Machine learning-based boiler energy efficiency anomaly early warning method

Abstract

This invention discloses a boiler energy efficiency anomaly early warning method based on machine learning, belonging to the field of machine learning technology. The method includes: reading multi-source data from a SIS boiler, completing time calibration and measurement point encoding; removing bad points and identifying frozen points to generate a reliable boiler operating status frame; identifying operating conditions and control events to form a boiler state sequence bound to operating conditions; constructing an improved MTGNN model to calculate multi-channel energy efficiency residuals; combining control adjustment event windows to remove the influence of control action residuals; performing confidence normalization according to boiler operating conditions to generate confidence residuals and risk residuals; executing the Page-Hinkley algorithm to calculate graph cooperative asymmetric drift; performing energy flow verification and contribution verification, and outputting energy efficiency anomaly early warning results. This invention can identify boiler energy efficiency degradation under conditions where parameters are not exceeded and adjustment actions are not disturbed, improving the accuracy of anomaly early warning and the interpretability of anomaly sources.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method for early warning of boiler energy efficiency anomalies based on machine learning. Background Technology

[0002] With the improvement of thermal power unit operation data acquisition capabilities, boiler energy efficiency monitoring is gradually shifting from manual inspections and periodic report analysis to real-time data acquisition via SIS, trend monitoring via DCS, and online alarms. Existing boiler monitoring methods primarily collect combustion parameters, steam-water parameters, flue gas parameters, auxiliary equipment parameters, environmental parameters, and load parameters. Alarm judgments are based on fixed limits, empirical rules, and changes at single measuring points, reflecting out-of-limit states of main steam temperature, flue gas temperature, furnace negative pressure, oxygen content, fan current, and desuperheating water volume. Some energy consumption analysis platforms incorporate boiler thermal efficiency calculations, coal consumption statistics, and operational indicator benchmarking to demonstrate changes in unit economic efficiency.

[0003] Current technologies for identifying boiler energy efficiency anomalies still have shortcomings. Boiler energy efficiency anomalies often manifest as coupled offsets of multiple measuring points within safe limits, and fixed thresholds are insufficient to identify energy efficiency degradation under conditions that do not exceed limits. Different loads, coal qualities, heating states, and combustion stability states correspond to different reasonable energy efficiency levels, and static benchmarks are difficult to adapt to changes in operating conditions. SIS data contains bad pixels, frozen points, supplementary data, and incomplete sampling, which can introduce false alarms if not processed in conjunction with data quality codes. Short-term residuals generated by operational adjustments are mixed with residuals of actual energy efficiency anomalies, and existing methods lack residual stripping processing. Existing machine learning early warning methods mostly focus on univariate prediction or single anomaly score output, lacking measuring point graph correlation, boiler energy flow constraints, counterfactual energy efficiency benchmark reconstruction, and drift contribution verification, making it difficult to provide anomaly sources and stability warning results.

[0004] Therefore, how to provide a machine learning-based method for early warning of boiler energy efficiency anomalies 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 machine learning-based method for early warning of boiler energy efficiency anomalies. This invention integrates SIS data processing, an improved MTGNN model, and the Page-Hinkley algorithm to complete reliable framing of multi-source operation correlation data of boilers, dynamic energy efficiency benchmark reconstruction, multi-channel energy efficiency residual calculation, graph collaborative asymmetric drift detection, and anomaly contribution verification. It has the advantages of strong early warning capability, low false alarm rate, clear anomaly source, and high engineering adaptability.

[0006] A method for early warning of boiler energy efficiency anomalies based on machine learning according to an embodiment of the present invention includes: Read multi-source operation correlation data of the boiler from the SIS server, perform time calibration and measurement point coding, and generate raw boiler operation data; Perform defect removal, freeze point identification, retransmission marking and sampling alignment on the raw boiler operation data to generate a reliable boiler operation status frame; Based on the trusted boiler operating status frame, identify operating conditions, control and regulation event records, and control and regulation event windows to generate boiler operating condition identification records and operating condition-bound boiler status sequences. An improved MTGNN model is constructed by sequentially inputting the boiler state sequence bound to the operating conditions into the operating condition routing expert graph temporal layer, the boiler energy flow constraint graph convolutional layer, and the counterfactual energy efficiency benchmark reconstruction layer to generate multi-channel energy efficiency residuals. Based on the control and regulation event record and control and regulation event window, control actions are performed to strip the residuals of multi-channel energy efficiency residuals, generating stripped energy efficiency residuals. Based on boiler operating condition identification records, residual confidence normalization is performed on the stripped energy efficiency residuals to generate multi-channel confidence residuals and comprehensive risk residuals. The multi-channel confidence residual and the comprehensive risk residual are input into the Page-Hinkley algorithm, and the graph cooperative asymmetric drift amount is calculated based on the directional asymmetric drift mechanism and the graph association cooperative drift mechanism. Perform energy flow consistency verification and counterfactual contribution verification operations on the graph cooperative asymmetric drift to generate boiler energy efficiency anomaly early warning results.

[0007] Optionally, generating the raw boiler operating data includes: Boiler combustion data, steam and water data, flue gas data, auxiliary machine data, operating condition data, and data quality codes are read through the SIS server read-only interface to generate original records of measurement points; Extract the measurement point name, measurement point number, sampling time, sampled value, and quality identifier to generate a measurement point sampling record; Perform unified time axis calibration on the sampling records of the measurement points based on the sampling time to generate time-calibrated sampling records; Based on the measurement point number, the measurement point code and data source mark are executed to generate the original boiler operation data.

[0008] Optionally, generating a reliable boiler operating status frame includes: Read the measurement point code, sampling time, sampling value and quality identifier from the original boiler operation data, and generate measurement point verification records based on the measurement point code matching measurement point boundary configuration; Based on the quality identifier, sampled value, and measurement point boundary configuration, perform bad point removal to generate valid measurement point records; Freeze point identification is performed based on the repetition status and rate of change of adjacent sampled values ​​in the effective measurement point records, and supplementary transmission marking is performed based on the adjacent sampling time interval to generate a reliable measurement point sequence. A unified sampling time axis is generated based on the sampling time, and sampling alignment is performed on the reliable measurement point sequence to generate a reliable boiler operating status frame.

[0009] Optionally, the generation of boiler operating condition identification records and operating condition-bound boiler state sequences includes: Read the load value, main steam flow, coal quality mark, heating status, combustion parameters and controllable execution quantities from the trusted boiler operating status frame, and generate operating condition identification input records; Based on load value, main steam flow rate, coal quality label and heating status, the operating conditions are segmented and the operating conditions are generated. The combustion stability state is identified based on the fluctuation amplitude and duration of combustion parameters, and the operating conditions and combustion stability state are associated and stored to generate a boiler operating condition identification record; Based on the change magnitude, change direction and occurrence time of the controllable execution quantity, control adjustment event records are identified, and control adjustment event windows are generated according to the occurrence time, sampling period and event duration; The boiler operating condition identification record, control and regulation event record, and control and regulation event window are bound to the trusted boiler operating status frame to generate an operating condition-bound boiler status sequence.

[0010] Optionally, generating multi-channel energy efficiency residuals includes: An improved MTGNN model is constructed, which includes a time-series layer of operating condition routing expert graph, a convolutional layer of boiler energy flow constraint graph, and a counterfactual energy efficiency benchmark reconstruction layer. Read the boiler operating condition identification record, trusted boiler operating status frame and measurement point code from the boiler state sequence bound to the operating condition, and generate measurement point timing embedding and operating condition embedding by inputting the encoding end; The measurement point timing embedding and the operating condition embedding are input into the operating condition routing expert diagram timing layer. The corresponding expert diagram timing branch is selected according to the operating condition to generate the operating condition timing representation. A boiler measurement point map is constructed based on measurement point coding and boiler energy flow relationship. The boiler measurement point map and the time-series representation of the operating condition map are input into the convolutional layer of the boiler energy flow constraint map to generate an energy flow constraint boiler representation. The energy flow-constrained boiler characterization is input into the counterfactual energy efficiency benchmark reconstruction layer. The counterfactual boiler characterization is generated according to the measurement point disturbance shielding and normal state replacement. A multi-channel energy efficiency benchmark is generated through the residual output terminal. Extract actual energy efficiency indicators from the reliable boiler operating status frame, perform channel-by-channel difference calculation between the actual energy efficiency indicators and the multi-channel energy efficiency benchmark, and generate multi-channel energy efficiency residuals. The improved MTGNN model was trained using a combination of energy efficiency benchmark prediction error, graph time series representation reconstruction error, operating condition routing allocation error, energy flow constraint deviation, and counterfactual benchmark reconstruction error as the joint training objective. The network parameters of the operating condition routing expert graph time series layer, the boiler energy flow constraint graph convolutional layer, and the counterfactual energy efficiency benchmark reconstruction layer were updated. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved MTGNN model was considered to have completed convergence training.

[0011] Optionally, the improved MTGNN model includes: The MTGNN model retains the core of adaptive graph learning structure, temporal convolution structure, hybrid hop number graph propagation structure, hop connection path, and prediction output structure. The input representation, operating condition routing, energy flow propagation, and counterfactual reconstruction are structurally modified to address the boiler energy efficiency anomaly early warning scenario. Set up an input encoding terminal at the model input end, read the reliable boiler operating status frame, boiler operating condition identification record and measurement point code in the boiler state sequence bound to the operating condition, and encode the measurement point sample value, measurement point type, sampling time and operating condition into measurement point time sequence embedding and operating condition embedding; A time-series layer of the operating condition routing expert graph is added at the time-series feature extraction location. The time-series branches of the expert graph are selected based on the operating conditions, combustion stability and heating status to form the time-series representation of the operating condition graph. The original graph convolution propagation structure is modified to form a boiler energy flow constraint graph convolution layer. The boiler measurement point graph and boiler energy flow relationship are embedded into the graph propagation path to constrain the energy efficiency coupling relationship between the combustion side, steam-water side, flue gas side and auxiliary machine side. A counterfactual energy efficiency benchmark reconstruction layer is added before the predicted output. The abnormal disturbance channel is shielded and replaced and the normal state is reconstructed to generate a counterfactual boiler characterization. The predicted output structure is transformed into a residual output terminal, outputting a multi-channel energy efficiency benchmark, and the difference between the benchmark and the actual energy efficiency index is calculated to generate a multi-channel energy efficiency residual.

[0012] Optionally, generating the energy efficiency residual after stripping includes: Read the multi-channel energy efficiency residuals, control and regulation event records, and control and regulation event windows; establish a residual event alignment table according to the sampling time; and associate and store each residual channel with the corresponding control and regulation event window. Based on the residual event alignment table, extract the residual peak value, residual start and end time, residual duration, residual change direction and residual channel identifier within the control and adjustment event window to generate window residual feature records; Based on the control and regulation event records, the direction of change, magnitude of change, start and end time of action, and duration of action of controllable execution quantity are extracted. The direction of change of controllable execution quantity is matched with the window residual feature records, and the start and end time of action is matched with the start and end time of residual to generate control-related residual tags. The residual stripping weight is determined based on the control-related residual label and residual duration. Weighted reduction is performed on the residual channels with control-related residual labels in the multi-channel energy efficiency residuals, and the residual channels without matching control-related residual labels are retained to generate stripped energy efficiency residuals.

[0013] Optionally, the generation of multi-channel confidence residuals and comprehensive risk residuals includes: Read the boiler operating condition identification record and the energy efficiency residual after stripping, and generate a residual benchmark matching record based on the operating condition matching operating condition residual benchmark table. Extract the mean of normal residuals, residual dispersion, and confidence interval boundaries for each residual channel from the residual benchmark matching record to generate the residual confidence benchmark record; The difference between the stripped energy efficiency residual and the mean of the normal residual is calculated, and the difference result is divided by the sum of the residual dispersion and the stability constant to generate multi-channel confidence residuals. Based on the boiler operating condition identification records, the energy efficiency weights of each residual channel are determined, and weighted fusion is performed on the confidence residuals of multiple channels to generate comprehensive risk residuals.

[0014] Optionally, the computational graph cooperative asymmetric drift includes: Read the multi-channel confidence residuals, comprehensive risk residuals, and boiler measuring point correlation matrix; extract the residual channel identifiers from the multi-channel confidence residuals; and establish a drift detection input record table according to sampling time, sampling period, residual channel identifiers, and measuring point codes. Based on the residual channel identifier matching offset weight configuration table, read the forward offset weight and reverse offset weight, perform directional asymmetric weighting on the multi-channel confidence residuals, and generate directional weighted confidence residuals; Input the direction-weighted confidence residual and the comprehensive risk residual into the Page-Hinkley algorithm to calculate the residual running mean, cumulative offset, historical minimum cumulative offset and single-channel drift statistic for each channel. Based on the boiler measuring point correlation matrix and measuring point coding, graph correlation propagation is performed on the drift statistics of single channels, and the drift statistics of the correlation residual channels are fused according to the measuring point correlation strength to generate channel collaborative drift statistics; Based on the boiler operating condition identification record and the residual channel identifier matching operating condition energy efficiency weight table, the channel collaborative drift statistics and the drift statistics corresponding to the comprehensive risk residual are weighted and fused to generate the graph collaborative asymmetric drift quantity and the complete drift statistical process record table.

[0015] Optionally, the generation of boiler energy efficiency anomaly early warning results includes: Read the graph collaborative asymmetric drift amount, complete drift statistical process record table, boiler operating condition identification record and multi-channel confidence residual, and generate an energy flow direction verification table based on the boiler operating condition identification record; Based on the energy flow direction verification table, an energy flow consistency verification operation is performed on the graph cooperative asymmetric drift, generating the verified drift and energy flow consistency flag; Based on the boiler operating condition identification record and the drift threshold record, the verified drift amount is compared with the drift threshold record to generate an energy efficiency anomaly trigger record; Based on the energy efficiency anomaly trigger record and the complete drift statistical process record table, perform counterfactual contribution verification operation, perform channel replacement and drift recalculation on the multi-channel confidence residuals, and generate an anomaly contribution ranking table. Boiler energy efficiency anomaly early warning results are generated based on energy efficiency anomaly trigger records, energy flow consistency markers, and anomaly contribution sorting tables.

[0016] The beneficial effects of this invention are: This invention proposes a machine learning-based early warning method for boiler energy efficiency anomalies. By constructing a multi-source boiler operation correlation data acquisition process, generating reliable boiler operating status frames, identifying operating conditions, reconstructing the energy efficiency benchmark using an improved MTGNN model, and calculating multi-channel energy efficiency residuals, the method improves the adaptability of boiler energy efficiency anomaly identification to different operating conditions. The improved MTGNN model introduces a time-series layer of operating condition routing expert graphs, a convolutional layer of boiler energy flow constraint graphs, and a counterfactual energy efficiency benchmark reconstruction layer. This enables the establishment of dynamic energy efficiency benchmarks under varying load, coal quality, heating conditions, and combustion stability conditions, reducing false alarms and missed alarms in complex operating conditions caused by fixed threshold alarms.

[0017] This invention provides stable early warning and anomaly source location for boiler energy efficiency residuals through control action residual stripping, residual confidence normalization, directional asymmetric drift detection, graph association collaborative drift detection, energy flow consistency verification, and counterfactual contribution verification. The Page-Hinkley algorithm, combined with measurement point graph association and energy efficiency risk direction, can identify continuous deviations of multiple measurement points in a non-limited state, reducing misjudgments caused by control adjustment actions, bad data points, and single-measurement point fluctuations. It outputs boiler energy efficiency anomaly early warning results, anomaly channels, and contribution ranking, possessing advantages such as strong early warning capability, clear anomaly location, and high field adaptability. Attached Figure Description

[0018] 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 1This is a flowchart of a machine learning-based early warning method for boiler energy efficiency anomalies proposed in this invention. Figure 2 This is a schematic diagram of the improved MTGNN model in the machine learning-based boiler energy efficiency anomaly early warning method proposed in this invention. Figure 3 The flowchart shows the graph collaborative asymmetric drift algorithm in the boiler energy efficiency anomaly early warning method based on machine learning proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A machine learning-based method for early warning of boiler energy efficiency anomalies includes: Read multi-source operation correlation data of the boiler from the SIS server, perform time calibration and measurement point coding, and generate raw boiler operation data; Perform defect removal, freeze point identification, retransmission marking and sampling alignment on the raw boiler operation data to generate a reliable boiler operation status frame; Based on the trusted boiler operating status frame, identify operating conditions, control and regulation event records, and control and regulation event windows to generate boiler operating condition identification records and operating condition-bound boiler status sequences. An improved MTGNN model is constructed by sequentially inputting the boiler state sequence bound to the operating conditions into the operating condition routing expert graph temporal layer, the boiler energy flow constraint graph convolutional layer, and the counterfactual energy efficiency benchmark reconstruction layer to generate multi-channel energy efficiency residuals. Based on the control and regulation event record and control and regulation event window, control actions are performed to strip the residuals of multi-channel energy efficiency residuals, generating stripped energy efficiency residuals. Based on boiler operating condition identification records, residual confidence normalization is performed on the stripped energy efficiency residuals to generate multi-channel confidence residuals and comprehensive risk residuals. The multi-channel confidence residual and the comprehensive risk residual are input into the Page-Hinkley algorithm, and the graph cooperative asymmetric drift amount is calculated based on the directional asymmetric drift mechanism and the graph association cooperative drift mechanism. Perform energy flow consistency verification and counterfactual contribution verification operations on the graph cooperative asymmetric drift to generate boiler energy efficiency anomaly early warning results.

[0021] In this embodiment, generating the raw boiler operating data includes: Boiler combustion data, steam and water data, flue gas data, auxiliary machine data, operating condition data, and data quality codes are read through the SIS server read-only interface to generate original records of measurement points; Extract the measurement point name, measurement point number, sampling time, sampled value, and quality identifier to generate a measurement point sampling record; Perform unified time axis calibration on the sampling records of the measurement points based on the sampling time to generate time-calibrated sampling records; Based on the measuring point number, measuring point coding and data source marking are performed to generate raw boiler operation data, including: The raw boiler operating data is generated as follows: Read the measurement point number, measurement point name, sampling time, sample value, quality identifier, and sampling period from the time calibration sampling record; A boiler measurement point configuration table is established based on the SIS measurement point list, boiler system classification configuration, measurement point boundary configuration, and sampling period configuration. The boiler measurement point configuration table includes measurement point number, SIS variable name, measurement point name, system to which it belongs, measurement point type, sampling period, upper limit of value, lower limit of value, maximum rate of change, minimum rate of change, freeze judgment duration, and missing sampling judgment duration. Each measuring point is assigned a unique measuring point code according to its system, measuring point type, and measuring point number. The measuring point code is then associated with and stored in the field of the boiler measuring point configuration table to generate a measuring point code table. Based on the system to which the sampling records of the measuring points belong, the sampling records of the measuring points belonging to the combustion system, steam-water system, flue gas system, auxiliary machine system and operating condition system are respectively marked as boiler combustion data, boiler steam-water data, boiler flue gas data, boiler auxiliary machine data and boiler operating condition data; The measurement point code, measurement point name, data source mark, sampling time, sampled value, quality mark, sampling cycle, boiler combustion data, boiler steam and water data, boiler flue gas data, boiler auxiliary equipment data, and boiler operating condition data are associated and written into the same operation record to generate the original boiler operation data.

[0022] In this embodiment, generating a reliable boiler operating status frame includes: The system reads the measurement point code, sampling time, sampled value, and quality identifier from the raw boiler operation data, and generates a measurement point verification record based on the measurement point code and the matching of measurement point boundary configurations. Generate measurement point verification records, specifically as follows: Read the measurement point code, measurement point name, data source mark, sampling time, sampling value, quality mark and sampling period from the original boiler operation data, and collect the sampling records of the same measurement point according to the measurement point code to generate a measurement point to be verified sequence; Based on the measurement point code, read the corresponding measurement point type, upper limit of value, lower limit of value, maximum rate of change, minimum rate of change, freeze judgment duration and missing sampling judgment duration from the measurement point code table, and generate the measurement point boundary configuration; Associate the sampled values ​​in the sequence to be verified of the measurement points with the boundary configuration of the measurement points, associate the sampling time with the sampling period, and associate the quality identifier with the measurement point code to generate a measurement point verification field group; Write the measurement point code, measurement point name, data source mark, sampling time, sampled value, quality mark, sampling period, upper limit of value, lower limit of value, maximum rate of change, minimum rate of change, freeze judgment duration and missing sampling judgment duration into the same verification record to generate a measurement point verification record; Based on quality identifiers, sampled values, and measurement point boundary configurations, defective point removal is performed to generate valid measurement point records, where: Defect removal is performed based on quality identifiers, sampled values, and measurement point boundary configurations, specifically as follows: Read the measurement point code, sampling time, sampled value, quality identifier, upper limit of value, lower limit of value, maximum rate of change and minimum rate of change from the measurement point calibration record; Based on the quality identifier, the validity status of the measurement point sampling record is determined, and the measurement point sampling record corresponding to the quality identifier in the invalid quality status, communication interruption status, and manual destruction status is marked as a quality defect point; The sampling records of measurement points with sampled values ​​greater than the upper limit are marked as upper boundary out-of-bounds bad points, and the sampling records of measurement points with sampled values ​​less than the lower limit are marked as lower boundary out-of-bounds bad points. The rate of change of the sampled values ​​is calculated according to the adjacent sampling time sequence corresponding to the same measurement point code. The sampling records of measurement points with a rate of change of the sampled values ​​greater than the maximum rate of change are marked as positive abrupt change bad points, and the sampling records of measurement points with a rate of change of the sampled values ​​less than the minimum rate of change are marked as reverse abrupt change bad points. Remove quality defects, upper boundary defects, lower boundary defects, positive mutation defects, and reverse mutation defects from the measurement point calibration record, and retain the measurement point sampling records that are not marked as defects to generate valid measurement point records; Freeze point identification is performed based on the repetition state and rate of change of adjacent sampled values ​​in the valid measurement point records, and supplementary marking is performed based on the time interval between adjacent samplings to generate a reliable measurement point sequence, wherein: Generate a reliable measurement point sequence, specifically as follows: Read the measurement point code, sampling time, sample value, sampling period, freeze judgment duration, and missing sample judgment duration from the valid measurement point record; The valid measurement point records are sorted by time according to the measurement point code, and the number of repetitions of adjacent sampled values, the duration of repetition, and the rate of change of sampled values ​​are calculated. The sampling records of measurement points whose repetition duration reaches the freeze determination time and whose sample value change rate is zero are marked as freeze points; Calculate the adjacent sampling time interval under the same measurement point code, and mark the measurement point sampling record with an adjacent sampling time interval greater than the sampling period and reaching the missing sampling judgment time as a supplementary transmission record; Remove frozen points, retain the retransmission markers and valid sampling records of the retransmission records, and generate a reliable measurement point sequence according to the sampling time order; A unified sampling time axis is generated based on the sampling time. Sampling alignment is performed on the reliable measurement point sequence to generate a reliable boiler operating status frame, wherein: Generate a reliable boiler operating status frame, specifically as follows: Read the measurement point code, sampling time, sample value, sampling period, and retransmission marker from the trusted measurement point sequence; The start and end times are determined according to the sampling time, and a unified sampling time axis is generated based on the sampling period; The reliable measurement point sequence is mapped to the unified sampling time axis according to the measurement point code, and the sampled value whose sampling time is consistent with the time point in the unified sampling time axis is written to the corresponding time point. Write the sampled value with the retransmission mark to the original corresponding time point, and retain the retransmission mark; Null value markers are written for time points with missing sample values ​​to generate a reliable boiler operating status frame. The reliable boiler operating status frame includes measurement point code, sampling time, sample value, quality identifier, retransmission marker, null value marker, boiler combustion data, boiler steam and water data, boiler flue gas data, boiler auxiliary equipment data, and boiler operating condition data. The boiler operating condition data includes load value, coal quality marker, and heating status. The boiler steam and water data includes main steam flow. The boiler combustion data includes combustion parameters. The boiler auxiliary equipment data and boiler combustion data include controllable execution quantities.

[0023] In this embodiment, generating the boiler operating condition identification record and the boiler state sequence bound to the operating condition includes: Read the load value, main steam flow rate, coal quality marker, heating status, combustion parameters, and controllable execution variables from the trusted boiler operating status frame, and generate an operating condition identification input record, including: The following steps are taken to generate the working condition identification input record: Read boiler operating status data, boiler combustion data, boiler steam-water data, and boiler auxiliary equipment data from the trusted boiler operating status frame; Extract load values, coal quality marks, and heating status from boiler operating data; extract main steam flow from boiler steam and water data; extract combustion parameters from boiler combustion data; and extract controllable execution quantities from boiler combustion data and boiler auxiliary equipment data. The load value, main steam flow, coal quality mark, heating status, combustion parameters and controllable execution quantity are stored in association according to the sampling time to generate working condition identification input record; Operating conditions are segmented based on load value, main steam flow rate, coal quality label, and heating status to generate operating conditions, including: The generated operating conditions are as follows: Read the sampling time, load value, main steam flow, coal quality mark and heating status from the operating condition identification input record; Read the rated load value and rated main steam flow rate from the boiler rated parameter table; Calculate the load rate by dividing the load value by the rated load value and then multiplying by 100; calculate the flow rate by dividing the main steam flow rate by the rated main steam flow rate and then multiplying by 100. Based on load rates of less than 40%, load rates greater than or equal to 40% and less than 60%, load rates greater than or equal to 60% and less than 80%, and load rates greater than or equal to 80%, low load segment, medium-low load segment, medium-high load segment, and high load segment are generated respectively. Based on the flow rate of less than 40%, the flow rate of greater than or equal to 40% and less than 60%, the flow rate of greater than or equal to 60% and less than 80%, and the flow rate of greater than or equal to 80%, low flow rate, low-medium flow rate, medium-high flow rate and high flow rate are generated respectively. The load segment, flow segment, coal quality label, and heating status are associated and coded according to the sampling time to generate operating conditions; The combustion steady state is identified based on the fluctuation amplitude and duration of combustion parameters, and the operating conditions and combustion steady state are associated and stored to generate a boiler operating condition identification record, in which: Generate boiler operating condition identification records, specifically: The system reads the sampling time, operating conditions, and combustion parameters from the input record of the operating condition identification system. The combustion parameters include primary air pressure, primary air flow rate, secondary air damper opening, furnace negative pressure, and flue gas oxygen content. The combustion identification window is divided into 30 consecutive sampling cycles based on the sampling time; Calculate the difference between the maximum and minimum values ​​of each combustion parameter within each combustion identification window to generate the fluctuation range of combustion parameters; Calculate the median of the fluctuation range of combustion parameters corresponding to each combustion identification window under the same operating condition, multiply the median by 1.2, and generate the combustion stability judgment threshold. The combustion identification window where the fluctuation of combustion parameters is less than the combustion stability judgment threshold and the duration reaches 30 sampling cycles is marked as a combustion stable state. The combustion identification window where the fluctuation amplitude of combustion parameters is greater than or equal to the combustion stability determination threshold is marked as a combustion fluctuation state; The operating conditions, combustion stability, combustion fluctuations, and corresponding sampling times are associated and stored to generate boiler operating condition identification records. Control and regulation event records are identified based on the magnitude, direction, and timing of changes in controllable execution quantities. A control and regulation event window is then generated according to the timing of occurrence, sampling period, and event duration. A control and adjustment event window is generated based on the occurrence time, sampling period, and event duration, specifically as follows: Read the sampling time and controllable execution quantity from the operating condition identification input record, and read the sampling period from the trusted boiler operating status frame; Based on the sampling time sequence of the same controllable execution quantity, calculate the absolute value of the difference between adjacent sample values ​​to generate the change range of the controllable execution quantity; Controllable execution quantities with a difference greater than 0 between adjacent sample values ​​are marked as positive adjustment, and controllable execution quantities with a difference less than 0 between adjacent sample values ​​are marked as negative adjustment; The sampling time when the change in the controllable execution quantity is greater than or equal to 2% of the action amplitude threshold is marked as the time when the control adjustment event occurs; Starting from the time of occurrence of the controllable adjustment event, the duration of continuous changes in the controllable execution quantity that are greater than or equal to the action amplitude threshold is counted, and the duration is multiplied by the sampling period to generate the event duration. The controllable execution quantity identifier, change range, change direction, occurrence time, and event duration are associated and stored to generate a control adjustment event record; A control and adjustment event window is generated, with the occurrence time as the starting point and the occurrence time plus the event duration plus two sampling periods as the ending point. The boiler operating condition identification record, control and regulation event record, and control and regulation event window are bound to the trusted boiler operating status frame to generate an operating condition-bound boiler status sequence.

[0024] In this embodiment, generating multi-channel energy efficiency residuals includes: An improved MTGNN model is constructed, which includes a time-series layer of operating condition routing expert graph, a convolutional layer of boiler energy flow constraint graph, and a counterfactual energy efficiency benchmark reconstruction layer, wherein: The timing layer of the working condition routing expert graph includes: Operating condition reading end: Reads the operating condition, combustion stability, heating status and measurement point timing embedding in the boiler status sequence bound to the operating condition, and encodes the operating condition, combustion stability and heating status into operating condition routing vectors; Expert branch index area: Based on the working condition routing vector, low load expert branches, medium-low load expert branches, medium-high load expert branches and high load expert branches are matched, and a corresponding graph time series parameter index is assigned to each expert branch; Routing weight gating channel: Input the working condition routing vector into the gating function, calculate the routing weight of each expert branch, and adjust the input intensity of the measurement point timing embedded in different expert branches according to the routing weight; Operating condition representation aggregation end: The time-series features of the graphs output by each expert branch are weighted and aggregated to generate a time-series representation of the operating condition graph that carries the differences in operating conditions, combustion stability and heating status. The boiler energy flow constraint diagram convolution layer includes: Energy flow relationship mapping area: Read the measuring point code, boiler combustion data, boiler steam and water data, boiler flue gas data and boiler auxiliary equipment data, and construct the boiler measuring point map and boiler measuring point association matrix according to the energy transfer relationship between the combustion side, steam and water side, flue gas side and auxiliary equipment side; Directional coupling propagation channel: Based on the boiler measuring point correlation matrix, the propagation direction and correlation strength between measuring points are determined, and the time series representation of the operating condition diagram is propagated by graph convolution along the energy flow path from the combustion side to the steam-water side, from the combustion side to the flue gas side, and from the auxiliary machine side to the combustion side; Energy efficiency constraint correction slot: Read the input energy term, output energy term and loss energy term in the boiler energy flow relationship, perform energy flow constraint correction on the measurement point representation after graph convolution propagation, and suppress propagation characteristics that are inconsistent with the boiler energy flow relationship; Energy flow characterization output: The measurement point characterization that has completed energy flow constraint correction is processed by channel splicing and normalization to generate an energy flow constraint boiler characterization. The counterfactual energy efficiency benchmark reconstruction layer includes: Disturbance channel screening end: Read the energy flow constrained boiler characterization and measurement point time series embedding, and screen abnormal disturbance channels based on the measurement point disturbance intensity, measurement point correlation intensity and operating condition stability. Normal state replacement zone: Based on the boiler operating condition identification record, the normal measuring point status under the same operating condition is matched, and the measuring point representation in the abnormal disturbance channel is replaced with the normal measuring point representation. Counterfactual reconstruction channel: The normal measurement point representation and the unreplaced measurement point representation are merged and reconstructed to generate a counterfactual boiler representation for representing the boiler energy efficiency status under no abnormal disturbance conditions; The benchmark output verification terminal: inputs the counterfactual boiler characterization into the prediction output structure, generates a multi-channel energy efficiency benchmark, and verifies the consistency between the multi-channel energy efficiency benchmark and the operating conditions, combustion stability, and boiler energy flow relationship. Read the boiler operating condition identification record, trusted boiler operating status frame, and measurement point code from the boiler state sequence bound to the operating condition. Generate measurement point time sequence embedding and operating condition embedding through the input encoding terminal, wherein: The timing embedding and operating condition embedding of the measurement points are generated through the input encoder, specifically as follows: Read the measurement point code, sampling time, sampled value, supplementary transmission mark, null mark, operating condition, combustion stability state, combustion fluctuation state and heating state from the boiler state sequence bound to the operating condition; Based on the measurement point code, read the measurement point type, system, data source marker, sampling period, upper limit of value and lower limit of value from the measurement point code table, and generate measurement point attribute fields; The measurement point type, system to which it belongs, and data source mark are input into the encoding terminal and then discretely mapped to generate the measurement point attribute code; The normalized sample value is generated by subtracting the lower limit from the sample value and then dividing by the sum of the difference between the upper and lower limits and the stability constant 0.000001. The normalized sampled values, supplementary transmission markers, and null value markers are input to the input encoding end and then concatenated and mapped to generate the measurement point numerical code. The time sequence position is calculated based on the sampling time and sampling period, and the time sequence position is encoded by sinusoidal position encoding and cosine position encoding through the input encoding terminal to generate the measurement point time code. The measurement point attribute code, measurement point numerical code, and measurement point time code are concatenated according to the measurement point code and sampling time to generate the measurement point time sequence embedding; The operating conditions, combustion stability, combustion fluctuation, and heating status are input to the encoding terminal and then discretized to generate an embedded operating condition. The measurement point time series embedding and operating condition embedding are input into the operating condition routing expert graph time series layer. The corresponding expert graph time series branch is selected according to the operating condition to generate the operating condition time series representation, where: The timing representation of the generated operating condition diagram is as follows: The timing embedding of the measurement points, the operating conditions, the combustion stability, and the heating status are read by the operating condition status reading terminal in the timing layer of the operating condition routing expert diagram. The operating condition embedding, operating condition, combustion stability state and heating state are spliced ​​together to generate an operating condition routing vector; The expert branch index corresponding to the operating condition is read from the expert branch index area, and the low load segment, medium-low load segment, medium-high load segment and high load segment are matched to the low load expert branch, medium-low load expert branch, medium-high load expert branch and high load expert branch respectively. By inputting the working condition routing vector into the linear mapping function through the routing weight gating channel, the routing score corresponding to each expert branch is generated, and the routing score is input into the Softmax function to generate the routing weight corresponding to each expert branch. The temporal embeddings of the measurement points are arranged into a temporal feature tensor based on the sampling time and measurement point encoding. The temporal feature tensor of the measurement point is input into the temporal convolution structure of each expert branch, and dilated temporal convolution is performed along the sampling time direction to extract the change features of adjacent sampling periods and cross periods, and generate branch temporal features. Based on the measurement point encoding, the corresponding node embedding is read, and the adjacency weight between measurement points is calculated through the adaptive graph learning structure in each expert branch to generate the branch measurement point adjacency matrix. The branch time features and the branch measurement point adjacency matrix are input into the hybrid hop number graph propagation structure. The associated measurement point features are aggregated according to the 1-hop, 2-hop and 3-hop propagation paths to generate multi-hop graph propagation features. By splicing and mapping the branch time features and multi-hop graph propagation features through skip connection paths, the branch graph time series features corresponding to each expert branch are generated. The routing weights corresponding to each expert branch are expanded to the same measurement point dimension and time dimension as the temporal features of the branch graph, generating a branch weight tensor. The branch weight tensor is multiplied element-wise with the corresponding branch graph temporal features to generate weighted branch graph temporal features; The initial working condition diagram time series representation is generated by summing the time series features of each weighted branch diagram according to the expert branch dimension through the working condition representation aggregation terminal; The initial working condition diagram time series representation is spliced ​​with the measurement point time series embedded execution residuals, and the working condition diagram time series representation is generated through linear mapping and normalization processing. The routing weight gate channel inputs the working condition routing vector into a linear mapping function to generate routing scores for each expert branch. Specifically: The linear mapping function takes the working condition routing vector as input, performs weighted mapping on each field in the working condition routing vector by the routing weight matrix, and superimposes the routing bias vector to generate the expert branch routing score vector. The expert branch routing score vector includes the low load expert branch routing score, the medium-low load expert branch routing score, the medium-high load expert branch routing score, and the high load expert branch routing score. Read the operating condition embedding, operating condition code, combustion stability state code and heating state code from the operating condition routing vector, and concatenate them into a routing input vector according to a fixed field order; The routing weight matrix and routing bias vector are read through the routing weight gating channel. The routing input vector is multiplied by the routing weight matrix, and the routing bias vector is superimposed to generate the expert branch routing score vector. Extract the low-load expert branch route score, medium-low load expert branch route score, medium-high load expert branch route score, and high load expert branch route score from the expert branch route score vector respectively. The route score is input into the Softmax function to generate the route weights corresponding to each expert branch, specifically: The Softmax function performs indexation processing on the expert branch routing scores for low load, medium-low load, medium-high load, and high load respectively, generating expert branch index values ​​for each. All expert branch index values ​​are added together to generate the expert branch index sum. Each expert branch index value is then divided by the expert branch index sum to generate the corresponding expert branch routing weight. Read the low-load expert branch route score, medium-low load expert branch route score, medium-high load expert branch route score, and high load expert branch route score from the expert branch route score vector; Perform index calculations on the low-load expert branch routing score, medium-low load expert branch routing score, medium-high load expert branch routing score, and high load expert branch routing score respectively to generate the low-load expert branch index value, medium-low load expert branch index value, medium-high load expert branch index value, and high load expert branch index value; The expert branch index values ​​for low load, medium-low load, medium-high load, and high load are summed to generate the expert branch index sum value. Divide each expert branch index value by the expert branch index and the sum of the values ​​to generate the low-load expert branch routing weight, medium-low load expert branch routing weight, medium-high load expert branch routing weight, and high load expert branch routing weight. Write the routing weights of low-load expert branches, medium-low-load expert branches, medium-high-load expert branches, and high-load expert branches into the routing weight record, so that the working condition characterization aggregation end can perform branch graph time-series feature weighted aggregation. A boiler measurement point map is constructed based on measurement point coding and boiler energy flow relationship. The boiler measurement point map and the time-series representation of the operating condition diagram are input into the convolutional layer of the boiler energy flow constraint map to generate an energy flow-constrained boiler representation, wherein: The generation of energy flow-constrained boiler characterization is as follows: The energy flow relationship mapping area in the convolutional layer of the boiler energy flow constraint diagram is used to read the measurement point code, system, measurement point type and data source mark in the measurement point code table, and to read the energy transfer relationship between the combustion side, steam-water side, flue gas side and auxiliary machine side in the boiler system energy flow topology configuration; A boiler energy flow relationship configuration table is established based on the measurement point code, system to which it belongs, measurement point type, data source mark and energy transfer relationship. The boiler energy flow relationship configuration table includes source side system, target side system, source measurement point code, target measurement point code, source measurement point type, target measurement point type, energy flow direction, energy flow relationship strength, input energy item field, output energy item field and loss energy item field. Set the measurement point code as the measurement point node in the boiler measurement point diagram, connect the measurement point node corresponding to the source measurement point code with the measurement point node corresponding to the target measurement point code to generate energy flow association edge; The energy flow relationship strength is generated by dividing the sum of all energy flow relationship strengths corresponding to the same source measurement point node by the sum of the stability constant 0.000001. Write the measurement point nodes, energy flow association edges, energy flow directions and measurement point association strengths into the graph structure record to generate the boiler measurement point map and the boiler measurement point association matrix. The boiler measuring point diagram, boiler measuring point association matrix and operating condition diagram time series representation are read through the directional coupling propagation channel. According to the energy flow direction from the combustion side to the steam-water side, the combustion side to the flue gas side, and the auxiliary machine side to the combustion side, the operating condition diagram time series representation of adjacent measuring point nodes is weighted and propagated to generate directional propagation measuring point representation. Based on the input energy item field, output energy item field, and loss energy item field in the boiler energy flow relationship configuration table, extract the corresponding combustion data, steam-water data, flue gas data, and auxiliary machine data from the trusted boiler operating status frame to generate input energy items, output energy items, and loss energy items; The input energy term, output energy term, loss energy term, measurement point correlation strength, and directional propagation measurement point characterization are read through the energy efficiency constraint correction slot; Summing the output energy term and the lost energy term generates a sum of energy destinations. Perform a difference calculation between the input energy item and the sum of energy destination values ​​to generate an energy flow difference; The normalized energy flow constraint deviation is generated by dividing the absolute value of the energy flow difference by the sum of the absolute value of the input energy term and the stability constant 0.000001. The propagation correction coefficient is calculated based on the normalized energy flow constraint deviation. The propagation correction coefficient is 1 divided by 1 and the sum of the normalized energy flow constraint deviation. The propagation correction coefficient and the measurement point association strength are multiplied to generate the corrected measurement point association strength. The corrected measurement point association strength is multiplied element-wise with the directional propagation measurement point characterization in the corresponding energy flow direction to generate the corrected propagation measurement point characterization. The calibration propagation measurement point representations received by the same target measurement point node are summed according to the energy flow correlation edge, and then added to the residual of the original working condition diagram time sequence representation of the target measurement point node to generate the calibration measurement point representation. By performing channel splicing, linear mapping and normalization on the corrected measurement point representation through the energy flow representation output terminal, an energy flow constrained boiler representation is generated. The energy flow-constrained boiler characterization is input into the counterfactual energy efficiency benchmark reconstruction layer. The counterfactual boiler characterization is generated according to the measurement point disturbance shielding and normal state replacement. A multi-channel energy efficiency benchmark is then generated through the residual output, where: The multi-channel energy efficiency benchmark is generated as follows: The disturbance channel filtering end in the counterfactual energy efficiency benchmark reconstruction layer reads the energy flow constraint boiler characterization, measurement point coding, boiler operating condition identification record and control regulation event window; Based on historical reliable boiler operating status frames, historical boiler operating condition identification records, and historical energy flow constrained boiler characterization, the mean and dispersion of normal characterization corresponding to each measuring point code are statistically analyzed according to the same operating conditions and the same combustion stability state, and a normal characterization benchmark record is generated. Based on the boiler operating condition identification record, match the normal characterization benchmark record under the same operating condition and the same combustion stability state, and read the normal characterization mean and normal characterization dispersion corresponding to each measurement point code. Subtract the mean value of the normal characterization from the characterization of the energy flow-constrained boiler to generate the deviation value of the characterization at the measuring point. The absolute value of the deviation value of the measurement point is divided by the sum of the normal characterization dispersion and the stability constant 0.000001 to generate the measurement point disturbance intensity. Measuring points whose disturbance intensity is greater than the disturbance intensity threshold 3 and do not fall within the control adjustment event window are coded and marked as abnormal disturbance channels; Read the measurement point code corresponding to the abnormal disturbance channel through the normal state replacement area, and extract the normal measurement point representation corresponding to the same measurement point code from the normal representation benchmark record. Replace the energy flow constrained boiler representation in the abnormal disturbance channel with the normal measurement point representation, and leave the energy flow constrained boiler representation that is not marked as an abnormal disturbance channel unchanged to generate the replaced boiler representation. Extract the energy flow constraint boiler characterization that was not marked as an anomalous disturbance channel and generate the characterization of the unreplaced measurement points; By reading the replaced boiler characterization, the unreplaced measuring point characterization, and the boiler measuring point correlation matrix through the counterfactual reconstruction channel, correlation compensation and channel fusion are performed on adjacent measuring point characterizations to generate counterfactual boiler characterization; The counterfactual boiler characterization is input into the residual output terminal through the benchmark output verification terminal. The residual output terminal performs linear mapping according to the flue gas heat loss channel, oxygen deviation channel, desuperheating water consumption channel, fan power consumption channel and comprehensive energy efficiency channel respectively to generate flue gas heat loss benchmark, oxygen deviation benchmark, desuperheating water consumption benchmark, fan power consumption benchmark and comprehensive energy efficiency benchmark. The benchmark output verification terminal performs energy flow consistency verification on the flue gas heat loss benchmark, oxygen deviation benchmark, desuperheating water consumption benchmark, fan power consumption benchmark and comprehensive energy efficiency benchmark, and eliminates benchmark output values ​​that do not meet the boiler energy flow relationship. The verified flue gas heat loss benchmark, oxygen deviation benchmark, desuperheating water consumption benchmark, fan power consumption benchmark and comprehensive energy efficiency benchmark are associated and written into the same output record through the residual output terminal to generate a multi-channel energy efficiency benchmark. Actual energy efficiency indicators are extracted from the reliable boiler operating status frame. Channel-by-channel difference calculations are then performed between the actual energy efficiency indicators and the multi-channel energy efficiency benchmark to generate multi-channel energy efficiency residuals, where: The multi-channel energy efficiency residual is generated as follows: Read boiler flue gas data, boiler steam and water data, boiler auxiliary equipment data and boiler operating condition data from the trusted boiler operating status frame, and read flue gas heat loss benchmark, oxygen deviation benchmark, desuperheating water consumption benchmark and fan power consumption benchmark from the multi-channel energy efficiency benchmark; Extract flue gas temperature and flue gas oxygen content from boiler flue gas data; extract desuperheating water flow rate and main steam flow rate from boiler steam-water data; extract fan active power from boiler auxiliary equipment data; and extract load value from boiler operating condition data. The flue gas temperature is used as the actual flue gas heat loss index, the flue gas oxygen content is used as the actual oxygen content deviation index, the desuperheating water flow rate is divided by the sum of the main steam flow rate and the stability constant 0.000001 to generate the actual desuperheating water consumption index, and the fan active power is divided by the sum of the load value and the stability constant 0.000001 to generate the actual fan power consumption index. The actual flue gas heat loss index, actual oxygen deviation index, actual desuperheating water consumption index, and actual fan power consumption index are respectively compared with the corresponding benchmarks to calculate the residuals of flue gas heat loss, oxygen deviation, desuperheating water consumption, and fan power consumption. The residuals of flue gas heat loss, oxygen deviation, desuperheating water consumption, and fan power consumption are divided by the sum of the corresponding absolute values ​​of the benchmarks and the stability constant 0.000001, respectively, to generate four normalized residuals. The four normalized residuals are weighted and summed to generate the comprehensive energy efficiency residual, with a weight of 0.35 for flue gas heat loss residual, 0.25 for oxygen deviation residual, 0.2 for desuperheating water consumption residual, and 0.2 for fan power consumption residual. The residuals of flue gas heat loss, oxygen content deviation, desuperheating water consumption, fan power consumption, and comprehensive energy efficiency are associated and stored to generate multi-channel energy efficiency residuals; The improved MTGNN model was trained using a combination of energy efficiency baseline prediction error, graph temporal representation reconstruction error, operating condition routing allocation error, energy flow constraint bias, and counterfactual baseline reconstruction error as the joint training objective. The network parameters of the operating condition routing expert graph temporal layer, the boiler energy flow constraint graph convolutional layer, and the counterfactual energy efficiency baseline reconstruction layer were updated accordingly. The improved MTGNN model was considered to have completed convergence training when the change in the joint loss value over five consecutive training rounds was less than 0.001. The improved MTGNN model is trained as follows: A training sample set was constructed from historical reliable boiler operating status frames, historical boiler operating condition identification records, historical control and regulation event records, and historical energy efficiency calculation records. The training sample set includes 8,000 sets of boiler operating samples. Each set of boiler operating samples contains a boiler state sequence bound to operating conditions for 60 consecutive sampling periods, multi-channel normal energy efficiency benchmark labels, measurement point reconstruction labels, operating condition branch labels, energy flow balance labels, and counterfactual benchmark labels. The boiler state sequence is input into the encoding terminal to bind the operating conditions, generating the time-series embedding of measurement points and the operating condition embedding. The operating condition state reading terminal, expert branch index area, routing weight gating channel and operating condition representation aggregation terminal generate the operating condition diagram time-series representation. The energy flow relationship mapping area, directional coupling propagation channel, energy efficiency constraint correction slot and energy flow representation output terminal generate the energy flow constraint boiler representation. The disturbance channel screening terminal, normal state replacement area, counterfactual reconstruction channel and benchmark output verification terminal generate the counterfactual boiler representation, and the multi-channel energy efficiency benchmark is output by the residual output terminal. The energy efficiency benchmark prediction error is generated by taking the average of the squared differences between the multi-channel energy efficiency benchmark and the multi-channel normal energy efficiency benchmark labels; the graph time series characterization reconstruction error is generated by taking the average of the squared differences between the reconstructed measurement point sequence and the measurement point reconstruction labels; the operating condition routing allocation error is generated by taking the average of the squared differences between the routing weight and the operating condition branch labels; the energy flow constraint deviation is generated by taking the average of the squared differences between the energy flow constraint deviation and the energy flow balance labels; and the counterfactual boiler characterization output benchmark is generated by taking the average of the squared differences between the counterfactual benchmark labels. The five types of errors were multiplied by 0.35, 0.2, 0.15, 0.15, and 0.15 respectively and then summed to generate a joint loss value. The network parameters of the operating condition routing expert graph temporal layer, the boiler energy flow constraint graph convolutional layer, and the counterfactual energy efficiency benchmark reconstruction layer were calculated and written back at a learning rate of 0.001. The training was repeated until the change in the joint loss value was less than 0.001 for five consecutive rounds, thus confirming that the improved MTGNN model had completed convergence training.

[0025] In this embodiment, the improved MTGNN model includes: The MTGNN model retains the core of adaptive graph learning structure, temporal convolution structure, hybrid hop number graph propagation structure, hop connection path, and prediction output structure. The input representation, operating condition routing, energy flow propagation, and counterfactual reconstruction are structurally modified to address the boiler energy efficiency anomaly early warning scenario. Set up an input encoding terminal at the model input end, read the reliable boiler operating status frame, boiler operating condition identification record and measurement point code in the boiler state sequence bound to the operating condition, and encode the measurement point sample value, measurement point type, sampling time and operating condition into measurement point time sequence embedding and operating condition embedding; A time-series layer of the operating condition routing expert graph is added at the time-series feature extraction location. The time-series branches of the expert graph are selected based on the operating conditions, combustion stability and heating status to form the time-series representation of the operating condition graph. The original graph convolution propagation structure is modified to form a boiler energy flow constraint graph convolution layer. The boiler measurement point graph and boiler energy flow relationship are embedded into the graph propagation path to constrain the energy efficiency coupling relationship between the combustion side, steam-water side, flue gas side and auxiliary machine side. A counterfactual energy efficiency benchmark reconstruction layer is added before the predicted output. The abnormal disturbance channel is shielded and replaced and the normal state is reconstructed to generate a counterfactual boiler characterization. The predicted output structure is transformed into a residual output terminal, outputting a multi-channel energy efficiency benchmark, and the difference between the benchmark and the actual energy efficiency index is calculated to generate a multi-channel energy efficiency residual.

[0026] In this embodiment, generating the energy efficiency residual after stripping includes: Read multi-channel energy efficiency residuals, control and regulation event records, and control and regulation event windows; establish a residual event alignment table according to sampling time; and associate and store each residual channel with its corresponding control and regulation event window. A residual event alignment table is established according to the sampling time, and each residual channel is associated with and stored with the corresponding control adjustment event window, specifically: Read the sampling time, residual channel identifier, and residual value from the multi-channel energy efficiency residual; Read the controllable execution quantity identifier, change direction, change magnitude, occurrence time, and event duration from the control and adjustment event log; Read the start and end points of the window in the control and adjustment event window; The residual values ​​are matched with the control adjustment event window according to the sampling time, and the residual channels whose sampling time falls between the start and end of the window are marked as the residual channels within the window. Write the residual channel, residual value, controllable execution quantity identifier, change direction, change magnitude, occurrence time, event duration, window start point and window end point within the window into the same record to generate a residual event alignment table; Each residual channel is associated with and stored in relation to its corresponding control and adjustment event window according to the residual channel identifier; Based on the residual event alignment table, the residual peak value, residual start and end time, residual duration, residual change direction, and residual channel identifier within the control and adjustment event window are extracted to generate window residual feature records, where: Generate window residual feature records, specifically: Read the residual channel identifier, residual value, sampling time, window start point, and window end point from the residual event alignment table; The residual values ​​whose sampling time falls between the start and end points of the window are selected to generate a residual sequence within the window. Calculate the maximum absolute value of the residuals in the residual sequence within the window to generate the residual peak value; The sampling time corresponding to the first residual value in the residual sequence within the window is taken as the residual start time, and the sampling time corresponding to the last residual value is taken as the residual end time. Subtract the residual start time from the residual end time to generate the residual duration. Subtract the first residual value from the last residual value in the residual sequence within the window. If the difference is greater than 0, it is marked as a positive change in the residual; if the difference is less than 0, it is marked as a negative change in the residual. The residual channel identifier, residual peak value, residual start time, residual end time, residual duration, and residual change direction are associated and stored to generate a window residual feature record; Based on the controllable execution variable event records, the direction, magnitude, start and end times, and duration of the action are extracted. The direction of change of the controllable execution variable is matched with the window residual feature records, and the start and end times of the action are matched with the start and end times of the residuals to generate control-related residual markers. Generate control-related residual markers, specifically as follows: Read the controllable execution quantity identifier, change direction, change magnitude, occurrence time and event duration from the control adjustment event record, and read the residual channel identifier, residual change direction, residual start time, residual end time and residual duration from the window residual feature record; Use the occurrence time as the start time of the action, and the occurrence time plus the event duration as the end time of the action; Window residual feature records whose controllable execution variable change direction is consistent with the residual change direction are marked as direction matching records; Directional matching records where the start time to end time of an action overlaps with the start time to end time of a residual are marked as temporal matching records; The residual channel where both direction matching record and timing matching record are true is marked as a control-related residual channel, and associated with the change amplitude, action start and end time and action duration to generate a control-related residual label; The residual stripping weights are determined based on the control-related residual labels and residual durations. Weighted reduction is performed on residual channels with control-related residual labels in the multi-channel energy efficiency residuals, while residual channels without matching control-related residual labels are retained, generating stripped energy efficiency residuals. The energy efficiency residual after stripping is generated as follows: Read the residual channel identifier, sampling time and residual value in the multi-channel energy efficiency residual, and read the control-related residual channel, action start time, action end time, action duration, residual start time, residual end time and residual duration in the control-related residual mark; The earlier of the action end time and the residual end time is taken as the overlap end time, and the later of the action start time and the residual start time is taken as the overlap start time. Subtract the start time of the overlap from the end time of the overlap to generate the time overlap length. When the time overlap length is less than 0, set the time overlap length to 0. Divide the time overlap length by the sum of the residual duration and the stability constant 0.000001 to generate the time overlap ratio; The continuous matching ratio is generated by dividing the smaller of the action duration and the residual duration by the sum of the larger of the action duration and the residual duration and the stability constant 0.000001. Multiply the time overlap ratio by 0.6, multiply the continuous matching ratio by 0.4, and add the two products to generate the residual stripping weight; When the residual stripping weight is greater than 0.9, the residual stripping weight is reset to 0.9; For residual channels with control-related residual labels, multiply the residual value by 1 and subtract the result after residual stripping weight to generate the stripped channel residual; For residual channels that do not match control-related residual markers, the residual values ​​are kept unchanged to generate retained channel residuals; The residual values ​​of the stripped channels and the residual values ​​of the retained channels are stored in association according to the sampling time to generate the energy efficiency residual value after stripping.

[0027] In this embodiment, generating multi-channel confidence residuals and comprehensive risk residuals includes: Read the boiler operating condition identification record and the energy efficiency residual after stripping, and generate a residual benchmark matching record based on the operating condition matching residual benchmark table. Specifically: Read the boiler operating condition identification record and the energy efficiency residual after stripping, and generate a residual benchmark matching record based on the operating condition matching residual benchmark table. Read the sampling time, operating conditions, combustion stability, and heating status from the boiler operating condition identification record; Read the sampling time, residual channel identifier, and residual value from the energy efficiency residual after stripping; Based on historical boiler operating condition identification records and historical energy efficiency residuals after stripping, historical residual values ​​are grouped according to operating conditions, combustion stability, heating status, and residual channel identifiers. Calculate the average value of each group of historical residuals to generate the mean of normal residuals, and calculate the standard deviation of each group of historical residuals to generate the residual dispersion. Subtract twice the residual dispersion from the mean of normal residuals to generate the lower boundary of the confidence interval, add twice the residual dispersion to the mean of normal residuals to generate the upper boundary of the confidence interval, and generate the operating condition residual benchmark table. The stripped energy efficiency residuals are associated with the boiler operating condition identification records according to the sampling time to generate residual records to be matched; The operating conditions, combustion stability, heating status, and residual channel identifiers in the residual records to be matched are matched with the operating condition residual benchmark table, and the corresponding normal residual mean, residual dispersion, lower boundary of confidence interval, and upper boundary of confidence interval are extracted. The sampling time, operating conditions, combustion stability, heating status, residual channel identifier, residual value, normal residual mean, residual dispersion, lower boundary of confidence interval and upper boundary of confidence interval are associated and stored to generate residual benchmark matching records. Extract the mean of normal residuals, residual dispersion, and confidence interval boundaries for each residual channel from the residual benchmark matching records to generate residual confidence benchmark records, where: Generate residual confidence baseline records, specifically as follows: Read the sampling time, operating conditions, combustion stability, heating status, residual channel identifier, mean of normal residuals, residual dispersion, lower boundary of confidence interval and upper boundary of confidence interval from the residual benchmark matching record; The mean of normal residuals, residual dispersion, lower boundary of confidence interval, and upper boundary of confidence interval are aggregated according to the residual channel identifier to generate the channel confidence benchmark field; The sampling time, operating conditions, combustion stability, heating status, residual channel identifier, and channel confidence benchmark field are associated and written into the same record to generate a residual confidence benchmark record; The difference between the stripped energy efficiency residual and the mean of the normal residual is calculated, and the difference result is divided by the sum of the residual dispersion and the stability constant to generate a multi-channel confidence residual, where: Generate multi-channel confidence residuals, specifically as follows: Read the sampling time, residual channel identifier and residual value in the energy efficiency residual after stripping, and read the normal residual mean and residual dispersion in the residual confidence benchmark record; According to the sampling time and residual channel identifier, the residual value is compared with the corresponding normal residual mean to generate the residual deviation value. Add the residual dispersion to the stability constant 0.000001 to generate the confidence normalized denominator, and divide the residual deviation value by the confidence normalized denominator to generate the confidence residual of the corresponding residual channel. The confidence residuals corresponding to each residual channel are stored in association according to the sampling time to generate multi-channel confidence residuals; Based on boiler operating condition identification records, the energy efficiency weights of each residual channel are determined. Weighted fusion is then performed on the multi-channel confidence residuals to generate a comprehensive risk residual, where: Generate comprehensive risk residuals, specifically as follows: Read the sampling time, operating conditions, combustion stability and heating status from the boiler operating condition identification record, and read the residual channel identifier and confidence residual value from the multi-channel confidence residual; When the operating conditions are low load or medium-low load, the weight of the flue gas heat loss channel is set to 0.3, the weight of the oxygen deviation channel is set to 0.3, the weight of the desuperheating water consumption channel is set to 0.15, and the weight of the fan power consumption channel is set to 0.25. When the operating conditions are medium-high load or high load, the weight of the flue gas heat loss channel is set to 0.35, the weight of the oxygen deviation channel is set to 0.25, the weight of the desuperheating water consumption channel is set to 0.2, and the weight of the fan power consumption channel is set to 0.2. The confidence residual value of each residual channel is multiplied by the corresponding energy efficiency weight to generate channel-weighted confidence residuals; The channel-weighted confidence residuals at the same sampling time are summed to generate a comprehensive risk residual. The sampling time, operating conditions, combustion stability, heating status, energy efficiency weights of each residual channel, and comprehensive risk residuals are associated and stored.

[0028] In this embodiment, the computational graph cooperative asymmetric drift includes: Read the multi-channel confidence residuals, comprehensive risk residuals, and boiler measuring point correlation matrix. Extract the residual channel identifiers from the multi-channel confidence residuals. Establish a drift detection input record table according to sampling time, sampling period, residual channel identifiers, and measuring point codes, where: Establish a drift detection input record table, specifically as follows: Read the sampling time, residual channel identifier, and confidence residual value from the multi-channel confidence residual; Read the sampling time and comprehensive risk residual value from the comprehensive risk residual; Read the sampling period in the trusted boiler operating status frame, and read the source measurement point code, target measurement point code, and measurement point association strength in the boiler measurement point association matrix; Based on the residual channel identifier matching the corresponding measurement point code, the confidence residual value is associated with the measurement point code in the boiler measurement point association matrix; The confidence residual value, comprehensive risk residual value, sampling period, residual channel identifier, measurement point code, and measurement point correlation strength are written into the same detection record according to the sampling time. Arrange the detection records in order of sampling time to generate a drift detection input record table; Based on the residual channel identifier matching offset weight configuration table, the forward offset weight and reverse offset weight are read, and directional asymmetric weighting is performed on the multi-channel confidence residuals to generate directional weighted confidence residuals, where: The generated direction-weighted confidence residuals are as follows: Read the sampling time, residual channel identifier, and confidence residual value from the drift detection input record table; Based on the residual channel identifiers, the forward and reverse offset weights are determined as follows: the flue gas heat loss channel corresponds to a forward offset weight of 1.2 and a reverse offset weight of 0.8; the oxygen deviation channel corresponds to a forward offset weight of 1.1 and a reverse offset weight of 0.9; the desuperheating water consumption channel corresponds to a forward offset weight of 1.15 and a reverse offset weight of 0.85; the fan power consumption channel corresponds to a forward offset weight of 1.1 and a reverse offset weight of 0.9; and the comprehensive risk channel corresponds to a forward offset weight of 1.2 and a reverse offset weight of 0.8. When the confidence residual value is greater than or equal to 0, the confidence residual value is multiplied by the corresponding positive offset weight to generate a positive weighted residual. When the confidence residual value is less than 0, the confidence residual value is multiplied by the corresponding reverse offset weight to generate the reverse weighted residual. The forward-weighted residuals and the reverse-weighted residuals are associated and stored according to the sampling time and residual channel identifier to generate directional-weighted confidence residuals; The direction-weighted confidence residuals and the overall risk residuals are input into the Page-Hinkley algorithm to calculate the residual running mean, cumulative offset, historical minimum cumulative offset, and single-channel drift statistic for each channel, where: The residual running mean, cumulative offset, historical minimum cumulative offset, and single-channel drift statistics are calculated channel by channel, specifically as follows: Read the sampling time, residual channel identifier, and weighted residual value from the directional weighted confidence residual, and read the comprehensive risk residual value from the comprehensive risk residual; The weighted residual values ​​and comprehensive risk residual values ​​of each residual channel are arranged according to the sampling time to generate a channel drift detection sequence; For each residual channel, set the initial residual running mean to 0, the initial cumulative offset to 0, and the initial historical minimum cumulative amount to 0; For each sampling time corresponding to the channel drift detection value, the residual running mean is recursively updated based on the residual running mean obtained in the previous sampling time and the current channel drift detection value, so that the updated residual running mean can reflect the overall change level of the residual before the current sampling time. The current channel drift detection value is compared with the updated residual running mean, the deviation of the current channel drift detection value from the residual running mean is calculated, and the drift tolerance constant 0.01 is subtracted from the deviation to generate the single-step offset of the current sampling time. Add the cumulative offset of the previous sampling time to the single-step offset to generate the cumulative offset of the current sampling time; Compare the cumulative offset at the current sampling time with the historical minimum cumulative offset at the previous sampling time, and take the smaller value as the historical minimum cumulative offset at the current sampling time; The cumulative offset at the current sampling time is subtracted from the historical minimum cumulative offset at the current sampling time to generate a single-channel drift statistic. The sampling time, residual channel identifier, residual running mean, cumulative offset, historical minimum cumulative amount and single-channel drift statistics are associated and stored to generate channel-by-channel drift statistics records; Based on the boiler measuring point correlation matrix and measuring point coding, graph correlation propagation is performed on the single-channel drift statistics, and the drift statistics of the correlated residual channels are fused according to the measuring point correlation strength to generate channel collaborative drift statistics, where: Generate channel cooperative drift statistics, specifically: Read the sampling time, residual channel identifier, measurement point code, and single-channel drift statistics from the channel-by-channel drift statistics record; Read the source measuring point code, target measuring point code, and measuring point association strength from the boiler measuring point association matrix; Using the measurement point code corresponding to the current residual channel as the target measurement point code, source measurement point codes that have energy flow correlation edges with the target measurement point code are selected to generate a set of related measurement points; Based on the matching of the corresponding residual channel identifiers of the associated measurement point set, the single-channel drift statistics of the associated residual channels under the same sampling time are read; The normalized association weight is generated by dividing the association strength of each source measurement point code by the sum of the association strengths of all measurement points in the associated measurement point set and the stability constant 0.000001. Multiply the single-channel drift statistic of the associated residual channel by the corresponding normalized association weight to generate the associated weighted drift statistic; Sum the associated weighted drift statistics corresponding to the same target measurement point code to generate graph association propagation drift; Multiply the single-channel drift statistic of the current residual channel by the self-channel preservation weight of 0.6, multiply the graph association propagation drift by the association propagation weight of 0.4, and add the two products to generate the channel cooperative drift statistic; When the set of associated measurement points is empty, the single-channel drift statistic of the current residual channel is used as the channel collaborative drift statistic. The sampling time, residual channel identifier, measurement point code, single-channel drift statistics, graph correlation propagation drift, and channel cooperative drift statistics are associated and stored to generate channel cooperative drift statistics records. Based on boiler operating condition identification records and residual channel identifier matching operating condition energy efficiency weight tables, weighted fusion is performed on the channel cooperative drift statistics and the drift statistics corresponding to the comprehensive risk residuals to generate graph cooperative asymmetric drift quantity and complete drift statistical process record tables, where: The table records the graph-coordinated asymmetric drift amount and the complete drift statistics process, as follows: Read the sampling time and operating conditions in the boiler operating condition identification record, read the residual channel identifier, measurement point code and channel collaborative drift statistics in the channel collaborative drift statistics record, and read the comprehensive risk drift statistics corresponding to the comprehensive risk channel; When the operating conditions are low load or medium-low load, the weights of the flue gas heat loss channel, oxygen deviation channel, desuperheating water consumption channel, fan power consumption channel and comprehensive risk channel are 0.2, 0.25, 0.15, 0.15 and 0.25, respectively. When the operating conditions are in the medium-high load range or high load range, the weights of the flue gas heat loss channel, oxygen deviation channel, desuperheating water consumption channel, fan power consumption channel and comprehensive risk channel are 0.25, 0.2, 0.15, 0.15 and 0.25, respectively. The channel-coordinated drift statistics of each residual channel are multiplied by their corresponding weights to generate channel-weighted drift statistics; Multiply the overall risk drift statistic by the overall risk channel weight to generate the overall risk weighted drift statistic; The channel-weighted drift statistic and the comprehensive risk-weighted drift statistic under the same sampling time are summed to generate a graph-coordinated asymmetric drift statistic. The sampling time, operating conditions, residual channel identifiers, measurement point codes, channel weights, channel collaborative drift statistics, comprehensive risk drift statistics, and graph collaborative asymmetric drift are associated and stored to generate a complete drift statistics process record table.

[0029] In this embodiment, the generation of boiler energy efficiency anomaly early warning results includes: Read the graph-cooperative asymmetric drift, the complete drift statistics record table, the boiler operating condition identification record, and the multi-channel confidence residuals. Generate an energy flow direction verification table based on the boiler operating condition identification record, where: An energy flow direction verification table is generated based on boiler operating condition identification records, specifically as follows: Read the sampling time and drift value in the graph cooperative asymmetric drift; Read the residual channel identifier, measurement point code, channel collaborative drift statistics and comprehensive risk drift statistics from the complete drift statistics process record table; Read the operating conditions, combustion stability and heating status from the boiler operating condition identification record, and read the confidence residual values ​​from the multi-channel confidence residuals; Based on residual channel identification, the direction of energy efficiency risk is determined. The positive confidence residuals of the flue gas heat loss channel, oxygen deviation channel, desuperheating water consumption channel and fan power consumption channel are marked as the direction of energy efficiency deterioration. When the operating conditions are low load or medium-low load, mark the oxygen deviation channel and the fan power consumption channel as the main verification channels. When the operating conditions are medium-high load or high load, the flue gas heat loss channel and the desuperheating water consumption channel are marked as the main verification channels. The sampling time, operating conditions, combustion stability, heating status, residual channel identifier, measurement point code, confidence residual value, energy efficiency risk direction and main verification channel mark are associated and stored to generate an energy flow direction verification table. Based on the energy flow direction verification table, an energy flow consistency verification operation is performed on the graph cooperative asymmetric drift, generating the verified drift and energy flow consistency flag, wherein: Generate a verified drift and energy flow consistency flag, specifically as follows: Read the sampling time, residual channel identifier, energy efficiency risk direction and main verification channel mark in the energy flow direction verification table, and read the drift value in the graph cooperative asymmetric drift. Filter residual channels marked with the master verification channel at the same sampling time to generate a set of master verification residual channels; The residual channels in the set of primary verification residual channels whose confidence residual direction is consistent with the energy efficiency risk direction are counted as consistent channels, and the number of consistent channels and the total number of primary verification residual channels are counted. The energy flow consistency ratio is generated by dividing the number of consistent channels by the sum of the total number of primary verification residual channels and the stability constant 0.000001. When the energy flow consistency ratio is greater than or equal to 0.5, the graph cooperative asymmetric drift is kept unchanged, the verified drift is generated, and the energy flow consistency mark is generated. When the energy flow consistency ratio is less than 0.5, the graph cooperative asymmetric drift is multiplied by 0.5 to generate the verified drift and generate an energy flow inconsistency flag. The sampling time, graph cooperative asymmetric drift, energy flow consistency ratio, post-verification drift, and energy flow consistency flag are associated and stored. Based on the boiler operating condition identification record and the drift threshold record, the verified drift amount is compared with the drift threshold record to generate an energy efficiency anomaly trigger record, where: Generate an energy efficiency anomaly trigger record, specifically: Read the sampling time, operating conditions, combustion stability and heating status from the boiler operating condition identification record, and read the drift value and energy flow consistency mark from the verified drift value; When the operating conditions are low load or medium-low load, set the warning drift threshold to 2.5 and the abnormal drift threshold to 3.5. When the operating condition is in the medium-high load section or high load section, set the early warning drift threshold 2 and the abnormal drift threshold 3; The sampling time, operating conditions, early warning drift threshold, and abnormal drift threshold are associated and stored to generate a drift threshold record; When the drift amount after verification is greater than or equal to the warning drift threshold and less than the abnormal drift threshold, and continues for 3 consecutive sampling cycles, an energy efficiency warning trigger flag is generated. When the drift amount after verification is greater than or equal to the abnormal drift threshold and continues for 3 consecutive sampling cycles, an energy efficiency abnormality trigger flag is generated. The sampling time, operating conditions, drift amount after verification, energy flow consistency flag, energy efficiency early warning trigger flag and energy efficiency anomaly trigger flag are associated and stored to generate an energy efficiency anomaly trigger record; Based on the energy efficiency anomaly trigger records and the complete drift statistical process record table, a counterfactual contribution verification operation is performed. Channel replacement and drift recalculation are performed on the multi-channel confidence residuals, generating an anomaly contribution ranking table, where: Generate an anomaly contribution ranking table, specifically as follows: Read the sampling time, energy efficiency warning trigger mark and energy efficiency anomaly trigger mark in the energy efficiency anomaly trigger record, and read the residual channel identifier, channel collaborative drift statistics, comprehensive risk drift statistics and graph collaborative asymmetric drift in the complete drift statistics process record table; Filter the sampling times that have energy efficiency warning trigger markers or energy efficiency anomaly trigger markers, and generate time points to be verified; Read the multi-channel confidence residuals corresponding to the time point to be verified, and select the residual channels one by one as the channels to be replaced. Set the confidence residual value of the channel to be replaced to 0, and keep the confidence residual values ​​of other residual channels unchanged to generate a counterfactual replacement residual sequence; Based on the counterfactual replacement residual sequence, directional asymmetric weighting, single-channel drift statistic calculation and graph association propagation are re-executed to generate counterfactual graph collaborative drift; Subtract the counterfactual graph collaborative drift from the original graph collaborative asymmetric drift to generate the abnormal contribution value of the channel to be replaced; Arrange the residual channel identifiers in descending order of abnormal contribution value to generate an abnormal contribution sorting table; Boiler energy efficiency anomaly early warning results are generated based on energy efficiency anomaly trigger records, energy flow consistency markers, and anomaly contribution ranking tables, where: Boiler energy efficiency anomaly early warning results are generated based on energy efficiency anomaly trigger records, energy flow consistency markers, and anomaly contribution ranking tables. Specifically: Read the sampling time, operating conditions, post-verification drift amount, energy efficiency warning trigger flag and energy efficiency anomaly trigger flag from the energy efficiency anomaly trigger record, and read the energy flow consistency flag; Read the residual channel identifier, abnormal contribution value, and sorting result from the abnormal contribution sorting table; When an energy efficiency anomaly trigger flag exists, a Level 1 anomaly warning is generated; when an energy efficiency warning trigger flag exists but no energy efficiency anomaly trigger flag exists, a Level 2 warning is generated. When the energy flow consistency flag is marked as energy flow inconsistency, the warning level will be downgraded by one level and written to the energy flow verification prompt field; The top 3 residual channels with the highest anomaly contribution values ​​are selected as the main anomaly channels, and the corresponding residual channel identifiers are written into the anomaly source field. The sampling time, operating conditions, drift amount after verification, energy flow consistency mark, early warning level, main abnormal channels, abnormal contribution value, sorting results and abnormal source fields are associated and stored to generate boiler energy efficiency abnormal early warning results.

[0030] Example 1: This example applies to an energy efficiency anomaly early warning scenario for a single ultra-supercritical coal-fired boiler. The boiler operation side includes the combustion system, steam-water system, flue gas system, auxiliary system, and operating condition system. On the monitoring screen, the flue gas temperature, flue gas oxygen content, furnace negative pressure, desuperheating water flow rate, and fan power are all within safe limits. Traditional fixed threshold methods do not trigger alarms. However, under similar loads, the unit coal consumption for power generation continues to increase, making it difficult for operators to determine the source of the anomaly.

[0031] This embodiment constructs a training sample set and a test sample set. The training sample set contains 8000 sets of boiler operation samples. Each set contains a sequence of boiler state conditions bound to operating conditions for 60 consecutive sampling periods. Fields include boiler combustion data, boiler steam-water data, boiler flue gas data, boiler auxiliary equipment data, boiler operating condition data, and data quality code. Among them, there are 1460 sets of samples for low load, 2035 sets for medium-low load, 2860 sets for medium-high load, 1645 sets for high load, 6200 sets for normal energy efficiency, 1180 sets for slight degradation, and 620 sets for significant degradation. The test sample set contains 2000 sets of boiler operation samples. The traditional fixed threshold method, the ordinary single-condition regression early warning method, and the method of this invention all use the same training and test data. The labels of the training samples are generated jointly by historical energy efficiency calculation records and manual review records, including multi-channel normal energy efficiency benchmark labels, measurement point reconstruction labels, operating condition branch labels, energy flow balance labels, and counterfactual benchmark labels, so that the model training objective is consistent with the on-site energy efficiency anomaly scenario.

[0032] During the simulated operation cycle, the system accessed 6280 measurement points through the SIS server read-only interface, with a sampling period of 1 second. The raw data included 37 quality defects, 14 communication interruption points, 8 manually added defects, 11 upper boundary out-of-bounds defects, 7 lower boundary out-of-bounds defects, 19 positive mutation defects, 16 reverse mutation defects, 13 frozen points, and 46 retransmission records. The system read the numerical boundaries, rate of change boundaries, frozen point judgment duration, and missing sampling judgment duration according to the measurement point coding table, and generated a reliable boiler operating status frame after removing defects and frozen points. The traditional fixed threshold method only removes quality defects, resulting in 23 false triggers caused by mutation points and frozen points in 2000 test samples, while the method of this invention resulted in 5 false triggers.

[0033] In a medium-to-high load operation segment, the boiler load is 452 MW, the rated load is 670 MW, the load factor is 67.46%, the main steam flow rate is 1346 t / h, the rated main steam flow rate is 1978 t / h, the flow rate is 68.05%, the coal quality is classified as medium calorific value coal, and the heating status is not in operation. The system divides the current state into a medium-to-high load segment and a medium-to-high flow segment. The combustion identification window uses 30 consecutive sampling cycles. The primary air pressure fluctuation range is 0.32 kPa, the primary air flow rate fluctuation range is 4.9 t / h, the secondary air damper opening fluctuation range is 1.8%, the furnace negative pressure fluctuation range is 22 Pa, and the flue gas oxygen fluctuation range is 0.19%. The median of the combustion parameter fluctuation range under the same operating condition is multiplied by 1.2 to form the combustion stability judgment threshold, and the current window is marked as a combustion stable state.

[0034] Subsequently, the secondary air damper opening increased from 46.2% to 49.8%, and the induced draft fan speed increased from 735 r / min to 756 r / min. The changes in controllable quantities all exceeded the action amplitude threshold by 2%, and the system generated control adjustment event records and control adjustment event windows. The flue gas temperature increased from 126.7 ℃ to 132.9 ℃, the flue gas oxygen content increased from 3.23% to 4.10%, the fan active power increased from 3890 kW to 4235 kW, and the desuperheating water flow rate increased from 30.6 t / h to 36.8 t / h. The traditional fixed threshold method has an alarm limit of 150 ℃ for flue gas temperature, 6.0% for flue gas oxygen content, and 4800 kW for fan power. All parameters did not exceed the limits, and the traditional method did not trigger an alarm.

[0035] This invention's method binds operating conditions to the boiler state sequence input into an improved MTGNN model. The operating condition routing expert graph temporal layer matches high-load expert branches, and the routing weight gated channel outputs weights of 0.71 for high-load expert branches, 0.18 for high-load expert branches, 0.08 for medium-low-load expert branches, and 0.03 for low-load expert branches. The boiler energy flow constraint graph convolutional layer establishes the boiler measuring point map and boiler measuring point association matrix, propagating along the combustion side to the steam-water side, combustion side to the flue gas side, and auxiliary machine side to the combustion side. The energy efficiency constraint correction slot calculates the normalized energy flow constraint deviation as 0.087. The counterfactual energy efficiency benchmark reconstruction layer calculates the perturbation intensity of the flue gas temperature measuring point as 3.42 and the perturbation intensity of the flue gas oxygen measuring point as 3.18, both greater than the perturbation intensity threshold of 3, completing the normal state replacement and generating a multi-channel energy efficiency benchmark.

[0036] Under the same operating conditions, the baseline for flue gas heat loss is 127.4 ℃, the oxygen content deviation from the baseline is 3.35%, the baseline for desuperheating water consumption is 0.023, and the baseline for fan power consumption is 8.62 kW / MW. The measured flue gas temperature is 133.1 ℃, the flue gas oxygen content is 4.08%, the desuperheating water consumption index is 0.027, and the fan power consumption index is 9.43 kW / MW. After channel-by-channel difference analysis, the residuals for flue gas heat loss are 5.7 ℃, oxygen content deviation is 0.73%, desuperheating water consumption is 0.004, and fan power consumption is 0.81 kW / MW. After residual confidence normalization, the multi-channel confidence residuals are 2.36, 2.11, 1.48, and 1.92, respectively, and the comprehensive risk residual is 2.05.

[0037] The system performs residual stripping on the control and regulation event window. The overlap time between the flue gas heat loss residual and the control window is 31 s, the residual duration is 65 s, and the time overlap ratio is 0.477; the continuous matching ratio between the action duration and the residual duration is 0.646; the residual stripping weight is 0.545. After stripping, the short-time residual peak value decreases from 2.91 to 1.32. The ordinary single-condition regression early warning method does not perform control action residual stripping, and in the test sample, the regulation action caused 31 false alarms, while the method of this invention caused 7.

[0038] The Page-Hinkley algorithm performs drift detection on the directional weighted confidence residual and the comprehensive risk residual. The positive offset weights for the flue gas heat loss channel are 1.20, 1.10, and 1.10 respectively, with a drift tolerance constant of 0.01. After graph correlation propagation, the collaborative drift statistics for the flue gas heat loss channel are 2.74, for the oxygen deviation channel 2.31, for the fan power consumption channel 1.98, and for the comprehensive risk drift 2.46. In the medium-to-high load segment, the collaborative asymmetric drift reaches 3.16, exceeding the abnormal drift threshold of 3.00 and persisting for three consecutive sampling periods. The energy flow consistency ratio is 0.75, higher than 0.50, generating a Level 1 anomaly warning.

[0039] In the counterfactual contribution verification, the system set the confidence residual of each channel to 0 and recalculated the drift. After replacement, the drift of the flue gas heat loss channel decreased from 3.16 to 1.92, with an abnormal contribution value of 1.24; the drift of the oxygen deviation channel decreased to 2.31, with an abnormal contribution value of 0.85; the drift of the fan power consumption channel decreased to 2.66, with an abnormal contribution value of 0.50; and the drift of the desuperheating water consumption channel decreased to 2.91, with an abnormal contribution value of 0.25. The abnormal contribution ranking results show that the main abnormal channels are the flue gas heat loss channel, the oxygen deviation channel, and the fan power consumption channel, and the abnormal sources correspond to combustion air distribution deviation and increased heat loss on the flue gas side.

[0040] Comparative test results show that, out of 2000 test samples, the traditional fixed threshold method identified 312 energy efficiency anomalies, with 214 missed detections and 18 false alarms, achieving an accuracy of 83.9%, a recall of 59.3%, and an average early warning time of 0 min; the ordinary single-condition regression early warning method identified 438 energy efficiency anomalies, with 88 missed detections and 49 false alarms, achieving an accuracy of 89.9%, a recall of 83.3%, and an average early warning time of 6.5 min; the method of this invention identified 512 energy efficiency anomalies, with 14 missed detections and 11 false alarms, achieving an accuracy of 97.9%, a recall of 97.3%, and an average early warning time of 18.3 min. After the simulated treatment, the secondary air damper opening was reduced by 2.1%, the induced draft fan speed was reduced by 1.8%, the flue gas oxygen content decreased from 4.08% to 3.49%, the flue gas temperature decreased from 133.1 ℃ to 128.4 ℃, the active power of the fan decreased from 4235 kW to 3971 kW, the comprehensive risk residual decreased from 2.05 to 0.84, the graph-cooperative asymmetric drift decreased from 3.16 to 1.12, and the unit coal consumption for power generation decreased from 293.8 g / kWh to 291.9 g / kWh.

[0041] The above description is only a preferred embodiment 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 method for early warning of boiler energy efficiency anomalies based on machine learning, characterized in that, include: Read multi-source operation correlation data of the boiler from the SIS server, perform time calibration and measurement point coding, and generate raw boiler operation data; Perform defect removal, freeze point identification, retransmission marking and sampling alignment on the raw boiler operation data to generate a reliable boiler operation status frame; Based on the trusted boiler operating status frame, identify operating conditions, control and regulation event records, and control and regulation event windows to generate boiler operating condition identification records and operating condition-bound boiler status sequences. An improved MTGNN model is constructed by sequentially inputting the boiler state sequence bound to the operating conditions into the operating condition routing expert graph temporal layer, the boiler energy flow constraint graph convolutional layer, and the counterfactual energy efficiency benchmark reconstruction layer to generate multi-channel energy efficiency residuals. Based on the control and regulation event record and control and regulation event window, control actions are performed to strip the residuals of multi-channel energy efficiency residuals, generating stripped energy efficiency residuals. Based on boiler operating condition identification records, residual confidence normalization is performed on the stripped energy efficiency residuals to generate multi-channel confidence residuals and comprehensive risk residuals. The multi-channel confidence residual and the comprehensive risk residual are input into the Page-Hinkley algorithm, and the graph cooperative asymmetric drift amount is calculated based on the directional asymmetric drift mechanism and the graph association cooperative drift mechanism. Perform energy flow consistency verification and counterfactual contribution verification operations on the graph cooperative asymmetric drift to generate boiler energy efficiency anomaly early warning results.

2. The method for early warning of boiler energy efficiency anomalies based on machine learning according to claim 1, characterized in that, The generation of raw boiler operating data includes: Boiler combustion data, steam and water data, flue gas data, auxiliary machine data, operating condition data, and data quality codes are read through the SIS server read-only interface to generate original records of measurement points; Extract the measurement point name, measurement point number, sampling time, sampled value, and quality identifier to generate a measurement point sampling record; Perform unified time axis calibration on the sampling records of the measurement points based on the sampling time to generate time-calibrated sampling records; Based on the measurement point number, the measurement point code and data source mark are executed to generate the original boiler operation data.

3. The method for early warning of boiler energy efficiency anomalies based on machine learning according to claim 1, characterized in that, The generation of a reliable boiler operating status frame includes: Read the measurement point code, sampling time, sampling value and quality identifier from the original boiler operation data, and generate measurement point verification records based on the measurement point code matching measurement point boundary configuration; Based on the quality identifier, sampled value, and measurement point boundary configuration, perform bad point removal to generate valid measurement point records; Freeze point identification is performed based on the repetition status and rate of change of adjacent sampled values ​​in the effective measurement point record, and supplementary transmission marking is performed based on the adjacent sampling time interval to generate a reliable measurement point sequence. A unified sampling time axis is generated based on the sampling time, and sampling alignment is performed on the reliable measurement point sequence to generate a reliable boiler operating status frame.

4. The method for early warning of boiler energy efficiency anomalies based on machine learning according to claim 1, characterized in that, The generation of boiler operating condition identification records and operating condition-bound boiler state sequences includes: Read the load value, main steam flow, coal quality mark, heating status, combustion parameters and controllable execution quantities from the trusted boiler operating status frame, and generate operating condition identification input records; Based on load value, main steam flow rate, coal quality label and heating status, the operating conditions are segmented and the operating conditions are generated. The combustion stability state is identified based on the fluctuation amplitude and duration of combustion parameters, and the operating conditions and combustion stability state are associated and stored to generate a boiler operating condition identification record; Based on the change magnitude, change direction and occurrence time of the controllable execution quantity, control adjustment event records are identified, and control adjustment event windows are generated according to the occurrence time, sampling period and event duration; The boiler operating condition identification record, control and regulation event record, and control and regulation event window are bound to the trusted boiler operating status frame to generate an operating condition-bound boiler status sequence.

5. The method for early warning of boiler energy efficiency anomalies based on machine learning according to claim 1, characterized in that, The generation of multi-channel energy efficiency residuals includes: An improved MTGNN model is constructed, which includes a time-series layer of operating condition routing expert graph, a convolutional layer of boiler energy flow constraint graph, and a counterfactual energy efficiency benchmark reconstruction layer. Read the boiler operating condition identification record, trusted boiler operating status frame and measurement point code from the boiler state sequence bound to the operating condition, and generate measurement point time sequence embedding and operating condition embedding by inputting the encoding end; The measurement point timing embedding and the operating condition embedding are input into the operating condition routing expert diagram timing layer. The corresponding expert diagram timing branch is selected according to the operating condition to generate the operating condition timing representation. A boiler measurement point map is constructed based on measurement point coding and boiler energy flow relationship. The boiler measurement point map and the time-series representation of the operating condition map are input into the convolutional layer of the boiler energy flow constraint map to generate an energy flow constraint boiler representation. The energy flow-constrained boiler characterization is input into the counterfactual energy efficiency benchmark reconstruction layer. The counterfactual boiler characterization is generated according to the measurement point disturbance shielding and normal state replacement. A multi-channel energy efficiency benchmark is generated through the residual output terminal. Extract actual energy efficiency indicators from the reliable boiler operating status frame, perform channel-by-channel difference calculation between the actual energy efficiency indicators and the multi-channel energy efficiency benchmark, and generate multi-channel energy efficiency residuals. The improved MTGNN model was trained using a combination of energy efficiency benchmark prediction error, graph time series representation reconstruction error, operating condition routing allocation error, energy flow constraint deviation, and counterfactual benchmark reconstruction error as the joint training objective. The network parameters of the operating condition routing expert graph time series layer, the boiler energy flow constraint graph convolutional layer, and the counterfactual energy efficiency benchmark reconstruction layer were updated. When the change in the joint loss value in five consecutive training rounds was less than 0.001, the improved MTGNN model was considered to have completed convergence training.

6. The method for early warning of boiler energy efficiency anomalies based on machine learning according to claim 1, characterized in that, The improved MTGNN model includes: The MTGNN model retains the core of adaptive graph learning structure, temporal convolution structure, hybrid hop number graph propagation structure, hop connection path, and prediction output structure. The input representation, operating condition routing, energy flow propagation, and counterfactual reconstruction are structurally modified to address the boiler energy efficiency anomaly early warning scenario. Set up an input encoding terminal at the model input end, read the reliable boiler operating status frame, boiler operating condition identification record and measurement point code in the boiler state sequence bound to the operating condition, and encode the measurement point sample value, measurement point type, sampling time and operating condition into measurement point time sequence embedding and operating condition embedding; A time-series layer of the operating condition routing expert graph is added at the time-series feature extraction location. The time-series branches of the expert graph are selected based on the operating conditions, combustion stability and heating status to form the time-series representation of the operating condition graph. The original graph convolution propagation structure is modified to form a boiler energy flow constraint graph convolution layer. The boiler measurement point graph and boiler energy flow relationship are embedded into the graph propagation path to constrain the energy efficiency coupling relationship between the combustion side, steam-water side, flue gas side and auxiliary machine side. A counterfactual energy efficiency benchmark reconstruction layer is added before the predicted output. The abnormal disturbance channel is shielded and replaced and the normal state is reconstructed to generate a counterfactual boiler characterization. The predicted output structure is transformed into a residual output terminal, outputting a multi-channel energy efficiency benchmark, and the difference between the benchmark and the actual energy efficiency index is calculated to generate a multi-channel energy efficiency residual.

7. The method for early warning of boiler energy efficiency anomalies based on machine learning according to claim 1, characterized in that, The generation of energy efficiency residuals after stripping includes: Read the multi-channel energy efficiency residuals, control and regulation event records, and control and regulation event windows; establish a residual event alignment table according to the sampling time; and associate and store each residual channel with the corresponding control and regulation event window. Based on the residual event alignment table, extract the residual peak value, residual start and end time, residual duration, residual change direction and residual channel identifier within the control and adjustment event window to generate window residual feature records; Based on the control and regulation event records, the direction of change, magnitude of change, start and end time of action, and duration of action of controllable execution quantity are extracted. The direction of change of controllable execution quantity is matched with the window residual feature records, and the start and end time of action is matched with the start and end time of residual to generate control-related residual tags. The residual stripping weight is determined based on the control-related residual label and residual duration. Weighted reduction is performed on the residual channels with control-related residual labels in the multi-channel energy efficiency residuals, and the residual channels without matching control-related residual labels are retained to generate stripped energy efficiency residuals.

8. The method for early warning of boiler energy efficiency anomalies based on machine learning according to claim 1, characterized in that, The generation of multi-channel confidence residuals and comprehensive risk residuals includes: Read the boiler operating condition identification record and the energy efficiency residual after stripping, and generate a residual benchmark matching record based on the operating condition matching operating condition residual benchmark table. Extract the mean of normal residuals, residual dispersion, and confidence interval boundaries for each residual channel from the residual benchmark matching record to generate the residual confidence benchmark record; The difference between the stripped energy efficiency residual and the mean of the normal residual is calculated, and the difference result is divided by the sum of the residual dispersion and the stability constant to generate multi-channel confidence residuals. Based on the boiler operating condition identification records, the energy efficiency weights of each residual channel are determined, and weighted fusion is performed on the confidence residuals of multiple channels to generate comprehensive risk residuals.

9. The method for early warning of boiler energy efficiency anomalies based on machine learning according to claim 1, characterized in that, The computational graph cooperative asymmetric drift includes: Read the multi-channel confidence residuals, comprehensive risk residuals, and boiler measuring point correlation matrix; extract the residual channel identifiers from the multi-channel confidence residuals; and establish a drift detection input record table according to sampling time, sampling period, residual channel identifiers, and measuring point codes. Based on the residual channel identifier matching offset weight configuration table, read the forward offset weight and reverse offset weight, perform directional asymmetric weighting on the multi-channel confidence residuals, and generate directional weighted confidence residuals; Input the direction-weighted confidence residual and the comprehensive risk residual into the Page-Hinkley algorithm to calculate the residual running mean, cumulative offset, historical minimum cumulative offset and single-channel drift statistic for each channel. Based on the boiler measuring point correlation matrix and measuring point coding, graph correlation propagation is performed on the drift statistics of single channels, and the drift statistics of the correlation residual channels are fused according to the measuring point correlation strength to generate channel collaborative drift statistics; Based on the boiler operating condition identification record and the residual channel identifier matching operating condition energy efficiency weight table, the channel collaborative drift statistics and the drift statistics corresponding to the comprehensive risk residual are weighted and fused to generate the graph collaborative asymmetric drift quantity and the complete drift statistical process record table.

10. A method for early warning of boiler energy efficiency anomalies based on machine learning according to claim 1, characterized in that, The generated boiler energy efficiency anomaly early warning results include: Read the graph collaborative asymmetric drift amount, complete drift statistical process record table, boiler operating condition identification record and multi-channel confidence residual, and generate an energy flow direction verification table based on the boiler operating condition identification record; Based on the energy flow direction verification table, an energy flow consistency verification operation is performed on the graph cooperative asymmetric drift, generating the verified drift and energy flow consistency flag; Based on the boiler operating condition identification record and the drift threshold record, the verified drift amount is compared with the drift threshold record to generate an energy efficiency anomaly trigger record; Based on the energy efficiency anomaly trigger record and the complete drift statistical process record table, perform counterfactual contribution verification operation, perform channel replacement and drift recalculation on the multi-channel confidence residuals, and generate an anomaly contribution ranking table. Boiler energy efficiency anomaly early warning results are generated based on energy efficiency anomaly trigger records, energy flow consistency markers, and anomaly contribution sorting tables.

Images