A self-adaptive switching decision system and method for multiple modes of working conditions of cutting cylinder and bypass heating

Abstract

The application discloses a kind of cut cylinder and bypass heat supply multi-mode working condition self-adaptive switching decision system and method, belong to steam turbine control technical field, it includes obtaining the real-time operating parameter and external demand parameter of unit, obtains current working condition state vector;Current working condition state vector and the demand prediction parameter of future time period are carried out multidimensional rolling optimization calculation, and optimization decision instruction is generated;Using current working condition state vector index pre-stored working condition control mapping relationship set, continuous control instruction is calculated;Current working condition state vector, the execution feedback data of continuous control instruction and the execution effect data of optimization decision instruction are associated with evaluation, and rule update instruction is generated;Rule update instruction is executed, and working condition control mapping relationship set is revised.The application uses the strategy that prospective rolling optimization and closed loop self-learning rule revision are combined, and the economic, stable and intelligent adaptive control of mode switching of steam turbine under variable working condition can be realized.

Description

Technical Field

[0001] This invention relates to the field of steam turbine control technology, and in particular to an adaptive switching decision system and method for multiple operating conditions, including cylinder cut-off and bypass heating. Background Technology

[0002] In the field of thermal power generation, especially for combined heat and power (CHP) units, the steam turbine, as the core power equipment, is crucial for its operational flexibility and economy. To meet the dual demands of grid peak shaving and heating network supply, steam turbines often need to switch between multiple operating modes, with cylinder cut-off operation and bypass heating being two typical conditions. Cylinder cut-off operation improves unit operating efficiency during low electrical loads by shutting off the steam intake of some low-pressure cylinders; bypass heating, on the other hand, extracts some steam from the turbine and directly feeds it into the heating system to meet peak-load heat demand. Effective management and smooth switching between these two modes are key to ensuring the safe, stable, and economical operation of the steam turbine.

[0003] Among related technologies, Chinese invention patent CN115421390A discloses a multi-condition adaptive control method for cogeneration units that incorporates deep reinforcement learning. This method considers the nonlinear changes in some state parameters of the CHP unit during random output operation across a wide range of conditions, and establishes a state operation model representing the CHP unit under different output conditions. Based on the established CHP unit state operation model, a multi-condition adaptive control model is established, taking into account the uncertain changes in system state parameters. For the parameter optimization problem of the control module in the multi-condition adaptive control model, a MADDPG algorithm-based multi-condition adaptive control parameter optimization strategy is designed. Through these steps, the load tracking of the cogeneration unit under multi-condition adaptive control is achieved.

[0004] However, existing technologies have significant drawbacks. First, control methods based on conventional PID control and operator experience are inherently lagging and reactive, making it difficult to cope with rapid and significant changes in future electrical and thermal loads. This can easily lead to parameter overshoot and long settling times during switching, affecting the unit's operational stability. Second, this decision-making approach lacks a global, forward-looking optimization perspective, typically focusing only on stability under current operating conditions. It fails to plan the optimal combination of operating modes over a longer timescale, resulting in suboptimal fuel consumption and equipment wear. Furthermore, once the parameters in the control system are tuned, they typically remain unchanged, unable to automatically adapt to performance degradation or boundary drift caused by long-term operation. The control effect gradually deteriorates over time. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an adaptive switching decision system and method for multiple operating modes of cylinder cut-off and bypass heating. It employs a strategy combining forward-looking rolling optimization and closed-loop self-learning rule correction, enabling economical, stable, and intelligent adaptive control of turbine operating mode switching under varying conditions.

[0006] The above objectives can be achieved through the following approach: A multi-mode adaptive switching decision-making method for cylinder cutting and bypass heating includes: acquiring real-time operating parameters and external demand parameters of the unit to obtain a current operating condition state vector; performing multi-dimensional rolling optimization calculations on the current operating condition state vector and demand prediction parameters for future periods to generate an optimized decision instruction containing a target operating mode and a setpoint; using the current operating condition state vector to index a pre-stored set of operating condition control mapping relationships, and calculating a continuous control instruction based on the optimized decision instruction; performing correlation evaluation on the current operating condition state vector, the execution feedback data of the continuous control instruction, and the execution effect data of the optimized decision instruction to generate a rule update instruction; and executing the rule update instruction to correct the set of operating condition control mapping relationships.

[0007] Optionally, obtaining the current operating condition state vector includes: collecting turbine connecting pipe pressure, low-pressure cylinder exhaust temperature data, and valve opening feedback signals from the bypass heating system to obtain real-time unit operating parameters; acquiring grid automatic generation control commands, real-time heat load demand commands from the heating network, and ambient temperature data to obtain external demand parameters; performing outlier cleaning and data alignment on the real-time unit operating parameters to generate a valid operating dataset; performing multi-scale feature extraction on the valid operating dataset to retain trend and fluctuation features, and concatenating it with the external demand parameters to obtain the current operating condition state vector.

[0008] Optionally, generating the optimization decision instruction containing the target operating mode and setpoint includes: processing the current operating condition state vector using thermodynamic mechanism evolution logic to perform performance deduction and obtain first prediction data; processing the current operating condition state vector and the first prediction data using a data-driven residual analysis algorithm to generate a deviation correction factor; using the deviation correction factor to perform online correction on the internal calculation coefficients of the thermodynamic mechanism evolution logic to obtain a corrected hybrid prediction logic; and using the corrected hybrid prediction logic and the demand prediction parameters for the future period, performing rolling optimization under the constraints of operating cost control indicators and equipment safety to generate an optimization decision instruction.

[0009] Optionally, the rolling optimization based on operating cost control indicators and equipment safety constraints includes: defining an optimization time window based on the demand forecast parameters for the future period; within the optimization time window, using the corrected hybrid forecast logic to simulate a combination of operating modes to generate a candidate operating trajectory data set; calculating the predicted operating cost and equipment safety margin for the candidate operating trajectory data set respectively to generate trajectory evaluation results; selecting operating trajectories that meet the equipment safety margin and whose predicted operating cost is lower than a preset threshold based on the trajectory evaluation results, and determining them as target operating trajectories; extracting operating parameters from the starting point of the target operating trajectory to generate optimization decision instructions.

[0010] Optionally, the method further includes: continuously calculating the difference between the first predicted data and the actual operating parameters of the unit to generate real-time residual data; using the real-time residual data to train the data-driven residual analysis algorithm online to generate an updated data-driven residual analysis algorithm; and using the updated data-driven residual analysis algorithm to dynamically adjust the output weight of the deviation correction factor.

[0011] Optionally, the calculation of the continuous control command includes: separating high-frequency disturbance parameter features from the current operating condition state vector; matching and extracting a subset of control rules from the operating condition control mapping relationship set according to the current operating mode determined by the optimization decision command; and using the subset of control rules to process the high-frequency disturbance parameter features to generate a continuous control command for adjusting the actuator.

[0012] Optionally, the generation of rule update instructions includes: quantifying the changes in safety parameters of key equipment after the execution of the optimization decision instruction to obtain equipment stress feature vectors; quantifying the execution feedback data of the continuous control instruction to obtain control action intensity feature vectors; correlating and analyzing the current operating condition state vector, the equipment stress feature vector, and the control action intensity feature vector to identify specific operating condition mode data that leads to a decrease in control performance; and constructing local optimization logic for the specific operating condition mode data to generate rule update instructions.

[0013] Optionally, the step of constructing local optimization logic for the specific operating condition mode data that leads to a decrease in control performance includes: extracting the difference between the setpoint given by the optimization decision instruction and the actual stable operating point of the unit under the specific operating condition mode data to obtain historical deviation data; fitting the historical deviation data to generate a setpoint correction function that can compensate for the deviation; and binding the setpoint correction function with the triggering conditions of the specific operating condition mode data to generate a rule update instruction.

[0014] Optionally, the step of modifying the operating condition control mapping relationship set includes: parsing the rule update instruction, extracting the target operating mode identifier, applicable operating condition data, and newly added local optimization rules; locating the control rule subset corresponding to the target operating mode identifier in the operating condition control mapping relationship set; and inserting the newly added local optimization rules and the applicable operating condition data into the control rule subset to update the operating condition control mapping relationship set.

[0015] Based on the same inventive concept, this invention also provides a multi-mode adaptive switching decision system for cylinder cutting and bypass heating, comprising: a state construction module for acquiring real-time operating parameters and external demand parameters of the unit to obtain a current operating condition state vector; an optimization decision module for performing multi-dimensional rolling optimization calculations on the current operating condition state vector and demand prediction parameters for future periods to generate optimization decision instructions containing target operating modes and setpoints; a control calculation module for using the current operating condition state vector to index a pre-stored set of operating condition control mapping relationships and calculating continuous control instructions based on the optimization decision instructions; an efficiency evaluation module for performing correlation evaluation on the current operating condition state vector, the execution feedback data of the continuous control instructions, and the execution effect data of the optimization decision instructions to generate rule update instructions; and a rule correction module for executing the rule update instructions to correct the set of operating condition control mapping relationships.

[0016] Compared with the prior art, the present invention has the following advantages: This invention achieves forward-looking prediction and global optimization of future operating conditions by constructing optimized decision instructions that include target operating modes and setpoints. This method can plan the optimal operating mode switching path and operation sequence that balances economy and safety before changes in electrical and thermal load demands occur, avoiding the passive response and lag of traditional control strategies. Thus, while meeting external demands, it improves the overall operational economy of the unit during dynamic switching between multiple modes.

[0017] This invention decouples optimization decision-making from continuous control in a layered manner and matches corresponding subsets of control rules according to different operating modes. This architecture allows the upper-level optimization to focus on macroeconomic and safety objectives, while the lower-level control focuses on rapid suppression of high-frequency disturbances and precise tracking of the setpoint. This method effectively ensures control stability during complex transient processes such as cylinder switching and bypass activation / deactivation, reduces fluctuations in key parameters, and improves the unit's operational stability and response quality to grid and heating network commands.

[0018] This invention establishes a closed-loop self-learning mechanism from performance evaluation to rule correction. Through continuous monitoring and correlation analysis of control effects and equipment status, the system can autonomously identify control performance bottlenecks under specific operating conditions and automatically generate local optimization logic to correct the control mapping relationship set. This enables the control strategy to continuously improve itself, automatically adapt to dynamic characteristic drift caused by factors such as equipment aging and changes in fuel characteristics, enhances the system's robustness and long-term adaptability, and reduces reliance on manual maintenance and parameter tuning.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating an adaptive switching decision-making method for multi-mode operating conditions of cylinder cutting and bypass heating according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of multi-scale feature extraction according to an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of the rolling optimization decision of the running trajectory according to an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of an adaptive switching decision system for multi-mode operating conditions of cylinder cutting and bypass heating according to an embodiment of the present invention. Detailed Implementation

[0025] 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.

[0026] Reference Figure 1One embodiment of the present invention proposes an adaptive switching decision method for multiple operating conditions of cylinder cutting and bypass heating. It adopts a strategy that combines forward rolling optimization and closed-loop self-learning rule correction, which can realize economical, stable and intelligent adaptive control of turbine operating mode switching under multiple operating conditions.

[0027] The method described in this embodiment specifically includes: S1. Obtain the real-time operating parameters and external demand parameters of the unit to obtain the current operating condition state vector; Optionally, obtaining the current operating condition state vector includes: Real-time operating parameters of the unit are obtained by collecting data on the pressure of the turbine connecting pipe, the exhaust temperature of the low-pressure cylinder, and the valve opening feedback signal of the bypass heating system. Obtain power grid automatic generation control commands, real-time heat load demand commands from the heating network, and ambient temperature data to obtain external demand parameters; The real-time operating parameters of the unit are cleaned of outliers and data aligned to generate a valid operating dataset; Multi-scale feature extraction is performed on the effective operating dataset to retain trend and fluctuation features, and then concatenated with the external demand parameters to obtain the current operating condition vector.

[0028] Specifically, through the distributed control system (DCS) interface, real-time operating parameters of the unit reflecting the core thermal state of the unit are continuously collected at a frequency of seconds or minutes. Key variables include the turbine connecting pipe pressure, which characterizes the turbine's work capacity; the low-pressure cylinder exhaust temperature data, which reflects the cooling status and vacuum of the low-pressure cylinder; and the valve opening feedback signal of the bypass heating system, which directly reflects the bypass heating flow rate.

[0029] The system obtains automatic power generation control commands for a specified power generation capacity from the power grid dispatch center, obtains real-time heat load demand commands for heating from the heating company's monitoring system, and collects ambient temperature data from the site, which together constitute the external demand parameters.

[0030] use Criteria or local outlier factor algorithms are used to clean outliers from the collected real-time operating parameters of the units, removing invalid data points caused by momentary sensor failures or communication interference. Through timestamp alignment and linear interpolation techniques, data from different sources are unified to a fixed time base, generating a time-synchronized and effective operating dataset.

[0031] For the effective operating dataset, a dual-window processing strategy is adopted. A longer time window, such as 10-15 minutes, is used to calculate the moving average to extract trend features reflecting slow changes in operating conditions. Simultaneously, a shorter time window, such as 1-3 minutes, is used to calculate the standard deviation or first difference to capture the fluctuation features of rapid responses. These features are concatenated with external demand parameters to ultimately form the current operating condition state vector. Its mathematical expression is: , in, This represents the current operating state vector. It is a set of trend features calculated from a valid running dataset. It is a set of fluctuation characteristics calculated from the effective running dataset. This is the value of the automatic generation control command for the power grid. This represents the real-time heat load demand command value for the heating network. This is the ambient temperature value. For example... Figure 2 As shown in the figure, the core process of data preprocessing is as follows: First, the collected raw signal is cleaned of outliers, and then a dual-window strategy is used to separate the signal into trend features that reflect slow changes and fluctuation features that reflect instantaneous disturbances.

[0032] For example, taking a 350MW supercritical cogeneration unit as the executing entity, the system collects real-time data at 1-second intervals through the distributed control system interface. At the current moment, the collected turbine connecting pipe pressure data is 0.9 MPa, the average low-pressure cylinder exhaust temperature data is 45 degrees Celsius, and the valve opening feedback signal of the bypass heating system is 40%. Simultaneously, the system obtains an automatic power generation control command of 280 MW from the power grid dispatch center, a real-time heat load demand command of 500 GJ / h from the heating company, and collects ambient temperature data of -10 degrees Celsius. The system performs outlier cleaning and timestamp alignment on the above-mentioned real-time operating parameters of the unit to generate a valid operating dataset. For the valid operating dataset, the system selects a 15-minute time window to calculate the moving average of the connecting pipe pressure to obtain trend characteristics. The value is 0.92 MPa. A 3-minute window is selected to calculate the standard deviation and obtain the fluctuation characteristics. The value is 0.015 MPa. Subsequently, the system concatenates these characteristics with external demand parameters, according to the formula... Construct the current operating condition state vector. After substituting the values, the current operating condition state vector at this moment is obtained. The values ​​are [0.92, 0.015, 280, 500, -10]. This method transforms heterogeneous time-series data into standardized state inputs through multi-scale feature extraction and vectorized concatenation, effectively preserving the historical evolution trend and instantaneous fluctuation information of the operating conditions, and providing a data foundation for subsequent models.

[0033] S2. Perform multi-dimensional rolling optimization calculations on the current operating condition vector and the demand prediction parameters for future periods to generate optimization decision instructions that include the target operating mode and setpoints; Optionally, the generation of optimization decision instructions including the target operating mode and setpoints includes: The current operating condition state vector is processed using thermodynamic mechanism evolution logic to perform performance deduction and obtain the first prediction data; The current operating condition vector and the first predicted data are processed using a data-driven residual analysis algorithm to generate a deviation correction factor. The internal calculation coefficients of the thermodynamic mechanism evolution logic are corrected online using the deviation correction factor to obtain the corrected hybrid prediction logic. Using the corrected hybrid forecasting logic and the demand forecasting parameters for the future period, rolling optimization is performed under the constraints of operating cost control indicators and equipment safety to generate optimization decision instructions.

[0034] Specifically, the current operating condition vector is input into a preset thermodynamic mechanism evolution logic. This logic is a unit component-level mathematical model built based on the laws of mass and energy conservation. It simulates physical processes such as turbine work and heat exchange by solving a system of differential equations. The output of this step is the first prediction data, which represents the predicted value of the unit performance response under ideal operating conditions and model parameters.

[0035] The current operating condition state vector and the first prediction data are input together into a pre-trained data-driven residual analysis algorithm, such as a Long Short-Term Memory (LSTM) network or a Gated Recurrent Unit (GRU). This algorithm learns the prediction error patterns of the mechanistic model from historical data and generates a deviation correction factor in real time. This factor quantifies the magnitude and direction of the prediction deviation of the mechanistic model under the current operating condition.

[0036] Perform online corrections to construct a corrected hybrid prediction logic. This involves not simply adding a bias correction factor to the final result, but rather using it to dynamically adjust key calculation coefficients within the thermodynamic mechanism evolution logic, such as turbine stage efficiency and heat exchanger heat transfer coefficients. The correction process can be expressed as: , in, These are the corrected internal calculation coefficients. This serves as the design or calibration reference value for this coefficient in the mechanistic model. This is the bias correction factor output by the data-driven residual analysis algorithm. The entire corrected model is called the corrected hybrid prediction logic, which combines the generalization ability of the mechanistic model with the accuracy of the data model.

[0037] Using a corrected hybrid forecasting logic as its core, and combining demand forecasting parameters for future periods, such as electrical and thermal load curves for the next 2-4 hours, multi-path simulation and optimization are performed within a rolling time window. The optimization goal is to find the optimal operating trajectory that balances economy and safety. Instead of relying solely on a single objective function, it generates a set of candidate operating trajectories by simulating the evolution of different combinations of operating modes within the future time window. Subsequently, the candidate trajectories undergo a dual evaluation: firstly, the predicted operating cost is calculated, encompassing fuel consumption and mode switching losses; secondly, the equipment safety margin is calculated, quantifying the distance between key indicators and limits. Among the trajectories that meet the safety margin requirements, the trajectory with the lowest cost is selected as the target operating trajectory. Finally, only the operation instructions for the first time step of this target trajectory are extracted, encapsulated into an optimization decision instruction containing the target operating mode and corresponding setpoint, and issued. The entire rolling simulation and evaluation process is then repeated in the next decision cycle to achieve dynamic, feedforward optimization control.

[0038] For example, based on the current operating condition state vector The system inputs this data into a pre-defined thermodynamic mechanism evolution logic. This logic uses a built-in turbine stage model and mass conservation equations to deduce the unit's power generation under ideal operating conditions as the first predicted data, which is 285.5 MW. Subsequently, the system inputs this operating condition state vector and the first predicted data into a pre-trained long short-term memory network for residual analysis. The algorithm identifies that the current extremely cold environment of -10 degrees Celsius leads to increased heat dissipation losses in the unit and a decrease in the efficiency of the flow path under high heat load, outputting a deviation correction factor. The value is -0.04. This is the system's benchmark value for turbine-level efficiency set in the mechanistic model. Set to 0.88, and use the correction formula. Perform online correction. The calculation process is as follows: Using a corrected internal calculation coefficient of 0.8448, the system constructs a corrected hybrid prediction logic. Based on this, the system combines demand forecast parameters indicating that the electricity load will rise to 300MW in the next two hours to simulate the evolution trajectory of two operating modes—"deep cylinder shaving" and "high-pressure bypass steam injection"—within a rolling time window. Evaluation revealed that while the "deep cylinder shaving" mode has lower fuel costs, its predicted excessive blade dynamic stress leads to insufficient equipment safety margin. Therefore, the system locks "high-pressure bypass steam injection" as the target operating trajectory and extracts the main steam pressure setpoint of 16.8MPa for the first step of this trajectory, encapsulating it into an optimization decision command for issuance. This method dynamically corrects the core internal parameters of the mechanistic model through a deviation correction factor, achieving a complementary advantage of interpretability of the physical mechanism and high precision driven by data, solving the problem of inaccuracy in traditional mechanistic models due to equipment aging. Simultaneously, the rolling optimization mechanism based on the hybrid prediction logic can proactively balance operating costs and equipment safety, ensuring that the unit always operates along the optimal trajectory under complex and variable operating conditions.

[0039] Optionally, the rolling optimization based on operating cost control indicators and equipment safety constraints includes: Based on the demand forecast parameters for the future period, define the optimization time window; Within the optimized time window, the corrected hybrid prediction logic is used to simulate the combined path of the running mode to generate a set of candidate running trajectory data. For each of the candidate trajectory datasets, the predicted operating cost and equipment safety margin are calculated to generate trajectory evaluation results; Based on the trajectory evaluation results, the operating trajectories that meet the equipment safety margin and whose predicted operating costs are lower than a preset threshold are selected and determined as the target operating trajectories. Operational parameters are extracted from the starting point of the target trajectory to generate optimization decision instructions.

[0040] Specifically, based on received demand forecast parameters for future periods, such as load curves provided by the power grid for the next 2 to 4 hours, a forward-looking optimization time window is dynamically defined. The length of the time window is adjustable and designed to cover the complete load change cycle, providing sufficient time span for forward-looking decision-making.

[0041] Using corrected hybrid prediction logic, starting from the current operating condition state vector, different combinations of operating modes are simulated. These paths include maintaining the current mode, switching to cylinder cut-off mode, switching to bypass heating mode, and various switching timing combinations between these modes. For each combination path, its state evolution is simulated step-by-step within the optimization time window, generating a detailed dataset containing time-varying key thermodynamic parameters such as pressure, temperature, and flow rate. All these datasets together constitute a candidate operating trajectory dataset.

[0042] For each trajectory in the candidate operating trajectory dataset, two core indicators are calculated independently. The first is the predicted operating cost, which is a weighted sum of fuel consumption cost and mode switching loss cost. Fuel consumption is calculated based on simulated coal consumption rate and load rate, while switching loss is empirically quantified based on the number and magnitude of switching. The second is the equipment safety margin, used to assess the impact of the trajectory on equipment lifespan. This margin is determined by calculating how close the thermal stress, fatigue damage accumulation, and other indicators of key equipment such as cylinders and rotors during the simulation process are to the design limits. The formula for calculating the equipment safety margin can be expressed as: , in, This represents the safety margin of the equipment, and its value is usually between 0 and 1. The larger the value, the safer it is. This represents the maximum value or rate of change of key equipment parameters predicted in the candidate operating trajectory, such as the maximum rate of temperature rise. This is the limit value specified in the design or operating procedures of this parameter. The evaluation results of all trajectories are integrated to generate the trajectory evaluation result.

[0043] From the trajectory evaluation results, all operating trajectories that meet the equipment safety margin are selected, i.e. The value exceeds a preset safety threshold, such as 0.2. Then, among these safe trajectories, the one with the lowest predicted operating cost is selected as the target operating trajectory. The entire target trajectory is not executed; instead, only the required operating parameters, such as the target operating mode identifier, main steam pressure setpoint, and bypass heating valve opening setpoint, are extracted from the starting point of the target operating trajectory, i.e., the first time step. These parameters are then encapsulated into immediate optimization decision instructions and sent to the control system. This process is continuously rolled out in each decision cycle to ensure that decisions are always based on the latest operating conditions and predictions. Figure 3 As shown, trajectories with a safety margin below the threshold are automatically identified as high-risk restricted areas and eliminated. Among the remaining safe candidate trajectories, the algorithm locks the point with the lowest predicted running cost as the target running trajectory.

[0044] For example, the system internally calculates coefficients based on the corrected hybrid prediction logic. Based on the demand forecast parameters indicating that the electricity load will rise to 300MW in the next two hours, a two-hour optimization time window is dynamically defined. Within this window, the system simulates and generates multiple candidate operating trajectory data sets, including "cylinder cut-in depth peak shaving" and "high-pressure bypass steam injection," starting from the current operating condition state vector. For the "cylinder cut-in depth peak shaving" trajectory, the model predicts the maximum dynamic stress value borne by the last-stage blades of the low-pressure cylinder during a large and rapid load change. It will reach 180 MPa, while the design limit value specified in the operating procedure for this component is... The pressure is 200 MPa. The system is based on the equipment safety margin formula. The trajectory is evaluated, and the calculation process is as follows: Because the calculated safety margin of 0.1 is less than the preset safety threshold of 0.2, the system determines that although the fuel cost of this trajectory is low, it carries an extremely high risk of blade fatigue fracture and is therefore removed from the candidate set. Subsequently, the system evaluates the remaining trajectories and selects the "high-pressure bypass steam replenishment" trajectory, which has a safety margin of 0.4 and the lowest predicted operating cost, as the target operating trajectory. Finally, the system extracts the operating parameters of the first time step of the target operating trajectory, encapsulates the main steam pressure setpoint of 16.8 MPa and the bypass heating valve opening setpoint of 15% into an optimization decision command and issues it.

[0045] By transforming qualitative safety requirements into quantifiable mathematical indicators, high-risk operations are forcibly intercepted during the decision-making stage, addressing the pain point of traditional control systems that neglect long-term equipment lifespan due to an excessive pursuit of economic indicators. Simultaneously, the mechanism of executing only the first step of the strategy within a rolling window ensures that control commands are always based on the latest operating condition predictions, achieving forward-looking, safe, and economically optimal operation of the unit under complex boundary conditions.

[0046] Optionally, the method further includes: The difference between the first predicted data and the actual operating parameters of the unit is continuously calculated to generate real-time residual data; The data-driven residual analysis algorithm is trained online using the real-time residual data to generate an updated data-driven residual analysis algorithm; The updated data-driven residual analysis algorithm is used to dynamically adjust the output weights of the deviation correction factor.

[0047] Specifically, in high-frequency cycles, for example every minute, the first predicted data output by the thermodynamic mechanism evolution logic is compared point-by-point with the actual operating parameters of the unit collected from the DCS at the same timestamp, and the difference between the two is calculated. This difference is the real-time residual data, which directly quantifies the prediction error of the pure mechanism model under the current operating conditions.

[0048] Online training of the data-driven residual analysis algorithm is performed. Instead of a full retraining in every cycle, incremental learning or mini-batch gradient descent is used. For example, real-time residual data from the past period, such as 24 hours, is cached, and every 1-2 hours, the weights and biases of the existing data-driven residual analysis algorithm, such as the LSTM network, are fine-tuned using this latest dataset. This online training process generates an updated data-driven residual analysis algorithm that more accurately reflects the current unit deviation characteristics.

[0049] The updated algorithm dynamically adjusts the output of the bias correction factor. This involves not only generating a new bias correction factor using the updated model, but also adjusting the factor's output weight based on the statistical characteristics of recent real-time residual data, such as variance or mean. This adjustment can be represented by a dynamic scaling factor. The logic is that when residual fluctuations are large, it indicates a severe transient or low model confidence in that region, and the weight of data correction should be appropriately reduced; conversely, it should be increased. The mathematical expression of this process is as follows: , in, This serves as the bias correction factor ultimately used to revise the mechanism model; These are the original correction values ​​directly output by the updated data-driven residual analysis algorithm; It is a dynamically adjusted output weight, the value of which is determined by the statistical analysis results of the real-time residual data, for example, It can be a function inversely proportional to the recent residual variance. In this way, not only is the bias prediction model itself updated, but also the adaptive adjustment of the correction strength is achieved, ensuring the robustness and accuracy of the entire hybrid prediction logic under various operating conditions.

[0050] For example, during continuous unit operation, the system executes an online learning and adaptive adjustment process. The system continuously reads the first predicted data output by the thermodynamic mechanism evolution logic and the actual active power of the unit at the same moment, collected by the DCS, at a 1-minute cycle. It calculates the real-time residual data for that moment as 2.5MW and stores this data in a historical residual cache. Every 2 hours, the system uses the real-time residual data accumulated over the past 24 hours in the cache to fine-tune the weights of the Long Short-Term Memory network using an incremental learning algorithm, generating an updated data-driven residual analysis algorithm. At the current decision moment, this updated algorithm analyzes the current operating conditions and outputs the original correction value. The value is 0.05. Simultaneously, the system analyzes the fluctuations in real-time residual data over the past 15 minutes, calculating a large residual variance, indicating that the current operating condition is in a non-steady-state region, and the model prediction confidence level is slightly reduced. The system determines the dynamically adjusted weights based on a preset variance inverse mapping relationship. It is 0.6. The system is based on the formula. The final deviation correction factor is calculated as follows: The system then uses this value of 0.03 to replace the bias correction factor of the mechanistic model, making a more conservative and robust correction to the mechanistic model parameters. This method, through closed-loop online incremental training, enables the data model to keep pace with the time-varying drift of equipment characteristics. More importantly, by introducing dynamic weights based on residual statistical characteristics, adaptive adjustment of the correction intensity is achieved, enhancing the robustness of the hybrid prediction logic during transient and steady-state switching.

[0051] S3. Using the current operating condition state vector index to access the pre-stored operating condition control mapping relationship set, and based on the optimization decision instruction, calculate the continuous control instruction; Optionally, the calculation of the continuous control command includes: High-frequency disturbance parameter features are extracted from the current operating condition state vector; Based on the current operating mode determined by the optimization decision instruction, a subset of control rules is matched and extracted from the set of operating condition control mapping relationships; The high-frequency disturbance parameter characteristics are processed using the subset of control rules to generate continuous control commands for adjusting the actuator.

[0052] Specifically, digital filtering techniques, such as Butterworth high-pass filters or moving average difference methods, are used to process key measurement variables in the current operating condition state vector in real time, such as turbine connecting pipe pressure and unit power. Low-frequency trend changes, driven by optimization decision commands, are filtered out, while rapid fluctuations with periods less than 1-2 minutes are extracted. These components are mainly caused by random load disturbances or internal parameter coupling, and together they constitute the high-frequency disturbance parameter characteristics.

[0053] The system receives instructions containing the target operating mode, such as "cylinder cut-off operating mode" or "bypass heating mode." Using this mode identifier as an index, it queries and matches against a pre-stored set of operating condition control mapping relationships. This mapping relationship set is a structured knowledge base that stores pre-tuned control parameters or control algorithm structures for different operating modes and load ranges. Upon successful matching, it extracts a subset of control rules specific to the current mode. This subset may contain a specific set of proportional-integral-derivative PID parameters, feedforward compensation coefficients, or a set of fuzzy control rules.

[0054] The high-frequency disturbance parameter characteristics are used as input, and the extracted subset of control rules is applied for calculation. For example, if the rule subset defines a PID controller, the high-frequency disturbance characteristics can be regarded as control deviation signals, which are then used to generate the adjustment quantity after PID calculation. This calculation process is expressed as follows: , in, yes The continuously generated control commands are analog or digital signals sent to the valve positioner or drive unit, such as a 4-20mA current signal. The value of the control command from the previous moment; , , These are the proportional, integral, and derivative gain coefficients extracted from the control rule subset for the current operating mode, respectively. This represents the high-frequency fluctuation value of the connecting pipe pressure at the current moment, i.e., a component in the high-frequency disturbance parameter characteristics. The algorithm calculates the control increment... This information is used to generate final commands that directly act on the valves, feedwater pumps, and other actuators of the bypass heating system, enabling rapid and smooth adjustment of the system's status and ensuring that the unit adheres to optimization objectives. For example, the system has issued an optimization decision command containing "high-pressure bypass steam replenishment" as the target operating mode. During subsequent real-time control, the system samples the current operating condition vector every 100 milliseconds. For the key variable—the turbine connecting pipe pressure—the system uses a Butterworth high-pass filter for real-time processing to filter out low-frequency setpoint trends determined by the optimization command, such as 1.2 MPa, and extract the current instantaneous high-frequency disturbance parameter characteristics. Current moment The high-frequency pressure fluctuation value was +0.02 MPa, the previous moment. It is +0.01MPa in the first two moments. The pressure is +0.005 MPa. Based on the current operating mode of "high-pressure bypass steam replenishment," the system performs matching within the operating condition control mapping set to extract a subset of PID control rules specific to this mode. The parameters set for this subset are: proportional coefficient... Integral coefficient Differential coefficients The system inputs high-frequency disturbance characteristics into the incremental PID algorithm formula to calculate control commands. Assuming the bypass control valve command at the previous moment... It is 45%. The calculations for each control item are as follows: Proportional item: Integral term: Differential term: Control increment The final generated continuous control commands The instruction is converted into a 4-20mA analog signal to directly drive the bypass regulating valve to micro-motion, thus smoothing and suppressing instantaneous pressure fluctuations. This method decouples millisecond-level instantaneous disturbance suppression from minute-level optimization strategy execution through frequency domain separation technology, avoiding erroneous responses of the optimization layer to high-frequency noise. Simultaneously, by dynamically invoking dedicated control rules based on the operating mode, it solves the problem of conventional controller instability caused by nonlinear changes in the characteristics of the controlled object under different operating conditions of the steam turbine, improving the system's anti-interference capability and regulation quality.

[0055] S4. Evaluate the current operating condition vector, the execution feedback data of the continuous control command, and the execution effect data of the optimization decision command, and generate a rule update command. Optionally, the rule update instruction includes: The changes in the safety parameters of key equipment after the execution of the optimization decision instruction are quantified to obtain the equipment stress feature vector; The execution feedback data of the continuous control commands are quantified to obtain the control action intensity feature vector; By correlating and analyzing the current operating condition vector, the equipment stress feature vector, and the control action intensity feature vector, specific operating condition mode data that leads to a decline in control performance are identified. For the specific operating condition data, local optimization logic is constructed to generate rule update instructions.

[0056] Specifically, after each optimization decision command is executed and a new stable operating condition is reached, key equipment safety parameters, such as the temperature difference between the inner and outer walls of the high-pressure cylinder and rotor stress monitoring data, are collected for a period of time, such as 15-30 minutes. By calculating the maximum values, average fluctuation amplitudes, and rates of change of these parameters during this period, they are integrated into an equipment stress feature vector. This vector directly quantifies the effect of the control action on equipment safety.

[0057] The actions of various actuators, such as regulating valves and baffles, driven by continuous control commands during this control process are recorded and analyzed. By collecting execution feedback data of the continuous control commands, such as valve position feedback signals, the total stroke, action frequency, and average rate of change per unit time are calculated throughout the process, ultimately forming a control action intensity feature vector. This vector reflects the activity level and stability of the control system.

[0058] The current operating condition vector, equipment stress feature vector, and control action intensity feature vector are used as data samples and stored in a historical database. Once the database reaches a certain size, a correlation analysis algorithm is triggered periodically or when control performance indicators, such as overshoot and settling time, deteriorate. This algorithm may use the Apriori association rule mining algorithm or a decision tree model. The algorithm aims to find strong association rules in the dataset, such as "when the load is in the 30-40% range and the ambient temperature is below 5 degrees Celsius, the temperature difference index in the equipment stress feature vector frequently exceeds limits, while the valve action frequency in the control action intensity feature vector increases significantly." In this way, specific operating condition patterns that lead to a decline in control performance are automatically identified.

[0059] Analyzing the deviation characteristics between historical optimization commands and actual unit responses under specific operating conditions aims to uncover the unit's "natural stability characteristics" under this condition. For example, it was found that although optimization commands require higher pressures, historical records for this condition show that thermal stress is minimized and the system is most stable only when the actual operating pressure is 0.2-0.5 MPa lower than the setpoint. Based on these findings, a local optimization logic based on deviation learning is automatically constructed. This logic includes triggering conditions and corresponding correction actions—namely, calling the setpoint correction function generated by fitting historical deviations to dynamically compensate for the original optimization commands. This complete local optimization logic is encapsulated into a structured rule update command, ready to be used to update the set of operating condition control mapping relationships.

[0060] For example, after the system executes an optimization decision command for the "cylinder cutting depth peak adjustment" mode, reducing the load rate to 35%, it enters a 30-minute stable observation period. During this period, the metal temperature and rotor stress data of the high-pressure cylinder are collected, and the maximum wall temperature difference is calculated to be 26℃, with a stress change rate of 0.8MPa / min. These statistical indicators are combined to generate a stress feature vector [26, 0.8]. Simultaneously, the system monitors the action feedback of the bypass regulating valve, calculating that the valve's total stroke within 30 minutes reaches 500%, with an average action frequency of 20 times per minute, generating a control action intensity feature vector [500, 20]. These two vector values ​​are relatively high, indicating that although the control maintained the load target, the equipment bore a significant thermal stress cost, and the actuator was in a violently fluctuating "vibration hunting" state, i.e., oscillation. The system stores the above feature vectors along with the current operating condition vector—load 35%, ambient temperature -5℃—into the historical database. That evening, the system triggered a decision tree-based association analysis algorithm to mine recent data, identifying a strong association rule: when the load was between 30-40% and the ambient temperature was below 5℃, both the wall temperature difference index in the equipment stress feature vector and the action frequency index in the control action intensity feature vector significantly exceeded normal levels. To address the performance degradation issue in this specific scenario, the system further analyzed historical data and found that under this specific low-temperature, low-load condition, when the main steam pressure setpoint was 0.3MPa lower than the theoretically optimized value, the wall temperature difference could be reduced to within 20℃, and the valve action frequency decreased by 60%, exhibiting the unit's "natural stability characteristics." Based on this finding, the system automatically constructed local optimization logic: setting the trigger condition as "load 30-40% AND ambient temperature < 5℃," and the corresponding correction action was to introduce a correction function. This logic was encapsulated into a rule update instruction, automatically issued and updated to the operating condition control mapping relationship set. This approach avoids the risk of overturning the global control strategy due to individual operating condition issues, achieving refined iteration of the control system while balancing the safety of long-term equipment operation and the stability of the actuators.

[0061] Optionally, the construction of local optimization logic for the specific operating condition data that leads to a decrease in control performance includes: Extract the difference between the setpoint given by the optimization decision instruction and the actual stable operating point of the unit under the specific operating mode data to obtain historical deviation data; Fit the historical deviation data to generate a setpoint correction function that can compensate for the deviation; The setpoint correction function is bound to the triggering conditions of the specific operating condition mode data to generate a rule update instruction.

[0062] Specifically, once a specific operating condition mode data is identified, the historical database is searched to filter out all operating records that occurred under that condition. For each record, two key data points are extracted: the relevant parameter setpoints given by the optimization decision command at that time, such as the main steam pressure setpoint. And the pressure value corresponding to the actual stable operating point that the unit can actually reach after the control system stabilizes, where the equipment stress is minimal. Calculate the difference Δ between the two. Then, all the differences calculated under that specific working condition are aggregated to form a historical deviation dataset.

[0063] Using historical deviation datasets as the target output, key variables from specific operating conditions that cause the deviation, such as load factor and ambient temperature, are used as inputs. Machine learning algorithms such as multiple nonlinear regression analysis or support vector regression (SVR) are employed for model fitting. Through training, a function capable of calculating the required setpoint compensation in real time based on current operating condition variables is obtained; this is the setpoint correction function. To balance computational efficiency and model interpretability, this function is concretized as a multiple multinomial regression model, with the following expression: , in, It is the calculated setpoint compensation value that should be applied; , It is a key input variable in the data of a specific working condition mode; , , These are regression weight coefficients determined based on the historical deviation dataset. For example, if the fitting results show that the pressure deviation is mainly affected by the load rate... and ambient temperature The specific formula is as follows: This quantifies the nonlinear coupling relationship between the deviation of the pressure setpoint and the ambient temperature and load rate under low load conditions.

[0064] The logical judgment conditions used to identify the specific operating condition data, such as "load rate L is between 30% and 40%, and ambient temperature T is less than 5 degrees Celsius," are defined as the trigger conditions for the setpoint correction function. Then, this trigger condition and the setpoint correction function itself are packaged into a complete, structured rule. This rule means that "when the system state meets the trigger condition, before executing the original setpoint given by the optimization decision instruction, the setpoint correction function should first be called to calculate the compensation amount, and the compensated value should be issued as the final setpoint." This complete logical package, containing the trigger condition and the compensation algorithm, is ultimately encapsulated into a rule update instruction, ready for precise, surgical updates to the operating condition control mapping relationship set.

[0065] For example, the system identifies the "load rate" Between 30% and 40% and ambient temperature The specific operating condition of "below 5 degrees Celsius" leads to a decrease in control performance. The system then initiates a local optimization logic construction process, first searching the historical database and filtering out all 120 operating records from the past three months that were under this condition. The system analyzes these records one by one, extracting the optimized command settings at that time, such as the main steam pressure setpoint. MPa and the actual operating value of the unit under subsequent stable and minimum stress conditions. MPa. The single-point deviation is calculated. MPa. These 120 data sets were compiled to form a historical deviation dataset. The system used a multivariate multinomial regression model to fit this dataset. The load factor was then... and ambient temperature As an input variable, the deviation value The target output is determined using the least squares method for training, which serves as the regression weight coefficient. to The value of is used to generate the following specific setpoint correction function: When the current operating condition is at the load factor Ambient temperature When the value is substituted into the function, the deviation to be compensated is calculated. MPa. Finally, the system generates a rule update instruction. This instruction explicitly defines the triggering condition as: The above correction function is then bound to this condition. The instruction logic stipulates that once the real-time monitoring data meets the trigger condition, the system must first calculate... The optimized setpoint, after deducting the compensation value, is then issued as the final instruction. This rule update instruction, containing a complete logic package, is sent to the control layer, completing the update of the operating condition control mapping relationship set. This method, through data mining and regression analysis, makes the implicit operating patterns of the unit under specific edge conditions explicit into a mathematical model. This not only corrects the control deviations in specific scenarios but also avoids interference with the global general control strategy, achieving continuous optimization and self-adaptation of the control system at minimal cost.

[0066] S5. Execute the rule update instruction to correct the working condition control mapping relationship set.

[0067] Optionally, the modification of the operating condition control mapping relationship set includes: Parse the rule update instruction to extract the target operating mode identifier, applicable operating condition data, and newly added local optimization rules; In the set of operating condition control mapping relationships, locate the subset of control rules corresponding to the target operating mode identifier; The newly added local optimization rules and the applicable operating condition data are implanted into the control rule subset to update the operating condition control mapping relationship set.

[0068] Specifically, the rule update instruction is decomposed by the parser, from which three core information units are extracted: target operating mode identifier, which is a code used to uniquely specify the operating mode to which the rule to be modified belongs, such as "cylinder cut-off low load mode"; applicable operating condition data, which is a set of logical conditions that define the effective boundary of the new rule, such as "unit load rate is between 35% and 45% and circulating water temperature is below 10 degrees Celsius"; and newly added local optimization rule, which is the constructed setpoint correction function or a set of adjusted control parameters.

[0069] Using the parsed target operating mode identifier as the primary key, an index lookup is performed in a pre-stored set of operating condition control mapping relationships. This mapping relationship set is a hierarchical, object-oriented database, with its top layer divided by operating mode. By matching the identifier, the subset of control rules associated with that mode can be quickly located. This subset contains all the control logic and parameters common to that mode.

[0070] The newly parsed local optimization rules and their associated applicable operating condition data are embedded as new, higher-priority logical units into the already located subset of control rules. This embedding is not a simple appending, but rather a conditional insertion. For example, a decision branch is added to the front end of the subset's execution logic: if the unit's real-time operating conditions meet the applicable operating condition data of the new rule, the new local optimization rule is executed first; otherwise, the existing general control logic in the subset continues to execute. After embedding, the modified subset of control rules is written back to the database, and the version number is updated, thus completing an atomic update of the entire operating condition control mapping set. Subsequently, whenever the unit enters that specific operating condition, this optimized, more targeted new rule will be automatically activated and executed.

[0071] For example, the system receives a generated rule update instruction. The parser first deconstructs the instruction's data packet, extracting three key fields: the target operating mode is identified as "cylinder cut-off low load mode" (ID: MODE_CYL_CUT_LOW); the applicable operating condition data is "unit load rate"; and the target operating mode is identified as "cylinder cut-off low load mode" (ID: MODE_CYL_CUT_LOW). Between 35% and 45% and circulating water temperature "Below 10 degrees Celsius"; the new local optimization rule is the setpoint correction function. The system then accesses a non-relational database storing the control condition mapping set. Using 'MODE_CYL_CUT_LOW' as the index key, the system quickly locates the control rule subset object corresponding to this mode. This object currently contains the common PID parameters and basic setpoint logic for this mode. Instead of directly overwriting the original logic, the system constructs a conditional decision node with the highest execution priority. The logic description of this node is: "IF ( THEN Execution The system inserts the new node at the logic execution flow entry point of this control rule subset, essentially adding a "pre-filter" before the general control logic of this mode. After implantation, the system re-serializes the updated control rule subset and writes it back to the database, while updating the version number of the subset from V2.1 to V2.2 and marking it as "active". The entire update process is completed atomically within milliseconds. Subsequently, when the unit operates in "cylinder-cutting low-load mode" again and encounters low-temperature conditions, the control system will automatically hit the pre-judgment condition, triggering the correction function to achieve precise optimization for specific operating conditions. This method not only ensures the security and stability of rule base updates but also ensures that the control system has the ability to continuously accumulate experience and constantly improve itself when facing complex and ever-changing actual operating conditions.

[0072] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides an adaptive switching decision system for multiple operating conditions of cylinder cutting and bypass heating, comprising: The state construction module is used to obtain the unit's real-time operating parameters and external demand parameters to obtain the current operating condition state vector; The optimization decision module is used to perform multi-dimensional rolling optimization calculations on the current operating condition vector and the demand forecast parameters for future periods, and generate optimization decision instructions that include the target operating mode and set points. The control calculation module is used to index the pre-stored set of operating condition control mapping relationships using the current operating condition state vector, and to calculate the continuous control command based on the optimization decision command. The performance evaluation module is used to correlate and evaluate the current operating condition vector, the execution feedback data of the continuous control command, and the execution effect data of the optimization decision command, and generate rule update commands. The rule correction module is used to execute the rule update instruction and correct the set of operating condition control mapping relationships.

[0073] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0074] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Other embodiments of the invention will be readily apparent to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A multi-mode adaptive switching decision-making method for cylinder cutting and bypass heating, characterized in that, The method includes: Obtain the unit's real-time operating parameters and external demand parameters to obtain the current operating condition state vector; The current operating condition vector and the demand prediction parameters for future periods are subjected to multi-dimensional rolling optimization calculations to generate optimization decision instructions that include the target operating mode and set points. The current operating condition state vector indexes the pre-stored set of operating condition control mapping relationships, and the continuous control command is calculated based on the optimization decision command. The current operating condition state vector, the execution feedback data of the continuous control command, and the execution effect data of the optimization decision command are correlated and evaluated to generate a rule update command; Execute the rule update instruction to correct the set of operating condition control mapping relationships.

2. The adaptive switching decision method for multi-mode operating conditions of cylinder cutting and bypass heating according to claim 1, characterized in that, The process of obtaining the current operating condition vector includes: Real-time operating parameters of the unit are obtained by collecting data on the pressure of the turbine connecting pipe, the exhaust temperature of the low-pressure cylinder, and the valve opening feedback signal of the bypass heating system. Obtain power grid automatic generation control commands, real-time heat load demand commands from the heating network, and ambient temperature data to obtain external demand parameters; The real-time operating parameters of the unit are cleaned of outliers and data aligned to generate a valid operating dataset; Multi-scale feature extraction is performed on the effective operating dataset to retain trend and fluctuation features, and then concatenated with the external demand parameters to obtain the current operating condition vector.

3. The adaptive switching decision method for multi-mode operating conditions of cylinder cutting and bypass heating according to claim 1, characterized in that, The generated optimization decision instructions, which include the target operating mode and setpoints, include: The current operating condition state vector is processed using thermodynamic mechanism evolution logic to perform performance deduction and obtain the first prediction data; The current operating condition vector and the first predicted data are processed using a data-driven residual analysis algorithm to generate a deviation correction factor. The internal calculation coefficients of the thermodynamic mechanism evolution logic are corrected online using the deviation correction factor to obtain the corrected hybrid prediction logic. Using the corrected hybrid forecasting logic and the demand forecasting parameters for the future period, rolling optimization is performed under the constraints of operating cost control indicators and equipment safety to generate optimization decision instructions.

4. The adaptive switching decision method for multi-mode operating conditions of cylinder cutting and bypass heating according to claim 3, characterized in that, The rolling optimization based on operating cost control indicators and equipment safety constraints includes: Based on the demand forecast parameters for the future period, define the optimization time window; Within the optimized time window, the corrected hybrid prediction logic is used to simulate the combined path of the running mode to generate a set of candidate running trajectory data. For each of the candidate trajectory datasets, the predicted operating cost and equipment safety margin are calculated to generate trajectory evaluation results; Based on the trajectory evaluation results, the operating trajectories that meet the equipment safety margin and whose predicted operating costs are lower than a preset threshold are selected and determined as the target operating trajectories. Operational parameters are extracted from the starting point of the target trajectory to generate optimization decision instructions.

5. The adaptive switching decision method for multi-mode operating conditions of cylinder cutting and bypass heating according to claim 3, characterized in that, The method further includes: The difference between the first predicted data and the actual operating parameters of the unit is continuously calculated to generate real-time residual data; The data-driven residual analysis algorithm is trained online using the real-time residual data to generate an updated data-driven residual analysis algorithm; The updated data-driven residual analysis algorithm is used to dynamically adjust the output weights of the deviation correction factor.

6. The adaptive switching decision method for multi-mode operating conditions of cylinder cutting and bypass heating according to claim 1, characterized in that, The calculated continuous control commands include: High-frequency disturbance parameter features are extracted from the current operating condition state vector; Based on the current operating mode determined by the optimization decision instruction, a subset of control rules is matched and extracted from the set of operating condition control mapping relationships; The high-frequency disturbance parameter characteristics are processed using the subset of control rules to generate continuous control commands for adjusting the actuator.

7. The adaptive switching decision method for multi-mode operating conditions of cylinder cutting and bypass heating according to claim 1, characterized in that, The generation rule update instruction includes: The changes in the safety parameters of key equipment after the execution of the optimization decision instruction are quantified to obtain the equipment stress feature vector; The execution feedback data of the continuous control commands are quantified to obtain the control action intensity feature vector; By correlating and analyzing the current operating condition vector, the equipment stress feature vector, and the control action intensity feature vector, specific operating condition mode data that leads to a decline in control performance are identified. For the specific operating condition data, local optimization logic is constructed to generate rule update instructions.

8. The adaptive switching decision method for multi-mode operating conditions of cylinder cutting and bypass heating according to claim 7, characterized in that, The local optimization logic constructed for the specific operating condition data that leads to a decrease in control performance includes: Extract the difference between the setpoint given by the optimization decision instruction and the actual stable operating point of the unit under the specific operating mode data to obtain historical deviation data; Fit the historical deviation data to generate a setpoint correction function that can compensate for the deviation; The setpoint correction function is bound to the triggering conditions of the specific operating condition mode data to generate a rule update instruction.

9. The adaptive switching decision method for multi-mode operating conditions of cylinder cutting and bypass heating according to claim 1, characterized in that, The modified set of operating condition control mapping relationships includes: Parse the rule update instruction to extract the target operating mode identifier, applicable operating condition data, and newly added local optimization rules; In the set of operating condition control mapping relationships, locate the subset of control rules corresponding to the target operating mode identifier; The newly added local optimization rules and the applicable operating condition data are implanted into the control rule subset to update the operating condition control mapping relationship set.

10. A multi-mode adaptive switching decision system for cylinder cutting and bypass heating, characterized in that, The system includes: The state construction module is used to obtain the unit's real-time operating parameters and external demand parameters to obtain the current operating condition state vector; The optimization decision module is used to perform multi-dimensional rolling optimization calculations on the current operating condition vector and the demand forecast parameters for future periods, and generate optimization decision instructions that include the target operating mode and set points. The control calculation module is used to index the pre-stored set of operating condition control mapping relationships using the current operating condition state vector, and to calculate the continuous control command based on the optimization decision command. The performance evaluation module is used to correlate and evaluate the current operating condition vector, the execution feedback data of the continuous control command, and the execution effect data of the optimization decision command, and generate rule update commands. The rule correction module is used to execute the rule update instruction and correct the set of operating condition control mapping relationships.

Images