Power load mutation identification algorithm based on multi-scale dynamic model

Abstract

This invention discloses an algorithm for identifying sudden changes in electricity load based on a multi-scale dynamic model, relating to the field of power load monitoring and anomaly identification technology. It performs time alignment, missing data correction, and anomaly smoothing on electricity consumption and power data within a continuous monitoring period, constructing a load feature vector containing load values, relative change rates, and relative change rate accelerations. Combining the temporal context of the current moment to be identified, it extracts recent trend samples, date-type samples, periodic similar samples, and seasonal same-period samples, establishing local feature models and forming a multi-scale dynamic reference model. Based on this, a target reference model matching the current scenario is generated. Finally, based on load value deviation, change rate deviation, and trend reversal deviation, it completes the identification of sudden changes in electricity load, quantifies the severity of anomalies, and outputs results. This algorithm is suitable for electricity consumption scenarios with complex load patterns and diverse fluctuation characteristics.

Description

Technical Field

[0001] This invention relates to the field of power load monitoring and anomaly identification technology, specifically to an algorithm for identifying sudden changes in power load based on a multi-scale dynamic model. Background Technology

[0002] Most existing methods for identifying abnormal electricity loads rely on fixed thresholds, single historical averages, or single-cycle statistical results as the basis for judgment. While these methods can identify some obvious anomalies, in actual operation, different date types, periodic patterns, and seasonal variations all affect load distribution patterns. If a static benchmark is still used for unified judgment, it is easy to cause a mismatch between the identification benchmark and the current scenario, leading to false alarms or missed alarms. Especially for objects such as commercial complexes and industrial users, whose load patterns are affected by production arrangements, operating hours, and environmental conditions, a single time-scale benchmark model is difficult to accurately reflect the true normal state at the current moment.

[0003] Furthermore, existing technologies typically focus more on whether the absolute value of electricity consumption or power exceeds limits, lacking the ability to collaboratively analyze the rate of change and trend reversals between adjacent time points. When the load value has not yet significantly exceeded the limit, but its rate of change or acceleration of change has become abnormal, traditional methods often fail to identify it in a timely manner, thus failing to effectively capture abnormal equipment start-ups and shutdowns, sudden changes in load patterns, or early signs of faults. Therefore, there is an urgent need for an algorithm for identifying sudden changes in electricity load that can dynamically construct reference benchmarks for multiple scenarios and jointly discriminate between load values, first-order change states, and second-order trend reversal states. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing an algorithm for identifying sudden changes in electricity load based on a multi-scale dynamic model.

[0005] The technical solution to achieve the purpose of this invention is as follows:

[0006] The algorithm for identifying sudden changes in electricity load based on a multi-scale dynamic model includes the following steps:

[0007] Collect electricity consumption and power consumption data of the target object within a continuous monitoring period, perform time alignment, missing data correction and anomaly smoothing, calculate the relative rate of change and relative rate of change acceleration between adjacent time points, and form the load feature vector corresponding to each time point;

[0008] Based on the temporal context of the current moment to be identified, recent trend samples, date type samples, periodic similar samples, and seasonal same-period samples are extracted from the historical load feature vector to form a multi-scale sample set.

[0009] For multi-scale sample sets, statistical analysis and dense region extraction are performed on various types of samples in each load feature dimension to construct local feature models corresponding to various types of samples and form a multi-scale dynamic model.

[0010] Based on the date attribute and temporal context of the current time to be identified, a local feature model that matches the current scene is selected from the multi-scale dynamic model, and consistency screening, weight allocation and fusion operation are performed to generate a target reference model;

[0011] Input the load feature vector at the current time to be identified into the target reference model, perform load value deviation judgment, relative rate of change deviation judgment and relative rate of change acceleration deviation judgment, and output the power load change change identification result based on the degree of deviation.

[0012] Furthermore, recent trend samples, date type samples, periodic similarity samples, and seasonal same-period samples are extracted from the historical load feature vector, including:

[0013] Using the day before the current time to be identified as the cutoff date, extend the recent trend time window forward, extract the historical load feature vectors that are consistent with the intraday time position of the current time to be identified within the time window, and form a recent trend sample;

[0014] Within the recent trend time window, historical data with date type attributes that are consistent with the current time to be identified are filtered out, and historical load feature vectors with consistent intraday time positions are extracted to form date type samples.

[0015] Using the day before the current time to be identified as the cutoff date, extend the periodic time window forward, filter historical data whose weekday attribute is consistent with the current time to be identified, extract historical load feature vectors with consistent intraday time positions, and form similar samples for the period.

[0016] Based on the date corresponding to the current time to be identified in the previous year, extend the seasonal time windows forward and backward, extract the historical load feature vectors that are consistent with the intraday time position of the current time to be identified within the time range, and form seasonal synchronous samples.

[0017] Furthermore, the load feature vector is a six-dimensional load feature vector, which includes, in sequence, electricity consumption, power consumption, relative rate of change of electricity consumption, relative rate of change of power consumption, acceleration of relative rate of change of electricity consumption, and acceleration of relative rate of change of power consumption.

[0018] Among them, the relative change rate of electricity consumption is the ratio of the difference between the current and previous electricity consumption to the electricity consumption of the previous time; the relative change rate of power consumption is the ratio of the difference between the current and previous power consumption to the power consumption of the previous time; the acceleration of the relative change rate of electricity consumption is the ratio of the difference between the current and previous relative change rates of electricity consumption to the relative change rate of electricity consumption of the previous time; and the acceleration of the relative change rate of power consumption is the ratio of the difference between the current and previous relative change rates of power consumption to the relative change rate of power consumption of the previous time.

[0019] In the calculation of each feature quantity, when the absolute value of the denominator is lower than the preset minimum threshold, the preset minimum threshold is used to replace the denominator in the calculation.

[0020] Furthermore, the load value deviation judgment, relative change rate deviation judgment, and relative change rate acceleration deviation judgment are all independent of each other, including:

[0021] Load deviation judgment is based on the actual observed values ​​of electricity consumption and power consumption. When the actual observed value falls outside the reasonable value range of the corresponding load characteristic dimension of the target reference model, it is judged as a sudden change in load value.

[0022] The relative rate of change deviation judgment is based on the actual observed values ​​of the relative rate of change of electricity consumption and the relative rate of change of power consumption. When the actual observed value falls outside the reasonable range of the corresponding load characteristic dimension, it is judged as a sudden change in the rate of change.

[0023] The relative rate of change acceleration deviation judgment is based on the actual observed values ​​of the relative rate of change acceleration of electricity consumption and the relative rate of change acceleration of power consumption. When the actual observed value falls outside the reasonable range of the corresponding load characteristic dimension, it is judged as a trend reversal abrupt change.

[0024] Furthermore, weight allocation and fusion operations are performed on each local feature model, including:

[0025] Basic weight coefficients are preset for the local feature models corresponding to recent trend samples, date type samples, periodic similar samples, and seasonal same period samples. The weight coefficient of the local feature model corresponding to the recent trend sample is higher than that of the local feature model corresponding to other samples. The weight coefficients of each local feature model are dynamically adjusted according to the seasonal stage and electricity consumption dynamic characteristics of the current time to be identified.

[0026] For valid candidate local feature models that pass the consistency screening, weighted fusion is performed on the typical feature values, sample mean, sample standard deviation and reasonable value range of each load feature dimension according to their respective weight coefficients to obtain the final typical feature values, final mean, final standard deviation and final reasonable value range of the corresponding load feature dimension, which constitute the core parameters of the target reference model.

[0027] Further consistency screening includes:

[0028] For each load characteristic dimension, calculate the ensemble mean and ensemble standard deviation of the typical characteristic value set corresponding to the candidate local characteristic model;

[0029] A dense screening interval is constructed with the set mean as the center and the set standard deviation as the boundary. Candidate local feature models whose typical feature values ​​fall within the dense screening interval are included in the effective candidate model set. Candidate local feature models whose typical feature values ​​fall outside the dense screening interval do not participate in subsequent weight allocation and fusion calculation.

[0030] When the set of effective candidate models is empty, all candidate local feature models will be included in the set of effective candidate models.

[0031] Furthermore, the determination of the final reasonable value range includes:

[0032] Calculate the average overlap between reasonable value ranges corresponding to effective candidate local feature models, and select the corresponding interval fusion strategy based on the level of the average overlap.

[0033] When the average overlap is at a high consistency level, the weighted average of the upper and lower bounds of the reasonable value range of each effective candidate local feature model is used as the lower and upper bounds of the final reasonable value range, respectively.

[0034] When the average overlap is at a medium consistency level, the upper and lower bounds of the final reasonable value range are determined with the final typical characteristic value as the center and the product of the final standard deviation and the preset expansion coefficient as the radius.

[0035] When the average overlap is at a low consistency level, the weighted quantiles of the reasonable value ranges of each effective candidate local feature model are used as the lower and upper bounds of the final reasonable value range, respectively.

[0036] Furthermore, in conjunction with the deviation degree output of the electrical load change identification results, the method also includes a step to quantify the severity of the anomaly:

[0037] For the load characteristic dimension that triggers deviation judgment, the mean parameter and standard deviation parameter of the corresponding dimension of the target reference model are used as the benchmark to calculate the standard deviation multiple of the deviation of the actual observed value from the mean parameter, which is used as the abnormality quantification index of that dimension.

[0038] Based on the range of the absolute value of the abnormal quantitative indicators, the severity of the abnormality is divided into three levels: mild, moderate and severe.

[0039] When multiple load characteristic dimensions or multiple discrimination levels are triggered at the same time, the highest anomaly level among all dimensions is taken as the overall anomaly level at that time, while retaining the independent anomaly markers and anomaly quantification indicators for each dimension.

[0040] Further, missing data correction and anomaly smoothing include:

[0041] Missing data correction is performed in a tiered manner based on the duration of the missing data: missing data with a duration of less than the preset short-term missing threshold is filled using linear interpolation of the adjacent valid data before and after the missing point; missing data with a duration of less than the preset short-term missing threshold is filled using typical values ​​of historical valid sampled data that are consistent with the date type attribute of the missing time and the same intraday time position.

[0042] The anomaly smoothing process identifies outliers as sampling points whose deviation from the mean of the corresponding historical sequence exceeds a preset standard deviation multiple, and replaces them with the moving average of a preset number of consecutive valid sampling points before and after the outlier.

[0043] Furthermore, the output of the electrical load change identification results also includes the following steps:

[0044] For moments that trigger a single discrimination level and a single load feature dimension deviation, and the degree of deviation is within a slight range, check whether a predetermined number of sampling moments before and after that moment trigger a deviation discrimination; if no deviation discrimination is triggered in adjacent sampling moments, mark that moment as an anomaly to be reviewed and do not include it in the valid mutation events;

[0045] When a deviation of the same type is triggered at a consecutive preset number of sampling times, when a single deviation of multiple load feature dimensions is triggered, or when a single deviation of multiple discrimination levels is triggered, it is directly identified as a valid mutation event.

[0046] Compared with the prior art, the advantages of this invention are as follows:

[0047] 1. By extracting recent trend samples, date type samples, periodic similar samples, and seasonal same-period samples around the current time to be identified, and constructing local feature models for each type of sample, and then generating a target reference model through consistency screening, weight allocation, and fusion operation, the identification benchmark no longer depends on a single fixed historical interval, but can be dynamically adjusted according to the current date attribute, time period location, and seasonal scenario, thereby improving the matching and stability of the identification benchmark in complex electricity consumption scenarios.

[0048] 2. Electricity consumption, power consumption, relative rate of change, and relative rate of change acceleration are uniformly constructed into a six-dimensional load feature vector, and load value deviation judgment, rate of change deviation judgment, and trend reversal deviation judgment are performed respectively. At the same time, the severity of anomalies is quantified by combining Z-Score. It can not only identify explicit abrupt changes in load level, but also identify implicit abrupt changes and early turning point anomalies where the value is still in the normal range but the change pattern is abnormal. Therefore, it improves the completeness, foresight, and practical value of electricity load abrupt change identification. Attached Figure Description

[0049] Figure 1 The flowchart shows the algorithm for identifying sudden changes in electricity load based on a multi-scale dynamic model.

[0050] Figure 2 This is a block diagram of the multi-scale dynamic model structure in this invention;

[0051] Figure 3 This is a flowchart of the target reference model generation and mutation discrimination process in this invention;

[0052] Figure 4 This is a visualization example of the anomaly detection results in this invention. Detailed Implementation

[0053] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0054] This invention discloses an algorithm for identifying sudden changes in electricity load based on a multi-scale dynamic model, comprising the following steps:

[0055] S1: Collect electricity consumption data and power consumption data of the target object within a continuous monitoring period, perform time alignment, missing data correction and anomaly smoothing on the data at each time point, and calculate the relative rate of change and relative rate of change acceleration based on the electricity consumption data and power consumption data at adjacent time points to form the load feature vector corresponding to each time point.

[0056] S2: Based on the time context of the current time to be identified, extract recent trend samples, date type samples, periodic similar samples, and seasonal same-period samples from the historical load feature vector to form a multi-scale sample set corresponding to the current time to be identified.

[0057] S3: For the multi-scale sample set, according to the time position corresponding to the current time to be identified, perform statistical analysis and dense region extraction on each type of sample in each load feature dimension to obtain the local feature model corresponding to each type of sample, and form a multi-scale dynamic model based on each local feature model.

[0058] S4: Based on the date attribute and temporal context of the current time to be identified, select a local feature model that matches the current scene from the multi-scale dynamic model, perform consistency screening, weight allocation and fusion operation on each local feature model, and generate the target reference model corresponding to the current time to be identified.

[0059] S5: Input the load feature vector at the current time to be identified into the target reference model, and perform load value deviation discrimination, relative rate of change deviation discrimination and relative rate of change acceleration deviation discrimination respectively, and output the power load change identification result based on the degree of deviation.

[0060] refer to Figure 1 , Figure 1 This is a flowchart of an algorithm for identifying sudden changes in electricity load based on a multi-scale dynamic model.

[0061] Figure 1 The overall implementation process is shown, from time-series data acquisition and preprocessing, multi-scale sample extraction, local feature modeling, target reference model generation to hierarchical mutation identification and result output.

[0062] In step S1, by collecting and preprocessing electricity consumption data and power consumption data, the rate of change and acceleration of the rate of change at adjacent time points are calculated, and finally combined to form a load feature vector, including:

[0063] S101: Time series data acquisition.

[0064] Collect electricity consumption time-series data of the target object within a continuous monitoring period. The electricity consumption time-series data includes electricity consumption sampled synchronously at a preset time granularity. With power consumption Electricity consumption The unit is kWh, power consumption. The unit is kW; the preset time granularity preferentially adopts 1-hour sampling, generating 24 sets of synchronous sampling data per day, which can be expanded to fine-grained sampling modes such as 15 minutes and 5 minutes according to the real-time requirements of the monitoring scenario; the duration of the continuous monitoring cycle is not less than 12 months to cover the complete annual seasonal change cycle, providing complete historical data support for subsequent multi-scale dynamic modeling.

[0065] S102: Time series data preprocessing.

[0066] The collected electricity consumption time-series data are sequentially processed with time alignment, missing data correction, and anomaly smoothing to eliminate the interference of data defects on subsequent feature calculations and modeling. Specifically, this includes:

[0067] Using the standard hourly time corresponding to the preset sampling granularity as the reference time axis, timestamp alignment is performed on the asynchronously sampled electricity consumption and power data. Linear interpolation is used to map the asynchronously sampled data to the corresponding reference time of the standard time axis, ensuring that the sampling time and time granularity of all data sequences are completely consistent, and eliminating feature calculation errors caused by timing misalignment.

[0068] For missing data points in time-series data, hierarchical completion processing is performed according to the duration of the missing data. For short-term missing data, i.e., missing data with a continuous missing duration of no more than 3 sampling points, linear interpolation of adjacent valid data before and after the missing point is used for completion. For long-term missing data, i.e., missing data with a continuous missing duration of more than 3 sampling points, typical values ​​of the same time period and type are used for completion. The same time period and type of time period refers to historical valid sampling data with the same weekday / non-weekday attribute and the same time period as the missing time. The completed data sequence must meet the requirements of temporal continuity and have no discontinuities or jump anomalies.

[0069] Smoothing corrections are applied to outliers in the original time series data, such as acquisition errors and impulse interference. Outliers are identified using the 3σ criterion, which states that when a sampled data point exceeds the mean of the corresponding historical sequence at that time by ±3 times the standard deviation, it is considered an outlier. For the identified outliers, the moving average of the two consecutive valid sampled points before and after the point is used to replace it, thereby eliminating random interference while preserving the true trend of electricity load changes.

[0070] S103: Calculation of load characteristic quantities.

[0071] For the preprocessed time-series data, each sampling time point is used as a calculation unit. Based on the electricity consumption and power consumption data of adjacent time points, the relative rate of change and the relative rate of change acceleration are calculated sequentially, specifically including:

[0072] Relative change rate of electricity consumption This represents the relative change in electricity consumption between adjacent time points, and is calculated using the following formula:

[0073] ,

[0074] in, For the current moment Electricity consumption For the previous moment Electricity consumption;

[0075] Relative change rate of electricity consumption This represents the relative change in power consumption between adjacent time points, and the calculation formula is:

[0076] ,

[0077] in, For the current moment Power consumption For the previous moment The power consumption;

[0078] Acceleration of the relative rate of change of electricity consumption The second-order magnitude of the relative rate of change in electricity consumption, i.e., the degree of inflection in the load change trend, is calculated using the following formula:

[0079] ,

[0080] in, For the previous moment The relative rate of change in electricity consumption;

[0081] Acceleration of relative change rate of electrical power The second-order magnitude of the relative rate of change of power consumption, i.e., the degree of inflection in the trend of power change, is represented by the following formula:

[0082] ,

[0083] in, For the previous moment The relative rate of change of power consumption.

[0084] S104: Numerical stability correction.

[0085] During the calculation of characteristic quantities, a smoothing correction is performed on the denominator to avoid division by zero errors and numerical overflow issues, thus ensuring the stability of the characteristic calculation. The correction rule is as follows:

[0086] When any denominator term in any calculation formula in S103 includes , , , Its absolute value is less than the preset minimum threshold. When, replace the corresponding denominator term with ;in, The default value is 0.001, which can be adaptively adjusted according to the magnitude of the sampled data and the monitoring accuracy requirements.

[0087] S105: Construction of load feature vector.

[0088] Current moment The corresponding electricity consumption, power consumption, relative rate of change of electricity consumption, relative rate of change of power consumption, acceleration of relative rate of change of electricity consumption, and acceleration of relative rate of change of power consumption are combined to construct the current moment. The corresponding six-dimensional load feature vector The expression is:

[0089] ,

[0090] The six-dimensional load feature vector corresponds sequentially to the current load's numerical state, first-order change state, and second-order trend reversal state, realizing a standardized full-dimensional representation of the operating characteristics of the electricity load and providing a unified feature input for subsequent multi-scale dynamic modeling and mutation identification.

[0091] In step S2, based on the temporal context of the current time to be identified, recent trend samples, date type samples, periodic similar samples, and seasonal same-period samples are extracted from the historical load feature vector to form a multi-scale sample set corresponding to the current time to be identified, including:

[0092] S201: Parsing of time context parameters for the moment to be identified.

[0093] Determine the standardized temporal context parameters of the current time to be identified to provide an accurate matching basis for subsequent sample extraction, specifically including:

[0094] Mark the current time to be identified as parse the corresponding date Intraday location ,in The value range matches the preset time granularity; when the preset time granularity is 1 hourly sampling... , representing the day's Every hour on the hour;

[0095] Date of parsing Corresponding week attribute , ;

[0096] Date of parsing Corresponding date type attribute , Working days are defined as Monday to Friday, and non-working days are defined as Saturday, Sunday, and statutory holidays;

[0097] Date of parsing The corresponding annual and seasonal attributes provide a basis for seasonal sample matching.

[0098] S202: Recent trend sample extraction.

[0099] Extract recent trend samples to characterize short-term load changes in the current stage from the historical load feature vector library, and construct a subset of recent trend samples. The specific extraction rules are as follows:

[0100] By date The previous day is the cutoff date, and the time range for extracting recent trend samples is tracing back 30 consecutive calendar days to form the time range for extracting recent trend samples.

[0101] Within the aforementioned time extraction range, extract all intraday time locations equal to... The historical load feature vector corresponding to the sampling time is included in the recent trend sample subset. ;

[0102] For high-dynamic electricity consumption scenarios where load fluctuations exceed preset fluctuation thresholds, the time window can be shortened to 15-20 calendar days to improve the sensitivity of tracking recent load change trends; for scenarios with stable loads, the time window can be extended to 45 calendar days to enhance the statistical stability of the sample.

[0103] The recent trend sample subset It is used to capture load pattern shifts caused by short-term changes in electricity consumption behavior, weather fluctuations, and temporary production plan adjustments before the time to be identified, providing core short-term trend references for subsequent dynamic modeling.

[0104] S203: Extraction of date type samples.

[0105] Extract date type samples from the historical load feature vector library to distinguish load patterns under different date attributes, and construct a subset of date type samples. The specific extraction rules are as follows:

[0106] Consistent with the time window of recent trend samples, i.e., based on date. The previous day is the cutoff date, and the countback period is 30 consecutive calendar days.

[0107] Within the aforementioned time extraction range, first filter out date type attributes and... For all consistent natural days, extract the intraday time positions from the filtered natural days. The historical load feature vector corresponding to the sampling time is included in the date type sample subset. ;

[0108] Date type sample subset The data is pre-divided into working days and non-working days. During sample extraction, only the subgroup data that matches the date type attribute of the current time to be identified is selected to separate the load characteristics of working mode and rest mode.

[0109] The date type sample subset This is used to eliminate systematic differences in electricity consumption behavior between weekdays and non-weekdays, and to avoid baseline modeling bias caused by date type mismatch.

[0110] S204: Extraction of similar samples within a periodic period.

[0111] Extract periodic similar samples from the historical load feature vector database to characterize repetitive variation patterns under a fixed period, and construct a subset of periodic similar samples. The specific extraction rules are as follows:

[0112] By date The day before the deadline is used as the cutoff date, and the time range for extracting similar samples in the periodic period is traced back 60 consecutive calendar days.

[0113] Within the aforementioned time extraction range, first filter out the weekday attribute and... For all consistent natural days, extract the intraday time positions from the filtered natural days. The historical load feature vector corresponding to the sampling time is included in the periodic sample subset. ;

[0114] For scenarios where electricity consumption exhibits strong cyclical characteristics, such as industrial scenarios with fixed weekly maintenance plans and fixed weekly production schedules, the time window can be extended to 90 calendar days to enhance the statistical significance of fixed periodic patterns; for scenarios with weaker cyclical characteristics, the time window can be shortened to 30 calendar days to reduce interference from irrelevant historical data.

[0115] The periodic similar sample subset It is used to capture fixed electricity consumption patterns in the weekly dimension, such as the periodic electricity consumption behavior of fixed shift schedules, equipment maintenance, and business activities on fixed dates each week, providing a periodic dimension feature reference for benchmark modeling.

[0116] S205: Seasonal sample extraction.

[0117] Extract seasonal samples from the historical load feature vector library to introduce annual seasonal similarity features, and construct a seasonal sample subset. The specific extraction rules are as follows:

[0118] By date Using the same date of the previous year as the base date, we trace back 15 calendar days and extend forward 14 calendar days to form a 30-day consecutive seasonal time range for the same period.

[0119] Within the aforementioned time extraction range, extract all intraday time locations equal to... The historical load feature vector corresponding to the sampling time is included in the seasonal same-period sample subset. ;

[0120] If the number of effective load feature vectors within the time range of the seasonal sample is less than the preset minimum sample size (default is 10 groups), the sample subset data is deemed invalid and will not be included in subsequent modeling processes. For scenarios with significant seasonal characteristics, the time window can be extended to 45-60 natural days to improve the coverage of seasonal characteristics.

[0121] The seasonal sample subset This is used to introduce annual seasonality and climate similarity benchmarks, eliminate the impact of annual periodic factors such as seasonal changes and temperature variations on load benchmark modeling, and improve detection accuracy in different seasons.

[0122] S206: Construction and validity verification of multi-scale sample sets.

[0123] The recent trend sample subset extracted above Date type sample subset Periodic similar sample subset Seasonal sample subset Combine them to construct a system that matches the current time to be identified. Corresponding multi-scale sample set ;

[0124] For each sample subset in the multi-scale sample set, an effective sample size check is performed. If the number of effective feature vectors of the sample subset is lower than the preset minimum statistical sample size (default value is 3 groups), the time window of the sample subset is dynamically expanded until the effective sample size meets the statistical requirements.

[0125] Ensure that all historical load feature vectors in each sample subset are six-dimensional feature vectors consistent with those defined in S105, with the physical meaning and calculation rules of each feature dimension being completely unified, providing consistent feature input for subsequent multi-dimensional statistical modeling.

[0126] In step S3, for the multi-scale sample set, statistical analysis and dense region extraction are performed on each type of sample according to the time position corresponding to the current time to be identified, in each load feature dimension, to obtain the local feature model corresponding to each type of sample and form a multi-scale dynamic model, including the following steps:

[0127] S301: Alignment splitting of sample subsets with feature dimensions.

[0128] The multi-scale sample set constructed for S206 Combined with the intraday time position of the current time to be identified This involves performing alignment and splitting of samples and feature dimensions to provide standardized input for subsequent parallel modeling, specifically including:

[0129] For each subset of samples in the multi-scale sample set , Filter out all corresponding intraday times with positions equal to The historical load feature vectors are used to form a modeling sample set that precisely matches the current location to be identified. This ensures that all samples in the modeling sample set are historical operational data from the same time period, eliminating statistical bias caused by time period differences;

[0130] For each modeling sample set According to the six-dimensional load characteristic dimension defined in S105, it is divided into 6 groups of single-characteristic dimension sample sequences, which are respectively used for the application of electricity. Power consumption Relative change rate of electricity consumption Relative change rate of power consumption Acceleration of relative change in electricity consumption The relative rate of change of electrical power and acceleration ; Record the first The first sample subset The sample sequence of each feature dimension is ,in ;

[0131] For a sample subset that includes subgroup partitioning, including Workday group / Non-workday group For each week's attribute subgroup, the above time alignment and feature dimension splitting operations are performed independently on each subgroup to form a subgroup-level modeling sample set, which provides a foundation for subsequent refined modeling.

[0132] S302: Statistical description calculation of a single feature dimension sample set.

[0133] For each subgroup / sample subset, the corresponding single-feature dimension sample sequence Perform standardized statistical analysis and calculate core statistics to provide a statistical basis for subsequent dense area extraction and modeling, specifically including:

[0134] Sample mean calculation: Calculate the sample sequence arithmetic mean This characterizes the central tendency of the sample subset under the corresponding feature dimension, and the calculation formula is:

[0135] ,

[0136] in, For sample sequences The number of valid samples in the sample. For sample sequences The first in Each sample value;

[0137] Sample standard deviation calculation: Calculate the sample sequence Standard deviation This characterizes the degree of dispersion of the sample subset under the corresponding feature dimension, and is calculated using the following formula:

[0138] ,

[0139] Statistic validity check: If the sample sequence Valid sample size If the sample size is less than the preset minimum modeling sample size (default value is 3 groups), the dynamic sample set expansion mechanism is triggered. The window expansion rules of the corresponding sample subset in S2 are used to supplement the sample until the statistical calculation meets the statistical significance requirements.

[0140] S303: Definition of dense sample areas and screening of effective samples.

[0141] Based on the calculated mean and standard deviation, dense sample regions are defined, and effective dense samples for modeling are selected. Historical outliers are removed to eliminate interference with the baseline modeling. Specifically, this includes:

[0142] Dense region interval definition: based on sample mean Centered on, with ±1 standard deviation Define the dense sample region corresponding to this sample sequence as the boundary. The calculation formula is:

[0143] ,

[0144] Based on the statistical properties of the normal distribution, approximately 68% of the normal operation data falls within this interval, which can effectively characterize the core distribution range of the normal operation status under the corresponding feature dimension.

[0145] Dense sample filtering: Traversing the sample sequence Of all the sample values, filter those falling within the densely populated sample region. The samples within constitute a dense sample set. ;

[0146] Boundary handling rules: For samples whose sample values ​​are equal to the endpoints of the dense region interval, they are directly included in the dense sample set to ensure that no samples are missed during the sample selection process.

[0147] S304: Typical eigenvalue calculation and local feature model parameter generation.

[0148] Based on the selected dense sample set, the typical feature values ​​of the sample subset under the corresponding feature dimension are calculated, and the complete local feature model parameters are generated by combining the core statistics, specifically including:

[0149] Typical eigenvalue calculation: Calculating dense sample sets The arithmetic mean of the values ​​is used as the typical feature value of the sample subset under the corresponding feature dimension. The calculation formula is:

[0150] ,

[0151] In the formula, For dense sample sets The number of valid samples; the typical feature values By eliminating the interference of historical outliers, the sample mean is a better representation of the characteristic level of electricity load under normal operating conditions compared to the original sample mean.

[0152] Determining the reasonable value range: based on the sample mean with standard deviation Based on this, determine the reasonable value range corresponding to this feature dimension. The lower bound of the range Upper bound of the range ;

[0153] Encapsulation of the single-dimensional local feature model: The sample mean calculated above is used as the model. Sample standard deviation Typical eigenvalues Reasonable value range Dense areas Valid sample size Combine them to construct the subset of samples in the first... Single-dimensional local feature model under multiple feature dimensions ;

[0154] Six-dimensional model integration: Integrating the single-dimensional local feature models of the six feature dimensions corresponding to the same modeling sample set to form a complete local feature model corresponding to that sample subset / subgroup. The expression is:

[0155] ,

[0156] in, This is a single-dimensional local feature model under the dimension of electricity consumption characteristics. This is a single-dimensional local feature model under the power consumption feature dimension. This is a single-dimensional local feature model under the feature dimension of relative change rate of electricity consumption. This is a single-dimensional local feature model under the feature dimension of relative change rate of electricity consumption. This is a single-dimensional local feature model under the feature dimension of the relative rate of change of electricity consumption acceleration. This is a single-dimensional local feature model under the feature dimension of the relative rate of change of power consumption acceleration; the local feature model It fully characterizes the power consumption operation mode reflected by the corresponding sample subset, providing standardized local model input for subsequent multi-scale model fusion.

[0157] S305: Parallel construction of local feature models for a subset of the full sample.

[0158] For all sample subsets and subgroups in the multi-scale sample set, the modeling process from S301 to S304 is executed in parallel to complete the construction of the full local feature model, specifically including:

[0159] A subset of recent trend samples Construct the corresponding local feature model This represents the normal operating mode corresponding to recent electricity consumption trends;

[0160] For a subset of date type samples The date type attribute of the current time to be identified For the matched subgroups, construct the corresponding local feature models. This indicates the normal operating mode under the same date type;

[0161] For periodic sample subsets The weekday attribute of the current time to be identified For each matched subgroup, a corresponding local feature model is constructed. , To correspond to the weekday attribute, it represents the normal operating mode under the same period type;

[0162] For seasonally synchronized sample subsets If it passes the data validity check in S205, then the corresponding local feature model is constructed. This represents the normal operating mode corresponding to the same period of the year and season.

[0163] Model validity verification: For each completed local feature model, verify the completeness and validity of its typical feature values ​​and reasonable value range parameters, eliminate invalid models with abnormal parameters, and ensure the stability of the subsequent fusion process.

[0164] S306: Combination encapsulation of multi-scale dynamic models.

[0165] All local feature models that have passed the validity check are structurally combined according to the temporal context of the current time to be identified, forming a multi-scale dynamic model representing the current electricity consumption scenario, specifically including:

[0166] The local feature model is divided into four layers: basic trend layer, date adaptation layer, cycle enhancement layer, and seasonal reference layer. The basic trend layer corresponds to... Date adaptation layer corresponding Periodic reinforcement layer corresponds to Seasonal reference layer corresponding Each level of the model corresponds one-to-one with the physical meaning of the multi-scale sample subsets defined in S2;

[0167] The layered local feature models are structurally combined to construct a model corresponding to the current time step to be identified. Corresponding multi-scale dynamic model The expression is:

[0168] ,

[0169] Supplement the multi-scale dynamic model with corresponding metadata, including the temporal context parameters of the current time to be identified, the sample subset information corresponding to each local feature model, the effective sample size, the modeling time window, etc., to achieve full-link traceability of model parameters;

[0170] A one-to-one mapping relationship is established between the multi-scale dynamic model and the temporal context of the current time to be identified, and the model is stored in the model library to provide a model input that can be quickly called for the generation of the subsequent target reference model.

[0171] refer to Figure 2 , Figure 2 This is a block diagram of the multi-scale dynamic model structure.

[0172] Figure 2 The structure of the historical six-dimensional load feature vectors, under the constraints of date, weekday, time location, and seasonal attributes, forms a basic trend layer, a date adaptation layer, a periodic reinforcement layer, and a seasonal reference layer, which are then combined into a multi-scale dynamic model after six-dimensional parallel modeling.

[0173] It should be noted that, Figure 2 This is merely a schematic diagram to help understand the mapping relationships between various sample subsets, local feature models, and multi-scale dynamic models. It does not represent the only parameter organization method in actual runtime.

[0174] In step S4, based on the date attribute and temporal context of the current time to be identified, a local feature model matching the current scene is selected from the multi-scale dynamic model to perform consistency screening, weight allocation, and fusion operations to generate the target reference model corresponding to the current time to be identified, including the following steps:

[0175] S401: Context-aware local feature model adaptation selection.

[0176] Based on the current time to be identified The date attribute and time-series context parameters are used to construct a multi-scale dynamic model from S306. In this process, effective local feature models that accurately match the current electricity consumption scenario are selected in a hierarchical manner, and a candidate model set is constructed. The specific selection rules are as follows:

[0177] Basic inclusion group : Local feature model corresponding to a subset of recent trend samples As a mandatory foundational model, it is included in the candidate model set, that is... This is to ensure the model's ability to track recent electricity consumption trends;

[0178] Date type adapter group Based on the date type attribute of the current time to be identified. Select the local feature model corresponding to the matching subgroup in the date type sample subset. If If it is a workday, then select the workday group corresponding to... ,like If it is a non-working day, then select the non-working day group corresponding to it. ;

[0179] Periodic reinforcement group : The set of local feature models corresponding to a periodic subset of similar samples In the middle, select the day of the week attribute that matches the current time to be identified. Matching all local feature models Incorporate them into the candidate model set;

[0180] Seasonal reference group If the local feature model corresponds to a subset of samples from the same seasonal period If the data passes the S305 validity check, it will be included in the candidate model set; if the data is invalid, it will not be included in this fusion calculation.

[0181] S402: Candidate feature set and statistical parameter extraction.

[0182] For candidate model set For each effective local feature model, according to the six-dimensional load feature dimension defined in S105, the corresponding core modeling parameters are extracted in parallel to construct a standardized candidate parameter set for each feature dimension, specifically including:

[0183] For the first Each load characteristic dimension From each selected local feature model, extract the typical feature values ​​corresponding to that dimension. Sample mean Sample standard deviation Reasonable lower bound Reasonable upper bound ,in This is the sample subset number corresponding to the local feature model;

[0184] Construct the first Candidate canonical value set for each feature dimension ,in For candidate model set The total number of effective local feature models;

[0185] Simultaneously construct candidate mean sets, candidate standard deviation sets, and candidate value range upper and lower bound sets that correspond one-to-one with the candidate typical value sets, providing full parameter input for subsequent fusion calculations.

[0186] S403: Candidate value consistency screening based on density clustering.

[0187] Candidate typical value set for each load feature dimension Perform statistical distribution analysis and dense area screening to remove outlier candidate values ​​with poor consistency and retain valid candidate values ​​with high matching degree to the current scene. The specific steps are as follows:

[0188] Calculate the candidate typical value set set mean with set standard deviation The calculation formulas are as follows:

[0189] ,

[0190] ,

[0191] In the formula, For the candidate typical value set The first in One candidate canonical feature value, The total number of candidate typical values;

[0192] Define the dense screening interval for candidate values This interval represents the core concentration range of candidate typical values;

[0193] Traverse the set of candidate typical values Filter all elements that fall within the densely selected range. The candidate values ​​within constitute the set of valid candidate typical values. Simultaneously retain its corresponding local feature model and full statistical parameters;

[0194] Boundary handling rules: For candidate typical values ​​whose values ​​are equal to the endpoints of the densely filtered interval, they are directly included in the set of valid candidate typical values; if the set of valid candidate typical values... If the value is empty, all candidate typical values ​​will be included in the valid set to ensure the continuity of the fusion calculation.

[0195] S404: Weight allocation and dynamic adjustment in multi-scale models.

[0196] For the set of valid candidate typical values For each candidate value in the local feature model, differentiated weight coefficients are assigned, and an adaptive weight optimization mechanism is provided. The specific weight allocation rules are as follows:

[0197] Preset basic weight allocation:

[0198] Local feature model corresponding to recent trends Weighting coefficient ;

[0199] Local feature model for date type adaptation Weighting coefficient ;

[0200] Local feature models corresponding to the same periodicity Weight coefficient of a single subgroup ,in The total number of valid subgroups of the same periodicity is 0.8, and the sum of the weights of all periodic subgroups is 0.8.

[0201] Local feature model corresponding to the same period of the season Weighting coefficient ;

[0202] Weight dynamic adjustment rules:

[0203] During seasonal transition periods, including the transitions between spring / summer and autumn / winter, the seasonal synchronization model will be used. The weighting coefficient was lowered to 0.6-0.8 to reduce the interference of historical data from the same period on the current seasonal transition scenario;

[0204] During the middle of typical seasons, including midsummer and midwinter, the adjustment will be maintained or increased. The weighting coefficients are increased to 0.9-1.1 to enhance the reference value of seasonal characteristics;

[0205] For highly dynamic electricity consumption scenarios, recent trend models can be used. The weighting coefficients were increased to 1.3-1.5 to improve the model's sensitivity to recent load changes.

[0206] S405: Calculation of the final characteristic parameters of the weighted fusion.

[0207] Based on the selected set of valid candidate typical values ​​and the assigned weight coefficients, the final typical characteristic value, final mean, and pooled standard deviation for each load characteristic dimension are calculated to complete the fusion calculation of core statistical parameters, specifically including:

[0208] Final canonical eigenvalue calculation: Calculate the weighted average of the effective candidate canonical values, which is used as the final canonical eigenvalue for that feature dimension. The calculation formula is:

[0209] ,

[0210] In the formula, The first in the set of valid candidate typical values One candidate value, These are the weight coefficients of the local feature model corresponding to the candidate value;

[0211] Final mean calculation: Calculate the weighted mean of the sample means corresponding to the effective candidate models, and use it as the final mean for that feature dimension. The calculation formula is:

[0212] ,

[0213] in, For the first The sample mean of each effective candidate model under the corresponding feature dimension;

[0214] Combined Standard Deviation Calculation: Calculate the weighted combined standard deviation of the effective candidate models, which serves as the final standard deviation for that feature dimension. The calculation formula is:

[0215] ,

[0216] in, For the first The standard deviation of the samples for each valid candidate model in the corresponding feature dimension.

[0217] S406: Adaptive reasonable value range fusion based on interval overlap.

[0218] Calculate the interval overlap of the reasonable value range of each effective local feature model, adaptively select the corresponding interval fusion strategy according to the overlap level, and determine the final reasonable value range for each load feature dimension. The specific steps are as follows:

[0219] Interval overlap calculation: Calculate the average overlap of all valid candidate models across their reasonable value ranges. The overlap is the ratio of the intersection length to the union length of any two value range intervals, and the average overlap is the arithmetic mean of the overlap of all pairwise combinations.

[0220] Adaptive fusion strategy selection:

[0221] like If the interval is determined to be highly consistent, the weighted interval intersection method is used to calculate the weighted mean of the upper and lower bounds of the value range of each valid candidate model, which is taken as the final reasonable value range. The calculation formula is as follows:

[0222] ,

[0223] ,

[0224] in, , The first The lower and upper bounds of the reasonable value range of the feature dimension corresponding to each valid candidate model;

[0225] like The interval is determined to be of moderate consistency. The confidence interval expansion method is used to calculate the final reasonable range based on the final typical eigenvalues ​​and the pooled standard deviation. The calculation formula is as follows:

[0226] ,

[0227] ,

[0228] in, This is the expansion factor, with a default value of 1.2 and a range of 1.0-1.5; for high-sensitivity detection scenarios, it is set to... High robustness detection scenarios ;

[0229] like If the interval is determined to be of low consistency, the quantile fusion method is adopted, and the upper and lower bounds of the value range of each effective candidate model are regarded as sample points. The weighted 16% quantile and the weighted 84% quantile are calculated and used as the lower and upper bounds of the final reasonable value range, respectively, corresponding to the statistical characteristics of the normal distribution ±1 standard deviation.

[0230] S407: Encapsulation and validity verification of the target reference model.

[0231] The full parameters of the fused six-dimensional load feature dimension are structured and encapsulated to construct the current time to be identified. Corresponding target reference model And perform model validity verification, specifically including:

[0232] For each load characteristic dimension The final typical eigenvalues Final mean Final standard deviation Final reasonable value range Dense Filtering Range Number of effective models Interval overlap The parameters of the reference model for this dimension are constructed by combining the fusion strategy identifiers; the parameters of the six feature dimensions are integrated to form a complete target reference model. The expression is:

[0233]

[0234] in, to The reference model parameters are listed in order for the six dimensions of the load characteristics.

[0235] Verify the completeness and numerical rationality of the parameters of each dimension of the target reference model to ensure that the lower bound of the final reasonable value range is less than the upper bound, the standard deviation is non-negative, and the typical feature values ​​fall within the reasonable value range. If the verification fails, trigger the fusion parameter rollback mechanism and use the equal-weighted fusion result of all candidate models as the final parameters to ensure the usability of the model.

[0236] Establish the target reference model with the current time to be identified. Intraday location The one-to-one mapping relationship is stored in the model library to provide a standardized discrimination benchmark that can be quickly invoked for subsequent identification of load mutations.

[0237] In step S5, the load feature vector at the current time to be identified is input into the target reference model, and load value deviation judgment, relative rate of change deviation judgment, and relative rate of change acceleration deviation judgment are performed respectively. The result of identifying sudden changes in electricity load is then output based on the degree of deviation. This includes the following steps:

[0238] S501: Standardized input and matching verification of the load feature vector to be identified.

[0239] Get the current time to be identified The corresponding six-dimensional load feature vector The load feature vector is completely consistent with the vector structure defined in S105, and its expression is:

[0240] ,

[0241] in, The current electricity consumption observation value, The current power consumption observation value, The observed relative rate of change of electricity consumption at the current moment. The observed value of the relative rate of change of power consumption at the current moment. The observed value of the relative rate of change of electricity consumption at the current moment. The observed value of the relative rate of change of power consumption at the current moment;

[0242] The feature vector of the load to be identified is compared with that generated by S407 at the current time. intraday position Corresponding target reference model Perform dimension matching verification to ensure a one-to-one correspondence between the six feature dimensions and the parameter dimensions of the target reference model, and verify the final reasonable value range for each feature dimension. Final mean Final standard deviation Completeness and numerical validity;

[0243] If the verification fails, the target reference model recalculation process is triggered, and the model selection, fusion and generation operations in step S4 are re-executed until the model parameters meet the discrimination requirements, ensuring the effectiveness and accuracy of the anomaly discrimination process.

[0244] S502: Hierarchical load mutation joint discrimination.

[0245] Based on the discrimination benchmarks of the target reference model in each dimension, a three-layer progressive anomaly discrimination is performed on the load feature vector at the current time to be identified, so as to achieve full coverage identification of explicit abrupt changes and implicit trend changes. The specific discrimination rules are as follows:

[0246] S5021: Load numerical change detection (Level-1).

[0247] Perform absolute value anomaly detection on the load numerical dimension to identify obvious sudden increases and decreases in load levels, specifically including:

[0248] From the target reference model In the process, the final reasonable value range corresponding to the electricity consumption dimension is extracted. The final reasonable value range corresponding to the power consumption dimension ;

[0249] Execute exception detection, the detection rule is: if or If so, it is determined that a sudden change in load value has occurred at the current moment and is marked as a Level-1 anomaly;

[0250] This level of discrimination is used to identify abnormal events where the absolute value of the load exceeds the normal fluctuation range of similar historical contexts, corresponding to explicit abnormal scenarios such as equipment overload, line fault, and start-up and shutdown of large-scale power-consuming equipment.

[0251] S5022: Abrupt change rate detection (Level-2).

[0252] For the relative load change rate dimension, rate anomaly detection is performed to identify latent abrupt events where the numerical value is normal but the change pattern is abnormal, specifically including:

[0253] From the target reference model In the process, the final reasonable value range corresponding to the relative change rate of electricity consumption is extracted. The final reasonable range of values ​​corresponding to the relative rate of change of power consumption. ;

[0254] Execute exception detection, the detection rule is: if or If the rate of change at the current moment is determined to be abrupt, it is marked as a Level-2 anomaly.

[0255] This level of discrimination is used to identify abnormal relative changes in load between adjacent times, and to capture hidden anomalies where the load value is still within the historical normal range but the rate of change deviates from the normal pattern. It corresponds to scenarios that traditional methods are prone to miss, such as progressive equipment failures, abnormal start-up and shutdown of power consumption behavior, and temporary changes in production plans.

[0256] S5023: Trend reversal and sudden change identification (Level-3).

[0257] Perform trend anomaly detection based on the relative rate of change acceleration of load to identify sudden turning points in load change trends and achieve early warning of anomalies, specifically including:

[0258] From the target reference model In the process, the final reasonable value range corresponding to the acceleration dimension of the relative rate of change of electricity consumption is extracted. The final reasonable range of values ​​corresponding to the acceleration dimension of the relative rate of change of power consumption. ;

[0259] Execute exception detection, the detection rule is: if or If so, it is determined that a trend reversal or abrupt change has occurred at the current moment and is marked as a Level-3 anomaly;

[0260] This level of discrimination is used to identify sudden reversals in load change trends and capture abnormal fluctuations in the rate of change itself. It corresponds to early warning scenarios such as sudden equipment start-up and shutdown, power grid impact events, and sudden changes in power consumption patterns, enabling proactive identification of abnormal events.

[0261] S503: Z-Score-based quantification of mutation severity.

[0262] For each feature dimension that triggers an anomaly, a quantitative indicator of its deviation from the normal operating mode is calculated, and the severity level of the anomaly is classified to provide a quantitative basis for operation and maintenance decisions. Specifically, this includes:

[0263] For the first For each feature dimension that triggers an anomaly, its Z-Score is calculated, representing the factor by which the observed value deviates from the standard deviation of the normal distribution. The calculation formula is as follows:

[0264] ,

[0265] In the formula, For the actual observed values ​​of this dimension, This represents the final mean for that dimension. This represents the final standard deviation for that dimension.

[0266] Based on the absolute value of the Z-Score The severity of the anomaly is divided into three levels:

[0267] Mild abnormality: meets the requirements This indicates that the observed value deviates slightly from the normal range, corresponding to a low-risk abnormal event;

[0268] Moderate abnormality: meets the requirements This indicates that the observed value deviates moderately from the normal range, corresponding to a medium-risk abnormal event;

[0269] Serious abnormality: meets the requirements This indicates that the observed value deviates significantly from the normal range, corresponding to a high-risk abnormal event;

[0270] For moments when multiple levels and dimensions of anomalies are triggered simultaneously, the highest anomaly level is taken as the overall anomaly level at that moment. Meanwhile, the independent anomaly markers, Z-Score values, and trigger levels of each dimension are retained to provide a comprehensive quantitative basis for subsequent anomaly tracing.

[0271] S504: Joint verification and false alarm suppression of anomaly detection results.

[0272] To reduce false alarms caused by random data fluctuations, a joint time-series verification is performed on the preliminary judgment results to confirm the validity of the abnormal events. The specific rules are as follows:

[0273] For moments that trigger only a single-level, single-dimensional anomaly and are of mild severity, continuous time-series verification is performed: if no anomaly of any level is triggered in the two adjacent sampling moments before and after this moment, it is determined to be a suspected false alarm caused by random fluctuations, marked as an anomaly to be verified, and not included in the valid mutation events.

[0274] If the same type and dimension of anomaly is triggered at two or more consecutive sampling times, or if a multi-dimensional and multi-level combined anomaly is triggered in a single instance, or if a severe anomaly level is triggered in a single instance, then the validity of the anomaly is directly confirmed and marked as a valid mutation event.

[0275] For scenarios with high robustness requirements, such as industrial production and critical power facilities, a threshold for consecutive time intervals for anomaly triggering can be configured. An anomaly must be triggered at three or more consecutive sampling times to be considered a valid mutation event, further reducing the false alarm rate.

[0276] S505: Structured output and storage of electrical load change identification results.

[0277] The identified anomaly results are structured, encapsulated, and output to construct a fully traceable anomaly event archive, specifically including:

[0278] The timestamp of the current time to be identified, the six-dimensional load characteristic observations, the triggered anomaly level and anomaly type, the Z-Score value of each anomaly dimension, the anomaly severity level, and the corresponding core parameters of the target reference model, including the final mean, the final standard deviation, and the final reasonable value range;

[0279] Effective mutation events are pushed out in a graded manner according to their severity level. Severe abnormal events are pushed to the power operation and maintenance monitoring system and management personnel in real time, while moderate and mild abnormal events are included in the regular inspection reports to achieve graded response to abnormal events.

[0280] The mutation identification results are associated and stored with the corresponding load feature vector, target reference model, and multi-scale sample set parameters to construct a full-link traceable archive of abnormal events, providing data support for subsequent anomaly cause analysis, algorithm optimization, and power consumption pattern research.

[0281] refer to Figure 3 , Figure 3 This is a flowchart of the target reference model generation and mutation discrimination process in this invention.

[0282] Figure 3 This paper demonstrates the complete process of selecting a matching local feature model from a multi-scale dynamic model, generating a target reference model through consistency screening, weight allocation, and fusion operation, and then performing three-layer deviation discrimination, severity quantification, and result output on the load feature vector at the current moment based on this model.

[0283] refer to Figure 4 , Figure 4 This is a visualization example of the anomaly detection results in this invention.

[0284] Figure 4 This diagram illustrates the visualization results of the actual observed curves, the typical value benchmark curve generated by the target reference model, and the corresponding reasonable value range during the intraday load detection process for the current date to be identified. The horizontal axis represents each sampling time of the day, and the vertical axis represents electricity consumption; the solid curve represents the actual observed value, the dashed line represents the typical value benchmark generated based on the multi-scale dynamic model fusion, and the light-shaded area represents the reasonable value range determined by the target reference model.

[0285] When the actual observed value at a certain sampling moment exceeds the reasonable range, an anomaly point is highlighted at the corresponding location, and the deviation quantization result for that moment is simultaneously provided to intuitively display the anomaly trigger location, deviation degree, and severity level. In the example shown in the figure, the actual electricity consumption at 10:00 on the same day is significantly higher than the upper limit of the reasonable range given by the target reference model. Based on this, the system determines that an anomaly has occurred at that moment and identifies it as a severe anomaly by combining the Z-Score quantization result. The actual observed values ​​at other times generally fall near the reasonable range, which demonstrates the ability of this invention to distinguish between normal fluctuation ranges and anomalous abrupt changes.

[0286] It should be noted that, Figure 4 This is merely a schematic diagram to help understand the display format of the anomaly detection results, the anomaly point location method, and the output logic of the judgment results in this invention. It does not represent the only interface style or display layout in the actual operating scenario.

[0287] For example:

[0288] Step 1: Data collection.

[0289] Taking the electricity consumption data of a commercial complex as an example, hourly electricity consumption data of the complex from January 1, 2025 to December 31, 2025 was collected, including hourly electricity consumption. Unit: kWh, hourly power consumption Unit: kW, sampling frequency: 1 hour / time, 24 sampling points per day.

[0290] Step 2: Construction of the six-dimensional load feature vector.

[0291] For each time point Calculate the relative rate of change in electricity consumption. Relative change rate of power consumption Acceleration of relative change in electricity consumption The relative rate of change of electrical power and acceleration When the absolute value of the denominator in the formula is less than 0.001, the denominator is assigned the value of 0.001 to avoid division by zero error.

[0292] Step 3: Constructing a multi-scale sample set.

[0293] Set the current testing date as June 15, 2025 (Wednesday, a weekday), and the testing time as 10:00 AM on that day. Construct 4 sample sets:

[0294] (Recent Trend Group): Data from May 16, 2025 to June 14, 2025, a total of 30 days, with a valid sample size of 30 groups;

[0295] (Date type group): From Data from weekdays (Monday to Friday) were separated, with a valid sample size of 22 groups;

[0296] (Periodic Similar Groups): Data from all Wednesdays from April 16, 2025 to June 14, 2025, totaling 8 days, with a valid sample size of 8 groups;

[0297] (Seasonal synchronous group): Data from May 16, 2024 to June 14, 2024, a total of 30 days, with a valid sample size of 30 groups.

[0298] Step 4: Calculate the local feature model of a single sample set.

[0299] Electricity consumption at 10:00 AM Taking dimension as an example, calculate the local feature model parameters for each sample set:

[0300] Sample set calculation: Sample mean kWh, sample standard deviation kWh, dense area: kWh, the sample mean (typical characteristic value) in a densely populated region. kWh, reasonable value range: kWh;

[0301] Similarly, calculate , , of value.

[0302] Step 5: Weighted fusion of multi-scale models and generation of target reference model.

[0303] Based on electricity consumption Taking dimension as an example, perform fusion computation:

[0304] Candidate typical value set and weight assignment: kWh, corresponding weight ; kWh, corresponding weight ; Effective subgroup weights: single group Total weight 0.8; kWh, corresponding weight .

[0305] Calculate the mean of the candidate value set kWh, ensemble standard deviation kWh, densely screened interval: kWh.

[0306] Valid candidate values ​​falling into dense intervals: All 8 sets of values ​​in the set, excluding outliers: kWh

[0307] Final typical eigenvalues:

[0308] ;

[0309] Final mean: kWh, final standard deviation: kWh;

[0310] Final reasonable range calculation:

[0311] Interval overlap The confidence interval expansion method is used, and the expansion coefficient is...

[0312] Reasonable lower bound of the range: kWh

[0313] Reasonable upper bound of the range: kWh

[0314] Target reference model (Electricity consumption) Dimensions):

[0315] Typical value 240.3 kWh, mean 239.8 kWh, standard deviation 4.2 kWh, reasonable value range kWh.

[0316] Step 6: Anomaly detection and result output.

[0317] Actual observations and anomaly detection at 10:00 AM on June 15, 2025:

[0318] Actual observed values: kWh

[0319] Hierarchical anomaly detection:

[0320] Level-1 discrimination: This triggers a sudden change in load values ​​and marks a Level-1 anomaly.

[0321] Level-2 discrimination: synchronous calculation Exceeding the reasonable range Triggering a rate-of-change mutation, marking a Level-2 anomaly.

[0322] Level-3 discrimination: synchronous calculation Exceeding the reasonable range Triggering a trend reversal and abrupt change, marking a Level-3 anomaly.

[0323] Quantification of anomaly severity:

[0324] Z-Score calculation: , It was determined to be a serious abnormality.

[0325] Results Verification: On-site investigation revealed that the central air conditioning system of the commercial complex experienced a sudden malfunction during that period, causing a surge in power, which perfectly matched the test results, thus verifying the effectiveness of the invention.

[0326] This invention discloses an algorithm for identifying sudden changes in electricity load based on a multi-scale dynamic model. By collecting electricity consumption and power data within a continuous monitoring period, the algorithm performs time alignment, missing data correction, and anomaly smoothing on the original time-series data to construct a load feature vector that characterizes the load's numerical state, changing state, and trend reversal state. Then, combining the date, day of the week, intraday location, and seasonal attributes of the current time to be identified, it extracts recent trends, date types, similar periodic data, and seasonal samples from the historical load feature vector to form a multi-scale dynamic model. Furthermore, it generates a target reference model that matches the current scenario. Finally, based on the target reference model, it performs hierarchical discrimination and severity quantification output for load numerical deviations, relative rate of change deviations, and relative rate of change acceleration deviations. This scheme enables the identification benchmark to be dynamically adjusted according to the actual electricity consumption scenario, improving the comprehensive identification capability of explicit changes, implicit changes, and trend reversal anomalies, and enhancing the accuracy and practicality of load anomaly detection in complex electricity consumption scenarios.

[0327] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An algorithm for identifying sudden changes in electricity load based on a multi-scale dynamic model, characterized in that, Includes the following steps: Collect electricity consumption and power consumption data of the target object within a continuous monitoring period, perform time alignment, missing data correction and anomaly smoothing, calculate the relative rate of change and relative rate of change acceleration between adjacent time points, and form the load feature vector corresponding to each time point; Based on the temporal context of the current moment to be identified, recent trend samples, date type samples, periodic similar samples, and seasonal same-period samples are extracted from the historical load feature vector to form a multi-scale sample set. For multi-scale sample sets, statistical analysis and dense region extraction are performed on various types of samples in each load feature dimension to construct local feature models corresponding to various types of samples and form a multi-scale dynamic model. Based on the date attribute and temporal context of the current time to be identified, a local feature model that matches the current scene is selected from the multi-scale dynamic model, and consistency screening, weight allocation and fusion operation are performed to generate a target reference model; Input the load feature vector at the current time to be identified into the target reference model, perform load value deviation judgment, relative rate of change deviation judgment and relative rate of change acceleration deviation judgment, and output the power load change change identification result based on the degree of deviation.

2. The power load mutation identification algorithm based on a multi-scale dynamic model as described in claim 1, characterized in that, Extract recent trend samples, date type samples, periodic similarity samples, and seasonal same period samples from historical load feature vectors, including: Using the day before the current time to be identified as the cutoff date, extend the recent trend time window forward, extract the historical load feature vectors that are consistent with the intraday time position of the current time to be identified within the time window, and form a recent trend sample; Within the recent trend time window, historical data with date type attributes that are consistent with the current time to be identified are filtered out, and historical load feature vectors with consistent intraday time positions are extracted to form date type samples. Using the day before the current time to be identified as the cutoff date, extend the periodic time window forward, filter historical data whose weekday attribute is consistent with the current time to be identified, extract historical load feature vectors with consistent intraday time positions, and form similar samples for the period. Based on the date corresponding to the current time to be identified in the previous year, extend the seasonal time windows forward and backward, extract the historical load feature vectors that are consistent with the intraday time position of the current time to be identified within the time range, and form seasonal synchronous samples.

3. The power load mutation identification algorithm based on a multi-scale dynamic model as described in claim 1, characterized in that, The load feature vector is a six-dimensional load feature vector, which includes, in order, electricity consumption, power consumption, relative rate of change of electricity consumption, relative rate of change of power consumption, acceleration of relative rate of change of electricity consumption, and acceleration of relative rate of change of power consumption. Among them, the relative change rate of electricity consumption is the ratio of the difference between the current and previous electricity consumption to the electricity consumption of the previous time; the relative change rate of power consumption is the ratio of the difference between the current and previous power consumption to the power consumption of the previous time; the acceleration of the relative change rate of electricity consumption is the ratio of the difference between the current and previous relative change rates of electricity consumption to the relative change rate of electricity consumption of the previous time; and the acceleration of the relative change rate of power consumption is the ratio of the difference between the current and previous relative change rates of power consumption to the relative change rate of power consumption of the previous time. In the calculation of each feature quantity, when the absolute value of the denominator is lower than the preset minimum threshold, the preset minimum threshold is used to replace the denominator in the calculation.

4. The power load mutation identification algorithm based on a multi-scale dynamic model as described in claim 1, characterized in that, Load value deviation judgment, relative change rate deviation judgment, and relative change rate acceleration deviation judgment are all independent of each other, including: Load deviation judgment is based on the actual observed values ​​of electricity consumption and power consumption. When the actual observed value falls outside the reasonable value range of the corresponding load characteristic dimension of the target reference model, it is judged as a sudden change in load value. The relative rate of change deviation judgment is based on the actual observed values ​​of the relative rate of change of electricity consumption and the relative rate of change of power consumption. When the actual observed value falls outside the reasonable range of the corresponding load characteristic dimension, it is judged as a sudden change in the rate of change. The relative rate of change acceleration deviation judgment is based on the actual observed values ​​of the relative rate of change acceleration of electricity consumption and the relative rate of change acceleration of power consumption. When the actual observed value falls outside the reasonable range of the corresponding load characteristic dimension, it is judged as a trend reversal abrupt change.

5. The power load mutation identification algorithm based on a multi-scale dynamic model as described in claim 1, characterized in that, Perform weight allocation and fusion operations on each local feature model, including: Basic weight coefficients are preset for the local feature models corresponding to recent trend samples, date type samples, periodic similar samples, and seasonal same period samples. The weight coefficient of the local feature model corresponding to the recent trend sample is higher than that of the local feature model corresponding to other samples. The weight coefficients of each local feature model are dynamically adjusted according to the seasonal stage and electricity consumption dynamic characteristics of the current time to be identified. For valid candidate local feature models that pass the consistency screening, weighted fusion is performed on the typical feature values, sample mean, sample standard deviation and reasonable value range of each load feature dimension according to their respective weight coefficients to obtain the final typical feature values, final mean, final standard deviation and final reasonable value range of the corresponding load feature dimension, which constitute the core parameters of the target reference model.

6. The power load mutation identification algorithm based on a multi-scale dynamic model as described in claim 5, characterized in that, Consistency screening includes: For each load characteristic dimension, calculate the ensemble mean and ensemble standard deviation of the typical characteristic value set corresponding to the candidate local characteristic model; A dense screening interval is constructed with the set mean as the center and the set standard deviation as the boundary. Candidate local feature models whose typical feature values ​​fall within the dense screening interval are included in the effective candidate model set. Candidate local feature models whose typical feature values ​​fall outside the dense screening interval do not participate in subsequent weight allocation and fusion calculation. When the set of effective candidate models is empty, all candidate local feature models will be included in the set of effective candidate models.

7. The power load mutation identification algorithm based on a multi-scale dynamic model as described in claim 5, characterized in that, The determination of the final reasonable value range includes: Calculate the average overlap between reasonable value ranges corresponding to effective candidate local feature models, and select the corresponding interval fusion strategy based on the level of the average overlap. When the average overlap is at a high consistency level, the weighted average of the upper and lower bounds of the reasonable value range of each effective candidate local feature model is used as the lower and upper bounds of the final reasonable value range, respectively. When the average overlap is at a medium consistency level, the upper and lower bounds of the final reasonable value range are determined with the final typical characteristic value as the center and the product of the final standard deviation and the preset expansion coefficient as the radius. When the average overlap is at a low consistency level, the weighted quantiles of the reasonable value ranges of each effective candidate local feature model are used as the lower and upper bounds of the final reasonable value range, respectively.

8. The power load mutation identification algorithm based on a multi-scale dynamic model as described in claim 1, characterized in that, In addition to identifying the degree of deviation from the output power load change, the results also include a step to quantify the severity of the anomaly. For the load characteristic dimension that triggers deviation judgment, the mean parameter and standard deviation parameter of the corresponding dimension of the target reference model are used as the benchmark to calculate the standard deviation multiple of the deviation of the actual observed value from the mean parameter, which is used as the abnormality quantification index of that dimension. Based on the range of the absolute value of the abnormal quantitative indicators, the severity of the abnormality is divided into three levels: mild, moderate, and severe. When multiple load characteristic dimensions or multiple discrimination levels are triggered at the same time, the highest anomaly level among all dimensions is taken as the overall anomaly level at that time, while retaining the independent anomaly markers and anomaly quantification indicators for each dimension.

9. The power load mutation identification algorithm based on a multi-scale dynamic model as described in claim 1, characterized in that, Missing data correction and anomaly smoothing include: Missing data correction is performed in a tiered manner based on the duration of the missing data: missing data with a duration of less than the preset short-term missing threshold is filled using linear interpolation of the adjacent valid data before and after the missing point; missing data with a duration of less than the preset short-term missing threshold is filled using typical values ​​of historical valid sampled data that are consistent with the date type attribute of the missing time and the same intraday time position. The anomaly smoothing process identifies outliers as sampling points whose deviation from the mean of the corresponding historical sequence exceeds a preset standard deviation multiple, and replaces them with the moving average of a preset number of consecutive valid sampling points before and after the outlier.

10. The power load mutation identification algorithm based on a multi-scale dynamic model as described in claim 1, characterized in that, The output of the electrical load change identification results also includes the following steps: For moments that trigger a single discrimination level and a single load feature dimension deviation, and the degree of deviation is within a slight range, check whether a predetermined number of sampling moments before and after that moment trigger a deviation discrimination; if no deviation discrimination is triggered in adjacent sampling moments, mark that moment as an anomaly to be reviewed and do not include it in the valid mutation events; When a deviation of the same type is triggered at a consecutive preset number of sampling times, when a single deviation of multiple load feature dimensions is triggered, or when a single deviation of multiple discrimination levels is triggered, it is directly identified as a valid mutation event.

Images