission intensity, and equipment health index;

[0020] S24: Trigger an early warning signal when the target parameter is detected to exceed the safety threshold.

[0021] In some preferred embodiments, in step S3, the predicted variables include the future load demand curve, the trend of fuel calorific value change, and the fluctuation of ambient temperature and humidity.

[0022] Step S3 specifically includes:

[0023] S31: Use an LSTM neural network to build a time series prediction model, and input a historical data window of a preset length;

[0024] S32: Generate short-term, medium-term, and long-term three-level forecast results;

[0025] S33: Quantify the range of uncertainty in the prediction results to provide robust constraints for optimizing decision-making.

[0026] In some preferred embodiments, the uncertainty quantification method in step S33 is as follows:

[0027] The Monte Carlo method is used to evaluate the confidence interval of the predicted values, specifically including:

[0028] During the forward propagation of the neural network, 10-20% of neurons are randomly discarded;

[0029] Repeat the prediction calculation 100-200 times to obtain the result distribution;

[0030] The 90%-95% confidence interval is taken as the boundary constraint for the optimization decision.

[0031] In some preferred embodiments, step S4 specifically includes:

[0032] S41: Select the optimal operating mode based on long-term forecasts and calculate the mode switching path;

[0033] S42: Employs model predictive control algorithms with the optimization objectives of maximizing unit thermal efficiency, minimizing pollutant emissions, and maximizing equipment lifespan.

[0034] S43: Dynamically adjust the model accuracy according to the degree of drastic change in operating conditions;

[0035] S44: Generate a sequence of control instructions containing timing constraints and add a safety margin correction factor.

[0036] In some preferred embodiments, in step S42, the multi-objective optimization function with the objectives of maximizing the unit's thermal efficiency, minimizing pollutant emissions, and maximizing equipment lifespan is:

[0037] ;

[0038] in, For multi-objective optimization functions, For the unit's thermal efficiency, For the nitrogen oxide emissions of the unit, This is the equipment life loss coefficient of the unit. , and These are the weighting coefficients.

[0039] In some preferred embodiments, step S5 specifically includes:

[0040] S51: Converts optimized instructions into control signals executable by the DCS system;

[0041] S52: Real-time monitoring of target device response speed and target tracking error;

[0042] S53: When the actual operating value deviates from the predicted value by more than a set threshold, the online learning mechanism is triggered;

[0043] S54: Update and optimize policy weights through reinforcement learning algorithms, and store new working condition features in the model library.

[0044] In some preferred embodiments, the online learning mechanism in step S53 specifically includes:

[0045] When the target parameter tracking error exceeds the set error value, the following learning process will be automatically initiated:

[0046] Collect the complete input and output dataset under the current operating conditions;

[0047] Freeze the original control strategy and switch to the backup PID controller;

[0048] Incremental training is performed on the prediction model to update the weights of the LSTM network;

[0049] After verifying that the new model's prediction accuracy meets the standard, optimize control is restored.

[0050] In a second aspect, the present invention also provides a dynamic adaptive control system for a gas-steam combined cycle unit, the system comprising:

[0051] The model library construction module is used to build a multimodal model library containing multiple operating modes, where each operating mode corresponds to a control model with different precision.

[0052] The data acquisition and evaluation module is used to collect the unit's operating parameters and environmental parameters in real time, and to perform multi-source data fusion and system status evaluation.

[0053] The prediction module is used to predict the unit's load demand, fuel characteristics, and environmental parameters at multiple time scales based on a deep learning model.

[0054] The instruction generation module is used to dynamically select the model accuracy based on the prediction results, implement multimodal collaborative optimization decision-making, and generate control instructions.

[0055] The execution and adjustment module is used to execute the control commands and monitor the execution effect, and to adaptively adjust the model parameters and strategies based on the feedback.

[0056] As can be seen from the above, the dynamic adaptive control method and system for a gas-steam combined cycle unit provided by the present invention has the following beneficial technical effects:

[0057] This invention significantly improves the comprehensive control performance of gas-steam combined cycle units under dynamic operating conditions by constructing a multimodal model library covering various operating modes, combining deep learning's ability to predict load demand, fuel characteristics, and environmental parameters across multiple time scales, and employing a variable-precision collaborative optimization decision-making mechanism. Specifically, the dynamic selection function of the multimodal model library can flexibly call different precision models based on actual operating conditions and prediction results, ensuring optimization accuracy under complex operating conditions while reducing computational resource consumption. Deep learning prediction technology provides more reliable forward control basis by processing multivariate nonlinear time-series data, effectively alleviating the response lag problem caused by insufficient prediction capabilities in traditional methods. The hierarchical optimization framework realizes collaborative decision-making from global operating mode planning to local control parameter optimization, balancing multiple objectives such as efficiency, emissions, and equipment lifespan. The online learning and adaptive mechanism can correct model parameters and optimization strategy weights in real time, enhancing the adaptability to long-term uncertainties such as fuel composition changes and equipment aging. The synergistic effect of these technologies enables the unit to maintain efficient and stable operation in scenarios such as grid peak shaving, rapid load fluctuations, and complex environmental changes, while reducing pollutant emissions and extending the service life of key equipment. Attached Figure Description

[0058] The features and advantages of the invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the drawings:

[0059] Figure 1 A flowchart of the dynamic adaptive control method for the gas-steam combined cycle unit of the present invention;

[0060] Figure 2 This is a schematic diagram of the dynamic adaptive control system of the gas-steam combined cycle unit of the present invention. Detailed Implementation

[0061] 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 with reference to the accompanying drawings. 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.

[0062] Based on the problems of reduced efficiency, excessive pollutant emissions, increased equipment wear and tear, and insufficient fuel adaptability of gas-steam combined cycle units under varying operating conditions as pointed out in the background art, this invention provides a dynamic adaptive control method and system for gas-steam combined cycle units. The aim is to enable the unit to maintain efficient and stable operation under scenarios such as grid peak shaving, rapid load fluctuations, and complex environmental changes, while reducing pollutant emissions and extending the service life of key equipment.

[0063] like Figure 1 As shown, the dynamic adaptive control method for gas-steam combined cycle units provided by this invention includes:

[0064] S1: Construct a multimodal model library containing multiple operating modes, where each operating mode corresponds to a control model with different precision;

[0065] In some optional implementations, the operating modes may include gas turbine stand-alone operating modes at different load levels, combined cycle modes (including single gas turbine combined cycle, dual gas turbine combined cycle, etc.), different steam parameter modes (such as high, medium and low pressure steam stand-alone or combined operation), different cooling system modes (such as air cooling and wet cooling), and rapid start-up / load change transition modes, etc. Those skilled in the art can customize the operating model based on actual needs and unit performance.

[0066] After the operating modes are defined, models of different precision are established for each operating mode. For example, high-precision mechanistic models are established based on detailed thermodynamic, fluid dynamic, and heat transfer principles to create dynamic physical models of the equipment. Examples include detailed characteristic curve models of the compressor, combustion chamber, and turbine of a gas turbine; heat balance and heat transfer models of the waste heat boiler sections (economizer, evaporator, and superheater); and expansion process models of the flow path of a steam turbine. These models involve large computational loads and are used for offline optimization and analysis of complex operating conditions. Medium-precision hybrid models can also be established, which combine mechanistic models with data-driven models to simplify or replace computationally intensive aspects with data models. For example, neural network models can be used to fit some characteristics of a gas turbine, or identification models can be used to describe the dynamic response of a waste heat boiler for core optimization calculations in online model predictive control (MPC). Low-precision simplified models can also be established, which are simplified linearized or reduced-order models for rapid response or mode switching decisions, used for rapid feasibility assessment and preliminary decision-making.

[0067] It should be noted that the above-mentioned data-driven model can utilize historical big data and employ methods such as system identification and machine learning (e.g., support vector machine, Gaussian process regression) to establish an input-output relationship model for each device or system as a whole, serving as a supplement or replacement for the mechanistic model, especially for aspects that are difficult to model accurately (e.g., fuel combustion efficiency is affected by composition).

[0068] In a preferred embodiment, step S1 specifically includes:

[0069] S11: The unit's operating modes are divided into steady-state operating mode, transient operating mode, and special environment mode;

[0070] S12: Establish a high-precision mechanism model, a medium-precision hybrid model, and a low-precision simplified model for each type of operating mode.

[0071] S2: Real-time acquisition of unit operating parameters and environmental parameters, and multi-source data fusion and system status assessment;

[0072] In the above, high-resolution sensors can be deployed in various parts of the unit. The unit's operating parameters can include gas turbine parameters, waste heat boiler parameters, steam turbine parameters, and control system parameters. Furthermore, gas turbine parameters can include inlet air temperature / pressure, compressor outlet temperature / pressure, combustion chamber temperature, turbine inlet temperature, exhaust temperature / flow rate, fuel flow rate / pressure / composition, and temperature / vibration values ​​of turbine blades at each stage. Waste heat boiler parameters can include inlet and outlet temperatures / pressures of each heating surface (economizer, evaporators at each stage, superheater), feedwater flow rate / temperature, steam flow rate / pressure / temperature at each stage, flue gas temperature, and flue gas composition (O2, NO). xSteam turbine parameters can include various parameters such as steam pressure / temperature / flow rate at each stage, extraction steam pressure / temperature, condenser vacuum / temperature, circulating water flow / temperature, bearing temperature / vibration value, and power output value. Control system parameters can include the opening degree of each regulating valve, pump speed, and fan frequency. Environmental parameters can include ambient temperature, humidity, atmospheric pressure, and wind speed and direction.

[0073] Preferably, step S2 specifically includes:

[0074] S21: Collect combustion parameters of the gas turbine unit, thermodynamic parameters of the steam cycle system, and environmental sensor data;

[0075] S22: Perform time synchronization, outlier removal, and noise filtering on the collected data;

[0076] In the above process, the raw data is cleaned, filtered, and calibrated to remove noise and missing data, and data from different subsystems are synchronized and fused in time.

[0077] S23: Calculate real-time thermal efficiency, pollutant emission intensity, and equipment health index;

[0078] S24: Trigger an early warning signal when the target parameter is detected to exceed the safety threshold.

[0079] The above can be used to assess the status of the system, such as unit performance assessment, emission assessment, and equipment health assessment. Unit performance assessment can be performed by calculating performance indicators such as real-time thermal efficiency, fuel utilization rate, steam production, power generation output, and auxiliary energy consumption. Emission assessment can be performed by calculating real-time pollutant emissions based on flue gas composition and flow rate. Equipment health assessment can be performed by using various collected parameters such as temperature, vibration, and pressure, combined with historical data and preset thresholds, to detect equipment anomalies, provide early warnings of faults, and calculate life loss based on physical models (such as creep / fatigue life loss of gas turbine thermal components and over-temperature monitoring of boiler heating surfaces). Safety thresholds can be set for the target parameters that need to be monitored, and an early warning signal will be issued if the safety threshold is exceeded.

[0080] It should be noted that those skilled in the art can use any method in the prior art to calculate thermal efficiency, pollutant emission intensity, and equipment health index, and such flexible selection of calculation methods does not constitute any limitation on the present invention.

[0081] S3: Based on deep learning models, multi-timescale predictions are made of unit load demand, fuel characteristics and environmental parameters;

[0082] Preferably, in step S3, the predicted variables include the future load demand curve, the trend of fuel calorific value change, and the fluctuation of ambient temperature and humidity. The future load demand curve can be obtained from power grid dispatch instructions or power market forecasts.

[0083] Step S3 specifically includes:

[0084] S31: Use an LSTM neural network to build a time series prediction model and input a historical data window of a preset length; of course, in addition to the LSTM neural network, GRU or Transformer predictors can also be used.

[0085] S32: Generate short-term, medium-term, and long-term three-level forecast results;

[0086] In the above, short-term forecasts can be for the next few minutes, medium-term forecasts can be for the next tens of minutes, and long-term forecasts can be for the next few hours. When setting specific time periods for short-term, medium-term, and long-term forecasts, the continuity of short-term, medium-term, and long-term forecasts should be ensured. For example, when short-term forecasts are within 10 minutes, medium-term forecasts can be from 10 minutes to 1 hour, and long-term forecasts can be more than 1 hour. Of course, the above time settings are merely exemplary and do not constitute any limitation on the present invention.

[0087] S33: Quantify the range of uncertainty in the prediction results to provide robust constraints for optimizing decision-making.

[0088] More preferably, the uncertainty quantification method in step S33 is as follows:

[0089] The Monte Carlo Dropout method is used to evaluate the confidence interval of the predicted values, specifically including:

[0090] During the forward propagation of the neural network, 10-20% of neurons are randomly discarded;

[0091] Repeat the prediction calculation 100-200 times to obtain the result distribution;

[0092] The 90%-95% confidence interval is taken as the boundary constraint for the optimization decision.

[0093] With this setup, when generating control commands, the optimization algorithm must ensure that the system still meets hard constraints such as efficiency, emissions, and equipment safety even under the most pessimistic boundary of the predicted values. By quantifying the uncertainty of predictions, it avoids the deviation of control commands from actual operating conditions caused by the accumulation of errors in traditional point predictions. For example, in the case of a sudden load drop, the confidence interval constraint can limit the upper limit of fuel supply in advance to prevent unstable combustion or excessive emissions. At the same time, it enhances the robustness of the optimization model, making the control strategy effective within the predicted fluctuation range. For example, when there is a deviation in the fuel calorific value prediction, the confidence interval constraint can automatically adjust the optimization target weights to ensure that the system operates within the safety boundary. Furthermore, Monte Carlo Dropout does not require modification of the model structure or the addition of additional calculation modules. It can achieve uncertainty assessment simply by random dropping and repeated sampling, balancing computational efficiency and engineering practicality. It is especially suitable for online optimization scenarios with high real-time requirements.

[0094] S4: Dynamically select the model accuracy based on the prediction results, implement multimodal collaborative optimization decision-making to generate control commands;

[0095] Optionally, based on long-term forecasts and the current unit status, a low- or medium-precision model library can be used to evaluate the overall performance of different operating modes within a future time window. Overall performance can include efficiency, emissions, start-up / shutdown costs, and the costs of mode switching (e.g., time, energy consumption, equipment wear and tear) can also be considered. The optimal mode switching sequence can then be planned (e.g., when to switch from single-unit mode to combined cycle mode, when to start / stop the standby gas turbine, etc.). This is the upper-level decision-making part. Once the current operating mode or mode switching path is determined, further lower-level optimization can be performed. In each control cycle, based on short-term forecasts and the current real-time status, a medium-precision hybrid model is used to perform online multi-objective optimization. Optimization variables may include: gas turbine fuel flow rate, inlet guide vane opening, steam turbine main steam valve / speed regulating valve opening, bypass valve opening at each stage, feedwater flow rate, air preheater or flue gas damper opening, cooling system parameters (such as circulating water pump speed, fan speed), etc. Optimization objectives may include maximizing real-time efficiency, minimizing pollutant emissions, optimizing the operating points of key equipment (such as controlling turbine inlet temperature to extend lifespan), maximizing response speed, etc. These objectives can be weighted according to actual operational needs. In addition, constraints can be set according to the optimization objectives, such as physical limitations of equipment (temperature, pressure, flow, speed limits), safety protection thresholds, control system output limits, emission standards, and power grid peak-shaving requirements. Then, real-time optimization algorithms, such as sequential quadratic programming (SQP), interior-point methods, or learning-based optimizers, are used. Finally, the model accuracy is dynamically selected based on the predicted severity of operating condition changes and control cycle. Simplified models can be used to speed up calculations under stable operating conditions, while more refined models are used for accurate prediction and optimization before rapid load changes or mode switching.

[0096] In a preferred embodiment, step S4 specifically includes:

[0097] S41: Select the optimal operating mode based on long-term forecasts and calculate the mode switching path;

[0098] S42: Employs model predictive control algorithms with the optimization objectives of maximizing unit thermal efficiency, minimizing pollutant emissions, and maximizing equipment lifespan.

[0099] In step S42, the multi-objective optimization function with the objectives of maximizing the unit's thermal efficiency, minimizing pollutant emissions, and maximizing equipment lifespan is as follows:

[0100] ;

[0101] in, For multi-objective optimization functions, For the unit's thermal efficiency, For the nitrogen oxide emissions of the unit, This is the equipment life loss coefficient of the unit. , and These are the weighting coefficients;

[0102] S43: Dynamically adjust the model accuracy according to the degree of drastic change in operating conditions;

[0103] S44: Generate a sequence of control instructions containing timing constraints and add a safety margin correction factor;

[0104] It should be noted that the safety margin correction factor is a dynamic adjustment parameter made by the unit's control system to the optimized command in the face of prediction errors, model uncertainties, or sudden disturbances to ensure the safe operation of the equipment. Essentially, it adds a safety buffer to the theoretically optimal control command, forming an actual execution command that balances efficiency and reliability. For example, in step S44 above, the safety margin correction factor is calculated as follows:

[0105] For temperature-sensitive equipment, setpoint correction amount ,in, To predict temperature deviation, This is a proportionality coefficient, with a value range of 0.5 ≤ ≤0.8, These are the differential coefficients, with values ​​ranging from 0.1 to... ≤0.3; Static deviation compensation Directly corrects current temperature forecast deviations and suppresses dynamic trends. To suppress excessively rapid temperature changes, set... It embodies a control logic that prioritizes the current error and uses the changing trend as a secondary factor.

[0106] For pressure-sensitive equipment, correction amount , The standard deviation of the predicted pressure is used to reflect uncertainty; the correction amount is related to the prediction uncertainty. The actual set pressure decreases by 0.7 MPa for every 1 MPa increase in standard deviation. The coefficient of 0.7 can be obtained by regression analysis of historical accident data, covering 95% of abnormal operating conditions.

[0107] S5: Execute the control command and monitor the execution effect, and adaptively adjust the model parameters and strategy based on the feedback;

[0108] In practical applications, the optimized setpoints or control sequences of each device can be transformed into analog or digital control signals recognizable by a specific DCS (Distributed Control System) or PLC (Programmable Logic Controller). The control strategy can employ setpoint tracking control or use the optimization results as feedforward signals, combined with underlying PID or tracking controllers to achieve precise control. The unit can also integrate an independent rapid safety protection system, which prioritizes interlocking protection actions when a critical condition is detected. Even if there are deviations in the optimization results or predictions, a certain safety margin is added when generating control commands, taking into account prediction errors. When the prediction model or optimization algorithm fails, it can quickly switch to a preset backup control strategy. Adaptive and online learning mechanisms can also be set up. Using online collected actual input-output data, algorithms such as recursive least squares and Kalman filtering can be used to correct the target parameters in medium-precision mixed models and data-driven models in real time, reducing model mismatch. The latest real-time data can be used to perform small-batch online training or fine-tuning of deep learning prediction models to capture short-term trends in operating conditions. In addition, reinforcement learning or adaptive dynamic programming can be combined to explore and learn better optimization strategies in actual operation, especially in complex operating conditions that are difficult to model precisely.

[0109] Preferably, step S5 specifically includes:

[0110] S51: Converts optimized instructions into control signals executable by the DCS system;

[0111] S52: Real-time monitoring of target device response speed and target tracking error;

[0112] S53: When the actual operating value deviates from the predicted value by more than a set threshold, the online learning mechanism is triggered;

[0113] More preferably, the online learning mechanism in step S53 specifically includes:

[0114] When the target parameter tracking error exceeds the set error value, the following learning process will be automatically initiated:

[0115] Collect the complete input and output dataset under the current operating conditions;

[0116] Freeze the original control strategy and switch to the backup PID controller;

[0117] Incremental training is performed on the prediction model to update the weights of the LSTM network;

[0118] After verifying that the new model's prediction accuracy meets the target, resume optimized control;

[0119] S54: Update and optimize policy weights through reinforcement learning algorithms, and store new working condition features in the model library.

[0120] For the same purpose, corresponding to the methods of any of the above embodiments, this invention also provides a dynamic adaptive control system for a gas-steam combined cycle unit, such as... Figure 2 As shown, the system includes:

[0121] The model library construction module 100 is used to construct a multimodal model library containing multiple operating modes, where each operating mode corresponds to a control model with different precision.

[0122] The data acquisition and evaluation module 200 is used to collect the unit's operating parameters and environmental parameters in real time, and to perform multi-source data fusion and system status evaluation.

[0123] Prediction module 300 is used to predict the unit's load demand, fuel characteristics and environmental parameters at multiple time scales based on a deep learning model.

[0124] The instruction generation module 400 is used to dynamically select the model accuracy based on the prediction results and implement multimodal collaborative optimization decision-making to generate control instructions.

[0125] The execution and adjustment module 500 is used to execute control commands and monitor the execution effect, and to adaptively adjust the model parameters and strategies based on feedback.

[0126] For ease of description, the above system is described by dividing it into various modules based on their functions. Of course, in implementing this invention, the functions of each module can be implemented in one or more software and / or hardware components.

[0127] The system described in the above embodiments is used to implement the corresponding control methods in the foregoing embodiments and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0128] For ease of description, the above system is described by dividing it into various functional units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0129] Those skilled in the art will understand that embodiments of the present invention can be provided as methods or systems. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0130] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0131] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0132] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present invention as described above, which are not provided in detail for the sake of brevity.

[0133] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications and variations of these embodiments will be apparent to those skilled in the art from the foregoing description.

[0134] One or more embodiments of the present invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of the present invention should be included within the scope of this disclosure.

Claims

1. A dynamic adaptive control method for a gas-steam combined cycle unit, characterized in that, The method includes the following steps: S1: Construct a multimodal model library containing multiple operating modes, where each operating mode corresponds to a control model with different precision; Step S1 specifically includes: S11: The unit's operating modes are divided into steady-state operating mode, transient operating mode, and special environment mode; S12: Establish a high-precision mechanism model, a medium-precision hybrid model, and a low-precision simplified model for each type of operating mode; S2: Real-time acquisition of unit operating parameters and environmental parameters, and multi-source data fusion and system status assessment; S3: Based on deep learning models, multi-timescale predictions are made of unit load demand, fuel characteristics and environmental parameters; In step S3, the predicted variables include the future load demand curve, the trend of fuel calorific value change, and the fluctuation of ambient temperature and humidity; Step S3 specifically includes: S31: Use an LSTM neural network to build a time series prediction model, and input a historical data window of a preset length; S32: Generate short-term, medium-term, and long-term three-level forecast results; S33: Quantify the range of uncertainty in the prediction results to provide robust constraints for optimal decision-making; S4: Dynamically select the model accuracy based on the prediction results, implement multimodal collaborative optimization decision-making to generate control commands; S5: Execute the control command and monitor the execution effect, and make adaptive adjustments to the model parameters and strategy based on the feedback.

2. The dynamic adaptive control method for a gas-steam combined cycle unit according to claim 1, characterized in that, Step S2 specifically includes: S21: Collect combustion parameters of the gas turbine unit, thermodynamic parameters of the steam cycle system, and environmental sensor data; S22: Perform time synchronization, outlier removal, and noise filtering on the collected data; S23: Calculate real-time thermal efficiency, pollutant emission intensity, and equipment health index; S24: Trigger an early warning signal when the target parameter is detected to exceed the safety threshold.

3. The dynamic adaptive control method for a gas-steam combined cycle unit according to claim 1, characterized in that, The uncertainty quantification method in step S33 is as follows: The Monte Carlo method is used to evaluate the confidence interval of the predicted values, specifically including: During the forward propagation of the neural network, 10-20% of neurons are randomly discarded; Repeat the prediction calculation 100-200 times to obtain the result distribution; The 90%-95% confidence interval is taken as the boundary constraint for the optimization decision.

4. The dynamic adaptive control method for a gas-steam combined cycle unit according to claim 1, characterized in that, Step S4 specifically includes: S41: Select the optimal operating mode based on long-term forecasts and calculate the mode switching path; S42: Employs model predictive control algorithms with the optimization objectives of maximizing unit thermal efficiency, minimizing pollutant emissions, and maximizing equipment lifespan. S43: Dynamically adjust the model accuracy according to the degree of drastic change in operating conditions; S44: Generate a sequence of control instructions containing timing constraints and add a safety margin correction factor.

5. The dynamic adaptive control method for a gas-steam combined cycle unit according to claim 4, characterized in that, In step S42, the multi-objective optimization function with the objectives of maximizing the unit's thermal efficiency, minimizing pollutant emissions, and maximizing equipment lifespan is as follows: in, For multi-objective optimization functions, For the unit's thermal efficiency, For the nitrogen oxide emissions of the unit, This is the equipment life loss coefficient of the unit. , and These are the weighting coefficients.

6. The dynamic adaptive control method for a gas-steam combined cycle unit according to claim 1, characterized in that, Step S5 specifically includes: S51: Converts optimized instructions into control signals executable by the DCS system; S52: Real-time monitoring of target device response speed and target tracking error; S53: When the actual operating value deviates from the predicted value by more than a set threshold, the online learning mechanism is triggered; S54: Update and optimize policy weights through reinforcement learning algorithms, and store new working condition features in the model library.

7. The dynamic adaptive control method for a gas-steam combined cycle unit according to claim 6, characterized in that, The online learning mechanism in step S53 specifically includes: When the target parameter tracking error exceeds the set error value, the following learning process will be automatically initiated: Collect the complete input and output dataset under the current operating conditions; Freeze the original control strategy and switch to the backup PID controller; Incremental training is performed on the prediction model to update the weights of the LSTM network; After verifying that the new model's prediction accuracy meets the standard, optimize control is restored.

8. A dynamic adaptive control system for a gas-steam combined cycle unit according to any one of claims 1 to 7, characterized in that, The system includes: The model library construction module is used to build a multimodal model library containing multiple operating modes, where each operating mode corresponds to a control model with different precision. The data acquisition and evaluation module is used to collect the unit's operating parameters and environmental parameters in real time, and to perform multi-source data fusion and system status evaluation. The prediction module is used to predict the unit's load demand, fuel characteristics, and environmental parameters at multiple time scales based on a deep learning model. The instruction generation module is used to dynamically select the model accuracy based on the prediction results, implement multimodal collaborative optimization decision-making, and generate control instructions. The execution and adjustment module is used to execute the control commands and monitor the execution effect, and to adaptively adjust the model parameters and strategies based on the feedback.

Images