A burner collaborative operation optimization method based on self-supervised learning

Abstract

The application discloses a kind of based on self-supervised learning's burner collaborative operation optimization method, it is related to industrial process control field, including the following steps: obtaining and preprocessing burner operation data, furnace operation data and natural disturbance record;Extract natural operation disturbance and its associated operating state and carry out continuous fragment division;Configuration self-supervised learning task and the joint time series representation learning of each burner operating state and furnace response state;Burner effect propagation chain is constructed;Unified dynamic state sequence is constructed and group oscillation mode is extracted based on copula operator;Abnormal group oscillation mode is identified and the participation degree and phase difference of each burner in abnormal group oscillation mode are extracted;Determine associated burner adjustment parameter and implement multi-burner reverse collaborative regulation, generate collaborative operation optimization result.The application utilizes self-supervised learning and dynamic mode decomposition, realizes disturbance positioning and collaborative regulation, improves regulation response speed and operating stability.

Description

Technical Field

[0001] This invention relates to the field of industrial process control, and in particular to a method for optimizing the cooperative operation of burners based on self-supervised learning. Background Technology

[0002] Existing boiler combustion control typically relies on fuel pressure, fuel flow, primary air volume, secondary air volume, burner load, flame temperature, and furnace operating parameters to independently monitor the status of each burner. It also employs fixed rules, empirical parameters, or conventional time-series models for load distribution and operational adjustment to reduce the impact of local fluctuations on combustion stability.

[0003] However, during the coordinated operation of multiple burners, natural disturbances generated by a single burner may continue to propagate through other burners or the common state of the furnace, resulting in synchronous oscillations, phase misalignment, and localized anomalous enhancements. Existing technologies struggle to accurately separate direct propagation, indirect propagation, and common furnace disturbances, and also find it difficult to determine the initiating burner, associated burners, and propagation path of abnormal oscillations. This leads to inaccurate configuration of the control target, direction, and sequence, easily resulting in control lag, redundant adjustments, and secondary fluctuations.

[0004] Therefore, how to provide a self-supervised learning-based method for optimizing the cooperative operation of burners is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a combustor cooperative operation optimization method based on self-supervised learning. This invention utilizes self-supervised learning and dynamic mode decomposition to achieve disturbance localization and cooperative adjustment, thereby improving positioning accuracy, response speed and operational stability.

[0006] A combustor cooperative operation optimization method based on self-supervised learning according to an embodiment of the present invention includes the following steps: Acquire burner operation data, furnace operation data, and natural disturbance records during the multi-burner collaborative operation process, and perform preprocessing to generate a basic burner dataset; Natural operating disturbances and their associated operating states are extracted from the burner basic dataset and divided into continuous segments to generate natural disturbance operating segment data; Based on the data of natural disturbance operation segments, a self-supervised learning task is configured to perform joint time-series characterization learning on the operating status of each burner and the furnace response status, generating burner collaborative state characterization data. Based on the burner cooperative state characterization data, conditional Granger causality analysis was used to determine the propagation direction and propagation delay of natural operating disturbances between each burner and the furnace, and the propagation effects were separated to construct the burner effect propagation chain; Based on the burner action propagation chain and burner basic dataset, a unified dynamic state sequence consisting of multiple burner operating states and furnace operating states is constructed. Based on the Koopman operator, the group oscillation mode in the unified dynamic state sequence is extracted to generate group oscillation mode data. Abnormal group oscillation modes are identified based on the group oscillation mode data, and the participation degree and phase difference of each burner in the abnormal group oscillation mode are extracted by combining the burner action propagation chain to generate abnormal oscillation location data; Based on abnormal oscillation location data, the adjustment parameters of the associated burners are determined and multi-burner reverse coordinated adjustment is implemented to generate coordinated operation optimization results.

[0007] Optionally, the preprocessing includes time alignment, burner attribution marking, operation phase division, collection anomaly removal, and data normalization.

[0008] Optionally, the generation of the natural disturbance runtime segment data specifically includes: Natural disturbance record data is read from the burner basic dataset. Disturbance information is marked according to disturbance record category, burner affiliation mark, and continuous recording time range to generate natural disturbance mark data. The natural disturbance mark data includes disturbance type, disturbance source burner, disturbance start time, and disturbance end time. Using the source burner of the disturbance, the start time of the disturbance, and the end time of the disturbance in the natural disturbance marker data as search criteria, the associated operating states within the corresponding operating time segment are extracted from the burner basic dataset to generate natural disturbance associated operating state data; Read the start and end times of disturbances and other burner and furnace response states from the natural disturbance-related operating status data, determine the start and recovery times of the state response, and generate segment boundary data. The natural disturbance-related operational state data is divided into continuous segments according to the segment boundary data. The start time of the state response is marked in the continuous operational segment, and the natural disturbance marking data is bound to each continuous segment to generate natural disturbance operational segment data. The natural disturbance operational segment data includes the operational segment before the disturbance, the segment in which the disturbance occurs, and the operational segment after the disturbance.

[0009] Optionally, the generation of the burner cooperative state characterization data specifically includes: Read the natural disturbance operation segment data, set the masking position for each burner operation state, and construct model input and state recovery labels using the unmasked operation state and the original operation state respectively to generate burner state masking recovery samples; Adjust the order of the running segments before the disturbance, the segment where the disturbance occurs, and the running segments after the disturbance, and use the original time order of each running segment as the time series discrimination label to generate disturbance time series discrimination samples; The operating status of the source burner of the disturbance is used as the prediction input, and the response status of other burners and the furnace response status are used as the prediction labels to generate cross-burner response prediction samples and furnace response reconstruction samples. The burner state masking recovery samples, disturbance time series discrimination samples, cross-burner response prediction samples, and furnace response reconstruction samples are jointly input into the time series representation network. The network parameters are jointly updated with the results of their respective tasks. The time series features obtained from the training are aggregated according to burner affiliation and sampling time to generate burner collaborative state representation data.

[0010] Optionally, the construction of the burner action propagation chain specifically includes: According to the burner affiliation and sampling time, the burner cooperative state characterization data are continuously reassembled into a sequence, and the analysis order and lag position between burners and between burners and furnace are configured to generate a conditional Granger causality analysis sequence. Based on the conditional Granger causality analysis sequence, the baseline prediction sequence of the response analysis object and the propagation analysis input prediction sequence are constructed respectively, and the state prediction deviations at different lag positions are compared to generate initial propagation causal data; Based on the initial propagation causal data configuration condition analysis data addition order, after adding the cross-burner propagation characterization data and furnace common response characterization data of the candidate intermediate burners item by item, the state prediction deviation is recalculated, and the three types of propagation effects are determined to generate propagation effect separation data. Based on the propagation direction and propagation delay, the propagation effect separation data are path-connected and propagation order marked to construct a burner effect propagation chain that includes the propagation effects of each burner.

[0011] Optionally, the generation of the group oscillation mode data specifically includes: Based on the burner action propagation chain, the operating states of multiple burners and furnace are extracted from the basic burner dataset and synchronously arranged according to the operating time and propagation order to construct a unified dynamic state sequence. The unified dynamic state sequence is extended by delay coordinates and combined with state changes. The extended operating states are then aggregated according to the propagation order in the burner action propagation chain to generate the Koopman observable state sequence. The Koopman observable state sequence is paired with preceding and following state snapshots at a fixed snapshot interval, and the paired preceding and following state snapshots are arranged according to the sampling time to generate state snapshot matrix data. Singular value decomposition and effective order truncation are performed on the state snapshot matrix data, and the low-order Koopman evolution operator is obtained based on the evolution mapping between the truncated state snapshots before and after truncation, generating Koopman operator decomposition data; Based on the Koopman operator decomposition data and state snapshot matrix data, the low-order Koopman evolution operator is eigenvalued and the decomposition results are mapped to the operating states of each burner and furnace in the unified dynamic state sequence to generate dynamic modal parameter data. Based on the participation amplitude of each burner mode, the phase of each burner mode, and the phase of the furnace mode in the dynamic modal parameter data, the dynamic modes are classified into group types to generate group oscillation mode data.

[0012] Optionally, the generation of the abnormal oscillation location data specifically includes: Read the group oscillation mode data, perform anomaly screening for each group oscillation mode according to the group type and oscillation amplitude, and retain the corresponding oscillation frequency, oscillation amplitude, amplitude of each burner mode participation, phase of each burner mode and phase of furnace mode to generate abnormal group oscillation mode data; Based on the abnormal population oscillation mode data, the participation degree of each burner in the corresponding abnormal population oscillation mode is calculated, and the phase of each burner mode and the phase of the furnace mode are compared in two and two directions to generate abnormal mode participation analysis data. By combining abnormal mode participation analysis data and burner action propagation chain, the initiating burner, dominant participating burner, associated burner and oscillation propagation path are determined and abnormal oscillation path location data is generated. The abnormal population oscillation mode data, abnormal mode participation analysis data, and abnormal oscillation path location data are collected according to the abnormal population oscillation mode, and the participation degree and phase difference of each burner are sorted out according to the oscillation propagation path to generate abnormal oscillation location data.

[0013] Optionally, the generation of the collaborative operation optimization results specifically includes: Read abnormal oscillation location data, extract associated operating status from the burner basic dataset based on associated burners, abnormal group oscillation modes and their oscillation frequencies, and filter adjustment objects to generate adjustment object configuration data; Based on the configuration data of the regulating object and the participation level and phase difference of each associated burner, the current phase position of the abnormal group oscillation mode is calculated, and the regulation direction and regulation amount are determined to generate reverse regulation parameter data. The reverse adjustment parameter data are arranged in reverse order according to the oscillation propagation path, and the adjustment interval of adjacent associated burners is configured according to the propagation delay to generate a multi-burner reverse coordinated adjustment sequence. The adjustment is implemented step by step according to the multi-burner reverse coordinated adjustment sequence, and the adjusted burner operating status, furnace operating status and group oscillation status are continuously verified to generate coordinated operation optimization results. The coordinated operation optimization results include the adjusted burner operating status, the adjusted furnace operating status and the adjusted group oscillation status.

[0014] The beneficial effects of this invention are: This invention unifies the collection and preprocessing of burner operation data, furnace operation data, and natural disturbance records during multi-burner coordinated operation. Based on the natural disturbance operation segments, it configures self-supervised learning tasks such as burner state masking recovery, disturbance timing discrimination, cross-burner response prediction, and furnace response reconstruction. This enables the time-series representation network to extract natural disturbance representations, cross-burner propagation representations, and furnace common response representations from continuous operation data lacking manual annotation. Compared to control methods that rely solely on fixed rules or the local state of a single burner, this invention fully utilizes the state change information contained in natural operational disturbances, improving the completeness of the representation of multi-burner associated states and furnace response characteristics. This provides a unified and continuous data foundation for subsequent disturbance propagation analysis and abnormal oscillation identification.

[0015] This invention employs conditional Granger causality analysis to determine the propagation direction and propagation delay between burners and between burners and the furnace. By sequentially incorporating candidate intermediate burner and furnace common response characterization data, it separates direct propagation effects, indirect propagation effects, and furnace common disturbance effects, thereby constructing a burner action propagation chain. This process reduces the interference of furnace common response and intermediate burner transmission on disturbance source determination, making the disturbance source burner, affected burners, propagation order, and propagation path clearer. It improves upon the problems of existing technologies that struggle to distinguish multiple propagation effects and are prone to misattribution.

[0016] This invention further employs dynamic mode decomposition based on the Koopman operator to extract common oscillation modes, reverse oscillation modes, and local abnormal modes from the unified dynamic state sequence of multiple burners and the furnace. Combining mode participation amplitude, mode phase, and burner propagation chain, it locates the initiating burner, the dominant participating burner, the associated burner, and the oscillation propagation path. Based on this, the adjustment targets are selected according to their contribution, and the adjustment direction and amount are configured according to the current phase position. Step-by-step coordinated adjustment is implemented in reverse order of the oscillation propagation path, thereby matching the adjustment targets, adjustment directions, adjustment order, and adjustment intervals with the propagation process of abnormal oscillations. This reduces the probability of adjustment lag, repeated adjustment, over-adjustment, and secondary fluctuations, improving the stability and targeted adjustment of multi-burner coordinated operation. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a combustor cooperative operation optimization method based on self-supervised learning proposed in this invention; Figure 2 This is a flowchart illustrating the group oscillation mode data generation process for a self-supervised learning-based combustor cooperative operation optimization method proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-2 A self-supervised learning-based method for optimizing the cooperative operation of burners includes the following steps: Acquire burner operation data, furnace operation data, and natural disturbance records during the multi-burner collaborative operation process, and perform preprocessing to generate a basic burner dataset; Natural operating disturbances and their associated operating states are extracted from the burner basic dataset and divided into continuous segments to generate natural disturbance operating segment data; Based on the data of natural disturbance operation segments, a self-supervised learning task is configured to perform joint time-series characterization learning on the operating status of each burner and the furnace response status, generating burner collaborative state characterization data. Based on the burner cooperative state characterization data, conditional Granger causality analysis was used to determine the propagation direction and propagation delay of natural operating disturbances between each burner and the furnace, and the propagation effects were separated to construct the burner effect propagation chain; Based on the burner action propagation chain and burner basic dataset, a unified dynamic state sequence consisting of multiple burner operating states and furnace operating states is constructed. Based on the Koopman operator, the group oscillation mode in the unified dynamic state sequence is extracted to generate group oscillation mode data. Abnormal group oscillation modes are identified based on the group oscillation mode data, and the participation degree and phase difference of each burner in the abnormal group oscillation mode are extracted by combining the burner action propagation chain to generate abnormal oscillation location data; Based on abnormal oscillation location data, the adjustment parameters of the associated burners are determined and multi-burner reverse coordinated adjustment is implemented to generate coordinated operation optimization results.

[0020] In this embodiment, the burner operating data includes fuel pressure, fuel flow rate, primary air volume, secondary air volume, burner load, and flame temperature; the furnace operating data includes furnace pressure, furnace temperature, and flue gas oxygen content; the natural disturbance records include fuel pressure fluctuation records, air volume change records, and load switching records; the preprocessing includes time alignment, burner attribution marking, operating stage division, collection anomaly removal, and data normalization.

[0021] In this embodiment, the generation of natural disturbance operation segment data specifically includes: Natural disturbance record data is read from the burner basic dataset. Disturbance information is marked according to disturbance record category, burner attribution label and continuous recording time range to generate natural disturbance label data. Natural disturbance label data includes disturbance type, disturbance source burner, disturbance start time and disturbance end time. The disturbance information marking includes marking fuel pressure fluctuation records, air volume change records and load switching records as corresponding disturbance types, determining the source burner of the disturbance based on the burner attribution mark, and determining the first and last recording times of consecutive disturbance records of the same type as the disturbance start time and disturbance end time, respectively. In practice: the time interval between adjacent records of the same type of disturbance is set to 1 to 3 sampling periods, and the number of consecutive records of the same type of disturbance is set to 3 to 10. Using the source burner of the disturbance, the start time of the disturbance, and the end time of the disturbance in the natural disturbance marker data as search criteria, the associated operating states within the corresponding operating time segment are extracted from the burner basic dataset to generate natural disturbance associated operating state data; The associated operating status extraction includes extracting the fuel pressure, fuel flow rate, primary air volume, secondary air volume, burner load, and flame temperature of the burner from which the disturbance originates, forming the operating status of the burner from which the disturbance originates; extracting the fuel pressure, fuel flow rate, primary air volume, secondary air volume, burner load, and flame temperature of other burners during the same operating period, forming the response status of other burners; and extracting the furnace pressure, furnace temperature, and flue gas oxygen content during the same operating period, forming the furnace response status. In specific implementation: the corresponding running time is set to 30 to 120 sampling periods before the start of the disturbance and 60 to 300 sampling periods after the end of the disturbance. When the response states of other burners and furnaces at the end of the corresponding running time have not all continuously entered their respective fluctuation ranges before the disturbance, the search range after the end of the disturbance is expanded by increasing by 60 to 300 sampling periods. Read the start and end times of disturbances and other burner and furnace response states from the natural disturbance-related operating status data, determine the start and recovery times of the state response, and generate segment boundary data. The state response start time is the earliest moment when the response states of other burners and the furnace first continuously exceed their respective pre-disturbance fluctuation ranges; the state response recovery time is the latest moment among the start times when the response states of other burners and the furnace first continuously re-enter their respective pre-disturbance fluctuation ranges. In specific implementation: the fluctuation range before the disturbance is set according to the operating status within 30 to 120 sampling periods before the start of the disturbance, the continuous exceedance time is set to 3 to 10 sampling periods, and the continuous entry time is set to 5 to 30 sampling periods; The natural disturbance-related operational status data is divided into continuous segments according to the segment boundary data. The start time of the state response is marked in the continuous operational segments, and the natural disturbance marking data is bound to each continuous segment to generate natural disturbance operational segment data. The natural disturbance operational segment data includes the operational segment before the disturbance, the segment in which the disturbance occurs, and the operational segment after the disturbance. The continuous segment division includes extracting the running segment before the disturbance according to 30 to 120 sampling periods before the disturbance start time, extracting the disturbance occurrence segment according to the disturbance start time to the disturbance end time, and extracting the running segment after the disturbance according to the disturbance end time to the state response recovery time.

[0022] In this embodiment, the generation of burner cooperative state characterization data specifically includes: Read the operating data of natural disturbance segments, set the masking position for each burner's operating state, and construct model input and state recovery labels using the unmasked operating state and the original operating state respectively to generate burner state masking recovery samples; specifically: arrange the operating states of each burner according to the sampling time, continuously combine the operating segments before disturbance, the disturbance occurrence segment, and the disturbance after disturbance in chronological order, and extract continuous operating sequences with an effective length of 60 to 600 sampling periods from the combination results as the masking processing objects; construct masking intervals with 3 to 10 consecutive sampling times, and follow the principle of 1 The masking interval is moved sequentially along each running segment with a moving step size of up to 5 sampling periods, and one masking interval is set for each move; the running state in the masking interval is replaced with zero value, and a masking mark is configured for each sampling time. Unmasked positions are recorded as 0, and masked positions are recorded as 1, forming the model input; the original running state before replacement is arranged according to the masking position and sampling time to form a state recovery label; the model input and state recovery label formed by each move are paired, and 10 to 200 groups are evenly selected from all the pairing results according to the sampling time to generate burner state masking recovery samples; The order of the pre-disturbance, disturbance-occurrence, and post-disturbance running segments is adjusted, and the original time sequence of each running segment is used as the time series discrimination label to generate disturbance time series discrimination samples. Specifically, the input length of the pre-disturbance, disturbance-occurrence, and post-disturbance running segments is set to 30 to 120 sampling periods respectively. When a running segment is less than the set length, it is padded to the set length according to the running state at the end of the running segment. When a running segment exceeds the set length, it is extracted to the set length at equal intervals according to the sampling time. The three running segments are combined according to 6 different arrangements. One combination arranged in the order of pre-disturbance, disturbance-occurrence, and post-disturbance running segments is marked as the original time sequence, and the remaining 5 combinations are marked as non-original time sequences. Each combination and its time series discrimination label are paired to generate disturbance time series discrimination samples. The operating status of the source burner of the disturbance is used as the prediction input, and the response status of other burners and the furnace response status are used as prediction labels to generate cross-burner response prediction samples and furnace response reconstruction samples. Specifically, the operating status of the source burner for 30 to 120 consecutive sampling periods is extracted according to the sampling time, and the prediction input interval is moved sequentially along the natural disturbance operation segment according to the moving step size of 1 to 5 sampling periods, and the prediction input interval with a complete prediction label length after the end of the prediction input interval is retained; the first sampling time after the end of the prediction input interval is used as the prediction label start time, and the response status of other burners in the following 5 to 30 sampling periods is extracted and arranged according to the burner attribution mark and sampling time to form cross-burner response prediction labels, and the furnace response status at the same time is extracted and arranged according to the sampling time to form furnace response reconstruction labels; the prediction input is paired with the cross-burner response prediction labels and the furnace response reconstruction labels respectively to generate cross-burner response prediction samples and furnace response reconstruction samples. The burner state masking recovery samples, disturbance time sequence discrimination samples, cross-burner response prediction samples and furnace response reconstruction samples are jointly input into the time sequence representation network. The network parameters are jointly updated with the results of their respective tasks. The time sequence features obtained from training are aggregated according to burner affiliation and sampling time to generate burner collaborative state representation data. Specifically, samples of the same perturbation type, perturbation source burner, perturbation start time, and perturbation end time are grouped into the same joint training sample group. Different joint training sample groups are then divided into training samples, validation samples, and test samples in an 8:1:1 ratio, with each batch of training samples consisting of 16 to 128 groups. The temporal representation network employs a shared coding structure, including an input mapping layer with an output dimension of 64 to 256, a bidirectional long short-term memory layer with 1 to 3 layers and 64 to 256 hidden units per direction per layer, and 2 to 8 attention heads, each with specific... The network consists of a temporal attention layer with feature dimensions ranging from 16 to 64, as well as a state recovery output layer, a temporal discrimination output layer, a cross-burner response prediction output layer, and a furnace response reconstruction output layer. The random deactivation ratio for each network layer is set to 0.1 to 0.5. The temporal representation network extracts shared temporal features at each sampling time and outputs the occlusion state recovery result, segment order discrimination result, other burner response prediction result, and furnace response reconstruction result, respectively. The output results of each task are compared with the corresponding labels item by item, and the mean square error is used to calculate the state recovery bias, response prediction bias, and furnace reconstruction bias, respectively. Cross-entropy loss is used to calculate the error. Temporal discrimination bias is defined, with the original temporal sequence samples weighted at 5 and non-original temporal sequence samples weighted at 1. The four types of bias are averaged according to their respective sample sizes, and then weighted and aggregated according to the proportion of each type's sample size to the total number of samples in the joint training sample group to form the joint training bias. The joint training bias is then propagated backwards from the output layer to the input layer of the temporal representation network. The gradient of each network parameter relative to the joint training bias is calculated, and an adaptive moment estimation optimization algorithm is used to adjust the connection weights and bias parameters of each layer. The initial learning rate is set to 0.00001 to 0.0. 01. Set the gradient norm limit to 1 to 5, and the maximum number of training rounds to 100 to 500 rounds, until the change in the joint training bias of five consecutive training rounds does not exceed 1% of the joint training bias of the previous training round; input the natural disturbance operation segment data into the trained time series representation network, extract the natural disturbance representation data with feature dimensions of 32 to 128, the cross-burner propagation representation data with feature dimensions of 32 to 128, and the furnace common response representation data with feature dimensions of 16 to 64, and arrange and aggregate them according to burner affiliation and sampling time to generate burner cooperative state representation data.

[0023] In this embodiment, the construction of the burner action propagation chain specifically includes: According to the burner affiliation and sampling time, the burner cooperative state characterization data are continuously reassembled into a sequence, and the analysis order and lag position between burners and between burners and furnace are configured to generate a conditional Granger causality analysis sequence. The continuous sequence reconstruction includes arranging the natural disturbance characterization data and the cross-burner propagation characterization data according to the burner affiliation, and arranging the furnace common response characterization data to the same sampling time. The basic sampling period is set to 0.1 to 10 seconds, and the length of a single continuous sequence is set to 180 to 1200 sampling periods. When the continuous missing length does not exceed 3 sampling periods, it is linearly filled according to the characterization data of the adjacent sampling times before and after the missing position. When the continuous missing length exceeds 3 sampling periods, the corresponding continuous sequence is discarded. The analysis sequence configuration includes selecting two different burners in sequence and swapping their positions, and selecting one burner and the furnace in sequence and swapping their positions. The characterization data in the preceding sequence is used as the input for propagation analysis, and the characterization data in the following sequence is used as the object of response analysis. The cross-burner propagation characterization data of the remaining burners that are not selected as the input for propagation analysis or the object of response analysis are used as the conditional analysis data. When both the input for propagation analysis and the object of response analysis are burners, the common response characterization data of the furnace at the same time is added to the conditional analysis data. The lag position configuration includes moving the sampling position of the propagation analysis input relative to the response analysis object in stages from 1 to 60 sampling periods. The maximum lag position is the smaller value between 60 sampling periods and 10% of the continuous sequence length. The effective number of samples in the continuous sequence is set to 5 to 20 times the maximum lag position, and analysis sequences with continuous sampling time and complete data are retained at each lag position. Based on the conditional Granger causality analysis sequence, the baseline prediction sequence of the response analysis object and the propagation analysis input prediction sequence are constructed respectively, and the state prediction deviations at different lag positions are compared to generate initial propagation causal data; The construction of the baseline prediction sequence involves reading the characterization data of the response analysis object from 1 to the maximum lag position before the current sampling time and arranging them according to the sampling time; The propagation analysis input participation prediction sequence construction includes adding the propagation analysis input characterization data at the corresponding lag position to the baseline prediction sequence; when the propagation analysis input is a burner, adding the natural disturbance characterization data of the corresponding burner and the cross-burner propagation characterization data; when the propagation analysis input is a furnace, adding the furnace common response characterization data, thus forming the propagation analysis input participation prediction sequence at each lag position. The state prediction bias comparison involves inputting the baseline prediction sequence and the propagation analysis input prediction sequence into a vector autoregressive prediction model of the same order, with the model order set from 1 to 60 and consistent with the current lag position. The least squares method is used to solve the model parameters. The first 70% to 90% of the data in the continuous sequence is used to solve the model parameters, and the remaining 10% to 30% of the data is used to calculate the prediction results. When the response analysis object is the burner, the mean square prediction bias is calculated by comparing the baseline prediction results and the input prediction results with the cross-burner propagation characterization data of the burner at the current sampling time. When the response analysis object is the furnace, the mean square prediction bias is calculated by comparing the baseline prediction results and the input prediction results with the furnace common response characterization data at the current sampling time. The variance ratio test is used to compare the two types of mean square prediction bias at each lag position. The difference between the baseline prediction bias and the input prediction bias is divided by the baseline prediction bias, and the result is recorded as the prediction bias reduction magnitude. When the prediction bias reduction magnitude reaches 5% to 20% and meets the statistical significance level of 0.01 to 0.10, the corresponding lag position is retained. The initial propagation causal data generation process includes: recording the direction from the propagation analysis input to the response analysis object at the retained lag position as the propagation direction; selecting the lag position with the largest reduction in prediction deviation from the lag positions that meet the statistical significance level; recording the time length formed by the selected lag position and the basic sampling period as the propagation delay; and binding the propagation analysis input, response analysis object, propagation direction, propagation delay, and reduction in prediction deviation to generate the initial propagation causal data. Based on the initial propagation causal data configuration condition analysis data addition order, after adding the cross-burner propagation characterization data and furnace common response characterization data of the candidate intermediate burners item by item, the state prediction deviation is recalculated, and the three types of propagation effects are determined to generate propagation effect separation data. The configuration of the order in which conditional analysis data is added includes extracting the propagation analysis input and response analysis object from the initial propagation causal data, both of which are the propagation direction of the burners. The propagation analysis input is denoted as the acting burner, and the response analysis object is denoted as the affected burner. The propagation direction of the acting burner pointing to the other burners is selected from the initial propagation causal data, and the corresponding burners are selected as candidate intermediate burners. The intermediate burners are arranged in order of propagation delay relative to the acting burner from shortest to longest, forming the order in which intermediate burners are added. The cross-burner propagation characterization data of one candidate intermediate burner is added at a time. The number of intermediate burners selected is set to 1 to 5. Then, the furnace common response characterization data is added separately for conditional analysis. The recalculation of the state prediction bias includes adding cross-burner propagation characterization data of each candidate intermediate burner to the baseline prediction sequence and the propagation analysis input prediction sequence used in the initial propagation causal data, and adding furnace common response characterization data after completing the analysis of each candidate intermediate burner. The baseline prediction results and the input prediction results after the condition addition are obtained by using a vector autoregression prediction model of the same order as the initial propagation causal data. The mean square prediction bias is calculated with the cross-burner propagation characterization data of the affected burner at the current sampling time. The variance ratio test is used to obtain the significance test probability after the addition of each condition analysis data. The three types of propagation effects are determined by sequentially adding cross-burner propagation characterization data and furnace common response characterization data to each candidate intermediate burner. When the propagation direction after each addition meets the statistical significance level, it is recorded as a direct propagation effect. When the propagation direction after adding a candidate intermediate burner no longer meets the statistical significance level, but both propagation directions meet the statistical significance level, and the difference between the sum of propagation delays and the initial propagation delay does not exceed 1 to 5 sampling periods, it is recorded as an indirect propagation effect. When the propagation direction after adding furnace common response characterization data no longer meets the statistical significance level, but the propagation directions from the acting burner to the furnace and from the furnace to the affected burner both meet the statistical significance level, and the difference between the sum of propagation delays and the initial propagation delay does not exceed 1 to 10 sampling periods, it is recorded as a furnace common disturbance effect. When both of the latter two conditions are met, it is recorded as a furnace common disturbance effect. Based on the propagation direction and propagation delay, the propagation effect separation data are path-connected and propagation order marked to construct a burner effect propagation chain that includes the propagation effect of each burner; The path connection includes taking the source burner of the disturbance as the starting node, filtering the propagation effects that are continuously connected to the source burner along the propagation direction from the propagation effect separation data step by step, connecting the direct propagation effects to the corresponding affected burners, connecting the indirect propagation effects in the order of the acting burner, intermediate burner and affected burner, and connecting the common disturbance effects in the furnace in the order of the acting burner, furnace and affected burner. The propagation order labeling includes calculating the cumulative propagation delay of each affected burner relative to the source burner of the disturbance. The cumulative propagation delay of direct propagation effects is the corresponding propagation delay, the cumulative propagation delay of indirect propagation effects is the sum of the propagation delays of each segment, and the cumulative propagation delay of the common disturbance effect in the furnace is the sum of the propagation delays from the acting burner to the furnace and from the furnace to the affected burner. The affected burners are arranged in ascending order of cumulative propagation delay to generate a burner action propagation chain that includes the source burner of the disturbance, the affected burner, the propagation order, the propagation direction, the propagation delay, and the type of propagation effect.

[0024] In this embodiment, the generation of group oscillation mode data specifically includes: Based on the burner action propagation chain, the operating states of multiple burners and furnace are extracted from the basic burner dataset and synchronously arranged according to the operating time and propagation order to construct a unified dynamic state sequence. The operation status extraction includes taking the disturbance source burner in the burner action propagation chain as the starting burner, reading each affected burner in sequence according to the propagation direction, extracting the disturbance source burner, the burner operation data of each affected burner in the same operating period and the furnace operation data at the same time from the burner basic dataset, and forming multi-burner operation status and furnace operation status according to burner affiliation and sampling time. Synchronous arrangement includes matching the operating status of each burner and furnace according to a basic sampling period of 0.1 to 10 seconds, grouping operating statuses with sampling time deviations not exceeding half of the basic sampling period into the same sampling time, placing the source of disturbance burner first, arranging each affected burner in the order of propagation, and arranging the furnace operating status after each burner operating status; combining the multiple burner operating statuses and furnace operating statuses arranged within a single sampling time into a unified dynamic state, and continuously arranging each unified dynamic state within 180 to 1200 sampling periods according to the operating time to construct a unified dynamic state sequence; The unified dynamic state sequence is extended by delay coordinates and combined with state changes. The extended operating states are then aggregated according to the propagation order in the burner action propagation chain to generate the Koopman observable state sequence. The delay coordinate extension includes extracting the unified dynamic state of the current sampling time and the 2 to 20 previous delayed sampling positions, setting the interval between adjacent delayed sampling positions to 1 to 5 sampling periods, and arranging the unified dynamic state of each delayed sampling position in order of running time. The state change combination includes calculating the state change of each burner operating state and furnace operating state between adjacent sampling times, and calculating the moving average and moving standard deviation over 5 to 30 consecutive sampling periods; combining the unified dynamic state at the current sampling time, the unified dynamic state at each delayed sampling position, the state change, the moving average, and the moving standard deviation according to the arrangement of the disturbance source burner, the affected burner, and the furnace, and discarding the combination results with missing operating states within the delay range to generate the Koopman observable state sequence; The Koopman observable state sequence is paired with preceding and following state snapshots at a fixed snapshot interval, and the paired preceding and following state snapshots are arranged according to the sampling time to generate state snapshot matrix data. The pairing of before and after state snapshots involves taking the observable state at the previous sampling position in the Koopman observable state sequence as the before state snapshot, taking the observable state after an interval of 1 to 5 sampling periods as the after state snapshot, and retaining the pairing results where the sampling time is continuous and the running phase is the same between the before and after state snapshots. The state snapshot matrix data includes a pre-state snapshot matrix formed by arranging each pre-state snapshot in sequence according to the sampling time, and a post-state snapshot matrix formed by arranging each post-state snapshot in sequence according to the same pairing order. The number of snapshots in the pre-state snapshot matrix and the post-state snapshot matrix is ​​consistent and does not exceed the number of valid pairing results. Singular value decomposition and effective order truncation are performed on the state snapshot matrix data, and the low-order Koopman evolution operator is obtained based on the evolution mapping between the truncated state snapshots before and after truncation, generating Koopman operator decomposition data; Singular value decomposition involves performing singular value decomposition on the previous state snapshot matrix to obtain the left singular vector, singular values, and right singular vector, and then arranging them from high to low according to the energy contribution represented by each singular value. Effective order truncation includes retaining left singular vectors, singular values, and right singular vectors whose cumulative energy contribution reaches 90% to 99% from the permutation results. The number of vectors retained is determined based on the cumulative energy contribution and is limited to 5 to 100, and does not exceed the actual rank of the previous state snapshot matrix. If the actual rank of the previous state snapshot matrix is ​​less than 5, it is retained according to the actual rank. The process of obtaining the low-order Koopman evolution operator involves constructing a low-dimensional state projection basis using the retained left singular vectors, combining the post-state snapshot matrix with the retained right singular vectors, scaling according to the reciprocal of the retained singular values, and then projecting using the transpose of the low-dimensional state projection basis to form a low-order Koopman evolution operator that evolves from the low-dimensional pre-state snapshot to the low-dimensional post-state snapshot. The Koopman operator decomposition data includes the retained left singular vectors, retained singular values, retained right singular vectors, and low-order Koopman evolution operators; Based on the Koopman operator decomposition data and state snapshot matrix data, the low-order Koopman evolution operator is eigenvalued and the decomposition results are mapped to the operating states of each burner and furnace in the unified dynamic state sequence to generate dynamic modal parameter data. Dynamic mode formation involves eigenvalue decomposition of the low-order Koopman evolution operator to obtain eigenvalues ​​and low-dimensional eigenvectors. The post-state snapshot matrix, the retained right singular vector, the reciprocals of the retained singular values, and the low-dimensional eigenvectors are combined according to the decomposition order. Each low-dimensional eigenvector is mapped to the state dimension of the Koopman observable state sequence. The original operating state component corresponding to the current sampling time is extracted from the mapping result to form dynamic modes corresponding to each burner operating state and furnace operating state. Dynamic modes with conjugate eigenvalues ​​are paired, and dynamic modes with positive oscillation frequencies are retained based on the oscillation frequency calculation results during the dynamic mode parameter data generation process. The generation of dynamic modal parameter data includes: calculating the oscillation frequency based on the phase change of each eigenvalue and the time interval between previous and subsequent state snapshots; performing least-squares decomposition on each dynamic mode using the first previous state snapshot; using the absolute value of each obtained modal coefficient as the oscillation amplitude of the corresponding dynamic mode; summing the squares of the component amplitudes of each operating state of the same burner in the corresponding dynamic mode and then taking the square root to generate the modal participation amplitude of each burner; weighting and synthesizing the complex components of each operating state of the same burner in the corresponding dynamic mode according to the component amplitudes; calculating the modal phase of each burner based on the real and imaginary parts of the synthesis result; and generating the furnace modal phase in the same way; taking the smaller value of the circumferential angle between any two modal phases; retaining dynamic modes with oscillation frequencies higher than the lowest resolvable frequency of the unified dynamic state sequence and not exceeding half the sampling frequency corresponding to the basic sampling period; and generating dynamic modal parameter data including oscillation frequency, oscillation amplitude, modal participation amplitude of each burner, modal phase of each burner, and furnace modal phase. Based on the participation amplitude of each burner mode, the phase of each burner mode, and the phase of the furnace mode in the dynamic modal parameter data, the dynamic modes are classified into group types to generate group oscillation mode data; Group classification includes common oscillation mode determination, reverse oscillation mode determination, and local abnormal mode determination. Burners whose modal participation amplitude is not lower than the average modal participation amplitude of all burners are categorized as participating burners. When the number of participating burners accounts for more than 60% of the total number of burners in the burner action propagation chain, and the modal phase difference between each participating burner does not exceed 30 degrees, and the modal phase difference between the furnace modal phase and the modal phase of each participating burner does not exceed 30 degrees, the corresponding dynamic mode is categorized as a common oscillation mode. Participating burners within a group whose modal phase difference does not exceed 30 degrees are classified into the same burner group. When two burner groups are formed, and the modal phase difference between any two participating burners belonging to the two burner groups does not exceed 30 degrees... When the position difference is between 150 and 180 degrees, the corresponding dynamic mode is recorded as the reverse oscillation mode. When the sum of the mode participation amplitudes of one or two burners with the largest mode participation amplitude accounts for more than 60% of the sum of the mode participation amplitudes of all burners, and one or two burners do not include all participating burners, the corresponding dynamic mode is recorded as the local abnormal mode. The group type is divided according to the order of local abnormal mode, reverse oscillation mode, and common oscillation mode. Dynamic modes with a determined group type are not divided again. The oscillation frequency, oscillation amplitude, mode participation amplitude of each burner, mode phase of each burner, and furnace mode phase of the common oscillation mode, reverse oscillation mode, local abnormal mode, and each group oscillation mode are bound to generate group oscillation mode data.

[0025] In this embodiment, the generation of abnormal oscillation location data specifically includes: Read the group oscillation mode data, perform anomaly screening for each group oscillation mode according to the group type and oscillation amplitude, and retain the corresponding oscillation frequency, oscillation amplitude, amplitude of each burner mode participation, phase of each burner mode and phase of furnace mode to generate abnormal group oscillation mode data; Anomaly screening includes arranging each group of oscillation modes from high to low according to the oscillation amplitude, and recording the reverse oscillation mode, local abnormal mode, and common oscillation mode whose oscillation amplitude ranks in the top 10% to 30% of all group oscillation modes as abnormal group oscillation modes, and binding each abnormal group oscillation mode with the corresponding burner mode participation amplitude, burner mode phase and furnace mode phase; Based on the abnormal population oscillation mode data, the participation degree of each burner in the corresponding abnormal population oscillation mode is calculated, and the phase of each burner mode and the phase of the furnace mode are compared in two and two directions to generate abnormal mode participation analysis data. The participation level calculation includes dividing the modal participation amplitude of a single burner by the sum of the modal participation amplitudes of all burners in the corresponding anomalous group oscillation mode, recording the resulting ratio as the participation level of the corresponding burner, and arranging the burners from high to low participation level. Circumferential comparison includes calculating the forward and reverse angles of the modal phases between each burner and between each burner and the furnace. The smaller of the two angles is recorded as the phase difference. The phase lead and phase lag order between each burner and between each burner and the furnace are recorded according to the sequential position of the modal phases along the oscillation period. The abnormal mode participation analysis data includes abnormal group oscillation modes, burner attribution, participation level, phase difference, phase lead order, and phase lag order; By combining abnormal mode participation analysis data and burner action propagation chain, the initiating burner, dominant participating burner, associated burner and oscillation propagation path are determined and abnormal oscillation path location data is generated. The determination of the starting burner includes selecting the burner that has the highest propagation order in the burner action propagation chain and whose phase advance order is located before the downstream burner from burners with a participation level not lower than the average participation level of all burners. When there are multiple selection results, the burner with the highest participation level is recorded as the starting burner. When there is no burner that meets the conditions at the same time, the burner with the highest propagation order and the highest participation level is recorded as the starting burner. The determination of the dominant burners involves selecting no more than three burners from the burners other than the initial burner, in descending order of their participation level. The selection stops when the cumulative participation level reaches 60%. If the cumulative participation level is still below 60% after selecting three burners, the three selected burners are recorded as the dominant burners. The identification of associated burners includes screening burners from burners other than the initiating burner and the dominant participating burner. These burners have a participation rate of 5% to 20%, are continuously connected to the initiating burner along the burner action propagation chain, and are arranged in sequence along the propagation direction in terms of phase lag order. The selected burners are recorded as associated burners. The determination of the oscillation propagation path includes taking the initiating burner as the starting point of the path, sequentially screening the propagation effects of the dominant and associated burners according to the propagation direction and order in the burner action propagation chain, and retaining the furnace involved in the corresponding propagation effects; multiplying the oscillation frequency of the abnormal group oscillation mode with the propagation delay between adjacent propagation nodes, and obtaining the initial propagation phase delay according to the conversion relationship of 360 degrees corresponding to one complete oscillation cycle, taking the remainder after dividing the initial propagation phase delay by 360 degrees, and recording the smaller value between the remainder and the angle obtained by subtracting the remainder from 360 degrees as the propagation phase delay; when the difference between the propagation phase delay and the phase difference between adjacent propagation nodes does not exceed 15 degrees to 45 degrees, the corresponding propagation effect is retained to form an oscillation propagation path; binding the initiating burner, dominant and associated burners, and oscillation propagation path to generate abnormal oscillation path location data; The abnormal group oscillation mode data, abnormal mode participation analysis data, and abnormal oscillation path location data are collected according to the abnormal group oscillation mode, and the participation degree and phase difference of each burner are sorted out according to the oscillation propagation path to generate abnormal oscillation location data. Data collection includes using the starting burner as the modal phase comparison benchmark, reading the phase difference between the dominant participating burner and the associated burner relative to the starting burner, and arranging each burner according to the propagation order in the oscillation propagation path; Abnormal oscillation localization data includes abnormal population oscillation modes, initiating burner, dominant participating burner, associated burner, oscillation propagation path, degree of participation, and phase difference.

[0026] In this embodiment, the generation of collaborative operation optimization results specifically includes: Read abnormal oscillation location data, extract associated operating status from the burner basic dataset based on associated burners, abnormal group oscillation modes and their oscillation frequencies, and filter adjustment objects to generate adjustment object configuration data; The associated operating status extraction includes extracting the burner operating data and concurrent furnace operating data of each associated burner according to the operating time segment corresponding to the abnormal group oscillation mode, and arranging them according to burner affiliation and sampling time; the operating time segment corresponding to the abnormal group oscillation mode takes the current sampling time as the end point and extracts 3 to 10 complete oscillation cycles corresponding to the oscillation frequency. The selection of control targets involves extracting operating state components with the same oscillation frequency as the abnormal group oscillation mode from the fuel flow, primary air volume, secondary air volume, and burner load of each associated burner, and calculating the oscillation amplitude of each operating state component. When the average fluctuation amplitude of the corresponding operating state in the operating segment before the disturbance is 0, the corresponding operating state component is excluded from the selection range of control targets. The oscillation amplitude of the remaining operating state components is divided by the average fluctuation amplitude of the corresponding operating state in the operating segment before the disturbance, and multiplied by the participation degree of the corresponding associated burner. The result is recorded as the contribution degree of the control target. The operating state components are arranged from high to low according to the contribution degree of the control target, and the operating state components with the top 1 to 2 contribution degrees of the control target are selected as control targets from each associated burner. The configuration data for the control object includes associated burner, control object, control object contribution, synchronous oscillation amplitude, current operating status, and control boundary. The control boundary uses the minimum and maximum values ​​of the control object within the operating segment before the disturbance as the lower and upper limits of the control, respectively. Based on the configuration data of the regulating object and the participation level and phase difference of each associated burner, the current phase position of the abnormal group oscillation mode is calculated, and the regulation direction and regulation amount are determined to generate reverse regulation parameter data. The determination of the current phase position includes reading the phase of the associated burner mode from the abnormal group oscillation mode, multiplying the oscillation frequency by the time length from the current moment to the first sampling moment of the unified dynamic state sequence corresponding to the abnormal group oscillation mode, converting it into an angle, adding it to the phase of the associated burner mode, and taking the remainder after dividing by 360 degrees as the current phase position; when the current phase position is not greater than 180 degrees, the current oscillation offset state is recorded as a positive offset state, and when the current phase position is greater than 180 degrees, the current oscillation offset state is recorded as a negative offset state; The adjustment direction determination includes setting the adjustment direction to the decreasing direction of the running state when the current oscillation offset state is a positive offset state, and setting the adjustment direction to the increasing direction of the running state when the current oscillation offset state is a negative offset state. Determining the adjustment amount involves multiplying the amplitude of the same-frequency oscillation of the object being adjusted by the contribution of the object being adjusted, scaling it according to an adjustment ratio of 0.2 to 0.8, and using the resulting value as the initial adjustment amount; limiting the single adjustment amount to 1% to 10% of the average operating state of the corresponding object being adjusted in the segment before the disturbance, and limiting the adjusted operating state within the adjustment boundary according to the adjustment direction; The reverse adjustment parameter data are arranged in reverse order according to the oscillation propagation path, and the adjustment interval of adjacent associated burners is configured according to the propagation delay to generate a multi-burner reverse coordinated adjustment sequence. The reverse order arrangement includes taking the last associated burner in the propagation sequence of the oscillation propagation path as the first regulating burner, and arranging the associated burners in the opposite direction of the propagation direction; when the oscillation propagation path contains multiple branches, the different branches are arranged from long to short according to the cumulative propagation delay of each associated burner relative to the starting burner; when the cumulative propagation delay is the same, the corresponding associated burners are arranged from high to low according to the degree of participation. The adjustment interval configuration includes reading the cumulative value of the propagation delay of each segment between adjacent associated burners in the oscillation propagation path, using 0.5 to 1.5 times the cumulative value as the adjustment interval between adjacent associated burners, and limiting the adjustment interval to 1 to 60 seconds; associated burners in the same propagation order and with a phase difference of no more than 30 degrees are configured with a synchronous adjustment mark; The adjustment is implemented step by step according to the multi-burner reverse coordinated adjustment sequence, and the adjusted burner operating status, furnace operating status and group oscillation status are continuously verified to generate coordinated operation optimization results. The coordinated operation optimization results include the adjusted burner operating status, the adjusted furnace operating status and the adjusted group oscillation status. Step-by-step adjustment involves reading the adjustment object, adjustment direction, and adjustment amount of each associated burner according to the adjustment sequence, dividing a single adjustment amount into 3 to 10 consecutive adjustment steps, and applying each adjustment step step step by step according to the basic sampling period. Each associated burner enters the corresponding adjustment time according to the adjustment interval; associated burners with synchronous adjustment marks implement corresponding adjustments at the same adjustment time. Continuous verification involves collecting burner and furnace operating status data for 30 to 120 consecutive sampling periods after each adjustment, constructing a unified dynamic state sequence after adjustment, and extracting the adjusted group oscillation state according to the generation process of the group oscillation mode data. When the oscillation amplitude of the abnormal group oscillation mode decreases by no less than 30% compared to before adjustment, and the furnace operating status continuously enters the fluctuation range before the disturbance, the reverse coordinated adjustment ends. When the oscillation amplitude of the abnormal group oscillation mode increases, or the furnace operating status continuously exceeds the fluctuation range before the disturbance, the most recent adjustment step is canceled, and the subsequent adjustment amount is reduced by 20% to 50%. The adjusted group oscillation state includes the oscillation frequency, oscillation amplitude, participation amplitude of each burner mode, and phase of each burner mode of the abnormal group oscillation mode.

[0027] Example 1: To verify the feasibility of the invention in practice, it was applied to a data playback scenario of coordinated operation of six burners in a coal-fired boiler. When the boiler load changes, the fuel pressure, air volume, and load fluctuations of a single burner may be transmitted to other burners through the furnace, forming synchronous oscillations and phase misalignment. Traditional single-burner sequential adjustment makes it difficult to identify the source and propagation path of the oscillations, which easily leads to adjustment lag and secondary fluctuations.

[0028] Using a 1-second sampling period, fuel pressure, fuel flow, primary air volume, secondary air volume, burner load, and flame temperature were collected for each burner, along with furnace pressure, furnace temperature, and flue gas oxygen content. 1200 sets of natural disturbance playback samples were constructed, including 400 sets each for fuel pressure fluctuations, air volume changes, and load switching. Each set contained 60 sampling periods before the disturbance, 8 to 20 sampling periods during the disturbance, and 180 sampling periods after the disturbance. These samples were divided into 960 training samples, 120 validation samples, and 120 test samples in an 8:1:1 ratio. 240 consecutive sampling periods were extracted from each combined sequence for time-series representation learning.

[0029] For the continuous running sequence, a masking interval of 5 sampling times was set. With a moving step size of 2 sampling periods, 100 sets of burner state masking recovery samples were evenly selected for each group. At the same time, perturbation time series discrimination samples, cross-burner response prediction samples, and furnace response reconstruction samples were generated. The time series representation network adopted a 2-layer bidirectional long short-term memory layer and 4 time series attention heads. 64 sets of samples were input in each batch, and the initial learning rate was set to 0.0001.

[0030] The burner coordinated state characterization data were reorganized according to burner affiliation and sampling time. Conditional Granger causality analysis was performed on the lag positions of 1 to 30 sampling periods to form the burner action propagation chain. Operating states for 600 consecutive sampling periods were selected, with 5 delayed sampling positions, a 2-sampling-per-cycle delay interval, and a 2-sampling-per-cycle snapshot interval. The group oscillation mode was extracted, and the initiating burner, dominant participating burner, associated burner, and oscillation propagation path were located. Fuel flow rate, primary air volume, secondary air volume, and burner load were selected based on the contribution of the regulated objects. The regulation ratio was set to 0.5, and the single regulation amount was limited to 5% of the average operating state before the disturbance, divided into 5 consecutive regulation steps. Reverse coordinated regulation was implemented in reverse order of the oscillation propagation path. After each regulation, data for 180 consecutive sampling periods were collected to verify the group oscillation state.

[0031] To ensure consistent comparison conditions, 120 test samples were input into both the traditional single-burner sequential adjustment method and the method of this invention, respectively, and playback calculations were performed using the same sampling period, operating status, disturbance input, adjustment boundary, and statistical time period. The source burner of the disturbance and the oscillation propagation path were verified based on the operating records and control logs.

[0032] Table 1. Performance Comparison of the Invention and the Traditional Single-Burner Sequential Adjustment Method

[0033] The accuracy of abnormal oscillation localization using the method of this invention is improved from 86.7% to 91.7%, an increase of 5.0 percentage points; the accuracy of oscillation propagation path identification is improved from 84.2% to 90.0%, an increase of 5.8 percentage points. This is because the invention combines burner modal participation amplitude, modal phase, propagation direction, and propagation delay to determine the initiating burner and associated burners, and separates the direct propagation effect, indirect propagation effect, and furnace common disturbance effect, reducing the interference of the furnace common response on the determination of the oscillation source.

[0034] The average adjustment response time was reduced from 38.7s to 34.6s, a reduction of 4.1s. This change is due to the arrangement of associated burners in reverse order of the oscillation propagation path, and the configuration of adjustment intervals in conjunction with the propagation delay. This avoids the traditional method of probing one burner at a time and repeatedly waiting, allowing different associated burners to enter the corresponding adjustment time according to the propagation relationship.

[0035] The standard deviation of furnace pressure fluctuation decreased from 42.6 Pa to 39.1 Pa, a reduction of 3.5 Pa; the standard deviation of flue gas oxygen content fluctuation decreased from 0.38% to 0.34%, a reduction of 0.04 percentage points. This is because the present invention selects only the operating state component with the same oscillation frequency as the abnormal group oscillation mode as the adjustment object, and divides the adjustment amount into 5 continuous adjustment steps, reducing the instantaneous impact of multiple burner adjustments on the furnace state.

[0036] The abnormal oscillation amplitude reduction rate increased from 31.7% to 35.9%, an increase of 4.2 percentage points; the regulation overshoot rate decreased from 8.8% to 7.9%, a decrease of 0.9 percentage points. This performance change is due to the regulation direction being set in the opposite direction to the current phase position, the single regulation amount being limited to 5% of the average operating state before the disturbance, and 180 consecutive sampling cycles being collected for verification after each regulation, thereby reducing the overshoot caused by the superposition of the regulation direction and the current oscillation offset.

[0037] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A combustor cooperative operation optimization method based on self-supervised learning, characterized in that, Includes the following steps: Acquire burner operation data, furnace operation data, and natural disturbance records during the multi-burner collaborative operation process, and perform preprocessing to generate a basic burner dataset; Natural operating disturbances and their associated operating states are extracted from the burner basic dataset and divided into continuous segments to generate natural disturbance operating segment data; Based on the data of natural disturbance operation segments, a self-supervised learning task is configured to perform joint time-series characterization learning on the operating status of each burner and the furnace response status, generating burner collaborative state characterization data. Based on the burner cooperative state characterization data, conditional Granger causality analysis was used to determine the propagation direction and propagation delay of natural operating disturbances between each burner and the furnace, and the propagation effects were separated to construct the burner effect propagation chain; Based on the burner action propagation chain and burner basic dataset, a unified dynamic state sequence consisting of multiple burner operating states and furnace operating states is constructed. Based on the Koopman operator, the group oscillation mode in the unified dynamic state sequence is extracted to generate group oscillation mode data. Abnormal group oscillation modes are identified based on the group oscillation mode data, and the participation degree and phase difference of each burner in the abnormal group oscillation mode are extracted by combining the burner action propagation chain to generate abnormal oscillation location data; Based on abnormal oscillation location data, the adjustment parameters of the associated burners are determined and multi-burner reverse coordinated adjustment is implemented to generate coordinated operation optimization results.

2. The combustor cooperative operation optimization method based on self-supervised learning according to claim 1, characterized in that, The preprocessing includes time alignment, burner attribution marking, operation phase division, collection anomaly removal, and data normalization.

3. The combustor cooperative operation optimization method based on self-supervised learning according to claim 1, characterized in that, The generation of the natural disturbance operation segment data specifically includes: Natural disturbance record data is read from the burner basic dataset. Disturbance information is marked according to disturbance record category, burner affiliation mark, and continuous recording time range to generate natural disturbance mark data. The natural disturbance mark data includes disturbance type, disturbance source burner, disturbance start time, and disturbance end time. Using the source burner of the disturbance, the start time of the disturbance, and the end time of the disturbance in the natural disturbance marker data as search criteria, the associated operating states within the corresponding operating time segment are extracted from the burner basic dataset to generate natural disturbance associated operating state data; Read the start and end times of disturbances and other burner and furnace response states from the natural disturbance-related operating status data, determine the start and recovery times of the state response, and generate segment boundary data. The natural disturbance-related operational state data is divided into continuous segments according to the segment boundary data. The start time of the state response is marked in the continuous operational segment, and the natural disturbance marking data is bound to each continuous segment to generate natural disturbance operational segment data. The natural disturbance operational segment data includes the operational segment before the disturbance, the segment in which the disturbance occurs, and the operational segment after the disturbance.

4. The combustor cooperative operation optimization method based on self-supervised learning according to claim 1, characterized in that, The generation of the burner cooperative state characterization data specifically includes: Read the natural disturbance operation segment data, set the masking position for each burner operation state, and construct model input and state recovery labels using the unmasked operation state and the original operation state respectively to generate burner state masking recovery samples; Adjust the order of the running segments before the disturbance, the segment where the disturbance occurs, and the running segments after the disturbance, and use the original time order of each running segment as the time series discrimination label to generate disturbance time series discrimination samples; The operating status of the source burner of the disturbance is used as the prediction input, and the response status of other burners and the furnace response status are used as the prediction labels to generate cross-burner response prediction samples and furnace response reconstruction samples. The burner state masking recovery samples, disturbance time series discrimination samples, cross-burner response prediction samples, and furnace response reconstruction samples are jointly input into the time series representation network. The network parameters are jointly updated with the results of their respective tasks. The time series features obtained from the training are aggregated according to burner affiliation and sampling time to generate burner collaborative state representation data.

5. The combustor cooperative operation optimization method based on self-supervised learning according to claim 1, characterized in that, The construction of the burner action propagation chain specifically includes: According to the burner affiliation and sampling time, the burner cooperative state characterization data are continuously reassembled into a sequence, and the analysis order and lag position between burners and between burners and furnace are configured to generate a conditional Granger causality analysis sequence. Based on the conditional Granger causality analysis sequence, the baseline prediction sequence of the response analysis object and the propagation analysis input prediction sequence are constructed respectively, and the state prediction deviations at different lag positions are compared to generate initial propagation causal data; Based on the initial propagation causal data configuration condition analysis data addition order, after adding the cross-burner propagation characterization data and furnace common response characterization data of the candidate intermediate burners item by item, the state prediction deviation is recalculated, and the three types of propagation effects are determined to generate propagation effect separation data. Based on the propagation direction and propagation delay, the propagation effect separation data are path-connected and propagation order marked to construct a burner effect propagation chain that includes the propagation effects of each burner.

6. The combustor cooperative operation optimization method based on self-supervised learning according to claim 1, characterized in that, The generation of the group oscillation mode data specifically includes: Based on the burner action propagation chain, the operating states of multiple burners and furnace are extracted from the basic burner dataset and synchronously arranged according to the operating time and propagation order to construct a unified dynamic state sequence. The unified dynamic state sequence is extended by delay coordinates and combined with state changes. The extended operating states are then aggregated according to the propagation order in the burner action propagation chain to generate the Koopman observable state sequence. The Koopman observable state sequence is paired with preceding and following state snapshots at a fixed snapshot interval, and the paired preceding and following state snapshots are arranged according to the sampling time to generate state snapshot matrix data. Singular value decomposition and effective order truncation are performed on the state snapshot matrix data, and the low-order Koopman evolution operator is obtained based on the evolution mapping between the truncated state snapshots before and after truncation, generating Koopman operator decomposition data; Based on the Koopman operator decomposition data and state snapshot matrix data, the low-order Koopman evolution operator is eigenvalued and the decomposition results are mapped to the operating states of each burner and furnace in the unified dynamic state sequence to generate dynamic modal parameter data. Based on the participation amplitude of each burner mode, the phase of each burner mode, and the phase of the furnace mode in the dynamic modal parameter data, the dynamic modes are classified into group types to generate group oscillation mode data.

7. The combustor cooperative operation optimization method based on self-supervised learning according to claim 1, characterized in that, The generation of the abnormal oscillation location data specifically includes: Read the group oscillation mode data, perform anomaly screening for each group oscillation mode according to the group type and oscillation amplitude, and retain the corresponding oscillation frequency, oscillation amplitude, amplitude of each burner mode participation, phase of each burner mode and phase of furnace mode to generate abnormal group oscillation mode data; Based on the abnormal population oscillation mode data, the participation degree of each burner in the corresponding abnormal population oscillation mode is calculated, and the phase of each burner mode and the phase of the furnace mode are compared in two and two directions to generate abnormal mode participation analysis data. By combining abnormal mode participation analysis data and burner action propagation chain, the initiating burner, dominant participating burner, associated burner and oscillation propagation path are determined and abnormal oscillation path location data is generated. The abnormal population oscillation mode data, abnormal mode participation analysis data, and abnormal oscillation path location data are collected according to the abnormal population oscillation mode, and the participation degree and phase difference of each burner are sorted out according to the oscillation propagation path to generate abnormal oscillation location data.

8. The combustor cooperative operation optimization method based on self-supervised learning according to claim 1, characterized in that, The generation of the collaborative operation optimization results specifically includes: Read abnormal oscillation location data, extract associated operating status from the burner basic dataset based on associated burners, abnormal group oscillation modes and their oscillation frequencies, and filter adjustment objects to generate adjustment object configuration data; Based on the configuration data of the regulating object and the participation level and phase difference of each associated burner, the current phase position of the abnormal group oscillation mode is calculated, and the regulation direction and regulation amount are determined to generate reverse regulation parameter data. The reverse adjustment parameter data are arranged in reverse order according to the oscillation propagation path, and the adjustment interval of adjacent associated burners is configured according to the propagation delay to generate a multi-burner reverse coordinated adjustment sequence. The adjustment is implemented step by step according to the multi-burner reverse coordinated adjustment sequence, and the adjusted burner operating status, furnace operating status and group oscillation status are continuously verified to generate coordinated operation optimization results. The coordinated operation optimization results include the adjusted burner operating status, the adjusted furnace operating status and the adjusted group oscillation status.

Images