A method and system for analyzing multi-parameter coupled data of a pipeline of a thermal power unit

Abstract

This invention relates to the field of thermal power unit operation monitoring technology, and discloses a method and system for multi-parameter coupled data analysis of thermal power unit pipelines. The method includes: acquiring time-series data; identifying whether the unit is in a rapid load increase transient condition based on data change characteristics and unit load command feedforward information; determining the physical response time-series offset when a transient condition is identified; using this offset to perform time-series alignment processing on the time-series data to construct a reconstructed sample set; calculating the correlation evaluation index between pressure and temperature based on the reconstructed sample set; judging the correlation evaluation index in conjunction with the operating condition context information, and outputting the analysis results. This invention solves the problem of correlation analysis distortion caused by parameter time-series misalignment during rapid load changes by dynamically identifying transient conditions and compensating for the physical response time lag between pressure and temperature, significantly improving the accuracy and reliability of the main steam pipeline condition assessment of thermal power units.

Description

Technical Field

[0001] This invention relates to the field of thermal power unit operation monitoring technology, and in particular to a method and system for multi-parameter coupled data analysis of thermal power unit pipelines. Background Technology

[0002] In the operational monitoring of main steam pipelines in thermal power units, existing technologies typically involve a distributed control system that collects temperature and pressure data at fixed timestamps and calculates correlation coefficients or cross-power spectral density using a sliding window. When the correlation index falls below a preset threshold, the system determines that the parameter correlation is abnormal and triggers thermal fatigue or stress alarms. However, under transient conditions of rapid load increases in the unit, the main steam pressure can respond to the new operating condition within seconds, while the main steam temperature lags behind by tens of seconds due to the thermal inertia of the pipeline metal. In this case, if data is still bound at the same timestamp, the pressure data that has responded to the new operating condition will be forcibly paired with the temperature data that is still in the previous steady state, resulting in data misalignment, calculating a false decrease in correlation strength, and thus triggering false alarms. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention discloses a multi-parameter coupled data analysis method and system for thermal power unit pipelines. It aims to solve the technical problem that, under the transient operating conditions of rapid load increase in thermal power units, the asynchronous physical response sequence of main steam pressure and temperature is ignored, leading to mismatched data pairing and false reductions in correlation strength and abnormal false alarms.

[0004] The technical solution of the present invention is as follows: In a first aspect, this invention discloses a method for multi-parameter coupled data analysis of pipelines in thermal power units, the method comprising: Acquire time-series data of main steam pressure and main steam temperature in the main steam pipeline of thermal power units; Based on the changing characteristics of time series data and combined with the feedforward information of unit load commands, it is possible to identify whether the main steam pipeline of the thermal power unit is in a rapid load increase transient condition. In the case of a rapid load increase transient condition, determine the physical response time offset between the main steam pressure and the main steam temperature as the condition changes dynamically. By utilizing the physical response time offset, time series data are time-series aligned to construct a reconstructed sample set reflecting the same physical evolution stage; The correlation evaluation index between main steam pressure and main steam temperature was calculated based on the reconstructed sample set. By combining the contextual information of the rapid load increase transient condition, the associated evaluation indicators are determined, and the output is a state analysis result including the current operating condition, physical response time offset, and anomaly determination conclusion.

[0005] This technical solution enables the dynamic identification of transient operating conditions and the timing alignment of pressure and temperature, eliminating the effects of physical response lag caused by thermal inertia, thereby accurately assessing parameter correlation and avoiding false alarms.

[0006] Furthermore, after acquiring the time-series data of the main steam pressure and main steam temperature of the main steam pipeline of the thermal power unit, the method also includes: An algorithm combining median filtering and physical extreme value constraints is used to clean time series data. When the value of a sampling point exceeds the physical extreme value range under the current working condition, the sampling point is determined to be an abnormal sampling point and is replaced by linear interpolation using adjacent valid sampling points. The rate of change and acceleration of change of time series data are calculated using a difference algorithm with a smoothing factor to characterize the physical evolution trend of the main steam pipeline of thermal power units.

[0007] Furthermore, based on the changing characteristics of time series data and combined with feedforward information from unit load commands, the steps to identify whether the main steam pipeline of a thermal power unit is in a rapid load increase transient condition include: Monitor the rate of change of unit load commands and the rate of change of main steam pressure; When the rate of change of the unit load command exceeds the preset load regulation rate threshold within a set period, and the rate of change of the main steam pressure exceeds the preset pressure response threshold, the main steam pipeline of the thermal power unit is determined to have entered a rapid load increase transient condition.

[0008] Furthermore, in the case of a rapidly increasing load transient condition, the steps for determining the physical response time offset between the main steam pressure and the main steam temperature as the condition dynamically changes include: Using a cross-correlation analysis algorithm with a restricted sliding window, the cross-correlation coefficient sequence between main steam pressure and main steam temperature within the current sliding window is calculated; The initial offset is determined based on the peak position of the cross-correlation coefficient sequence; The Kalman filter algorithm is used to smooth the initial offset and obtain the physical response timing offset that changes dynamically with the operating conditions.

[0009] Furthermore, the step of determining the initial offset based on the peak position of the cross-correlation coefficient sequence includes: Extract local maxima points in the cross-correlation coefficient sequence whose amplitudes exceed the baseline threshold, and construct a candidate peak set; The comprehensive confidence score of each candidate peak is calculated based on the normalized amplitude of each candidate peak and the difference between the delay time corresponding to each candidate peak and the historical continuity benchmark; the historical continuity benchmark is the physical response time offset after smoothing and tracking at the previous moment. The delay time corresponding to the candidate peak with the highest overall confidence score is selected as the initial offset.

[0010] Furthermore, the method also includes: If the overall confidence score of all candidate peaks is lower than the safety lower limit, the initial offset is set to the physical response time series offset of the previous time step, and a low observation confidence indicator is generated. The steps for smoothing the initial offset using the Kalman filter algorithm include: When a low observation confidence flag is received, the observation noise covariance matrix is ​​increased to enhance the dependence of the filtering update process on the state prediction model.

[0011] Furthermore, the steps of using physical response time-series offsets to perform time-series alignment processing on time-series data to construct a reconstructed sample set reflecting the same physical evolution stage include: For each sampling moment within the sliding time window, the main steam temperature at that sampling moment is used as the alignment anchor point. Based on the time difference between temperature and pressure, represented by the physical response time offset, the sampling moment is moved forward along the time axis by the duration corresponding to the time difference, thus obtaining the target pressure sampling moment that is in the same physical evolution stage as the main steam temperature. The main steam pressure channel in the time series data is fitted using the cubic spline interpolation algorithm. The main steam pressure value corresponding to the target pressure sampling time is calculated. The main steam temperature at the sampling time is bound to the main steam pressure value to reconstruct a pair of sample data in the same physical evolution stage. The reconstructed sample set, aligned with the physical evolution stages, is composed of sample data corresponding to each sampling time.

[0012] Furthermore, the correlation evaluation index is the residual between the actual observed value and the theoretical prediction value obtained based on the pre-trained long short-term memory network; The steps for determining the relevant evaluation indicators by combining the contextual information of the rapid load increase transient condition, and outputting the state analysis results including the current operating condition, physical response time series offset, and anomaly determination conclusions, include: The category identifier of the rapid load increase transient condition and the physical response time offset are used as inputs for fuzzy logic reasoning to determine the current anomaly judgment threshold correction coefficient. The preset baseline threshold is adjusted using a correction coefficient to obtain an adaptive judgment threshold; Based on the comparison between the residual and the adaptive decision threshold, the state analysis results are output.

[0013] Furthermore, the method also includes: After identifying a rapid load increase transient condition, determine whether the estimated confidence level of the physical response timing offset is lower than a preset threshold. If the estimated confidence level is lower than the preset threshold, the deterministic anomaly determination in the current time window is paused, and the unconfirmed associated status flag affected by transient delay is output.

[0014] Secondly, the present invention also discloses a multi-parameter coupled data analysis system for thermal power unit pipelines, comprising: The data acquisition module is used to acquire time-series data of the main steam pressure and main steam temperature of the main steam pipeline of the thermal power unit; The operating condition identification module is used to identify whether the main steam pipeline of the thermal power unit is in a rapid load increase transient operating condition based on the changing characteristics of time series data and the feedforward information of the unit load command. The offset determination module is used to determine the physical response timing offset between the main steam pressure and the main steam temperature as the operating condition changes dynamically when the condition is identified as a rapid load increase transient condition. The time-series alignment module is used to perform time-series alignment processing on time-series data using physical response time-series offsets, so as to construct a reconstructed sample set that reflects the same physical evolution stage; The index calculation module is used to calculate the correlation evaluation index between main steam pressure and main steam temperature based on the reconstructed sample set. The status analysis module is used to determine the relevant evaluation indicators by combining the context information of the rapid load increase transient condition, and output the status analysis results including the current status, physical response time offset and anomaly determination conclusion.

[0015] In summary, this invention provides a multi-parameter coupled data analysis method and system for thermal power unit pipelines. The method addresses the data misalignment problem caused by differences in the main steam pressure and temperature response rates during rapid load increases in thermal power units by introducing a dynamic physical response time-series offset calculation and time-series alignment mechanism. This method can identify transient conditions in real time, obtain dynamic time delays through cross-correlation analysis and Kalman filtering smoothing, and reconstruct the data using cubic spline interpolation, ensuring the synchronization of correlation analysis samples in the physical evolution stage. Combined with adaptive threshold determination using fuzzy logic reasoning, it effectively eliminates the false correlation strength reduction caused by time-series asynchrony, significantly improving the accuracy and reliability of pipeline thermal fatigue and stress anomaly monitoring, and avoiding false alarms. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a multi-parameter coupled data analysis method for thermal power unit pipelines provided in an embodiment of the present invention.

[0017] Figure 2This is a schematic diagram of a multi-parameter coupled data analysis system for thermal power unit pipelines provided in an embodiment of the present invention.

[0018] Labeling Explanation: 210, Data Acquisition Module; 220, Operating Condition Identification Module; 230, Offset Determination Module; 240, Timing Alignment Module; 250, Index Calculation Module; 260, Status Analysis Module. Detailed Implementation

[0019] The technical solutions of this invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some, not all, of the embodiments of this invention. The components of this invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0020] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] In the operation monitoring of the main steam pipeline of thermal power units, a distributed control system typically collects temperature and pressure data at fixed timestamps and calculates correlation coefficients using a sliding window to determine if the parameter correlation is abnormal. When the correlation is lower than a preset threshold, the system triggers an alarm for thermal fatigue or abnormal stress. However, under transient conditions of rapid load increase in the unit, the main steam pressure responds quickly, while the main steam temperature lags significantly due to the thermal inertia of the pipeline metal. In this case, if data is still bound at the same timestamp, the pressure data that has responded to the new operating condition will be forcibly paired with the temperature data that is still in the previous stage, resulting in data misalignment, a false decrease in correlation strength, and thus false alarms. To solve the problem of false alarms caused by the above-mentioned time sequence misalignment, an analysis mechanism that can adapt to the physical response delay needs to be introduced. That is, the system can dynamically identify the actual time difference between pressure and temperature according to the operating condition and reconstruct the data pairing relationship accordingly, so that the correlation evaluation results truly correspond to the same thermodynamic evolution stage, rather than just the same sampling time.

[0022] Firstly, please see Figure 1 This invention discloses a multi-parameter coupled data analysis method for pipelines in thermal power units, the method comprising: S1. Obtain time series data of main steam pressure and main steam temperature of the main steam pipeline of the thermal power unit; S2. Based on the changing characteristics of time series data and combined with the feedforward information of unit load command, identify whether the main steam pipeline of thermal power unit is in a rapid load increase transient condition. S3. When the condition is identified as a rapid load increase transient condition, determine the physical response time offset between the main steam pressure and the main steam temperature as the condition changes dynamically. S4. Using the physical response time offset, perform time series data alignment processing to construct a reconstructed sample set that reflects the same physical evolution stage; S5. Calculate the correlation evaluation index between main steam pressure and main steam temperature based on the reconstructed sample set; S6. Combine the context information of the rapid load increase transient condition to determine the associated evaluation indicators, and output the state analysis results including the current operating condition, physical response time offset, and anomaly determination conclusion.

[0023] Specifically, the rapid load increase transient condition refers to the non-steady-state transition process in which, after receiving an instruction to increase power generation, the fuel, feedwater, and other regulating mechanisms deployed within the system rapidly activate, causing significant changes in unit operating parameters within a short period. This process typically has two characteristics: first, there is a clear intention to increase load on the control side; and second, an observable response has already appeared on the physical side. The physical response time series offset refers to the time difference between two physical quantities (such as pressure and temperature) reaching the same evolutionary stage on the time axis under the same external instruction or operating condition change trigger, due to the different transmission mechanisms and inertial characteristics of these quantities. This time difference does not require the two parameters to have the same value, but rather that they are adjacent corresponding stages in the same round of operating condition change. For example, if the pressure has already entered the plateau transition phase after rising, while the temperature is still in the initial stage of rising, it indicates that they are not yet in the same evolutionary stage; only by shifting the pressure sequence back to the historical direction for an appropriate period can pressure data corresponding to the current temperature change stage be found. This shift period is the physical response time series offset.

[0024] A reconstructed sample set refers to a data set that breaks down the original data structure aligned by absolute timestamps, shifts and misaligns the time series of different parameters based on the calculated time difference, and then recombines them to form a data set that reflects the same physical change stage. The core of this data set is not that it is collected at the same time, but that it matches the same physical stage. For example, a temperature sample at a certain time can be combined with a pressure sample from an earlier time to form a new set of analytical samples, as long as both reflect the corresponding change stage in the same round of load increase process, they can be included in the reconstructed sample set.

[0025] In the operational logic of the above overall technical solution, the system first needs to continuously acquire basic operational data from the underlying control system. Specifically, a communication connection can be established with the distributed control system (DCS) of the thermal power unit via industrial Ethernet. The measured values ​​from the main steam pressure sensor and the main steam temperature sensor are read in real time at a fixed sampling frequency, and these timestamped values ​​are stored in a time-series database in chronological order, forming continuous time-series data. To ensure the usability of subsequent time-series analysis, preferably, the main steam pressure and main steam temperature should be time-stamped using a unified clock source, or a unified time synchronization correction should be performed by the data acquisition service before entering the time-series database, to avoid additional human-induced timing deviations caused by clock drift from different acquisition channels. Furthermore, in addition to the measured values ​​and timestamps, the time-series data can also simultaneously record the measurement point identifier, data quality flag, acquisition status word, and source system identifier. This is because common occurrences in industrial settings such as communication interruptions, measurement point maintenance, instrument switching, and signal freezing can all affect the reliability of subsequent offset identification. By retaining these auxiliary fields during the data entry stage, the system can remove data segments with abnormal quality indicators before subsequent analysis, thus avoiding the misuse of invalid data for operating condition identification and time series alignment.

[0026] After acquiring the basic data, the system needs to determine the current operating status of the unit. Specifically, it can read the unit load command values ​​issued by the Automatic Generation Control (AGC) system in real time and make logical judgments based on the current absolute values ​​of the main steam pressure time series data. For example, when the unit load command value is detected to jump to a certain set high load range, and the main steam pressure value also crosses the corresponding steady-state baseline, the system triggers a state switch, identifying that the thermal power unit's main steam pipeline is currently in a rapid load increase transient condition. To make this identification process more feasible, the high load range and steady-state baseline are jointly determined based on the unit's historical operating data, design operating condition segments, and operating procedures. Specifically, multiple stable operating periods can be extracted from the unit's historical database, and the stable distribution range of main steam pressure under different load levels can be statistically analyzed. These ranges are then organized into a steady-state reference table divided by load range. During operation, when the AGC load command enters a target range, the system calls the main steam pressure steady-state reference range corresponding to that target range and determines whether the current pressure has significantly deviated from the original steady-state range and is evolving towards the target range. If this condition is met, it indicates that not only has the control command changed, but the main steam system has also begun to deviate from the original equilibrium point. Therefore, it can be used as an auxiliary criterion for entering a rapid load increase transient operating condition.

[0027] Once the transient operating condition is confirmed, the core step lies in quantifying the asynchronous characteristics between pressure and temperature. More specifically, the system extracts a fixed-length time series data segment after the current operating condition occurs. For the main steam pressure and temperature sequences, it calculates their correlation at different delay time steps, generating a curve showing the correlation index changing with the delay time. The fixed length preferably satisfies two conditions: first, it should cover the complete transition period from the start of load increase to when both pressure and temperature show significant responses; second, it should not be too long to avoid mixing too much irrelevant steady-state data into the analysis window, weakening its ability to characterize the current transient features. In a feasible processing flow, the system first uses the operating condition identification time as a starting point, retaining a small segment of baseline data to characterize the original steady-state background, and then continuously accumulating the current transient data to form an analysis window for offset search. Subsequently, the system sequentially shifts the main steam pressure sequence towards the historical direction (i.e., shifts it to the left, matching the current temperature with the pressure at an earlier time) according to multiple candidate delay amounts. At each candidate delay amount, the shifted pressure sequence is paired and compared with the unshifted temperature sequence to obtain the corresponding correlation index. To avoid misjudging local extrema due to occasional noise, preferably, not only is the correlation magnitude of a single delay point examined, but also the continuity and stability of the correlation changes among several adjacent delay values ​​near that delay point. If a candidate delay value corresponds to the highest correlation, and the correlation within its vicinity remains relatively smooth without drastic jumps, then that candidate delay value can be determined as the physical response time offset of the main steam temperature lagging behind the main steam pressure under the current operating condition. Furthermore, during continuous operation, this offset is not permanently fixed after being determined, but can be updated as the analysis window scrolls. When the system detects a new load increase stage, a change in the load increase rate, or a significant change in the boiler heating state, the above search process can be re-executed to obtain a dynamic offset that matches the current operating condition. The significance of this approach is that the thermal inertia performance during different load increase processes may not be completely consistent, and the dynamic update mechanism can more realistically reflect the actual response characteristics of the field equipment.

[0028] After obtaining the offset, the system performs a crucial time-series alignment operation. Since temperature lags behind pressure, the current temperature at time T corresponds to the pressure at an earlier time T-τ. Therefore, the pressure sequence is shifted historically by a time interval of τ and paired with the current temperature, where τ is the physical response time-series offset, representing the time difference between the main steam temperature and the main steam pressure. In some implementations, the system uses the main steam temperature timeline as a reference and shifts the entire main steam pressure time series historically by the aforementioned determined physical response time-series offset. For example, the main steam temperature data at time T is paired and bound with the main steam pressure data at time T minus the physical response time-series offset. After this shifting and pairing process, the data that were originally misaligned in absolute time are recombined, constructing a reconstructed sample set where pressure and temperature are synchronized in the physical change process. To ensure the quality of the reconstructed sample set, the following constraints are preferably implemented during the actual execution: First, for the first and last boundary data of sequences that cannot find corresponding pairing points after translation, they are not directly involved in the association analysis but are removed as invalid boundary samples; Second, when the sampling period is fixed but there are some missing points, the principle of temporal proximity can be used to supplement points or interpolate, provided that the quality of adjacent data is confirmed to be valid, in order to maintain the continuity of the reconstructed sample set; Third, if the number of valid paired samples in a certain analysis window is too small, the association evaluation results of that window are temporarily suspended, and the analysis is carried out after subsequent data is supplemented. The reason for this setting is that temporal translation essentially changes the pairing relationship of the original samples. If the missing boundary and missing point situations are not handled, new analytical biases can easily be introduced when the number of samples is insufficient or the pairing is incomplete.

[0029] Subsequently, the system performs deep correlation analysis based on this reconstructed sample set. Specifically, the main steam pressure sequence and unit load command sequence from the reconstructed sample set are input into a pre-trained Long Short-Term Memory (LSTM) network. This network has been pre-trained under historical steady-state and normal transient operating conditions, learning the nonlinear dynamic mapping relationship between main steam pressure, unit load command, and main steam temperature under healthy conditions, and outputting the theoretical prediction value of the current main steam temperature. The system then calculates the dynamic residual sequence between the actual main steam temperature and the theoretical prediction value, using this residual as a correlation evaluation index. If the pipeline is in a healthy state, the actual temperature has a high fit with the predicted value and the residual is small because the input samples have already achieved physical stage alignment; if there are real health problems in the pipeline, the inherent mapping relationship between pressure and temperature is disrupted, and the residual increases significantly. Through calculation based on aligned samples and a deep learning model, the system can remove the interference of normal transient time delays, allowing subtle real anomalies to clearly appear in complex dynamic operating conditions.

[0030] Furthermore, the system can use the associated evaluation indicators in conjunction with the current operating condition context, rather than drawing anomaly conclusions based on a single indicator in isolation. Finally, the system extracts the contextual information of the rapid load increase transient condition, such as the current load target range, the duration of the load increase, whether it is currently in the initial or middle stage of the load increase, and whether the offset is in a stable state, and calls the preset judgment rule base. If the calculated associated evaluation indicator is within the normal fluctuation range corresponding to the context information, it is judged as normal; otherwise, it is judged as abnormal. The preset judgment rule base can be obtained by sorting through historical operation samples. Specifically, it can filter out normal load increase processes confirmed by both operation records and maintenance records from past operation records, statistically analyze the common fluctuation ranges of associated evaluation indicators under different operating condition contexts, and store the operating condition label—indicator normal range—judgment conclusion in the form of rule entries. For example, in a certain type of operating condition with a fast load increase rate and a large offset, a wider normal fluctuation band for the associated evaluation indicator is allowed than in a steady-state operating condition; while in another type of operating condition with a slower load increase rate and a more stable offset, a stricter judgment boundary is adopted. Finally, the system encapsulates the current operating condition identifier, the calculated offset, and the judgment conclusion into a structured data packet and outputs it to the operation monitoring terminal interface. It can also include the start and end time of the current analysis window, the number of samples involved in the analysis, and the data quality status, which is convenient for subsequent traceability and verification.

[0031] In real-world industrial environments, time-series data acquired by sensors are often inevitably affected by electromagnetic interference or equipment vibration, resulting in instantaneous spike noise or invalid glitch data. Directly using these raw, unprocessed sequences for subsequent trend analysis and offset calculations can easily lead to serious biases in the analysis results.

[0032] In some preferred embodiments, after acquiring time-series data of the main steam pressure and main steam temperature of the main steam pipeline of the thermal power unit, the method further includes: An algorithm combining median filtering and physical extreme value constraints is used to clean time series data. When the value of a sampling point exceeds the physical extreme value range under the current working condition, the sampling point is determined to be an abnormal sampling point and is replaced by linear interpolation using adjacent valid sampling points. The rate of change and acceleration of change of time series data are calculated using a difference algorithm with a smoothing factor to characterize the physical evolution trend of the main steam pipeline of thermal power units.

[0033] In this specific implementation scheme, the system first constructs a two-layer data cleaning mechanism. The first layer is physical extreme value constraint: based on the unit design specifications and operating procedures, the absolute physical boundaries of the main steam pressure and temperature are preset. When the sampling point value exceeds this boundary or exhibits abrupt jumps that do not conform to physical continuity, it is marked as an abnormal sampling point, and replaced by linear interpolation using adjacent valid sampling points. The physical extreme value range is determined jointly by design parameters and historical operating data, which can simultaneously retain two levels of constraints: absolute physical boundaries and operating condition reference boundaries. The second layer is median filtering: using a sliding window method, the median value of the data within the window is used to replace the center point value, eliminating random noise that does not exceed the extreme values ​​but exhibits local spikes. The window length is determined based on the sampling frequency and noise duration characteristics, balancing the suppression of glitches with the preservation of true transient change edge characteristics.

[0034] After cleaning, a differential algorithm with a smoothing factor is used to calculate the rate of change and acceleration. In practice, local smoothing fitting is first performed on continuous sampling points, and then the first-order and second-order trends are extracted from the fitted curves. This avoids directly amplifying high-frequency noise through differential calculation, providing a stable basis for subsequent identification of the time-series relationship where pressure precedes and temperature lags.

[0035] Furthermore, in the actual operation of thermal power units, sometimes the dispatching system issues load commands, but due to jammed bottom valves or sluggish combustion system, the unit does not actually execute the load increase action; or during steady-state operation, local disturbances may cause brief pressure fluctuations. If the operating condition is judged solely based on a single command value or parameter exceeding the limit, it is very easy to misidentify transient operating conditions.

[0036] To more rigorously define the operating condition boundaries, optionally, based on the changing characteristics of time series data and combined with feedforward information from unit load commands, the steps to identify whether the main steam pipeline of a thermal power unit is in a rapid load increase transient operating condition include: Monitor the rate of change of unit load commands and the rate of change of main steam pressure; When the rate of change of the unit load command exceeds the preset load regulation rate threshold within a set period, and the rate of change of the main steam pressure exceeds the preset pressure response threshold, the main steam pipeline of the thermal power unit is determined to have entered a rapid load increase transient condition.

[0037] In this specific implementation scheme, the system strictly binds the feedforward control intent with the feedback physical response logically. In practice, the system reads the AGC load command sequence in real time and uses the previously calculated rate of change data to set an evaluation period and two key thresholds: the load regulation rate threshold and the pressure response threshold. The system continuously monitors these two rate indicators. Only when the load command change rate consistently exceeds the load regulation rate threshold within a continuous evaluation period, and the main steam pressure change rate simultaneously exceeds the pressure response threshold, is the AND gate triggered, formally determining that the unit has entered a rapid load increase transient condition. The preset thresholds are obtained through historical sample calibration: three types of samples are extracted from historical operating records: real rapid load increase processes, invalid command processes, and steady-state disturbance processes. Typical distribution ranges of their load command change rates and pressure change rates are statistically analyzed, and the boundary value that best distinguishes the first type from the latter two types of samples is selected as the initial threshold value. After the initial threshold value is determined, online verification is performed during the initial commissioning phase: if invalid commands are frequently misidentified as rapid load increases, the pressure response threshold is increased; if real load increase processes are missed, the corresponding threshold is decreased or the evaluation period is extended. By combining offline processing with online verification, a threshold configuration adapted to the current unit characteristics is formed. Furthermore, to avoid misjudgments triggered by single-point, accidental fluctuations, the two rate conditions mentioned above are required to remain true for a consecutive number of sampling moments, or to remain true continuously within the evaluation period before outputting the operating condition switching result. Through this processing, the system identifies not instantaneous spikes, but rather continuous load increase intentions and execution responses.

[0038] After accurately identifying the rapid load increase transient condition, the system needs to further determine the physical response time series offset between the main steam pressure and the main steam temperature. In practical applications, due to the complex operating environment of thermal power units, the data collected by sensors is often accompanied by high-frequency noise and local fluctuations. If only simple static time difference calculations or single correlation analyses are used, it is easily affected by local noise, causing drastic jumps in the calculated offset. Such jumps do not conform to the physical law of the slow change in the thermal inertia of pipeline metal, which in turn affects the accuracy of subsequent time series alignment.

[0039] To address this issue, a dynamic offset tracking mechanism that can balance local correlation characteristics with global physical continuity is needed.

[0040] Optionally, in the case of a rapidly increasing load transient condition, the step of determining the physical response time offset between the main steam pressure and the main steam temperature as the condition dynamically changes includes: Using a cross-correlation analysis algorithm with a restricted sliding window, the cross-correlation coefficient sequence between main steam pressure and main steam temperature within the current sliding window is calculated; The initial offset is determined based on the peak position of the cross-correlation coefficient sequence; The Kalman filter algorithm is used to smooth the initial offset and obtain the physical response timing offset that changes dynamically with the operating conditions.

[0041] Specifically, the physical response time-series offset refers to the time difference between the measured change in main steam pressure and the corresponding change in main steam temperature during a rapid load increase process. The main steam pressure is measured first, while the main steam temperature changes over time after steam flow, heat exchange, and the release of stored heat in the pipe metal. This time difference is not a fixed constant but changes slowly with variations in boiler combustion intensity, main steam flow rate, desuperheating water participation, pipe wall temperature, and the unit load ramp-up rate. Therefore, obtaining this offset cannot rely solely on correlation analysis at a single moment but requires dynamic tracking over a continuous time axis.

[0042] In this specific implementation scheme, the system first employs a restricted sliding window cross-correlation analysis algorithm to capture local temporal correlations. A restricted sliding window means that the window length is strictly controlled within a range that covers the typical transient response period without introducing excessive steady-state historical data. To more clearly and completely describe the implementation of this step, in one specific embodiment, the main steam pressure and main steam temperature data can be directly obtained from the same time series channel in the distributed control system, historical database, or real-time data acquisition platform. During acquisition, it should be ensured that the two signals have a unified time stamp, preferably using historical archived values ​​under the same sampling period; if there is a clock deviation at different measurement points on site, synchronization can be achieved first by using a unified time stamp or resampling according to the sampling timestamp. For data with isolated missing points, adjacent valid sampled values ​​can be used to fill in the gaps; for data segments with consecutive missing durations exceeding the allowable range, the corresponding window can be directly marked as a low-availability data window and the correlation calculation can be skipped, thereby avoiding contamination of offset identification by missing data.

[0043] Within the current sliding window, the system shifts the main steam pressure sequence at different time steps and calculates the cross-correlation coefficient between the main steam pressure sequence and the main steam temperature sequence after each shift, thus obtaining a cross-correlation coefficient sequence reflecting the change in correlation with the delay time. These different time steps are not unlimited traversals, but rather searches within a pre-defined allowable delay range. This allowable delay range can be determined based on the unit's historical operating experience, the length of the main steam pipeline, the location of measuring points, and the sequential relationship of temperature and pressure responses during typical load increases. For example, several historical rapid load increase records can be retrieved first to observe the approximate time difference between the inflection point of the main steam pressure rise and the inflection point of the main steam temperature rise within a visually identifiable response segment. Then, the ends of this interval can be appropriately widened to form a restricted search range. The purpose of this setting is to avoid including excessively large delays or negative delays that clearly do not conform to physical laws in the candidates, thereby reducing the probability of false peaks.

[0044] Subsequently, the system searches for the peak position in the cross-correlation coefficient sequence. The delay time corresponding to this peak position represents the time difference with the highest correlation between the two components within the current window, and the system uses this as the initial offset. However, due to the presence of local noise, this initial offset may experience unreasonable jumps between adjacent sliding windows. To address this, the system introduces a Kalman filter algorithm to smoothly track the initial offset. The Kalman filter algorithm is a recursive filtering algorithm based on a state-space model. Its core idea is to combine the system's physical prediction model with the current observations to minimize the covariance of the estimation error. In the context of this scenario, the initial offset given by the cross-correlation analysis is equivalent to the time difference observed in the current window, while the result output by the Kalman filter is equivalent to a more reliable final time difference after considering the continuous change law of thermal inertia. Therefore, this filtering step does not change the original physical phenomenon, but rather suppresses non-physical jitter caused by noise and occasional disturbances while preserving the true trend of change.

[0045] In a specific embodiment, the system models the time delay change process caused by the thermal inertia of the pipeline metal as a Kalman filter state transition equation, assuming that the rate of change of the offset is gradual in a short time. Simultaneously, the initial offset calculated above is used as the observation input for the Kalman filter. At each time step, the Kalman filter first predicts the offset at the current time based on the optimal estimate from the previous time step, then calculates the residual between the predicted value and the currently observed initial offset, and corrects the predicted value using the Kalman gain, finally outputting the optimal estimate at the current time step, i.e., the physical response time-series offset that dynamically changes with the operating conditions. For ease of implementation, in an executable example, the state variables of the Kalman filter can be defined as two parts: the current offset and the short-term trend of the offset change. During the initialization phase, the system can use the initial offset obtained through cross-correlation analysis of the first effective window as the initial state value, and set the initial trend of change to zero or the average trend of the offset changes in the previous few windows. The state prediction covariance matrix and the observation noise covariance matrix can be obtained by adjusting based on historical data: if the historical offset changes are relatively stable, the state prediction covariance can be set relatively small; if the cross-correlation output is frequently disturbed, the observation noise covariance can be appropriately increased to enhance the filter's dependence on physical continuity. In actual deployment, multiple segments of historical rapid load increases can be selected for offline playback to observe whether the offset curve after filtering can follow the real slow change trend without excessive lag. Then, the above parameters can be corrected and solidified into the operating parameter table.

[0046] In the process of determining the initial offset using cross-correlation coefficients, real-world data often exhibits multiple similar peaks due to multi-period fluctuations. Simply selecting the peak with the largest amplitude may lead to significant offset deviations due to noise spurious peaks. Therefore, it is necessary to comprehensively consider both the salience of the peak (its prominence relative to the background) and its continuity in temporal evolution (its connection to the offset at the previous moment). Real physical delays drift slowly, while noise spurious peaks jump randomly; introducing continuity constraints can significantly improve the reliability of peak selection.

[0047] In some preferred embodiments, the step of determining the initial offset based on the peak position of the cross-correlation coefficient sequence includes: Extract local maxima points in the cross-correlation coefficient sequence whose amplitudes exceed the baseline threshold, and construct a candidate peak set; The comprehensive confidence score of each candidate peak is calculated based on the normalized amplitude of each candidate peak and the difference between the delay time corresponding to each candidate peak and the historical continuity benchmark; the historical continuity benchmark is the physical response time offset after smoothing and tracking at the previous moment. The delay time corresponding to the candidate peak with the highest overall confidence score is selected as the initial offset.

[0048] In this specific implementation scheme, the system no longer relies solely on the maximum amplitude but establishes a multi-dimensional peak evaluation mechanism. First, the system traverses the cross-correlation coefficient sequence, extracting all local maxima points whose amplitudes exceed a preset benchmark threshold, and includes these points in the candidate peak set. This step effectively filters out small fluctuations with excessively low amplitudes that are clearly background noise. Regarding the method of obtaining the benchmark threshold, in one specific embodiment, multiple cross-correlation coefficient sequences can be extracted from historical normal rapid load increase samples. The distribution range of the true effective peak and noise peak in amplitude can be statistically analyzed, and then the threshold that can better separate the two can be selected as the preset benchmark threshold. If the correlation strength varies among different units and different measurement point combinations, parameter configuration tables can also be established separately according to unit type or measurement point pair. For example, the parameter configuration table can record entries such as unit number, measurement point combination, allowable delay range, benchmark threshold, and safety lower limit. The system automatically reads the corresponding configuration based on the current unit and signal channel during operation. After this processing, the so-called preset is not an abstract given but rather an executable parameter solidified after historical operating condition playback and manual verification.

[0049] Next, the system calculates a comprehensive confidence score for each peak in the candidate peak set. The calculation of the comprehensive confidence score integrates information from two dimensions: first, the significance of the current observation, i.e., the normalized amplitude of the candidate peak; a larger amplitude indicates a stronger mathematical correlation between the two at that delay time; second, the continuity of physical evolution, i.e., the difference between the delay time corresponding to the candidate peak and the historical continuity benchmark. Here, the historical continuity benchmark is set as the physical response time-series offset after smoothing and tracking at the previous moment. Since the thermodynamic process of thermal power units has inertia, the actual physical delay cannot abruptly change between adjacent moments. Therefore, the closer the delay time corresponding to the candidate peak is to the historical continuity benchmark, the greater the likelihood that it represents the actual physical delay. To illustrate this process more clearly and completely, in one implementation example, the system can first perform uniform scaling on the amplitudes of the candidate peaks to ensure they fall within the same evaluation range, and then classify the continuity levels according to the distance between the candidate peak and the historical continuity benchmark, for example, into four categories: highly close, basically close, significantly deviating, and severely deviating, and assign different continuity evaluation values ​​to each. Subsequently, the significance evaluation value and the continuity evaluation value are combined according to preset weights to form a comprehensive confidence score. The weights here can also be obtained by adjusting parameters under historical operating conditions: if the on-site noise is heavy, the continuity weight can be appropriately increased; if the operating conditions change rapidly and the actual offset is prone to significant drift, the significance weight can be appropriately increased to prevent the filtering from being too conservative.

[0050] In a specific embodiment, the system can calculate the overall confidence score using a weighted summation method. For example, a positive weight is assigned to the normalized amplitude, and a negative penalty weight is assigned to the absolute value of the delay time difference. When two peaks with similar amplitudes appear in the cross-correlation coefficient sequence, the system calculates their respective distances from the previous time offset. If one peak has a slightly higher amplitude but its corresponding delay time differs significantly from the historical benchmark, its overall confidence score will be greatly reduced due to the large negative penalty; conversely, the other peak with a slightly lower amplitude but closely following the historical evolution trajectory will obtain a higher overall confidence score. Finally, the system selects the delay time corresponding to the candidate peak with the highest overall confidence score as the initial offset. To facilitate understanding, a typical scenario can be considered: during a rapid load increase phase, due to the simultaneous existence of combustion adjustment and desuperheating water disturbance, two strong peaks may appear in the cross-correlation coefficient sequence, one corresponding to a shorter delay time and the other to a longer delay time. If the timing offset of the physical response after smooth tracking in the previous moment is closer to the shorter delay time, the system will tend to identify the peak value corresponding to the shorter delay time as the true peak value, even if the other peak value is slightly higher. This can effectively suppress the risk of occasional noise dominating peak selection.

[0051] In the process of selecting the initial offset based on the comprehensive confidence score, an extreme situation may be encountered: due to sensor failure, large-scale data loss, or drastic changes in operating conditions, the data quality within the current sliding window is extremely poor, and no peak in the cross-correlation coefficient sequence meets the confidence standard. In this case, if the peak with the highest score, which is still unreliable, is forcibly used as the observation input for Kalman filtering, it will severely pollute the state update process, causing subsequent offset predictions to deviate from the true trajectory.

[0052] To address this extreme observation failure scenario, a filtering update mechanism that adaptively adjusts observation confidence is needed. Low observation confidence does not mean the system has stopped working, but rather that the system recognizes that it cannot extract sufficiently reliable new information within the current window. Therefore, it temporarily increases its trust in historical physical evolution results and reduces its reliance on current observation results.

[0053] In practical applications, this method also includes: If the overall confidence score of all candidate peaks is lower than the safety lower limit, the initial offset is set to the physical response time series offset of the previous time step, and a low observation confidence indicator is generated. The steps for smoothing the initial offset using the Kalman filter algorithm include: When a low observation confidence flag is received, the observation noise covariance matrix is ​​increased to enhance the dependence of the filtering update process on the state prediction model.

[0054] In this specific implementation scheme, the system introduces the concept of a safety lower limit to define the usability boundary of the observation data. After the system calculates the comprehensive confidence score of all candidate peaks, it compares it with the preset safety lower limit. If all scores are lower than the safety lower limit, it indicates that the cross-correlation analysis results within the current window are extremely unreliable and cannot provide valuable observation information. At this time, the system decisively abandons the current observation results, forcibly sets the initial offset to the physical response time-series offset of the previous moment, and simultaneously generates a low observation confidence flag. This flag is equivalent to sending a warning signal of current observation failure to the subsequent Kalman filter module. Regarding the setting method of the safety lower limit, in one specific embodiment, multiple historical fault samples, missing samples, and strong disturbance samples can be selected, and the distribution of the comprehensive confidence scores of candidate peaks under these poor windows can be statistically analyzed. Combined with the score distribution of the normal window, a boundary value that can effectively distinguish between usable and unusable observations can be determined. If many misjudgments are found during field operation, the safety lower limit can be periodically reviewed and adjusted according to the operation records to adapt to seasonal operating conditions or signal quality changes after the measurement points age.

[0055] In the smoothing tracking process of the Kalman filter algorithm, the core of state update lies in balancing the weights of the prediction model and the actual observations. This balance is adjusted through the observation noise covariance matrix and the state prediction covariance matrix. In a specific embodiment, when the Kalman filter module receives a low observation confidence flag, the system dynamically and significantly increases the value of the observation noise covariance matrix. In the mathematical derivation of Kalman filtering, increasing the observation noise covariance matrix means that the system considers the current observation to be full of noise and uncertainty. Therefore, when calculating the Kalman gain, the system automatically reduces the weight of the observation residuals, making the filtered and updated state value more inclined to the prediction value given by the state prediction model, rather than being pulled by unreliable observations. To describe this update process more specifically, in an executable example, the system can set two sets of observation noise levels: a normal level and a low observation confidence level. The normal level is used in the normal window so that the filter can absorb the true offset changes reflected by the cross-correlation analysis in a timely manner; when a low observation confidence flag is received, the system immediately switches to the low observation confidence level and continues until a reliable peak reappears in the next window. If multiple consecutive windows are in a low observation confidence state, the system can also synchronously record the duration of the anomaly and upload the anomaly segment to the operation monitoring interface or write it to the alarm log, so that subsequent maintenance personnel can check whether there are any anomalies in the sensors, network links or data archiving modules.

[0056] After accurately obtaining the physical response time-series offset, which dynamically changes with operating conditions and has been smoothly tracked, the next core task is to use this offset to time-align the original time series. However, in practical discrete sampling systems, the calculated dynamic offset is often a continuous floating-point time value, while the sampling timestamps of the original data are discrete fixed intervals. If simply rounding is used to align the offset to the nearest sampling point, truncation errors will be introduced; especially under rapid load-increasing transient conditions, parameters change drastically, and even small time truncation errors can lead to large deviations in paired pressure values, thus affecting the accuracy of the reconstructed sample set. To eliminate this alignment error caused by discrete sampling, high-precision interpolation reconstruction methods are needed to achieve sub-sampling level time-series alignment.

[0057] After obtaining the smoothly tracked dynamic time series offset, this offset needs to be used to align the original time series. However, in practical discrete sampling systems, the offset is a continuous floating-point number, while the sampling timestamps are discrete fixed intervals. Simply rounding to the nearest sampling point will introduce truncation errors; especially under rapid load increase transient conditions, parameters change drastically, and even a small time truncation error can lead to a large deviation in the paired pressure values, affecting the accuracy of the reconstructed sample set. To eliminate this error, high-precision interpolation reconstruction is required to achieve sub-sampling level time series alignment.

[0058] In a preferred embodiment of the present invention, the step of using physical response time-series offset to perform time-series alignment processing on time-series data to construct a reconstructed sample set reflecting the same physical evolution stage includes: For each sampling moment within the sliding time window, the main steam temperature at that sampling moment is used as the alignment anchor point. Based on the time difference between temperature and pressure, represented by the physical response time offset, the sampling moment is moved forward and backward along the time axis to obtain the target pressure sampling moment that is in the same physical evolution stage as the main steam temperature. The main steam pressure channel in the time series data is fitted using the cubic spline interpolation algorithm. The main steam pressure value corresponding to the target pressure sampling time is calculated. The main steam temperature at the sampling time is bound to the main steam pressure value to reconstruct a pair of sample data in the same physical evolution stage. The reconstructed sample set, aligned with the physical evolution stages, is composed of sample data corresponding to each sampling time.

[0059] In this specific implementation, the system first establishes alignment logic based on the main steam temperature. Since the main steam temperature response typically lags behind the main steam pressure during rapid load increases, the system uses the corresponding main steam temperature as the alignment anchor point for any sampling moment within the sliding time window. Next, the system uses the physical response time-series offset calculated at the current moment—the specific time difference between temperature and pressure—to backtrack the sampling moment along the historical time axis (i.e., forward) by the duration corresponding to that time difference. The resulting time point after backtracking is the target pressure sampling moment. Since the offset is a continuously changing floating-point number, this target pressure sampling moment will in most cases fall between two actual discrete sampling points. To ensure stable implementation of this step, in one specific embodiment, the system can first determine whether the target pressure sampling moment falls within the effective time range of the currently available pressure sequence before proceeding with the interpolation calculation. If the target pressure sampling time is earlier than the earliest valid sampling time available for fitting within the window, a small number of historical pressure points can be read forward. If the field historical cache is insufficient to provide enough interpolation support points, the sampling time is marked as an unreconstructable point and skipped, thus avoiding extrapolation errors that exceed the boundaries. Correspondingly, the same valid range verification strategy can be applied to sample points at the end of the window.

[0060] To obtain extremely accurate pressure values ​​at the target pressure sampling time, the system abandons simple nearest-neighbor or linear interpolation methods, and instead employs a cubic spline interpolation algorithm to fit the time series data of the main steam pressure channel. The core advantage of the cubic spline interpolation algorithm lies in its piecewise construction of a cubic polynomial, which not only ensures the continuity of the function values ​​of the fitted curve at each discrete sampling point, but also strictly guarantees the continuity of the curve's first derivative (rate of change) and second derivative (acceleration of change). In a specific embodiment, the system extracts several actual pressure sampling points before and after the target pressure sampling time to construct a smooth cubic spline fitting curve. Then, the target pressure sampling time is substituted into the fitting curve equation to calculate the corresponding main steam pressure value. Finally, the system binds the main steam temperature at the original sampling time to the calculated high-precision main steam pressure value one-to-one, thereby reconstructing a pair of sample data truly in the same physical evolution stage. Repeating the above operation for all sampling times within the sliding time window allows these sample data to form a reconstructed sample set with completely aligned physical evolution stages. To more clearly and completely describe the data preparation method for cubic spline interpolation, in one implementation example, the system can select multiple actual pressure sampling points adjacent to the target pressure sampling time as local fitting support points. The number of support points can be set according to the sampling period and the smoothness of pressure changes. If the sampling frequency is high and the data is dense, fewer support points are sufficient to form a smooth fit; if the sampling frequency is low, the coverage of support points can be appropriately expanded, but it is still preferable to limit it to a local time period to avoid bringing irrelevant distant trends into the current interpolation. For pressure points with spikes, a lightweight outlier removal process can be performed before fitting. For example, sampling values ​​that deviate significantly from the preceding and following trends and only show single-point jumps can be marked as outliers and replaced with nearby valid points to prevent outliers from disrupting the local smoothness of the spline curve.

[0061] After constructing a reconstructed sample set reflecting the same physical evolution stage, the system needs to calculate the correlation evaluation index based on the sample set and combine it with the operating context to determine anomalies. In order to more accurately quantify the dynamic following relationship between pressure and temperature and adapt to the normal fluctuation differences under different transient operating conditions, in a preferred embodiment of the present invention, the correlation evaluation index is the residual between the actual observed value and the theoretical predicted value obtained based on the pre-trained long short-term memory network; The steps for determining the relevant evaluation indicators by combining the contextual information of the rapid load increase transient condition, and outputting the state analysis results including the current operating condition, physical response time series offset, and anomaly determination conclusions, include: The category identifier of the rapid load increase transient condition and the physical response time offset are used as inputs for fuzzy logic reasoning to determine the current anomaly judgment threshold correction coefficient. The preset baseline threshold is adjusted using a correction coefficient to obtain an adaptive judgment threshold; Based on the comparison between the residual and the adaptive decision threshold, the state analysis results are output.

[0062] In this specific implementation scheme, the system first introduces a Long Short-Term Memory (LSTM) network to deeply mine the nonlinear coupling characteristics between parameters after time-series alignment. In the reconstructed sample set, the main steam pressure and main steam temperature are already at the same physical evolution stage. The system uses an LSTM model pre-trained with a large amount of normal historical operating data, taking the main steam pressure sequence (or historical temperature and pressure sequence) in the reconstructed sample set as input, to predict the theoretical value of the main steam temperature at the current moment. Since the LSTM network is good at processing time series data with long-term dependencies, it can accurately learn the dynamic following law between pressure and temperature under normal operating conditions. Subsequently, the system calculates the residual between the actual observed main steam temperature and the theoretical prediction value, and uses it as a correlation evaluation index. The larger the absolute value of the residual, the more likely the actual temperature deviates from the theoretically expected following state, and the more likely the inherent correlation between the two is abnormal.

[0063] Furthermore, under rapid load increase transient conditions, even if timing alignment is achieved, the normal fluctuation range of the residuals will vary due to different load increase rates, different initial load levels (i.e., category identifiers), and different degrees of thermal inertia delay, i.e., physical response timing offsets. If a fixed baseline threshold is used for judgment, false alarms are likely to occur under severe transients, or missed alarms under smooth transients. Therefore, the system introduces a fuzzy logic inference mechanism to achieve dynamic adaptive adjustment of the judgment threshold. In specific implementation, the system uses the category identifier of the current operating condition, such as large-scale rapid load increase or small-scale rapid load increase, and the currently calculated physical response timing offset as input variables, inputting them into a preset fuzzy logic inference engine. The fuzzy logic inference engine contains a fuzzy rule base built based on expert experience and historical data statistics. For example, if the load increase range is large and the offset is large, the allowable fluctuation range is widened, and the correction coefficient is increased. Through fuzzy inference, the system outputs an anomaly judgment threshold correction coefficient for the current specific operating condition. Then, the system uses this correction coefficient to scale and adjust the preset baseline threshold to obtain an adaptive judgment threshold. Finally, the calculated residual is compared with the adaptive judgment threshold. If the residual exceeds the adaptive judgment threshold, it is judged as abnormal, and the state analysis result containing the current working condition, offset and abnormal conclusion is output.

[0064] Furthermore, in actual operation, some extremely complex transient conditions may be encountered, leading to significant uncertainty in the calculated physical response timing offset. If timing alignment and anomaly detection are based on unreliable offsets, it can easily lead to misleading conclusions. Therefore, this invention further proposes that the method also includes: After identifying a rapid load increase transient condition, determine whether the estimated confidence level of the physical response timing offset is lower than a preset threshold. If the estimated confidence level is lower than the preset threshold, the deterministic anomaly determination in the current time window is paused, and the unconfirmed associated status flag affected by transient delay is output.

[0065] In this specific implementation, when the system calculates the physical response timing offset in the offset determination module, it simultaneously evaluates the estimated confidence level of the offset, such as the comprehensive confidence score mentioned earlier, or the trace of the covariance matrix output by the Kalman filter. When the system identifies a rapid load increase transient condition, but the calculated offset estimate confidence level is lower than a preset safety threshold due to severe electromagnetic interference from sensors, drastic changes in operating conditions, or large-scale data loss, it indicates that the underlying data used for timing alignment is unreliable. At this point, the system no longer forcibly reconstructs the sample set based on this unreliable offset and performs a deterministic anomaly determination; that is, it does not directly output a normal or abnormal conclusion. Instead, it adopts a conservative defensive strategy, proactively suspending the determination process for the current time window. Simultaneously, the system outputs a pending confirmation status flag indicating that the data segment is significantly affected by transient delays and external interference, and the system is waiting for higher-quality data to restore normal determination logic.

[0066] Secondly, please see Figure 2 The present invention also proposes a multi-parameter coupled data analysis system for thermal power unit pipelines, used to perform the steps in the above method, including: Data acquisition module 210 is used to acquire time series data of main steam pressure and main steam temperature of main steam pipeline of thermal power unit; The operating condition identification module 220 is used to identify whether the main steam pipeline of the thermal power unit is in a rapid load increase transient operating condition based on the changing characteristics of time series data and the feedforward information of the unit load command. The offset determination module 230 is used to determine the physical response timing offset between the main steam pressure and the main steam temperature as the operating condition changes dynamically when the condition is identified as a rapid load increase transient condition. The time-series alignment module 240 is used to perform time-series alignment processing on time-series data using physical response time-series offsets, so as to construct a reconstructed sample set that reflects the same physical evolution stage; The index calculation module 250 is used to calculate the correlation evaluation index between main steam pressure and main steam temperature based on the reconstructed sample set. The status analysis module 260 is used to determine the associated evaluation indicators by combining the context information of the rapid load increase transient condition, and output the status analysis results including the current condition status, physical response time offset and anomaly determination conclusion.

[0067] Through the above technical solution, the system modularizes and structures the various steps in the aforementioned method, forming a complete automated monitoring and analysis architecture. The data acquisition module 210 and the operating condition identification module 220 work together to accurately locate transient operating conditions requiring special handling; the offset determination module 230 and the timing alignment module 240 work closely together to fundamentally eliminate timing misalignment problems caused by physical thermal inertia; finally, the index calculation module 250 and the state analysis module 260 complete high-fidelity coupling relationship evaluation and adaptive judgment. The data flow between modules is clear, and the logic is interconnected, jointly realizing an intelligent analysis mechanism capable of adapting to physical response delays. This significantly reduces the false alarm rate caused by data mismatch during rapid load increases in thermal power units, providing reliable system-level support for accurate monitoring of pipeline thermal fatigue and stress anomalies.

[0068] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for multi-parameter coupled data analysis of pipelines in thermal power units, characterized in that, The method includes: Acquire time-series data of main steam pressure and main steam temperature in the main steam pipeline of thermal power units; Based on the changing characteristics of the time series data and combined with the feedforward information of the unit load command, it is identified whether the main steam pipeline of the thermal power unit is in a rapid load increase transient condition. In the case of identifying the rapid load increase transient condition, determine the physical response timing offset between the main steam pressure and the main steam temperature as the condition dynamically changes; Using the physical response time offset, the time series data is time-aligned to construct a reconstructed sample set reflecting the same physical evolution stage; Based on the reconstructed sample set, calculate the correlation evaluation index between the main steam pressure and the main steam temperature; The associated evaluation indicators are determined by combining the context information of the rapid load increase transient condition, and the state analysis results, including the current operating condition, physical response time offset, and anomaly determination conclusion, are output.

2. The method for multi-parameter coupled data analysis of thermal power unit pipelines according to claim 1, characterized in that, After acquiring the time-series data of the main steam pressure and main steam temperature of the main steam pipeline of the thermal power unit, the method further includes: The time series data is cleaned using an algorithm that combines median filtering and physical extreme value constraints. When the value of a sampling point exceeds the physical extreme value range under the current operating condition, the sampling point is determined to be an abnormal sampling point and is replaced by linear interpolation using adjacent valid sampling points. The rate of change and acceleration of change of the time series data are calculated using a difference algorithm with a smoothing factor to characterize the physical evolution trend of the main steam pipeline of the thermal power unit.

3. The method for multi-parameter coupled data analysis of thermal power unit pipelines according to claim 1, characterized in that, The step of identifying whether the main steam pipeline of the thermal power unit is in a rapid load increase transient condition based on the changing characteristics of the time series data and the feedforward information of the unit load command includes: Monitor the rate of change of the unit load command and the rate of change of the main steam pressure; When the rate of change of the unit load command exceeds the preset load regulation rate threshold within a set period, and the rate of change of the main steam pressure exceeds the preset pressure response threshold, it is determined that the main steam pipeline of the thermal power unit has entered a rapid load increase transient condition.

4. The method for multi-parameter coupled data analysis of thermal power unit pipelines according to claim 1, characterized in that, The step of determining the physical response time offset between the main steam pressure and the main steam temperature as the operating condition dynamically changes, when the condition is identified as a rapid load increase transient condition, includes: Using a cross-correlation analysis algorithm with a restricted sliding window, the cross-correlation coefficient sequence between the main steam pressure and the main steam temperature within the current sliding window is calculated; The initial offset is determined based on the peak position of the cross-correlation coefficient sequence; The initial offset is smoothed using a Kalman filter algorithm to obtain the physical response timing offset that changes dynamically with the operating conditions.

5. The method for multi-parameter coupled data analysis of thermal power unit pipelines according to claim 4, characterized in that, The step of determining the initial offset based on the peak position of the cross-correlation coefficient sequence includes: Extract local maxima points in the cross-correlation coefficient sequence whose amplitudes exceed a benchmark threshold, and construct a candidate peak set; The comprehensive confidence score of each candidate peak is calculated based on the normalized amplitude of each candidate peak and the difference between the delay time corresponding to each candidate peak and the historical continuity benchmark; the historical continuity benchmark is the physical response time offset after smoothing tracking processing at the previous moment. The delay time corresponding to the candidate peak with the highest comprehensive confidence score is selected as the initial offset.

6. The method for multi-parameter coupled data analysis of thermal power unit pipelines according to claim 5, characterized in that, The method also includes: If the overall confidence score of all candidate peaks is lower than the safety lower limit, the initial offset is set to the physical response time-series offset of the previous moment, and a low observation confidence indicator is generated. The step of using the Kalman filter algorithm to perform smooth tracking processing on the initial offset includes: Upon receiving the low observation confidence flag, the observation noise covariance matrix is ​​increased to enhance the dependence of the filtering update process on the state prediction model.

7. The method for multi-parameter coupled data analysis of thermal power unit pipelines according to claim 1, characterized in that, The step of using the physical response time offset to perform time-series alignment processing on the time series data to construct a reconstructed sample set reflecting the same physical evolution stage includes: For each sampling moment within the sliding time window, the main steam temperature at that sampling moment is used as the alignment anchor point. Based on the time difference between temperature and pressure, represented by the physical response timing offset, the sampling moment is moved forward along the time axis by the duration corresponding to the time difference, to obtain the target pressure sampling moment that is in the same physical evolution stage as the main steam temperature. The main steam pressure channel in the time series data is fitted using a cubic spline interpolation algorithm. The main steam pressure value corresponding to the target pressure sampling time is calculated. The main steam temperature at the sampling time is bound to the main steam pressure value to reconstruct a pair of sample data in the same physical evolution stage. The sample data corresponding to each sampling time constitutes a reconstructed sample set aligned with the physical evolution stage.

8. The method for multi-parameter coupled data analysis of thermal power unit pipelines according to claim 1, characterized in that, The correlation evaluation index is the residual between the actual observed value and the theoretical prediction value obtained based on the pre-trained long short-term memory network; The step of determining the associated evaluation indicators by combining the context information of the rapid load increase transient condition, and outputting a state analysis result including the current operating condition, physical response time series offset, and anomaly determination conclusion includes: The category identifier of the rapid load increase transient condition and the physical response time offset are used as inputs for fuzzy logic reasoning to determine the current anomaly judgment threshold correction coefficient. The preset benchmark threshold is adjusted using the correction coefficient to obtain an adaptive judgment threshold; The state analysis result is output based on the comparison between the residual and the adaptive decision threshold.

9. The method for multi-parameter coupled data analysis of thermal power unit pipelines according to claim 1, characterized in that, The method also includes: After identifying the rapid load increase transient condition, it is determined whether the estimated confidence level of the physical response timing offset is lower than a preset threshold. If the estimated confidence level is lower than the preset threshold, the deterministic anomaly determination for the current time window is paused, and the unconfirmed associated status identifier affected by transient delay is output.

10. A multi-parameter coupled data analysis system for thermal power unit pipelines, used to execute the steps of the method according to any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire time-series data of the main steam pressure and main steam temperature of the main steam pipeline of the thermal power unit; The operating condition identification module is used to identify whether the main steam pipeline of the thermal power unit is in a rapid load increase transient operating condition based on the changing characteristics of the time series data and the feedforward information of the unit load command. The offset determination module is used to determine the physical response timing offset between the main steam pressure and the main steam temperature as the operating condition dynamically changes, when the condition is identified as the rapid load increase transient operating condition. The time-series alignment module is used to perform time-series alignment processing on the time-series data using the physical response time-series offset, so as to construct a reconstructed sample set reflecting the same physical evolution stage; The index calculation module is used to calculate the correlation evaluation index between the main steam pressure and the main steam temperature based on the reconstructed sample set. The status analysis module is used to determine the associated evaluation indicators by combining the context information of the rapid load increase transient condition, and output the status analysis results including the current condition status, physical response time offset, and anomaly determination conclusion.

Images