Gas turbine combustion stability online regulation and control method based on deep learning

Abstract

The invention discloses a gas turbine combustion stability online regulation and control method based on deep learning, and aims to solve the problem of combustion instability prevention and control under the scene that fuel components / heat values are variable and are constrained by emission. The method comprises the following steps: firstly, fusing unit load, fuel characteristics and dynamic pressure of a combustion chamber, and constructing a ten-dimensional fuel characteristic package and a 2,000-448-dimensional pressure sequence; then, through a self-adaptive stability margin estimation network containing a stability judgment trunk and a fuel condition calibration layer, in combination with comparison event online calibration, self-adaptive stability margin of cross-fuel alignment is output; matching the risk level with the threshold group, screening regulation and control actions in an emission qualified adjustable range, calculating the amplitude, and generating a regulation and control instruction draft; and finally, through closed-loop verification, outputting a final regulation and control instruction or a rollback instruction according to the stable response quantity. According to the method, online self-adaptive evaluation of the stability margin and closed-loop regulation and control under emission constraint are realized, the combustion instability risk is effectively inhibited, and the operation safety and continuity of the unit are improved.

Description

Technical Field

[0001] This invention relates to the field of flow monitoring and energy efficiency analysis of gas turbine generator sets, and in particular to an online control method for combustion stability of gas turbines based on deep learning. Background Technology

[0002] Gas turbines, as core equipment in energy and power systems, are widely used in power generation, industrial heating, and distributed energy sectors. With the increasing prevalence of new fuels such as hydrogen-blended fuels and renewable energy coupling, combustion stability control faces challenges from complex operating conditions, including fuel composition fluctuations and calorific value drift. Among related technologies, a combustion dynamic pressure monitoring system, combined with signal processing algorithms, constructs a combustion instability early warning system. At the control strategy level, existing systems mostly employ test-calibrated combustion control maps, load-segmented PID control, and expert systems, achieving combustion organization adjustments through rule-based association of "event type—action." The combustion stability event database, as a core supporting technology, stores records of stable / instability events under historical operating conditions and maps them to control parameters, providing decision-making references for operators.

[0003] However, existing technologies suffer from systemic flaws in scenarios with variable fuels. Specifically, fixed-threshold systems are prone to discrimination drift when fuel composition changes, resulting in completely opposite stability conclusions for the same pressure amplitude under different fuel conditions. While experience-based control strategies can achieve basic regulation, they lack a feasibility verification mechanism within the NOx / CO emission compliance boundary, posing a risk of control quantities exceeding limits. Furthermore, traditional methods lack a closed-loop verification system for dynamic pressure feedback after execution. When fuel switching causes a shift in the stability boundary, control commands may trigger increased combustion oscillations or flame shedding. The absence of a fuel condition calibration layer means that the core stability discrimination features cannot adapt to changes in fuel characteristics, while the separation of the event-driven control framework from emission constraints makes it difficult for control actions to simultaneously meet environmental and stability requirements. These technological limitations, particularly in practical scenarios such as fluctuations in hydrogen-blended fuels and pipeline gas source switching, may lead to unplanned unit shutdowns, increased annual maintenance costs, and an inability to meet the dual demands of real-time performance and reliability for new fuel applications. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose an online control method for combustion stability of gas turbines based on deep learning.

[0006] Another objective of this invention is to propose an online control device for combustion stability of gas turbines based on deep learning.

[0007] The third objective of this invention is to provide a computer device.

[0008] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, a first aspect of the present invention proposes an online control method for combustion stability of a gas turbine based on deep learning, comprising:

[0010] S1: Obtain unit load, fuel composition and calorific value, and combustion chamber dynamic pressure; retrieve fuel composition and calorific value database and dynamic pressure history database; and fuse them to generate a fuel characteristic package for the current operating condition. S2 retrieves stable / instable comparison events based on the current operating condition fuel feature package, inputs them into the runtime adaptive stability margin assessment network, calibrates the fuel condition calibration layer online, and outputs the adaptive stability margin; S3, based on the adaptive stability margin, matches the event tags of the combustion stability event database, determines the risk level, and generates a risk threshold group containing trigger, release, and escalation thresholds; S4. Define the adjustable range of emissions compliance in the emission-operating condition mapping matrix, retrieve the corresponding control action family according to the risk level to obtain the control action, calculate the control amplitude by combining the adaptive stability margin and risk threshold group, and generate a draft control instruction. S5, execute the draft control command, collect the dynamic pressure of the combustion chamber after execution, input the dynamic pressure of the combustion chamber after execution and the current operating condition fuel characteristic package into the evaluation network, and calculate and output the stable response quantity; S6, when the stable response reaches the release threshold, the draft control instruction will be output as an online control instruction; when it reaches the upgrade threshold, a backoff control instruction will be generated and output as an online control instruction under the constraint of the emission compliance adjustable range.

[0011] In one embodiment of the present invention, S1 includes: Collect the unit load and determine the unit load sampling time. At the same time, collect the dynamic pressure of the combustion chamber and use the unit load sampling time as the operating time reference. Access the fuel composition and calorific value database based on the operating condition time reference to obtain the fuel's lower heating value and the volume fraction of the fuel's main components, which include methane, ethane, propane, hydrogen, carbon monoxide, carbon dioxide, nitrogen, and inert gases. The unit load, the lower heating value of the fuel, and the volume fraction of the main components of the fuel are spliced ​​together in a fixed dimensional order to form a ten-dimensional fuel feature package under the current operating condition. A continuous sampling sequence of length 2,048 is extracted from the dynamic pressure of the combustion chamber to form a 2,048-dimensional pressure sequence for subsequent input steps.

[0012] In one embodiment of the present invention, S2 includes: The event retrieval conditions are determined based on the current operating condition fuel feature package. The event retrieval conditions include the unit load difference not exceeding a first load threshold, the fuel lower heating value difference not exceeding a first calorific value threshold, and the volume fraction differences of the main fuel components not exceeding a first component threshold. The event retrieval criteria are input into the combustion stability event database to retrieve stable and unstable events that meet the event retrieval criteria. The stable and unstable events are then arranged into a control event in chronological order. Based on the aforementioned control event, the fuel condition calibration layer is calibrated online. The online calibration is limited to updating the parameters of the fuel condition calibration layer while keeping the parameters of the stable discrimination backbone unchanged. The parameters of the fuel condition calibration layer are iteratively updated until the calibration judgment condition is met or the preset iteration limit is reached, and the adaptive stability margin after online calibration is output.

[0013] In one embodiment of the present invention, S3 includes: The adaptive stability margin and the fuel characteristic package under the current operating condition are input into the combustion stability event database. Candidate event records are filtered according to the following criteria: the difference between the unit load and the fuel load does not exceed the second load threshold, the difference between the fuel lower heating value and the fuel volume fraction of each dimension does not exceed the second component threshold. In the candidate event records, determine the event label corresponding to the margin interval to which the adaptive stabilization margin belongs, and when there are multiple event labels, select the event label with the smallest difference in adaptive stabilization margin as the matching event label.

[0014] In one embodiment of the present invention, S4 includes: Under the constraint of the adjustable range of emission compliance, the feasibility of the candidate set of control actions is screened, and the control actions with remaining adjustment space for the adjustable initial value corresponding to the dimension of target control quantity are selected as control actions. The adaptive stability margin and risk threshold are grouped together to calculate the amplitude. When the adaptive stability margin does not meet the trigger threshold, the control amplitude is set to zero. When the adaptive stability margin meets the trigger threshold but does not meet the upgrade threshold, the control amplitude is determined by linear interpolation. When the adaptive stability margin meets the upgrade threshold, the control amplitude is set to the preset maximum amplitude. The adjustment amplitude is applied to the target control quantity dimension of the adjustment action and the boundary is clipped. When the applied target control quantity exceeds the lower boundary or upper boundary of the control quantity, the target control quantity is clipped to the corresponding boundary and the adjustment amplitude is corrected simultaneously. The control actions, the revised control range, and the target control quantity are output as a draft control instruction.

[0015] In one embodiment of the present invention, S5 includes: The draft control instructions are sent to the unit control system and the control actions are executed to obtain the operating status after execution. After a preset response delay, the dynamic pressure of the combustion chamber after execution is collected, and a continuous sampling sequence of length 2,048 is extracted to form a 2,048-dimensional pressure sequence after execution. The 2,048-dimensional pressure sequence after execution and the fuel characteristic package under the current operating condition are input into the adaptive stability margin estimation network during runtime. The adaptive stability margin after execution is calculated and output using the online calibrated fuel condition calibration layer parameters and stability discrimination backbone parameters. The stable response is obtained by subtracting the adaptive stability margin after execution from the adaptive stability margin before the execution of the draft control instruction.

[0016] In one embodiment of the present invention, S6 includes: The stable response quantity is compared with the escalation threshold in the risk threshold group. When the stable response quantity meets the escalation threshold, the rollback generation process is initiated. Before entering the rollback generation process, the stable response quantity is compared with the release threshold in the risk threshold group. When the stable response quantity meets the release threshold, the draft control instruction is output as an online control instruction. In the rollback generation process, the rollback control quantity dimension is determined according to the target control quantity dimension of the draft control instruction, the rollback adjustment direction is set to be opposite to the target control quantity adjustment direction of the draft control instruction, and the rollback magnitude is set to a preset rollback magnitude. Under the constraint of the adjustable range of emission compliance, the backoff amplitude is applied to the backoff control quantity dimension to obtain the backoff target control quantity, and the backoff target control quantity is subjected to boundary trimming to generate a backoff control command, which is then output as an online control command.

[0017] To achieve the above objectives, a second aspect of the present invention provides an online control device for combustion stability of a gas turbine based on deep learning, comprising: The fuel feature package fusion generation module is used to obtain unit load, fuel composition and calorific value, and combustion chamber dynamic pressure, retrieve fuel composition and calorific value database and dynamic pressure history database, and fuse them to generate the current operating condition fuel feature package. The adaptive stability margin assessment and calibration module is used to retrieve stable / instability comparison events based on the fuel feature package under the current operating conditions, input the runtime adaptive stability margin assessment network, calibrate the fuel condition calibration layer online, and output the adaptive stability margin. The risk level determination and threshold generation module is used to determine the risk level and generate a risk threshold group containing trigger, release, and escalation thresholds by matching event tags in the combustion stability event database with adaptive stability margin. The module for generating draft control instructions is used to define the acceptable and adjustable range of emissions in the emission-operating condition mapping matrix, retrieve the corresponding control action family according to the risk level to obtain the control action, calculate the control amplitude by combining the adaptive stability margin and risk threshold group, and generate a draft control instruction. The stable response calculation module is used to execute the draft control command, collect the dynamic pressure of the combustion chamber after execution, input the dynamic pressure of the combustion chamber after execution and the current operating condition fuel characteristic package into the evaluation network, and calculate and output the stable response. The stable response quantity assessment and instruction output module is used to output the draft control instruction as an online control instruction when the stable response quantity reaches the release threshold; and to generate and output the back-off control instruction as an online control instruction under the constraint of the emission compliance adjustable range when the stable response quantity reaches the upgrade threshold.

[0018] This invention discloses a deep learning-based online control method and apparatus for gas turbine combustion stability, which enables online adaptive assessment and control of gas turbine combustion stability under changes in fuel composition and calorific value. It ensures the generation of executable control commands while meeting emission constraints, and improves control reliability and unit operation safety through closed-loop verification and backoff mechanisms.

[0019] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, for implementing a deep learning-based online control method for gas turbine combustion stability as described in the first aspect embodiment.

[0020] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a deep learning-based online control method for gas turbine combustion stability as described in the first aspect embodiment.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] Figure 1 This is a flowchart of an online control method for combustion stability of a gas turbine based on deep learning, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the process of acquiring and constructing input data according to an embodiment of the present invention; Figure 3 This is a flowchart of adaptive stability margin calculation and online calibration according to an embodiment of the present invention; Figure 4This is a flowchart illustrating the determination of risk level and risk threshold according to an embodiment of the present invention; Figure 5 This is a flowchart illustrating the generation process of a draft control instruction according to an embodiment of the present invention; Figure 6 This is a flowchart of the post-execution stability closed-loop evaluation according to an embodiment of the present invention; Figure 7 This is a flowchart of the online control command release and rollback process according to an embodiment of the present invention; Figure 8 This is a comparison diagram of conventional control and online stability regulation according to an embodiment of the present invention; Figure 9 This is a network structure diagram according to an embodiment of the present invention; Figure 10 This is a structural diagram of an online control device for combustion stability of a gas turbine based on deep learning, according to an embodiment of the present invention. Figure 11 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] The following description, with reference to the accompanying drawings, describes an online control method and apparatus for combustion stability of a gas turbine based on deep learning, according to an embodiment of the present invention.

[0026] Figure 1 This is a flowchart of an online control method for combustion stability of a gas turbine based on deep learning, according to an embodiment of the present invention. Figure 1 As shown, it includes: S1 acquires unit load, fuel composition and calorific value, and combustion chamber dynamic pressure, retrieves the fuel composition and calorific value database and the dynamic pressure history database, and integrates them to generate the current operating condition fuel characteristic package.

[0027] Specifically, such as Figure 1As shown, under variable fuel conditions, the input data for the adaptive stability margin estimation network during runtime is constructed, which can be directly input. The current fuel feature package serves as the input to the fuel condition calibration layer, and the 2048-dimensional pressure sequence serves as the input to the stability discrimination backbone. To ensure consistency between subsequent network inference and event retrieval, this embodiment defines the data acquisition time, feature dimension order, and missing data completion rules in a fixed manner.

[0028] The unit load value read from the unit control system is recorded as follows: The time when the unit load is read is defined as the unit load sampling time. Simultaneously, a continuous sampling data stream of combustion chamber dynamic pressure is acquired from the combustion chamber dynamic pressure measurement point, denoted as... .Will It serves as a time reference for operating conditions, used for accessing and locating fuel composition and calorific value databases and dynamic pressure history databases, and for subsequently capturing continuous sampling sequences of combustion chamber dynamic pressure.

[0029] Based on working condition time benchmark Access the fuel composition and calorific value database to read and A corresponding fuel record must include at least the lower heating value of the fuel, denoted as... And the volume fraction of the eight-dimensional fuel's main components, denoted as The eight-dimensional fuel's main components, in volume fraction, include methane, ethane, propane, hydrogen, carbon monoxide, carbon dioxide, nitrogen, and inert gases in a fixed order. This fixed order remains unchanged in subsequent steps to ensure that the input dimension semantics received by the fuel condition calibration layer are constant.

[0030] When fuel composition and calorific value databases lack operating time references When the corresponding fuel record is accessed, the fuel record is completed. Specifically, the dynamic pressure history database is accessed and fuel switching event records associated with the combustion chamber dynamic pressure history data are retrieved. These fuel switching event records include the fuel switching time. Determine the distance to the center in the fuel switching event log. Recently and earlier than The fuel switching time is denoted as and determine the distance Recently and later than The fuel switching time is denoted as ;Will and The search is limited to a fuel record retrieval range. Then, within this range, the fuel composition and calorific value database is accessed, a usable fuel record is read, and the lower heating value of the fuel in that record is determined as... The volume fraction of the eight-dimensional fuel main components in the fuel record is determined as follows: Thus, a time reference for operating conditions is obtained. The corresponding lower heating value of the fuel and the volume fraction of the main components of the fuel.

[0031] Unit load Lower heating value of fuel Volume fraction of eight-dimensional fuel main components The fuel feature package under the current operating condition is assembled in a fixed dimensional order and denoted as follows: ,in The dimension order is fixed as follows: unit load, lower heating value of fuel, methane volume fraction, ethane volume fraction, propane volume fraction, hydrogen gas integral, carbon monoxide volume fraction, carbon dioxide volume fraction, nitrogen gas integral, and inert gas volume fraction. Subsequently, based on the operating condition time base... From combustion chamber dynamic pressure A continuous sampling sequence of length 2048 is extracted from the data to form a 2048-dimensional pressure sequence, denoted as […]. The ten-dimensional current operating condition fuel characteristic package With 2048-dimensional pressure sequence This serves as unified input data for subsequent steps, supporting the collaborative processing of the stability discrimination backbone and fuel condition calibration layer in the runtime adaptive stability margin estimation network.

[0032] S2 retrieves stable / instable comparison events based on the current operating condition fuel feature package, inputs them into the runtime adaptive stability margin assessment network, calibrates the fuel condition calibration layer online, and outputs the adaptive stability margin.

[0033] Specifically, such as Figure 2 As shown, a one-dimensional adaptive stability margin that can be interpreted across fuels is generated under the current fuel conditions, and the fuel condition calibration layer is calibrated online using a combustion stability event database to form control events. The current operating condition fuel characteristic package is denoted as... The Includes unit load Lower heating value of fuel Volume fraction of eight-dimensional fuel main components The dynamic pressure of the combustion chamber is extracted to form a 2048-dimensional pressure sequence, denoted as ; .

[0034] according to Determine the event retrieval criteria. Let the first load threshold be... The first calorific value threshold is The threshold for the first component is For each event record in the combustion stability event database, the unit load at which that event record is read is denoted as... The lower heating value of fuel is denoted as The volume fraction of the main components of the eight-dimensional fuel is denoted as: ;calculate and Take the absolute value of the difference and determine whether it does not exceed [the specified value]. ,calculate and Take the absolute value of the difference and determine whether it does not exceed [the specified value]. and to and Calculate the difference dimension by dimension and take the absolute value to determine whether each dimension does not exceed the limit. The above criteria will be used together as conditions for event retrieval.

[0035] The event retrieval criteria are input into the combustion stability event database for filtering, resulting in a candidate event set. The candidate event set is then divided into a stable event set and an unstable event set based on the stability markers in the event records. The occurrence time of each event record is then recorded as follows: ;by A continuous sampling sequence of length 2048 was extracted from the combustion chamber dynamic pressure data associated with this event record, centered on the event record, and this sequence was taken as the 2048-dimensional pressure sequence corresponding to that event record. The stable event set and the unstable event set were then separated into... After sorting, each stable event is paired with the most recent unstable event in chronological order to form a control event, so that each control event group contains one stable event and one unstable event, and the two 2048-dimensional pressure sequences corresponding to the control event group are retained for subsequent online calibration.

[0036] Will The stable discriminant backbone of the runtime adaptive stable residual estimation network is input. The stable discriminant backbone sequentially consists of a first 1D convolutional layer and a downsampling layer, a second 1D convolutional layer and a downsampling layer, a third 1D convolutional layer and a downsampling layer, a flattening operation, and a fully connected layer. The first 1D convolutional layer has 32 convolutional kernels and outputs 32 channels of stress features; the second 1D convolutional layer has 64 convolutional kernels and outputs 64 channels of stress features; and the third 1D convolutional layer has 128 convolutional kernels and outputs 128 channels of stress features. The 128-channel stress features are flattened and input into the fully connected layer with 128 neurons, outputting a 64-dimensional stable discriminant representation vector, denoted as [vector name missing]. .

[0037] Will The input is the fuel condition calibration layer of the runtime adaptive stable margin estimation network. The fuel condition calibration layer consists of a first fully connected layer and a second fully connected layer, where the first fully connected layer has 64 neurons and the second fully connected layer has 128 neurons. The output of the second fully connected layer is mapped to 128-dimensional calibration parameters, denoted as... and will The 64-dimensional scaling parameter is denoted as follows: (It is then divided into sixty-four dimensions in a fixed order.) The 64-dimensional offset parameter is denoted as .

[0038] Will and Acting on ,right Each dimension is first scaled and then offset to obtain a 64-dimensional calibrated stable discriminative representation vector, denoted as . The dimensional correspondence between scaling and offset is as follows: The dimensions are matched one by one in order.

[0039] Will Input stabilization margin output layer. The stabilization margin output layer consists of a fully connected layer with 32 neurons and an output layer. The output layer has one neuron and outputs a one-dimensional adaptive stabilization margin, denoted as . The This serves as the input for subsequent risk level assessment and adjustment calculation.

[0040] Online calibration of the fuel condition calibration layer is performed based on control events. Online calibration is limited to updating the parameters of the fuel condition calibration layer while maintaining the parameters of the stable discrimination backbone unchanged. Let the first calibration threshold be... The second calibration threshold is The preset maximum number of iterations is The learning rate is In each iteration, for each set of control events, the 2048-dimensional pressure sequence corresponding to its stable event is taken and denoted as... The 2048-dimensional pressure sequence corresponding to its instability event is denoted as ,Will and The adaptive stabilization margin estimation network at runtime is used to obtain the adaptive stabilization margin for stable events. ,Will and The adaptive stabilization margin estimation network at runtime is used to obtain the adaptive stabilization margin for instability events. ;by Meets the first calibration threshold and The second calibration threshold is used as the calibration judgment condition. When a control event fails to meet the calibration judgment condition, the fuel condition calibration layer is updated in a targeted manner using the control event-driven calibration loss. The calibration loss is defined as: ; in, For calibration loss; The number of control groups; For comparison event index; For the first Adaptive stabilization margin for stable events in a control group; For the first Adaptive stabilization margin for unstable events in a control group; The first calibration threshold; This is the second calibration threshold; This is used to restrict the result within parentheses to a non-negative value for operations that take the larger value. It is calculated via backpropagation. Gradients of the fuel condition calibration layer parameters, and according to the learning rate. Update fuel condition calibration layer parameters; when all control events meet the calibration judgment criteria or the number of iterations reaches [a certain threshold]. Stop iteration when it stops, and assign the time of stopping iteration to and Received The output is the adaptive stabilization margin after online calibration.

[0041] S3, based on the adaptive stability margin, matches the event tags of the combustion stability event database, determines the risk level, and generates a risk threshold group containing trigger, release, and escalation thresholds.

[0042] Specifically, such as Figure 3 As shown, the one-dimensional adaptive stability margin output by the runtime adaptive stability margin estimation network is mapped to an executable risk level and risk threshold group to support the segmented determination and control magnitude calculation of the subsequent event-driven emission compliance control framework. The adaptive stability margin is denoted as... The current operating condition fuel characteristic package is denoted as The unit load is denoted as The lower heating value of fuel is denoted as The volume fraction of the main components of the eight-dimensional fuel is denoted as: The eight-dimensional fuel's main components, in a fixed volume fraction, include methane, ethane, propane, hydrogen, carbon monoxide, carbon dioxide, nitrogen, and inert gases. The event record index in the combustion stability event database is denoted as... The unit load corresponding to the event log is recorded as The lower heating value of fuel is denoted as The volume fraction of the main components of the eight-dimensional fuel is denoted as: The event tag corresponding to the event record is denoted as The adaptive stability margin corresponding to the event record is denoted as... ,in The dynamic pressure of the combustion chamber at the time of the event and the fuel characteristic data at the time of the event are input into the runtime adaptive stability margin estimation network with the same structure and then written into the event record for cross-operating condition margin alignment comparison.

[0043] Will and Input the combustion stability event database to filter candidate event records. Set the second load threshold to... The second calorific value threshold is The threshold for the second component is Record each event. ,right and Calculate the difference and take its absolute value to determine if it does not exceed [the specified value]. ;right and Calculate the difference and take its absolute value to determine if it does not exceed [the specified value]. ;right and Calculate the difference dimension by dimension and take the absolute value to determine whether each dimension does not exceed the limit. Event records that simultaneously meet all three criteria are combined to form a candidate event record set, and the validity period of each record in the candidate event record set is retained. and Used for subsequent matching.

[0044] The event label corresponding to the margin interval to which the adaptive stability margin belongs is determined from the candidate event record set. The combustion stability event database pre-configures a set of margin interval boundaries for the margin interval, and the margin interval index is denoted as... The lower boundary of the remaining interval is denoted as The upper boundary of the remaining interval is denoted as And configure an event label for each margin interval as follows: .Will Compare with the boundaries of each remaining interval to determine if it meets the requirements. and Remaining interval index Configure the event label for this margin range The event labels are used as the corresponding event tags for the remaining range; subsequently, the event tags are selected from the candidate event record set. A subset of candidate event records; when multiple event records exist in the subset of candidate event records, a calculation is performed for each event record. and The difference between the event records is taken, and the absolute value of the difference is used. The event tag corresponding to the event record with the smallest difference is selected as the matching event tag, and the matching event tag is denoted as [missing information]. .

[0045] Based on matching event tags The risk level is determined by calling the event tag-to-risk level mapping table. This mapping table is stored as key-value pairs in the combustion stability event database, with the event tag key denoted as... The risk level value is recorded as ;Will Input the event tag key into the mapping table, and read the result. The corresponding risk level is denoted as .

[0046] According to risk level Matching event tags Extract risk threshold groups from the combustion stability event database. The risk threshold groups are denoted as... The trigger threshold is denoted as The threshold for release is denoted as The upgrade threshold is denoted as The combustion stability event database maintains a risk threshold group record for each pair of risk level and event label combinations, setting the risk level field equal to... And the event label field equals Risk threshold group records are used as target records and their data are read. , , composition When there are multiple target records, use them sequentially. Minimum, then use The minimum rule identifies a unique target record and outputs the result. This serves as input for subsequent steps, representing a risk threshold group.

[0047] S4. Define the acceptable and adjustable range of emissions in the emission-operating condition mapping matrix, retrieve the corresponding control action family according to the risk level to obtain the control action, calculate the control amplitude by combining the adaptive stability margin and risk threshold group, and generate a draft control instruction.

[0048] Specifically, such as Figure 4 As shown, a draft control instruction is generated under the event-driven emission compliance control framework, enabling control actions to complete the risk level-driven amplitude calculation and boundary trimming within the hard constraints of the adjustable emission compliance range. The current operating condition fuel characteristic package is denoted as... The unit load is denoted as The lower heating value of fuel is denoted as The volume fraction of the main components of the eight-dimensional fuel is denoted as: The The gases, in a fixed order, include methane, ethane, propane, hydrogen, carbon monoxide, carbon dioxide, nitrogen, and inert gases. The current control parameters for the unit are denoted as... ,in It consists of the current setpoints of multiple control variables in a fixed order; the adjustable range for emission compliance is denoted as... ,in Output the lower and upper boundaries of the control variables one by one according to the aforementioned control variable dimensions. The risk level is denoted as... Matching event tags are denoted as The adaptive stability margin is denoted as Risk threshold groups are denoted as The trigger threshold is denoted as The threshold for release is denoted as The upgrade threshold is denoted as .

[0049] Will Search the emission-operational condition mapping database. The emission-operational condition mapping database uses... Using the unit load, lower heating value of fuel, and volume fraction of major fuel components as search keys, the system returns results related to the above. A corresponding record of the adjustable emission compliance range; the record of the adjustable emission compliance range is indexed by the control quantity dimension, and gives the lower boundary and upper boundary of the control quantity for each corresponding dimension, forming... And maintain the order of control dimensions with Consistent.

[0050] Will and Perform boundary checks. For each control dimension, the current setpoint is read, along with the corresponding lower and upper control boundaries. When the current setpoint is less than the lower control boundary, it is corrected to the lower boundary; when the current setpoint is greater than the upper control boundary, it is corrected to the upper boundary; and when the current setpoint is between the lower and upper control boundaries, it remains unchanged. The corrected value is then... It is determined to be an adjustable initial value, denoted as . .

[0051] according to and The control action families in the combustion stability event database are invoked to obtain a candidate set of control actions. The combustion stability event database pre-establishes association records between event tags and historical actions, and these association records are grouped into control action families according to event tags; each control action includes a target control quantity dimension and a target control quantity adjustment direction. As a search key and with As a screening criterion, a candidate set of regulatory actions matching the current risk level is extracted.

[0052] Under the constraint of adjustable emission compliance range, the feasibility of the candidate set of control actions is screened and the control actions are determined. For each control action in the candidate set, its target control variable dimension is read and... Read the adjustable initial value of this dimension from [the source], and simultaneously from [the source]. The lower and upper boundaries of the control quantity for this dimension are read. When the target control quantity adjustment direction points to the upper boundary, the remaining adjustment space is determined as the difference between the upper boundary and the adjustable initial value. When the target control quantity adjustment direction points to the lower boundary, the remaining adjustment space is determined as the difference between the adjustable initial value and the lower boundary. When the remaining adjustment space is greater than zero, the adjustment action is deemed feasible. Feasible adjustment actions are used as the screening results. When multiple feasible adjustment actions exist, they are selected according to the combustion stability event database. The highest priority control action is selected from the pre-configured action priority order as the control action.

[0053] Will and Calculate the control amplitude using the input amplitude calculation rules. Read the trigger threshold. With upgrade threshold And read the preset maximum amplitude corresponding to the target control quantity dimension of the adjustment action, denoted as ,in This is the upper limit of the amplitude that can be directly applied to the target control quantity; the calculated control amplitude is denoted as... The Adopt relatively and The position is obtained by linear interpolation followed by saturation clipping: ; in, The adjustment range; This is the preset maximum amplitude; This is an adaptive stability margin; This is the trigger threshold; To upgrade the threshold; To calculate the smaller value; To calculate the larger value; Greater than Risk threshold group Given, to ensure that the denominator of the linear interpolation is positive.

[0054] Adjustment range Apply the target control variable dimension to the control action and perform boundary clipping. Read the target control variable dimension and the target control variable adjustment direction of the control action, from... Read the adjustable initial value of this dimension as the application reference, and adjust the control amplitude in the direction according to the target control value. The applied target control quantity is obtained by superimposing it onto the application reference. Then, the applied target control quantity is compared with the lower boundary and upper boundary of the control quantity in this dimension. When the applied target control quantity exceeds the lower boundary or upper boundary of the control quantity, the applied target control quantity is clipped to the corresponding boundary. The adjustment amplitude is then synchronously corrected to the difference between the adjustable initial value before clipping and the target control quantity after clipping, and the absolute value is taken. This ensures that the corrected adjustment amplitude remains a non-negative amplitude and is consistent with the final executable target control quantity.

[0055] The control action, the revised control range, and the target control quantity are combined and output as a draft control instruction. The draft control instruction must include at least the target control quantity dimension, the target control quantity adjustment direction, the target control quantity, and the control range, and must keep the target control quantity between the lower boundary and the upper boundary of the control quantity corresponding to the emission compliance adjustable range.

[0056] S5, execute the draft control command, collect the dynamic pressure of the combustion chamber after execution, input the dynamic pressure of the combustion chamber after execution and the current operating condition fuel characteristic package into the evaluation network, and calculate and output the stable response quantity.

[0057] Specifically, such as Figure 5 As shown, a closed-loop stability evaluation is performed on the execution results of the draft control directive, and the stable response is formed by estimating the network's output difference using the runtime adaptive stability margin. The adaptive stability margin before the execution of the draft control directive is denoted as... The current operating condition fuel characteristic package is denoted as... The unit load is denoted as The lower heating value of fuel is denoted as The volume fraction of the main components of the eight-dimensional fuel is denoted as: The The pressure sequence, following a fixed order, includes methane, ethane, propane, hydrogen, carbon monoxide, carbon dioxide, nitrogen, and inert gases; the resulting 2048-dimensional pressure sequence is denoted as... The adaptive stabilization margin after execution is denoted as... The steady-state response is denoted as The preset response delay is denoted as The fuel condition calibration layer parameters calibrated online are denoted as follows: The stable discrimination backbone parameters are denoted as ,in and It is a set of parameters for the weights and biases of each layer of the network, and is stored in a fixed manner when the online calibration is completed.

[0058] The draft control instructions are sent to the unit control system. The unit control system parses the target control quantity dimension, target control quantity adjustment direction, target control quantity, and control amplitude contained in the draft control instructions, and writes the target control quantity into the setpoint register of the corresponding control loop to trigger execution; the moment of triggering execution is recorded as... The control loop setpoint after execution is used as the control input for the post-execution operating condition, and is used for time alignment with the subsequent dynamic pressure acquisition window.

[0059] Preset response delay Upon arrival, the dynamic pressure signal of the combustion chamber is acquired and executed. At the starting moment, a continuous sampling data stream is read from the dynamic pressure measurement point in the combustion chamber, and a continuous sampling sequence of length 2048 is extracted in chronological order; the continuous sampling sequence is then used to form a 2048-dimensional pressure sequence after execution. and will As a pressure input after execution.

[0060] Will and Input: Runtime adaptive stabilizing margin estimation network; Output: The calculations consistently use online calibrated fuel condition calibration layer parameters. Stability discrimination backbone parameters Among them, the stable discrimination of the main trunk is based on The input is sequentially passed through three one-dimensional convolutional layers and downsampling layers, flattening layers, and fully connected layers to obtain a 64-dimensional stable discriminant representation vector; the fuel condition calibration layer... The input is processed through two fully connected layers to generate 64-dimensional scaling and offset parameters, which are then applied to the 64-dimensional stable discriminant representation vector to obtain a 64-dimensional calibrated stable discriminant representation vector. The stability margin output layer inputs the 64-dimensional calibrated stable discriminant representation vector into a fully connected layer and the output layer to obtain a one-dimensional adaptive stability margin. .

[0061] Will and The steady-state response is obtained by taking the difference. The difference calculation uses an adaptive stabilization margin after execution. Subtract the adaptive stability margin before the implementation of the draft regulatory directive obtain in this way and will It is stored in association with the draft regulatory directives for subsequent quantitative evaluation of the effectiveness of regulatory actions and for writing back event records.

[0062] S6, when the stable response reaches the release threshold, the draft control instruction will be output as an online control instruction; when it reaches the upgrade threshold, a backoff control instruction will be generated and output as an online control instruction under the constraint of the emission compliance adjustable range.

[0063] Specifically, such as Figure 6 As shown, based on the stable response quantity, the draft control instruction is either released online or a rollback control instruction is generated, thereby outputting the online control instruction. The stable response quantity is denoted as... Risk threshold groups are denoted as The threshold for release is denoted as The upgrade threshold is denoted as The draft regulatory directive is referred to as The adjustable range of emission standards is recorded as follows: Online control commands are recorded as follows: The rollback control instruction is recorded as The aforementioned The dimension including the target control quantity is denoted as The direction of adjustment of the target control quantity is denoted as The target control quantity is denoted as The adjustment range is denoted as ,in The values ​​are discrete and used to indicate the direction of adjustment of the target control quantity. This indicates that the target control quantity adjustment direction points to the upper boundary of the control quantity. This indicates that the target control variable adjustment direction points towards the lower boundary of the control variable. For each control dimension The lower and upper boundaries of the output control quantity, where the dimension The lower boundary of the control quantity is denoted as The upper boundary of the control variable is denoted as .

[0064] When performing an upgrade determination, and Upgrade threshold in A comparison is made to determine whether to proceed with the rollback generation process. A decision rule is employed. The stable response quantity is used as a criterion for determining whether the upgrade threshold is met. When the criterion is met, the rollback generation process begins, and the comparison of the release threshold is stopped. When the criterion is not met, the rollback generation process begins, and the process switches to the release determination. The comparison used... The adaptive stability margin is obtained by subtracting the adaptive stability margin before and after the implementation of the draft control instruction. Risk threshold group It can be read directly.

[0065] If the release decision is made before the rollback generation process has begun, then... and The release threshold A comparison is made to determine whether to approve the draft regulatory directive. A decision-making rule is adopted. As a criterion for determining whether a stable response quantity meets the release threshold; when the criterion is met, Output without modification When the condition is not met, no new output is generated. It maintains the current setpoints of the unit's control system unchanged, and reacquires and updates the dynamic pressure of the combustion chamber after execution in the next control cycle. Then, an upgrade and de-upgrade determination are performed again. Risk threshold group It can be read directly.

[0066] In the rollback generation process, such as Figure 7 As shown, according to Target control dimension Determine the dimension of the rollback control amount, and set the rollback adjustment direction and rollback magnitude. Denote the dimension of the rollback control amount as... and will Set as with Same; denoted as the direction of rollback adjustment. and will Set as Set the rollback range to the preset rollback range, denoted as . ,in This refers to a value that can be directly applied to the control quantity setpoint, and is calculated according to the fallback control quantity dimension. Read after being pre-configured in the parameter table.

[0067] Within the adjustable range of emission standards The backoff range under constraints Acting on the dimension of the backoff control quantity The target control quantity for the rollback is obtained, and boundary trimming is performed on the target control quantity to generate a rollback control command. This rollback control command is then output as the online control command. First, the dimensions in the unit control system are read. The current setting value is used as the basis for rollback application, denoted as ;when At that time, the rollback target control quantity is calculated as follows: ,when At that time, the rollback target control quantity is calculated as follows: Then read and And perform boundary clipping when the rollback target control value is less than The rollback target control quantity is then trimmed to When the rollback target control quantity is greater than The rollback target control quantity is then trimmed to When the rollback target control quantity is located and The time remains unchanged; the target control value for the retraction after trimming is denoted as And write And the rollback range will be adjusted accordingly. ,in This represents absolute value operations, used to convert the difference within parentheses into a non-negative magnitude, and finally outputs the result. As .

[0068] Figure 8 The combustion stability effect is demonstrated by comparing the left and right sides. Figure 9 This is a network structure diagram of the present invention. Figure 8 As shown, the left side represents traditional control: under changes in fuel composition or load disturbances, stronger pressure waves and local oscillations occur in the combustion chamber, causing flame morphology fluctuations and posing a risk of instability / backfire. The right side represents the present invention: real-time acquisition of dynamic combustion chamber pressure and a 10-dimensional fuel characteristic package is input into an adaptive stability margin estimation network, which outputs an "adaptive stability margin." This margin is then combined with event database thresholds for risk classification, generating control actions constrained within the adjustable range of emission compliance. After execution, pressure feedback is used to calculate the stability response; if the threshold for release is met, the operation is allowed; otherwise, a rollback is triggered. This significantly reduces pressure waves and maintains flame stability under the same disturbances, lowering the risk of instability, reducing reliance on empirical tuning, and improving unit safety and continuous operation capability.

[0069] The embodiments of the present invention also have the following technical effects: This invention decouples the discriminative capability of combustion chamber dynamic pressure from the adaptive calibration of fuel conditions, achieving runtime alignment of the stable margin under variable fuel conditions. Specifically, the scheme extracts a stable discriminative representation from the dynamic pressure as a fixed-length sequence input to a one-dimensional convolutional network backbone. It then introduces a fuel feature package consisting of "unit load + fuel lower heating value + main component volume fraction" as conditional input. A fuel condition calibration layer generates scaling and offset parameters to modulate the representation vector dimensionally, thereby ensuring a consistent interpretation benchmark for the same backbone feature under different fuels. Compared to approaches relying on fixed thresholds or offline calibration models, this structure, without altering the backbone discriminative mechanism, achieves stable margin adaptation under fuel switching, calorific value drift, and other conditions through conditional calibration, reducing the risk of discriminative calibrator drift caused by fuel changes.

[0070] This invention enables not only online estimation of the stability margin, but also allows it to be constrained by an event database to a decision-making range under current fuel conditions. The scheme retrieves stable and unstable events from a combustion stability event database for similar loads and fuels, pairs them to form control events, updates only the fuel condition calibration layer parameters, and freezes the core stability discrimination parameters. Iterative calibration is driven by the criterion that "the stable event margin meets the first calibration threshold and the unstable event margin meets the second calibration threshold," ensuring that the calibration direction directly serves the online alignment of the stability / instability boundaries, avoiding operational uncertainties caused by network-wide updates. Based on this, the online adaptive stability margin is associated with the event database margin range, event tags, risk levels, and trigger / release / escalation threshold groups to form segmented judgment criteria for on-site handling. This distinguishes it from monitoring algorithms that only provide alarms, establishing a traceable mapping between "margin—event semantics—risk threshold" under current operating conditions.

[0071] This invention provides a qualified adjustable range for the control quantity dimension through an "emission-operating condition mapping library," and completes action screening, amplitude calculation, and closed-loop release / reverse under this hard constraint. The scheme uses the family of control actions corresponding to risk levels and event labels as a candidate set, and performs feasibility screening based on the current control quantity and adjustable boundaries to avoid selecting actions with no remaining adjustment space within the boundaries. Then, based on the adaptive stability margin and trigger / upgrade thresholds, linear interpolation and saturation pruning are performed to determine the control amplitude, and boundary pruning is performed on the target control quantity to ensure that the command is executable and does not exceed the limits. After the command is executed, new dynamic pressure is collected after a preset response delay, and the margin is recalculated. The difference between the margins before and after execution constitutes the stable response quantity. The command is released based on the release threshold or a reverse reverse command is generated based on the upgrade threshold, realizing online verification of the control effect and closed-loop handling of adverse responses. Compared to strategies that directly issue experience-based actions or lack feedback verification, this method integrates fuel variability, emission compliance boundaries, and post-control verification into a single online process to solve the stability and controllability problem in scenarios with variable fuel and emission constraints.

[0072] To achieve the above embodiments, such as Figure 10 As shown, this embodiment also provides a deep learning-based online control device 10 for gas turbine combustion stability, comprising: The fuel feature package fusion generation module 100 is used to acquire unit load, fuel composition and calorific value, and combustion chamber dynamic pressure, retrieve fuel composition and calorific value database and dynamic pressure history database, and fuse them to generate the current operating condition fuel feature package. The adaptive stability margin assessment and calibration module 200 is used to retrieve stable / instability comparison events based on the fuel feature package under the current operating conditions, input the runtime adaptive stability margin assessment network, calibrate the fuel condition calibration layer online, and output the adaptive stability margin. The risk level determination and threshold generation module 300 is used to determine the risk level and generate a risk threshold group containing trigger, release, and escalation thresholds by matching event tags in the combustion stability event database with adaptive stability margin. The draft control instruction generation module 400 is used to define the acceptable adjustable range of emissions in the emission-operating condition mapping matrix, retrieve the corresponding control action family according to the risk level to obtain the control action, calculate the control amplitude by combining the adaptive stability margin and risk threshold group, and generate a draft control instruction. The stable response calculation module 500 is used to execute the draft control command, collect the dynamic pressure of the combustion chamber after execution, input the dynamic pressure of the combustion chamber after execution and the current operating condition fuel characteristic package into the evaluation network, and calculate and output the stable response. The stable response quantity assessment and command output module 600 is used to output the draft control command as an online control command when the stable response quantity reaches the release threshold; and to generate and output the back-off control command as an online control command under the constraint of the emission compliance adjustable range when the stable response quantity reaches the upgrade threshold.

[0073] An embodiment of the present invention provides an online control device for combustion stability of a gas turbine based on deep learning, which can realize online adaptive assessment and control of combustion stability of the gas turbine under changes in fuel composition and calorific value, ensure the generation of executable control commands under the premise of meeting emission constraints, and improve control reliability and unit operation safety through closed-loop verification and backoff mechanisms.

[0074] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 11 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the online control method for combustion stability of gas turbine based on deep learning described above.

[0075] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a deep learning-based online control method for gas turbine combustion stability as described in the foregoing embodiments.

[0076] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0077] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for online control of combustion stability in gas turbines based on deep learning, characterized in that, Includes the following steps: S1: Obtain unit load, fuel composition and calorific value, and combustion chamber dynamic pressure; retrieve fuel composition and calorific value database and dynamic pressure history database; and fuse them to generate a fuel characteristic package for the current operating condition. S2 retrieves stable / instable comparison events based on the current operating condition fuel feature package, inputs them into the runtime adaptive stability margin assessment network, calibrates the fuel condition calibration layer online, and outputs the adaptive stability margin; S3, based on the adaptive stability margin, matches the event tags of the combustion stability event database, determines the risk level, and generates a risk threshold group containing trigger, release, and escalation thresholds; S4. Define the adjustable range of emissions compliance in the emission-operating condition mapping matrix, retrieve the corresponding control action family according to the risk level to obtain the control action, calculate the control amplitude by combining the adaptive stability margin and risk threshold group, and generate a draft control instruction. S5, execute the draft control command, collect the dynamic pressure of the combustion chamber after execution, input the dynamic pressure of the combustion chamber after execution and the current operating condition fuel characteristic package into the evaluation network, and calculate and output the stable response quantity; S6, when the stable response reaches the release threshold, the draft control instruction will be output as an online control instruction; when it reaches the upgrade threshold, a backoff control instruction will be generated and output as an online control instruction under the constraint of the emission compliance adjustable range.

2. The method as described in claim 1, characterized in that, S1 includes: Collect the unit load and determine the unit load sampling time. At the same time, collect the dynamic pressure of the combustion chamber and use the unit load sampling time as the operating time reference. Access the fuel composition and calorific value database based on the operating condition time reference to obtain the fuel's lower heating value and the volume fraction of the fuel's main components, which include methane, ethane, propane, hydrogen, carbon monoxide, carbon dioxide, nitrogen, and inert gases. The unit load, the lower heating value of the fuel, and the volume fraction of the main components of the fuel are spliced ​​together in a fixed dimensional order to form a ten-dimensional fuel feature package under the current operating condition. A continuous sampling sequence of length 2,048 is extracted from the dynamic pressure of the combustion chamber to form a 2,048-dimensional pressure sequence for subsequent input steps.

3. The method as described in claim 1, characterized in that, The S2 includes: The event retrieval conditions are determined based on the current operating condition fuel feature package. The event retrieval conditions include the unit load difference not exceeding a first load threshold, the fuel lower heating value difference not exceeding a first calorific value threshold, and the volume fraction differences of the main fuel components not exceeding a first component threshold. The event retrieval criteria are input into the combustion stability event database to retrieve stable and unstable events that meet the event retrieval criteria. The stable and unstable events are then arranged into a control event in chronological order. Based on the aforementioned control event, the fuel condition calibration layer is calibrated online. The online calibration is limited to updating the parameters of the fuel condition calibration layer while keeping the parameters of the stable discrimination backbone unchanged. The parameters of the fuel condition calibration layer are iteratively updated until the calibration judgment condition is met or the preset iteration limit is reached, and the adaptive stability margin after online calibration is output.

4. The method as described in claim 1, characterized in that, The S3 includes: The adaptive stability margin and the fuel characteristic package under the current operating condition are input into the combustion stability event database. Candidate event records are filtered according to the following criteria: the difference between the unit load and the fuel load does not exceed the second load threshold, the difference between the fuel lower heating value and the fuel volume fraction of each dimension does not exceed the second component threshold. In the candidate event records, determine the event label corresponding to the margin interval to which the adaptive stabilization margin belongs, and when there are multiple event labels, select the event label with the smallest difference in adaptive stabilization margin as the matching event label.

5. The method as described in claim 1, characterized in that, The S4 includes: Under the constraint of the adjustable range of emission compliance, the feasibility of the candidate set of control actions is screened, and the control actions with remaining adjustment space for the adjustable initial value corresponding to the dimension of target control quantity are selected as control actions. The adaptive stability margin and risk threshold are grouped together to calculate the amplitude. When the adaptive stability margin does not meet the trigger threshold, the control amplitude is set to zero. When the adaptive stability margin meets the trigger threshold but does not meet the upgrade threshold, the control amplitude is determined by linear interpolation. When the adaptive stability margin meets the upgrade threshold, the control amplitude is set to the preset maximum amplitude. The adjustment amplitude is applied to the target control quantity dimension of the adjustment action and the boundary is clipped. When the applied target control quantity exceeds the lower boundary or upper boundary of the control quantity, the target control quantity is clipped to the corresponding boundary and the adjustment amplitude is corrected simultaneously. The control actions, the revised control range, and the target control quantity are output as a draft control instruction.

6. The method as described in claim 1, characterized in that, The S5 includes: The draft control instructions are sent to the unit control system and the control actions are executed to obtain the operating status after execution. After a preset response delay, the dynamic pressure of the combustion chamber after execution is collected, and a continuous sampling sequence of length 2,048 is extracted to form a 2,048-dimensional pressure sequence after execution. The 2,048-dimensional pressure sequence after execution and the fuel characteristic package under the current operating condition are input into the adaptive stability margin estimation network during runtime. The adaptive stability margin after execution is calculated and output using the online calibrated fuel condition calibration layer parameters and stability discrimination backbone parameters. The stable response is obtained by subtracting the adaptive stability margin after execution from the adaptive stability margin before the execution of the draft control instruction.

7. The method as described in claim 1, characterized in that, The S6 includes: The stable response quantity is compared with the escalation threshold in the risk threshold group. When the stable response quantity meets the escalation threshold, the rollback generation process is initiated. Before entering the rollback generation process, the stable response quantity is compared with the release threshold in the risk threshold group. When the stable response quantity meets the release threshold, the draft control instruction is output as an online control instruction. In the rollback generation process, the rollback control quantity dimension is determined according to the target control quantity dimension of the draft control instruction, the rollback adjustment direction is set to be opposite to the target control quantity adjustment direction of the draft control instruction, and the rollback magnitude is set to a preset rollback magnitude. Under the constraint of the adjustable range of emission compliance, the backoff amplitude is applied to the backoff control quantity dimension to obtain the backoff target control quantity, and the backoff target control quantity is subjected to boundary trimming to generate a backoff control command, which is then output as an online control command.

8. A deep learning-based online control device for combustion stability of a gas turbine, characterized in that, include: The fuel feature package fusion generation module is used to obtain unit load, fuel composition and calorific value, and combustion chamber dynamic pressure, retrieve fuel composition and calorific value database and dynamic pressure history database, and fuse them to generate the current operating condition fuel feature package. The adaptive stability margin assessment and calibration module is used to retrieve stable / instability comparison events based on the fuel feature package under the current operating conditions, input the runtime adaptive stability margin assessment network, calibrate the fuel condition calibration layer online, and output the adaptive stability margin. The risk level determination and threshold generation module is used to determine the risk level and generate a risk threshold group containing trigger, release, and escalation thresholds by matching event tags in the combustion stability event database with adaptive stability margin. The module for generating draft control instructions is used to define the acceptable and adjustable range of emissions in the emission-operating condition mapping matrix, retrieve the corresponding control action family according to the risk level to obtain the control action, calculate the control amplitude by combining the adaptive stability margin and risk threshold group, and generate a draft control instruction. The stable response calculation module is used to execute the draft control command, collect the dynamic pressure of the combustion chamber after execution, input the dynamic pressure of the combustion chamber after execution and the current operating condition fuel characteristic package into the evaluation network, and calculate and output the stable response. The stable response quantity assessment and instruction output module is used to output the draft control instruction as an online control instruction when the stable response quantity reaches the release threshold; and to generate and output the back-off control instruction as an online control instruction under the constraint of the emission compliance adjustable range when the stable response quantity reaches the upgrade threshold.

9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the online control method for combustion stability of gas turbine based on deep learning as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements a deep learning-based online control method for combustion stability of gas turbines as described in any one of claims 1-7.

Images