A power system operation state comprehensive evaluation method and system

Abstract

This invention relates to the field of power distribution system condition assessment technology, and discloses a comprehensive assessment method and system for the operating status of power distribution systems. The method includes: dividing historical data into time period subsets based on default time period characteristics that characterize the thermodynamic state of equipment; independently constructing load loss models for each time period subset to obtain a healthy baseline model cluster; after acquiring real-time data, calling the model corresponding to its assigned time period to calculate the normalized loss residual, and applying multidimensional statistical signature analysis of mean drift, instantaneous mutation, and increased volatility in parallel to the residual sequence for comprehensive assessment. This invention utilizes time periods as proxy variables for thermal state, thereby avoiding interference from equipment thermal inertia on the assessment, no longer relying on ambient temperature data, and obtaining a purer characteristic signal reflecting the health status of equipment.

Description

Technical Field

[0001] This invention relates to a method and system for comprehensive evaluation of the operating status of power distribution systems, belonging to the technical field of power distribution system status evaluation. Background Technology

[0002] Ensuring the operational reliability of massive amounts of equipment is currently a core issue. To achieve this goal, the industry generally seeks to utilize existing basic measurement data for low-cost condition assessment and fault early warning, replacing high-cost dedicated sensor solutions that are only applicable to critical nodes. In this technical approach, the conventional practice is to build a correlation model between equipment operating losses and its workload, as well as the external ambient temperature. By introducing the ambient temperature variable, the loss changes caused by temperature fluctuations are corrected, and the early health degradation of the equipment is identified by analyzing the model residuals. This modeling method, which relies on the external ambient temperature, has limitations: what truly determines equipment losses, especially load losses, is the actual internal operating temperature, such as the winding temperature, rather than the external ambient temperature. The internal operating temperature is a variable with huge thermal inertia. Its current value is not only related to the current load and ambient temperature, but also depends on the historical load sequence of the past few hours or even tens of hours. At two times with the same load and ambient temperature, such as when the load rises in the morning and when the load falls in the evening, the equipment may be in completely different internal thermal states due to its thermal inertia, thus exhibiting completely different loss characteristics.

[0003] Existing technologies attempt to fit complex dynamic processes dominated by unmeasurable, time-dependent internal thermal states using static models that rely solely on the current ambient temperature. This is fundamentally inconsistent. The errors caused by this model mismatch result in residuals containing significant model fitting noise rather than pure equipment health degradation signals, leading to the following technical challenges: 1. The assessment system cannot extract early, weak sub-health state signals from a high-noise background, making fault prediction difficult to achieve; 2. The model has poor effectiveness in adapting to changes in operating conditions, easily misjudging normal loss fluctuations caused by thermal inertia as abnormalities, generating numerous false alarms; 3. Due to the lack of effective characterization of the equipment's true thermal state, in-depth fault mode diagnosis is impossible, such as distinguishing between heat dissipation defects strongly correlated with high temperatures and global increases in fixed losses. Faced with this high-noise background caused by model mismatch, some existing technologies have shifted... Attempts to use more complex data-driven models, such as deep learning algorithms, to directly identify states from high-dimensional time-series data are being made. For example, Chinese invention patent CN115033984B discloses a health assessment method for integrated modular avionics systems based on LSTM. This method is applied to avionics systems, and its core is to use Long Short-Term Memory (LSTM) networks to process time-series data. Such methods are essentially black-box models, and their evaluation performance is highly dependent on the selected feature parameters (such as the memory size required to complete the function selected in this patent) and massive amounts of high-quality training data. They do not solve the aforementioned thermal inertia problem from a mechanistic (i.e., physical process) perspective, but instead attempt to forcibly fit this dynamic process through complex network structures. This not only increases the complexity and uninterpretability of the model, but also poses a serious challenge to its generalization ability and evaluation robustness when operating conditions, such as load characteristics, change.

[0004] Therefore, the technical problem to be solved by this invention is how to rely solely on existing basic measurement data, avoid inaccurate explicit modeling of complex thermal inertial processes, and find a more reliable benchmark that can characterize the true thermodynamic state of the device in order to build a health status assessment model with a high signal-to-noise ratio. Summary of the Invention

[0005] This invention provides a comprehensive evaluation method for the operating status of a power distribution system. Its main purpose is to solve the problem of how to rely solely on existing basic measurement data, avoid inaccurate modeling of thermal inertia, and find a more reliable benchmark that can characterize the true thermodynamic state of the equipment in order to construct a high signal-to-noise ratio health assessment model.

[0006] To achieve the above objectives, the present invention provides a comprehensive evaluation method for the operating status of a power distribution system, comprising the following steps:

[0007] Obtain time-series data of load power and corresponding power loss of the evaluation object over multiple operating days within the historical health cycle; based on the default time period characteristics of the operating days, divide the time-series data into at least two time period subsets, and the time period characteristics are used to characterize the preset thermodynamic state of the evaluation object within the time period;

[0008] For each time period subset, an independent load-loss health baseline model is constructed for that time period subset, resulting in a set of health baseline model clusters; the real-time load power and real-time power loss of the evaluation object at the current moment are obtained; and the time period to which the current moment belongs is determined.

[0009] Call the health baseline model that uniquely corresponds to the time period in the health baseline model cluster, and calculate the theoretical health loss based on the real-time load power; calculate the normalized loss residual between the real-time power loss and the theoretical health loss to obtain the normalized loss residual sequence.

[0010] Multidimensional statistical signature analysis is performed on the normalized loss residual sequence, and in parallel: the first statistical monitoring is performed to obtain the first monitoring result representing the persistent drift of the mean of the sequence; the second statistical monitoring is performed to obtain the second monitoring result representing the instantaneous change of the sequence; the third statistical monitoring is performed to obtain the third monitoring result representing the continuous increase of the volatility of the sequence; based on the combination of the first monitoring result, the second monitoring result and the third monitoring result, the operating status of the evaluation object is comprehensively evaluated.

[0011] Preferably, a comprehensive evaluation is performed based on a combination of the first monitoring result, the second monitoring result, and the third monitoring result, including: when the second monitoring result indicates that a transient change occurs in the normalized loss residual sequence, it is determined to be a non-faulty systemic change in operating parameters, and the output of the evaluation conclusion based on the first monitoring result is suppressed.

[0012] Preferably, the comprehensive evaluation based on the combination of the first monitoring result, the second monitoring result, and the third monitoring result further includes: distinguishing the failure modes of the evaluated object based on the combination of the first monitoring result and the third monitoring result; when the first monitoring result alarms but the third monitoring result does not alarm, it is determined to be a progressive degradation mode; when the first monitoring result does not alarm but the third monitoring result alarms, it is determined to be an intermittent connection failure mode. The default time period characteristics include: the deep cooling period corresponding to the lowest daily operating load, and the high temperature steady state period corresponding to the peak daily operating load or daytime high temperature.

[0013] Preferably, the step of performing third statistical monitoring to obtain third monitoring results characterizing the continuous increase in sequence volatility includes: based on the normalized loss residual sequence Generate a volatility characteristic sequence that characterizes its short-term volatility. ,in, Then, an exponentially weighted moving average control chart is used to analyze the fluctuation characteristic sequence. Implement statistical monitoring and obtain third-party monitoring results.

[0014] Preferably, after obtaining the real-time load power and real-time power loss of the evaluation object at the current moment, and before determining the time period to which the current moment belongs, the process includes: determining whether the real-time load power is lower than the default low-confidence load threshold; if the real-time load power is lower than the low-confidence load threshold, then in the step of calculating the normalized loss residual, the process of including the calculation result of the current moment into the normalized loss residual sequence is stopped, and the steps of performing multidimensional statistical signature analysis on the normalized loss residual sequence and comprehensively evaluating the operating status of the evaluation object based on the combination are stopped.

[0015] Preferably, the method further includes: when it is determined that the real-time load power is lower than the low confidence load threshold, the steps of determining the time period to which the current time belongs, calling the health baseline model in the health baseline model cluster that uniquely corresponds to the time period, and calculating the normalized loss residual are still performed to calculate the pseudo residual data point, and the pseudo residual data point is incorporated into an independent model confidence monitoring sequence to perform self-diagnosis of the stability of the health baseline model cluster itself.

[0016] Preferably, the method further includes: calculating the mean persistence drift of the normalized loss residual sequence at different time periods; and diagnosing the potential failure modes of the evaluated object based on the differences in mean persistence drift at different time periods.

[0017] Preferably, the diagnosis is based on the differences in mean persistence drift at different time periods, including: when the normalized loss residual sequence only shows mean persistence drift during the high-temperature steady-state period and does not show mean persistence drift during the deep cooling period, the assessment object is determined to have a state anomaly related to high-temperature operation; when the normalized loss residual sequence shows an equal amount of drift in all time periods, the assessment object is determined to have a state anomaly related to fixed loss.

[0018] Preferably, the step of performing first statistical monitoring and obtaining a first monitoring result representing the persistent drift of the mean of the sequence includes: using an exponentially weighted moving average control chart to monitor whether there is a statistical persistent drift in the mean of the normalized loss residual sequence and obtaining the first monitoring result.

[0019] A comprehensive evaluation system for the operating status of a power distribution system, the system comprising:

[0020] The data acquisition unit is used to acquire time series data of load power and power loss at corresponding times for the evaluated object within a historical health cycle covering multiple operating days, and to acquire the real-time load power and real-time power loss of the evaluated object at the current time.

[0021] The data segmentation unit, based on the default time period characteristics of the operating day, divides the time series data into at least two time period subsets. The time period characteristics are used to characterize the preset thermodynamic state of the evaluation object within the time period.

[0022] The model building unit is used to independently build the load-loss health baseline model corresponding to each time period subset, thereby obtaining a set of health baseline model clusters;

[0023] The residual calculation unit is used to determine the time period to which the current moment belongs, call the health baseline model in the health baseline model cluster that uniquely corresponds to the time period, calculate the theoretical health loss based on the real-time load power, calculate the normalized loss residual between the real-time power loss and the theoretical health loss, and obtain the normalized loss residual sequence.

[0024] The multidimensional analysis unit is used to perform multidimensional statistical signature analysis on the normalized loss residual sequence, and executes the first statistical monitoring, the second statistical monitoring and the third statistical monitoring in parallel to obtain the first monitoring result representing the persistent drift of the sequence mean, the second monitoring result representing the instantaneous change of the sequence, and the third monitoring result representing the continuous increase of the sequence volatility, respectively.

[0025] The comprehensive evaluation unit is used to comprehensively evaluate the operational status of the evaluation object based on a combination of the first monitoring result, the second monitoring result, and the third monitoring result.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] 1. By segmenting the operating data according to the time period characteristics that can characterize the thermodynamic state of the equipment by default, and independently constructing a corresponding load-loss health baseline model for each time period subset, a set of model clusters is formed. This approach utilizes the daily cyclical regularity of equipment operation and uses the time period as a fixed classification variable to replace the direct modeling or indirect estimation of the complex thermal inertia process inside the equipment that has a strong time-series dependence. As a result, by calling the normalized loss residual generated by the corresponding model in the corresponding time period, the baseline fluctuations caused by load fluctuations and thermal state changes are stripped away, and a purer characteristic signal reflecting the health status of the equipment itself is obtained, providing a solid foundation for subsequent high-sensitivity early condition assessment.

[0028] 2. Based on the main scheme, a dual-channel parallel analysis mechanism for the same normalized loss residual sequence is introduced. One channel, such as exponentially weighted moving average, is used to monitor the slow mean drift of the sequence to capture gradual physical degradation. The other channel uses a fast detection algorithm to monitor whether the sequence has instantaneous abrupt changes. Through collaborative adjudication logic, the system only determines health degradation when the slow drift channel is triggered alone. When the fast abrupt change channel is triggered, the system determines it as a non-faulty step caused by changes in external operating parameters (such as grid voltage adjustment) and actively suppresses the output of the evaluation conclusion. This mechanism distinction enables the evaluation method to distinguish between internal lesions and external shocks, avoiding systematic misjudgments caused by the failure of the baseline model. At the same time, the triggered adaptive calibration process also ensures the long-term effectiveness of the baseline model after changes in the external environment.

[0029] 3. This invention also provides another dual-channel parallel monitoring method for residual sequences, which is no longer limited to the single analysis of the sequence mean. While retaining the first statistical monitoring of the sequence mean to identify progressive degradation, this method extracts the fluctuation feature sequence that characterizes the short-term volatility of the sequence in parallel and applies a second statistical monitoring to the fluctuation feature sequence. Since intermittent connection faults in the residual sequence are mainly manifested as increased volatility rather than mean drift, by identifying the combined results of mean monitoring and volatility monitoring, the evaluation conclusion is elevated from a single state alarm to a diagnostic level of fault mode differentiation, covering the blind spot of intermittent faults that cannot be detected by mean monitoring alone. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the overall architecture and data processing of the evaluation system of this invention.

[0031] Figure 2 This is a comparison chart showing the signal-to-noise ratio improvement effect of the residuals in the model of this invention;

[0032] Figure 3 This is a timing diagram for constructing the health baseline model cluster during the system initialization phase of this invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the scope of protection of the present invention.

[0034] This invention discloses a comprehensive evaluation method and system for the operating status of power distribution systems. The method utilizes time periods as proxy variables for the thermodynamic state of equipment, avoiding interference from thermal inertia in the evaluation. In system implementation, a data acquisition unit acquires time-series data of load power and power loss of the evaluated object within its health cycle. A data segmentation unit divides the data into multiple time period subsets based on default time period characteristics representing the thermodynamic state. A model building unit independently constructs a load-loss health baseline model for each time period subset, obtaining a set of health baseline model clusters. During system operation, the data acquisition unit also acquires real-time data. A residual calculation unit determines the time period to which the residual belongs and calls the unique model corresponding to that time period to calculate the normalized loss residual, forming a residual sequence. A multidimensional analysis unit applies multidimensional statistical signature analysis to the residual sequence in parallel, obtaining multidimensional monitoring results of mean drift, instantaneous mutations, and increased volatility. The comprehensive evaluation unit performs a comprehensive evaluation of the operating status of the evaluated object based on this combined result. In the application of fault prediction and health management (PHM) for power distribution systems, the power loss of equipment (taking distribution transformers as an example) is considered. Not only with load Relatedly, it is also strongly affected by its actual internal operating temperature. Since internal temperature has thermal inertia, under the same external ambient temperature and load, the equipment may be in different thermal states, exhibiting different loss characteristics. This causes errors in conventional models that rely on ambient temperature. This method uses a data acquisition unit to collect data on the evaluation object within its historical health cycle, such as the first year after commissioning, covering multiple operating days. and Time series data, of which The input power can be obtained through a high-precision energy meter. With export power Subtraction yields the result. The data segmentation unit then segments the time series data based on the default time period characteristics. These time period characteristics characterize the preset thermodynamic state. The determination of these characteristics can be based on the analysis of the daily operating load curve of the equipment to identify time windows with stable and reproducible thermal states. As a specific, non-limiting implementation method, these time period characteristics include at least two typical time periods: a deep cooling period corresponding to the lowest daily operating load, which can be determined as 03:00 to 05:00 every day. During this period, the equipment has been fully cooled after a long period of low-load operation, and its internal thermal state is close to the cold state baseline; and a high-temperature steady-state period corresponding to the peak daily operating load or daytime high temperature, which can be determined as 14:00 to 16:00 every day. During this period, the equipment is in the peak thermal state of the whole day. This segmentation based on operating patterns replaces the reliance on ambient temperature measurement and complex thermal inertia modeling.

[0035] In a specific calibration implementation, the steps for determining the default time period characteristics include: acquiring load power time series data for at least 30 complete operating days of the evaluation object, calculating the 24-hour average daily load curve, identifying the global minimum load period of the curve, and expanding outwards from the minimum point to determine a continuous window where the load is consistently at a stable low point for at least 90 minutes, which is calibrated as the 'deep cooling' period; similarly, identifying the global maximum load period of the curve, determining a continuous window where the load is consistently at a stable high point for at least 120 minutes, which is calibrated as the high-temperature steady-state period, and the calibration result is called by the model building unit; after the data is divided into at least two time period subsets such as the deep cooling subset and the high-temperature steady-state subset, the model building unit independently constructs the load-loss health baseline model corresponding to each time period subset. Taking the deep cooling time period subset as an example, the model building unit only uses data points from 03:00 to 05:00 in all historical operating days. The cold-state health baseline model was obtained through training. ,Right now Similarly, a thermal health baseline model is obtained by training a subset of data from the high-temperature steady-state period. ,Right now Due to the different physical loss characteristics (such as winding resistance) in cold and hot states, and This set of models will have different model parameters. Together, they form a health baseline model cluster that characterizes the health characteristics of the equipment under different thermal states. Compared to a single global model, this model cluster can more accurately depict the health characteristics of the equipment. When the system enters the real-time PHM assessment phase, the data acquisition unit obtains the current time... Real-time load power With real-time power loss Residual calculation unit judgment The time period to which it belongs, if If the time is 14:30, the system determines that it belongs to the high-temperature steady-state period. Subsequently, the residual calculation unit calls the thermal health baseline model that uniquely corresponds to this period from the health baseline model cluster. According to real-time load power Theoretical health loss was calculated. The residual calculation unit calculates the real-time power loss. Compared with theoretical health loss Normalized loss residuals between This normalization can be achieved using statistically standardized residual calculations, such as... ,in It is a model The standard deviation of the residuals on the healthy training set is obtained in this way. It is a characteristic signal that has been stripped of the effects of load and thermal state to characterize additional losses, and is continuously fed into the normalized loss residual sequence for subsequent analysis.

[0036] The multidimensional analysis unit receives the normalized loss residual sequence. Furthermore, multidimensional statistical signature analysis is applied in parallel to differentiate failure modes. This analysis includes at least three parallel statistical monitoring channels: First, a first statistical monitoring is performed, which can be achieved through an exponentially weighted moving average (EWMA) control chart used for monitoring... The first monitoring result is obtained by assessing whether the mean of the sequence exhibits statistical persistence drift, and by ensuring high sensitivity to gradual, small mean increases (such as additional losses caused by slow degradation of equipment insulation). The second monitoring involves performing a second statistical monitoring process, which can be implemented using algorithms for detecting transient mutations, such as CUSUM control charts or Shewhart-based rapid detection rules, to capture... The second monitoring result is obtained by detecting instantaneous step fluctuations in the sequence caused by non-fault parameter changes such as grid voltage adjustments; third, a third statistical monitoring is performed to capture the continuous increase in sequence volatility, such as random spike pulses caused by intermittent connection faults. To achieve this monitoring, the multidimensional analysis unit is based on... Sequence generation characterizes the volatility feature sequences of its short-term volatility. One way to implement it is That is, taking the absolute value of the residuals and using a second exponentially weighted moving average (EWMA) control chart for this. Statistical monitoring is applied to the sequence to monitor... The mean (i.e.) The third monitoring result is obtained by determining whether the volatility of the data continues to increase. The comprehensive evaluation unit receives the first, second, and third monitoring results and performs a comprehensive evaluation of the operating status of the evaluation object based on the combination of these results: Firstly, when the second monitoring result indicates... When a sudden change occurs in the sequence, the system determines that this is a non-faulty systemic change in operating parameters. At this time, even if the mean of the first monitoring result drifts, an alarm will also be triggered. The comprehensive evaluation unit will suppress the output of the evaluation conclusion based on the first monitoring result, namely, progressive degradation, thereby avoiding systemic misjudgment caused by changes in external benchmarks. Secondly, if the second monitoring result does not trigger an alarm, the comprehensive evaluation unit will distinguish the fault modes based on the combination of the first and third monitoring results. When the mean of the first monitoring result triggers an alarm and drifts, and the third monitoring result does not trigger an alarm and has normal fluctuations, it is determined to be a progressive degradation mode. When the mean of the first monitoring result does not trigger an alarm and has normal fluctuations, and the third monitoring result triggers an alarm and has increased fluctuations, it is determined to be an intermittent connection fault mode. This diagnostic logic based on the orthogonal analysis of mean and fluctuation can be used to identify intermittent faults that cannot be detected by mean monitoring alone.

[0037] This method can also utilize the time-period dimension for diagnosis. The multi-dimensional analysis unit can be used to calculate and store the mean persistence drift of the normalized loss residual sequence in different time periods (such as the deep cooling period and the high-temperature steady-state period). The comprehensive evaluation unit diagnoses potential fault modes based on the differences in mean persistence drift across different time periods. As a specific diagnostic rule: when the system diagnoses that the mean persistence drift only occurs in the residual sequence during the high-temperature steady-state period, but not during the deep cooling period, the system determines that the evaluated object has a state anomaly related to high-temperature operation, such as radiator blockage or cooling fan failure. When the system diagnoses that the residual sequence has approximately equal drift in all time periods (including the deep cooling period), the system determines that the evaluated object has a state anomaly related to fixed losses, such as core aging. In addition, this method also includes a data confidence discrimination mechanism, which obtains real-time load power in the residual calculation unit. Then, it will determine whether the load is below the preset low confidence threshold. ,Should The procedure for determining the low-confidence load threshold is based on the pre-calibration of the accuracy characteristics of the measuring transformer under low range conditions, such as setting it to 5% of the rated load. This includes: obtaining the accuracy class specifications of the measuring transformer used by the evaluation object, such as 0.2S class, and its technical specifications; consulting the measurement error characteristic curves at different load percentages provided in the document; and defining the load power value corresponding to the inflection point on the curve where the measurement error begins to deviate from its nominal accuracy range, or where the measurement uncertainty begins to increase non-linearly, as the low-confidence load threshold. This threshold is used by the residual calculation unit to identify the validity of the real-time data. Below that This indicates and If the measured value is no longer reliable, the system will stop including the calculation result of the current moment into the normalized loss residual sequence and stop executing the subsequent multidimensional statistical signature analysis and comprehensive evaluation steps. This is to prevent false alarms caused by measurement data under low load contaminating the statistical model. As an alternative implementation method, when it is determined that... Below At this time, the system still performs subsequent judgment, invocation and calculation steps to obtain pseudo residual data points. However, these pseudo residual data points are not incorporated into the residual sequence used to assess the health of the equipment. Instead, they are incorporated into an independent model confidence monitoring sequence. This sequence should exhibit stable measurement background noise in a healthy state. Therefore, the system can monitor the statistical characteristics of this model confidence monitoring sequence to perform self-diagnosis on the stability of the healthy baseline model cluster itself.

[0038] Example 1: In a specific operational scenario, a comprehensive evaluation system for the operation status of a power distribution system continuously monitors a 10kV distribution transformer. The evaluation object has been modeled using historical data, including deep cooling periods. With high temperature steady-state period model The system includes a cluster of health baseline models; at the beginning of the monitoring period, the system calls the model corresponding to its time period and continuously calculates the normalized loss residual sequence. The system consistently fluctuated steadily near zero. The first, second, and third statistical monitoring operations executed in parallel by the multi-dimensional analysis unit did not trigger any alarms, and the system was assessed as healthy. On the 103rd day of the monitoring period, the upstream power grid, for the purpose of overall reactive power balance and voltage regulation, performed a region-wide voltage regulation operation on the tap changers of the 110kV main substation in this area. This operation caused a momentary 2% increase in the reference voltage at the input of the 10kV distribution transformer. Since the transformer's iron loss is related to the square of the voltage, this non-faulty systemic change in operating parameters resulted in a real-time power loss for the assessed object. An immediate, continuous step increase unrelated to load occurs. Conventional assessment methods relying solely on mean drift monitoring would immediately trigger a major false alarm indicating severe equipment degradation due to this step increase. In this embodiment, the multidimensional analysis unit performs analysis in parallel. On one hand, the first statistical monitoring unit, used to monitor continuous mean drift, does indeed rapidly exceed its control limits due to this step increase, triggering a mean drift alarm. However, at the same moment, the second statistical monitoring unit, dedicated to monitoring transient changes, is also triggered, identifying the transient change characteristics of this change. The comprehensive assessment unit receives this combination of the first and second monitoring result alarms and, based on the collaborative decision-making logic, determines that a transient change has occurred in the second monitoring result indication sequence. The system then identifies this as a non-faulty systemic change in operating parameters and actively suppresses the output of the assessment conclusion based on the first monitoring result. Therefore, the system does not push any warnings about equipment degradation to the maintenance personnel, but may instead indicate a change in the system operating baseline.

[0039] On another monitored evaluation object, a different diagnostic scenario emerged, where the device's normalized loss residual sequence over several weeks... The mean value remained stable, and the first statistical monitoring showed no abnormalities; however, maintenance personnel noticed occasional, difficult-to-reproduce instantaneous voltage drops at the end of the equipment, suspecting an intermittent high-impedance connection fault. In this scenario, the method relying on mean monitoring would fail to detect this hidden danger because the additional losses caused by the intermittent connection fault manifest as random, brief pulse spikes, which contribute almost nothing to the sequence mean. In the method of this embodiment, the third statistical monitoring channel of the multidimensional analysis unit continuously monitors based on... Sequence generation of fluctuation feature sequences and to Statistical monitoring is applied, and as pulse spikes appear, The mean of the sequence begins to rise continuously, eventually triggering an alarm in the third monitoring result. The comprehensive evaluation unit receives a specific combination of no alarm in the first monitoring result and an alarm in the third monitoring result, and based on this combination, determines the evaluation conclusion as an intermittent connection failure mode, thereby capturing this high-risk false negative failure without any drift in the mean.

[0040] Example 2: This example is used to verify the detection sensitivity of the method based on the health baseline model cluster of this invention in fault prediction and health management (PHM) applications, compared with conventional methods that depend on ambient temperature (control group). The experimental data comes from two years of continuous operation SCADA data of a 10kV distribution transformer, with a sampling interval of 15 minutes. The data fields include inlet power. Export power And the ambient temperature measured by the weather station deployed near the transformer. Among them, power loss pass Calculations were performed; the operational data from the first year (Year 1) was used as the health baseline training set, at which point the equipment was in a healthy state; for the slow-developing early progressive degradation that the simulation (PHM) focuses on, the experiment was conducted in the second year of health... Based on the data, a synthesized fault signal is injected digitally, starting from 0 kW on the first day of the second year (Day 366) and increasing over time. The signal increased linearly, reaching 0.2kW by the end of the second year (Day 730). The amplitude of this signal was smaller than the normal loss fluctuations caused by load and operating condition fluctuations. The experiment was set up with two groups: a control group and a conventional model dependent on ambient temperature. A nonlinear regression model was trained using health data from Year 1. The sample data of this invention uses the health baseline model cluster method of this invention to divide the health data of Year 1 into two time period subsets: a deep cooling period (03:00-05:00) and a high-temperature steady-state period (14:00-16:00), and independently constructs a cold-state model for each period. With thermal model Obtain model clusters Two different models were used to fit the Year 1 healthy baseline training set, and their respective normalized loss residuals were calculated. The sequence was analyzed by examining the standard deviation of the healthy residual sequence. This allows us to assess the background noise levels of the two models, i.e., the model mismatch noise caused by the model's inability to accurately fit the operating conditions, as shown in Table 1.

[0041] Table 1: Comparison of Residual Noise in Health Data for the Two Methods

[0042]

[0043] Table 1 shows that the control group's conventional method, due to its The variables cannot accurately represent the internal thermal state with thermal inertia, and the residual sequence has a standard deviation of 0.82 kW; while the sample time period model cluster of this invention uses the time period as a proxy variable for the thermal state, and the standard deviation of the residual sequence is 0.09 kW, reducing the background noise level. The two trained models are applied to the second year (Year 2) data injected with weak fault signals, and the normalized loss residual sequences generated by each model are analyzed. and The same first statistical monitoring, namely the exponentially weighted moving average (EWMA) control chart, was applied in parallel to monitor persistent mean drift; in the control group, a weak fault signal (0 to 0.2 kW) was applied to a background noise of 0.82 kW. Above, no statistically persistent drift was detected in the EWMA control chart throughout the second-year monitoring period; in the sample of this invention, the weak fault signal (0 to 0.2 kW) was applied to a low background noise of 0.09 kW. Above, the EWMA control chart detected a positive shift in the statistical mean in month 4.3 and triggered a persistent (PHM) warning.

[0044] Example 3: This example combines Figures 1 to 3 A comprehensive evaluation method and system description for the operating status of a power distribution system, such as... Figure 1 As shown in the figure, the data acquisition unit obtains data from historical time series data on load power and power loss, as well as real-time load and loss. Historical data is sent to the data segmentation unit to be segmented according to default time period characteristics. The model building unit independently builds models for each time period subset, forming a cluster of healthy baseline models representing different thermodynamic states. The residual calculation unit calls the time period models and combines them with real-time data to calculate the normalized loss residuals. This residual sequence is delivered to the multidimensional analysis unit, which applies multidimensional statistical signature analysis to the residual sequence in parallel. Specifically, this includes a first statistical monitoring that represents the continuous drift of the mean, a second statistical monitoring that represents instantaneous changes, and a third statistical monitoring that represents the continuous increase of volatility. Finally, the comprehensive evaluation unit evaluates based on the combination of the three types of monitoring results and outputs a comprehensive evaluation conclusion such as gradual degradation, intermittent failure, or non-failure change.

[0045] like Figure 2As shown, the vertical axis represents the normalized loss residual in kW, and the horizontal axis represents the time from 0:00 to 23:00. The residual of the conventional model, represented by the dashed line, exhibits violent periodic fluctuations with significant amplitude, while the residual of the model of this invention, represented by the solid line, remains highly stable near zero. This demonstrates that this method, after removing operating condition interference, obtains a characteristic signal with a higher signal-to-noise ratio. Figure 3 As shown, the process begins when the user / system administrator sends a system initialization command to the data acquisition unit. The data acquisition unit then performs the operation of acquiring historical health cycle data, collecting load power and power loss data for multiple operating days, and transmitting this time series data to the data segmentation unit. The data segmentation unit performs data segmentation based on time period characteristics, including identifying deep cooling periods such as 03:00-05:00 and identifying high-temperature steady-state periods such as 14:00-16:00. After segmenting the data into at least two time period subsets, the data of each time period subset is transmitted to the model building unit. The model building unit performs the core task of independently building a model for each time period subset, including training a cold-state model using deep cooling data and training a hot-state model using high-temperature steady-state data. Finally, the trained cold-state health baseline model and hot-state health baseline model are stored in the health baseline model cluster. At this point, the model building is complete, and the system is ready.

[0046] Example 4: This example illustrates the built-in defense mechanism of the present invention's evaluation system when dealing with common boundary conditions and data compliance challenges in power distribution management (PHM) applications. In a deployment scenario where the evaluation object is a distribution transformer supplying power to residential areas, the system faces a dual boundary challenge: First, the load power on which the evaluation depends. The data contains highly sensitive user privacy information, and its transmission and processing are subject to strict legal regulations; secondly, during late-night hours, the total load... It may drop to an extremely low level, at which point the error of the current transformer will be amplified non-linearly, leading to... Data distortion; to address the primary compliance challenge, the system architecture of this invention is deployed using edge computing, with the data acquisition unit, data segmentation unit, and residual calculation unit all running on a local controller at the transformer substation. This local controller directly processes the raw data. and The data stream, based on the time period it belongs to, calls the corresponding health baseline model to calculate the normalized loss residual. In this process, the highly sensitive primitive The data never leaves the local site; the system only stores anonymized data that does not contain user behavior characteristics. The multi-dimensional analysis unit and comprehensive evaluation unit that upload sequences to the cloud ensure user data privacy and security from the source; to address the second data integrity challenge, the system executes a low-confidence load threshold. The discrimination mechanism, during the late-night period, determines the real-time load power by the residual calculation unit. Below that When this data point is not found, the system marks it as unreliable and proactively stops including the residual calculated at that time into the normalized loss residual sequence used for (PHM) health assessment. This defensive measure can avoid false positive alarms caused by measurement noise and ensure the reliability of subsequent first, second and third statistical monitoring results.

[0047] Example 5: When the evaluation system is first deployed to a new evaluation object or when performing periodic maintenance, the default time period characteristics need to be calibrated. The system's data segmentation unit obtains the historical load power of the object for at least 30 consecutive operating days. The system calculates the average daily load curve covering 24 hours, analyzes the statistical characteristics of the curve, and determines the load trough period (e.g., 03:00 to 05:00) characterizing the deep cooling thermodynamic state and the load peak period (e.g., 14:00 to 16:00) characterizing the high-temperature steady-state thermodynamic state. These determined periods are set as parameters for the evaluation object for subsequent steps. The system also executes a periodic adaptive reconstruction mechanism, independent of the recalibration triggered by the second statistical monitoring. The comprehensive evaluation unit retrieves the normalized loss residual sequence of the most recent 30 days every 12 months. The multidimensional analysis unit performs statistical evaluation on the sequence. If the first, second, and third monitoring results of the sequence are all determined to be healthy, the system confirms that the equipment has not deteriorated in the past period and automatically triggers the model building unit to rebuild the health baseline model cluster using the data of the most recent period to adapt to the slow time-dependent drift of normal non-faulty equipment.

[0048] Example 6: The first statistical monitoring uses an Exponentially Weighted Moving Average (EWMA) control chart, and its statistical measures... The calculation logic is as follows ,in For smoothing coefficients, The residual at the current moment, The value of the parameter balances detection sensitivity and noise smoothness in this embodiment. The calibration is set to 0.2; the control limit for this channel is... and Residual standard deviation based on the training set of a healthy baseline Confirmed, calculated as ,in To control the width limit, it is set to 3. Exceeding or At that time, the first monitoring result is determined as an alarm; the second statistical monitoring uses a fast mutation detection algorithm based on sliding window mean comparison. The system maintains two adjacent, non-overlapping data windows, window A containing arrive Moment Each residual value, window B contains arrive Moment Each residual value, window length In this embodiment, the value is set to 10; the system continuously calculates the average of the two windows. and And calculate the absolute value of their difference. At the same time, a mutation threshold is set. This threshold is based on the standard deviation of the healthy baseline residuals. The setting is that its calculation is ,when At that time, the second monitoring result determined that a transient change had occurred; the third statistical monitoring obtained the fluctuation characteristic sequence. Subsequently, EWMA control charts were used, with parameters calibrated for volatility monitoring and smoothing coefficients. In this embodiment, the control limit is set to 0.4; and Calculated based on health baseline data mean of the sequence and standard deviation Confirmed, calculated as ,in Set to 3, volatility EWMA statistic exceeds At that time, the third monitoring result determined that the volatility continued to increase.

[0049] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for comprehensive evaluation of the operation state of a power distribution system, characterized by, Includes the following steps: Obtain time-series data of load power and corresponding power loss of the assessment object covering multiple operating days within the historical health cycle; Based on the default time period characteristics of the operating day, the time series data is divided into at least two time period subsets. The time period characteristics are used to characterize the preset thermodynamic state of the evaluation object within the time period. For each time period subset, an independent load-loss health baseline model is constructed for that time period subset, resulting in a set of health baseline model clusters; the real-time load power and real-time power loss of the evaluation object at the current moment are obtained; and the time period to which the current moment belongs is determined. The theoretical health loss is calculated based on the real-time load power by calling the health baseline model that uniquely corresponds to the time period in the health baseline model cluster. Calculate the normalized loss residual between real-time power loss and theoretical health loss to obtain the normalized loss residual sequence; Multidimensional statistical signature analysis was performed on the normalized loss residual sequence, and in parallel: the first statistical monitoring was performed to obtain the first monitoring results characterizing the persistent drift of the sequence mean; Perform a second statistical monitoring to obtain second monitoring results characterizing transient mutations in the sequence; Perform third statistical monitoring to obtain third monitoring results characterizing the continuous increase in sequence volatility; The operational status of the evaluated object is comprehensively assessed based on a combination of the first, second, and third monitoring results.

2. The method of claim 1, wherein, A comprehensive evaluation is conducted based on a combination of the first, second, and third monitoring results, including: when the second monitoring result indicates a transient change in the normalized loss residual sequence, it is determined to be a non-faulty systemic change in operating parameters, and the output of the evaluation conclusion based on the first monitoring result is suppressed.

3. The method of claim 1, wherein the method further comprises: The comprehensive evaluation based on the combination of the first, second, and third monitoring results also includes: distinguishing the failure modes of the evaluated object based on the combination of the first and third monitoring results; when the first monitoring result alarms but the third monitoring result does not alarm, it is determined to be a progressive degradation mode; when the first monitoring result does not alarm but the third monitoring result alarms, it is determined to be an intermittent connection failure mode. The default time period characteristics include: the deep cooling period corresponding to the lowest daily operating load, and the high temperature steady state period corresponding to the peak daily operating load or daytime high temperature.

4. The method for comprehensive evaluation of the operating status of a power distribution system according to claim 3, characterized in that, The steps for performing third statistical monitoring to obtain third monitoring results characterizing the continued increase in sequence volatility include: based on the normalized loss residual sequence. Generate a volatility characteristic sequence that characterizes its short-term volatility. ,in, Then, an exponentially weighted moving average control chart is used to analyze the fluctuation characteristic sequence. Implement statistical monitoring to obtain third-party monitoring results.

5. The method for comprehensive evaluation of the operating status of a power distribution system according to claim 1, characterized in that, After obtaining the real-time load power and real-time power loss of the evaluation object at the current moment, and before determining the time period to which the current moment belongs, the process includes: determining whether the real-time load power is lower than the default low-confidence load threshold; if the real-time load power is lower than the low-confidence load threshold, then in the step of calculating the normalized loss residual, the calculation result of the current moment is stopped from being included in the normalized loss residual sequence, and the steps of performing multidimensional statistical signature analysis on the normalized loss residual sequence and comprehensively evaluating the operating status of the evaluation object based on the combination are also stopped.

6. The method for comprehensive evaluation of the operating status of a power distribution system according to claim 5, characterized in that, The method also includes: when the real-time load power is determined to be lower than the low confidence load threshold, the steps of determining the time period to which the current time belongs, calling the health baseline model in the health baseline model cluster that uniquely corresponds to the time period, and calculating the normalized loss residual are still performed to calculate pseudo residual data points, and the pseudo residual data points are incorporated into an independent model confidence monitoring sequence to perform self-diagnosis of the stability of the health baseline model cluster itself.

7. The method for comprehensive evaluation of the operating status of a power distribution system according to claim 1, characterized in that, The method also includes: calculating the mean persistence drift of the normalized loss residual sequence at different time periods; and diagnosing the potential failure modes of the evaluated object based on the differences in mean persistence drift at different time periods.

8. The method for comprehensive evaluation of the operating status of a power distribution system according to claim 7, characterized in that, Diagnosis is based on the differences in mean persistence drift across different time periods, including: when it is diagnosed that mean persistence drift only occurs in the normalized loss residual sequence during the high-temperature steady-state period, but not during the deep cooling period, the assessed object is determined to have a state anomaly related to high-temperature operation; when it is diagnosed that the normalized loss residual sequence has an equal amount of drift across all time periods, the assessed object is determined to have a state anomaly related to fixed loss.

9. The method for comprehensive evaluation of the operating status of a power distribution system according to claim 1, characterized in that, The step of performing first statistical monitoring to obtain first monitoring results characterizing the persistence of the mean of the sequence includes: using an exponentially weighted moving average control chart to monitor whether there is statistical persistence of the mean of the normalized loss residual sequence to obtain the first monitoring results.

10. A comprehensive evaluation system for the operating status of a power distribution system, used to implement the comprehensive evaluation method for the operating status of a power distribution system as described in claim 1, characterized in that, The system includes: The data acquisition unit is used to acquire time series data of load power and corresponding power loss of the evaluation object covering multiple operating days within the historical health cycle, and to acquire the real-time load power and real-time power loss of the evaluation object at the current moment. The data segmentation unit is used to divide time series data into at least two time period subsets based on the default time period characteristics of the running day. The time period characteristics are used to characterize the preset thermodynamic state of the evaluation object within the time period. The model building unit is used to independently build the load-loss health baseline model corresponding to each time period subset, thereby obtaining a set of health baseline model clusters; The residual calculation unit is used to determine the time period to which the current moment belongs, call the health baseline model in the health baseline model cluster that uniquely corresponds to the time period, calculate the theoretical health loss based on the real-time load power, and calculate the normalized loss residual between the real-time power loss and the theoretical health loss to obtain the normalized loss residual sequence. The multidimensional analysis unit is used to perform multidimensional statistical signature analysis on the normalized loss residual sequence, and executes the first statistical monitoring, the second statistical monitoring and the third statistical monitoring in parallel to obtain the first monitoring result representing the persistent drift of the sequence mean, the second monitoring result representing the instantaneous change of the sequence, and the third monitoring result representing the continuous increase of the sequence volatility, respectively. The comprehensive evaluation unit is used to comprehensively evaluate the operational status of the evaluation object based on a combination of the first monitoring result, the second monitoring result, and the third monitoring result.

Images