Multi-source data fusion power distribution network operation state identification method based on intelligent fusion terminal

Abstract

The present application relates to the technical field of smart grid, and discloses a multi-source data fusion power distribution network operation state identification method based on a smart fusion terminal, comprising the following steps: step S1, generating a static asset map; step S2, calculating node disturbance energy; step S3, calculating thermal and power mislocation quantity; step S4, generating section shear curvature; step S5, calculating node pseudo-state amplitude; step S6, calculating network state probability; and step S7, outputting state category code and dominant section code. The present application constructs the mislocation characteristics of the power disturbance front and the thermal and power response front, forms thermal and power double-front shear curvature distributed along the feeder, realizes the differentiation of the apparent state deviation caused by thermal inertia and the real structure change of the power grid, completes state reasoning in combination with a dynamic graph neural network of physical constraints, makes the identification result fit the physical law of power distribution network operation, and finally outputs a result suitable for the on-site application requirements of power distribution network operation control.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, and more specifically, to a method for identifying the operating status of distribution networks based on multi-source data fusion using smart fusion terminals. Background Technology

[0002] With the continuous integration of distributed photovoltaic power, energy storage, electric vehicles, and flexible loads into the distribution network, the operating state of the distribution network is gradually changing from a traditional single-load state to a complex state of multi-source injection, multi-timescale fluctuations, and multi-device collaborative action. At the same time, smart converged terminals, feeder terminal devices, smart meters, distribution transformer monitoring devices, and distributed power source interfaces are continuously deployed in the distribution field, significantly increasing the types of data available on-site. However, the sampling periods, transmission delays, and timing accuracy of various data types are inconsistent, resulting in the same physical event often being scattered across multiple data sources with different time labels and different dimensions.

[0003] In existing technologies, distribution network state estimation and operational status assessment mostly focus on single-moment measurement fitting, steady-state power flow relationships, or extraction of single-type measurement features. These methods can achieve basic state perception in scenarios where measurements are relatively synchronous and line parameters are relatively stable. However, in feeders with mixed overhead and underground cable connections, changes in electrical state and changes in electrothermal impedance are not synchronous. After load surges or photovoltaic fluctuations, voltage and power often change first, while the temperature and impedance of the lines and cables continue to evolve according to their respective thermal inertia. This difference in thermal response speed and duration is particularly pronounced at the junction of overhead and cable sections.

[0004] Under the aforementioned conditions, if state judgment is based solely on voltage, current, power, or local structural characteristics at a single moment, it is easy to misjudge apparent impedance changes caused by thermal tailing as actual topology switching, load migration, or equipment malfunction. Especially when multi-source data arrives asynchronously, the preceding electrical disturbance and the subsequent electrothermal response may be compressed into the same identification window, causing pseudo-state drift. Existing solutions generally lack a unified mechanism to simultaneously characterize electrical fronts, electrothermal fronts, spatial bending intensity, and segment-level dominant positions, thus making it difficult to simultaneously address network-wide category judgment, dominant segment location, and state intensity quantification. Summary of the Invention

[0005] This invention provides a method for identifying the operating status of a distribution network based on multi-source data fusion using an intelligent fusion terminal, thereby solving the technical problems mentioned in the background.

[0006] This invention provides a method for identifying the operating status of a distribution network based on multi-source data fusion using an intelligent fusion terminal, comprising the following steps: Step S1: Smoothly align the multi-source raw data to the event time axis according to the collection interval to generate the aligned measurement tensor and static asset graph; Step S2: Calculate the change amplitude of the alignment measurement tensor to obtain the charge perturbation density, take the time centroid of the charge perturbation density as the charge perturbation frontier quantity, and calculate the nodal perturbation energy. Step S3: Calculate the equivalent current based on the alignment measurement tensor, and recursively calculate the electrothermal response density along the static asset diagram. Take the time centroid of the electrothermal response density as the electrothermal response frontier quantity, and subtract the charge disturbance frontier quantity from the electrothermal response frontier quantity to obtain the thermoelectric misalignment quantity. Step S4: Calculate the second-order difference generation node shear curvature of thermoelectric misalignment on the branch electrical distance of the static asset map, and use the node perturbation energy to weight the node shear curvature to generate the segment shear curvature. Step S5: Input the alignment measurement tensor into the convolutional layer to extract temporal features, incorporate node shear curvature to construct attention weights to aggregate temporal features, and apply physical constraints to output node embedding vectors and node pseudo-state amplitudes. Step S6: Multiply the node perturbation energy, the node pseudo-state amplitude, and the node shear curvature to obtain the segment state intensity; input the node embedding vector and the segment shear curvature into the classification layer to obtain the overall network state probability. Step S7: Take the highest probability item of the entire network state as the state category code, take the highest state intensity item of the segment as the dominant segment code, and simultaneously output the segment state intensity corresponding to the dominant segment code. The beneficial effects of this invention are: By constructing an aligned measurement tensor and a static asset graph, this invention unifies asynchronous heterogeneous multi-source measurements onto the event time axis, then sequentially extracts the frontier quantities of electrical disturbances, electrothermal response frontier quantities, and thermoelectric misalignment quantities, and further forms node shear curvature, segment shear curvature, node pseudo-state amplitude, and segment state intensity, thus forming a continuous closed loop in the time dimension, spatial dimension, and physical constraint dimension. Compared with schemes that rely solely on single-point measurements or a single steady-state model, this invention can uniformly identify normal steady state, new energy disturbance-dominated state, thermal memory pseudo-state drift state, and abnormal entity structure state in scenarios where distributed power fluctuations, mixed feeder thermal inertia differences, and multi-source asynchronous data transmission coexist, and simultaneously outputs state category codes, dominant segment codes, and pseudo-state intensities, thereby providing a more complete basis for operation monitoring, condition-based maintenance, event review, and disposal prioritization. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating the calculation process of the multi-source data fusion distribution network operation status identification method based on intelligent fusion terminals of the present invention. Detailed Implementation

[0008] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0009] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0010] like Figure 1 As shown, the method for identifying the operating status of a distribution network based on multi-source data fusion using intelligent fusion terminals includes the following steps: Step S1: Smoothly align the multi-source raw data to the event time axis according to the collection interval to generate the aligned measurement tensor and static asset graph; Step S2: Calculate the change amplitude of the alignment measurement tensor to obtain the charge perturbation density, take the time centroid of the charge perturbation density as the charge perturbation frontier quantity, and calculate the nodal perturbation energy. Step S3: Calculate the equivalent current based on the alignment measurement tensor, and recursively calculate the electrothermal response density along the static asset diagram. Take the time centroid of the electrothermal response density as the electrothermal response frontier quantity, and subtract the charge disturbance frontier quantity from the electrothermal response frontier quantity to obtain the thermoelectric misalignment quantity. Step S4: Calculate the second-order difference generation node shear curvature of thermoelectric misalignment on the branch electrical distance of the static asset map, and use the node perturbation energy to weight the node shear curvature to generate the segment shear curvature. Step S5: Input the alignment measurement tensor into the convolutional layer to extract temporal features, incorporate node shear curvature to construct attention weights to aggregate temporal features, and apply physical constraints to output node embedding vectors and node pseudo-state amplitudes. Step S6: Multiply the node perturbation energy, the node pseudo-state amplitude, and the node shear curvature to obtain the segment state intensity; input the node embedding vector and the segment shear curvature into the classification layer to obtain the overall network state probability. Step S7: Take the item with the highest probability of state across the entire network as the state category code, take the item with the highest state intensity of a segment as the dominant segment code, and output the segment state intensity corresponding to the dominant segment code simultaneously.

[0011] In one embodiment of the present invention, multi-source raw data is smoothly aligned to the event time axis according to the acquisition interval to generate an aligned measurement tensor and a static asset graph, including: Based on the unified discrete event time, the actual timestamp, the time smoothing scale, and the multi-source raw data, the calculation formula for physical measurement data is as follows: ; in To unify the timing of discrete events, Real timestamp For time smoothing scale, For multi-source raw data, For physical measurement data, Number the nodes. Number the data source type. The original reporting sequence number; Based on physical measurement data, channel time mean, and channel time standard deviation, the formula for calculating normalized measurement data is as follows: ; in This represents the channel time average. The standard deviation of the channel time. To prevent zero positive numbers, Normalized measurement data; Based on the physical measurement data and the normalized measurement data, the calculation formula for the aligned measurement tensor is extracted as follows: ; in A combined tensor composed of physical measurement data. For the combined tensor composed of normalized measurement data, For aligning the measurement tensor; The formula for generating a static asset map based on the set of physical network nodes and the set of physical network branches in the distribution network is as follows: ; in It is the set of physical network nodes of the distribution network. For the set of physical network branches, This is a static asset diagram.

[0012] It should be noted that the unified discrete event time is a unified calculation point on the event timeline, used to map data with different sampling rhythms to the same identification window. Preferred values ​​are 1 second, 2 seconds, 5 seconds, 15 seconds, 30 seconds, or 60 seconds, meaning it should not exceed the fastest main measurement cycle on site, and should be able to balance smooth mapping for slow measurements and the computational burden of the window. The actual timestamp is the actual time marker when each piece of multi-source raw data is sampled or uploaded. It can be obtained through the local clock of the intelligent fusion terminal, satellite timing module, feeder terminal device message timestamp, smart meter sampling timestamp, or master station receiving timestamp. The time difference is the offset between the unified discrete event time and the actual timestamp, used to characterize the time proximity between the raw data and the target event time. The time smoothing scale is a smoothing parameter that controls the rate at which time difference weights decay. It determines how far along a time range of raw data can participate in the alignment of the same event point. The preferred value is the median arrival interval of the most recent complete upload cycle of the corresponding data source. Common values ​​range from 1 second to 300 seconds; that is, the more stable the upload and the faster the sampling, the smaller the time smoothing scale should be. The squared value is the square of the time difference after normalization by the time smoothing scale. It is used to construct the exponential term of the Gaussian smoothing weight, causing larger deviations in time points to decay rapidly. The Gaussian smoothing weight is a smoothing weight assigned to the raw data based on the size of the time difference, used to highlight measurement samples closer to the unified discrete event time. Multi-source raw data consists of raw measurement records from different devices and interfaces, including raw sequences of voltage, current, power, photovoltaic output, switch status, and environmental quantities. This data can be obtained through smart meters, feeder terminal devices, distribution transformer monitoring devices, inverter communication interfaces, meteorological acquisition devices, and the main station historical database connected to the intelligent converged terminal. The physical measurement data are measurement results with physical units obtained by time alignment of multi-source raw data using Gaussian smoothing weights, and are used for subsequent calculations of variation amplitude, electrothermal response and physical constraints.

[0013] It should be noted that the node number is a unique identifier used in the distribution network to distinguish different physical nodes, equipment access points, or equivalent measurement points. The data source type number is an index identifier used to distinguish different measurement source categories, indicating the measurement type of different channels under the same node. The original reporting sequence number is the sequential identifier of the same data source in the original time series, used to traverse reporting records by time. The channel time mean is the average level of physical measurement data for a certain channel within the current identification window, used to eliminate differences in channel DC bias and long-term static dimensions. The channel time standard deviation is the dispersion of physical measurement data for a certain channel within the current identification window, used to compress channels with different fluctuation ranges to a comparable scale. Zero-prevention positive numbers are stabilization constants introduced to avoid zero or excessively small denominators; their preferred value is 0.000001 to 0.001, meaning it should be significantly smaller than the normal variation scale of the target quantity, while being sufficient to suppress numerical divergence caused by abnormally small denominators. Normalized measurement data is the dimensionless measurement result obtained after mean decentering and standard deviation scaling of physical measurement data, used as a unified input for the neural network. The combined tensor composed of physical measurement data is a multidimensional structure formed by arranging physical measurement data of all source data types according to nodes, channels, and time, used to preserve the original physical quantity information. The combined tensor composed of normalized measurement data is a multidimensional structure formed by arranging normalized measurement data of all source data types according to nodes, channels, and time, used to preserve comparable statistical characteristics. The aligned measurement tensor is a unified input structure composed of the combined tensor composed of physical measurement data and the combined tensor composed of normalized measurement data, used to support subsequent disturbance, electrothermal, and graph network calculations. The distribution network physical network node set is a set describing physical entities such as feeder heads, segment nodes, switch nodes, transformer area nodes, and terminal nodes, which can be obtained by importing from the distribution automation master station asset ledger, geographic information ledger, distribution network diagram file, or equipment file. The physical network branch set is a set describing the connection relationships between nodes in terms of lines, switches, tie points, or cable segments, which can be obtained from the geographic information ledger, line parameter ledger, primary wiring diagram model, or distribution network diagram file. A static asset graph is a graph structure constructed using the set of physical network nodes and physical network branches of a distribution network. It is used to represent the physical connectivity between devices and the propagation relationships of subsequent graph models.

[0014] It should be noted that the method for setting the unified discrete event time is as follows: First, the nominal sampling period, actual arrival interval, and timing accuracy of all acquisition sources are statistically analyzed. A time step not greater than the fastest main measurement period and compatible with the full-cycle mapping of most slow channels is selected as the event time step. Then, the event time axis is continuously generated according to the current identification window. When 1-second current, 15-second power, and 300-second meteorological data are simultaneously accessed, 1 second can be used as the event time step, and all unified discrete event times are generated sequentially from the start to the end of the window. The method for selecting the time smoothing scale is as follows: For each type of data source, the actual arrival interval within the most recent complete observation period is statistically analyzed. Outliers caused by communication interruptions and duplicate reporting are eliminated. Then, the median arrival interval or quantile stable value is used as the time smoothing scale. For feeder terminal device current with stable uploading cycles, 1 to 2 seconds can be directly used. For 15-second smart meter data, 10 to 20 seconds can be used. For 5-minute meteorological data, 120 to 300 seconds can be used. The data source types and field definitions for multi-source raw data are as follows: the data source is divided into at least eight categories of fields: voltage, current, active power, reactive power, photovoltaic output, load power, switch status and environmental quantity, and unified into five basic attributes: node identifier, time tag, measurement value, quality bit and source type. During on-site implementation, the intelligent fusion terminal can first perform protocol parsing on the protocol messages of different manufacturers, and then rewrite them into a unified field format.

[0015] It should be noted that the handling method for time anomalies and missing data is as follows: records with a true timestamp earlier than the start of the window or later than the end of the window are directly excluded; records reported repeatedly at the same time are retained in the order of quality bit priority, time priority, and most recent valid value priority; short-term missing measurements are allowed to be supplemented by neighboring samples with weights within the time smoothing window; long-term missing measurements are marked as invalid and the missing identifier is synchronously written into the subsequent tensor. The node granularity, branch granularity, and breakpoint mapping method of the static asset map are as follows: nodes should at least cover power supply points, sectionalizing switches, tie switches, transformer area buses, distributed power supply access points, and terminal load points; branches should be divided according to the smallest physical segment with continuous line parameters and consistent equipment attributes; for multiple branches on the same pole or cable junction box locations, a separate node should be formed at the breakpoint and the relationship between adjacent branches should be recorded. The calculation window and update method for the channel time mean and channel time standard deviation are as follows: the mean and standard deviation are calculated separately for each node and channel within each online identification window. The window can be a fixed-length sliding window or an event-triggered window. When implementing a fixed window, the data from the most recent 60 seconds, 300 seconds, or 900 seconds can be taken. When implementing an event-triggered window, continuous samples covering the entire process before and after the disturbance are taken, and the calculation is recalculated after the window slides. It is prohibited to directly reuse old statistics across different windows.

[0016] In one embodiment of the present invention, the charge perturbation density is obtained by calculating the amplitude of the change in the alignment measurement tensor, the time centroid of the charge perturbation density is taken as the charge perturbation frontier, and the node perturbation energy is calculated, including: Based on the physical measurement data within the alignment measurement tensor, the formula for calculating the magnitude of change is as follows: ; in The physical measurement data at the current moment. For physical measurement data at a specific moment in history, The range of change; The formula for calculating the energy adaptive weight, based on the magnitude of the change, is as follows: ; in The total number of discrete times. For power channel set, To iterate through the changes in the variables within the set, Energy-adaptive weights; Based on the energy adaptive weights and the magnitude of change, the formula for calculating the charge disturbance density is as follows: ; in For electrical disturbance density; Based on the discrete time and the charge disturbance density, the formula for calculating the charge disturbance frontier quantity is as follows: ; in For discrete time, For the leading edge of the electrical disturbance; Based on the charge disturbance density, the formula for calculating the nodal disturbance energy is as follows: ; in This refers to the energy of node perturbation.

[0017] It should be noted that the magnitude of change is the absolute difference between physical measurement data at two adjacent discrete moments, used to describe the strength of change of a certain channel at the current moment. The total number of discrete times is the number of discrete time points included in the current identification window, used to limit the range of disturbance accumulation and time centroid calculation. A preferred value is 60 to 900 discrete time points, meaning it should completely cover the occurrence, expansion, and decay process of a single state disturbance. The set of electrical channels is the set of channels used to characterize electrical disturbances, typically composed of voltage amplitude, current amplitude, active power, reactive power, photovoltaic output, and load power. A preferred value is a set that includes at least voltage amplitude, active power, and reactive power, preferably including 6 types of electrical channels, meaning it should simultaneously reflect changes in source-side injection and grid-side response. The magnitude of change of variables within the set is an ergonomic representation of the magnitude of change of each channel in the electrical channel set at each discrete moment, used to complete the accumulation of the total change between channels. The total change of independent channels is the cumulative sum of the magnitude of change of a certain electrical channel within the entire identification window, used to characterize the overall contribution strength of that channel to the current disturbance. The energy-adaptive weights are channel weights normalized based on the total change in independent channels, used to ensure that channels with more significant changes have a higher proportion in the charge perturbation density. The charge perturbation density is the comprehensive perturbation intensity obtained by weighting and summing the energy-adaptive weights and the corresponding channel change amplitudes at each time step, used to describe the perturbation distribution of the node's charge state on the time axis. The weighted total perturbation is a time-weighted sum obtained by multiplying the discrete time by the charge perturbation density and then accumulating the sums on the time axis. The time centroid is the center time extracted based on the centroid of the charge perturbation density distribution on the time axis, used to avoid defining the perturbation time using only a single abrupt change point. The charge perturbation frontier is the main occurrence time of the node's charge perturbation obtained from the time centroid, used for comparison with the electrothermal response frontier. The node perturbation energy is the accumulated result of the charge perturbation density within the entire identification window, used to characterize the total node perturbation intensity and provide a basis for subsequent segment weighting.

[0018] It should be noted that the power channel set is set as follows: at least six types of channels that can directly reflect changes in source load and grid-side status are selected: voltage amplitude, current amplitude, active power, reactive power, photovoltaic output, and load power. Specific channels can be added or removed based on on-site availability. When a certain type of terminal lacks current measurement capabilities, the remaining five types of channels can be retained, but only the effective channels should be summed during weight normalization. The handling of the first moment and window boundary of the change amplitude is as follows: at the first discrete moment of the identification window, the change amplitude can be set to 0 or the last effective physical measurement data of the previous window can be used as a historical reference value. When the window crosses the offline / online boundary of equipment, validity should be determined first, and then a decision should be made on whether differential calculation is allowed. The handling of the power disturbance frontier when the node disturbance energy is zero is as follows: when the power disturbance density within the entire identification window is all 0 or less than a set small threshold, the power disturbance frontier is directly set to the power disturbance frontier of the window center moment or the previous effective window, and the current window is simultaneously marked as a window without significant power disturbance. The window length for power disturbance density is set as follows: it should be determined jointly based on the event duration and the on-site sampling period. For online rapid identification, the window length can be the most recent 60 to 300 seconds, and for offline review, it can be 300 to 1800 seconds. When there is a slow thermal tailing process in the feeder, the window length should be no shorter than one complete power change process. The method for multi-channel missing measurements to participate in the calculation of power disturbance density is as follows: first, remove failed channels according to the effective position of the channel, and then calculate the total change of independent channels and the adaptive weight of energy only for the effective channels. For short-term missing values, the value held at the previous effective moment can be used to participate in the difference; for long-term missing values, they are directly removed from the channel set in this window to prevent invalid zero values ​​from diluting the true disturbance density.

[0019] In one embodiment of the present invention, the equivalent current is calculated based on the alignment measurement tensor, and the electrothermal response density is obtained by recursively calculating the equivalent current along the static asset diagram. The time centroid of the electrothermal response density is taken as the electrothermal response frontier quantity. The thermoelectric misalignment quantity is obtained by subtracting the charge perturbation frontier quantity from the electrothermal response frontier quantity. Based on active power, reactive power, and voltage amplitude, the formula for calculating the equivalent current is as follows: ; in Active power Reactive power Voltage amplitude, To prevent zero positive numbers, Equivalent current; The formula for calculating the path reference resistance is as follows, based on the branch reference resistance of the power supply path: ; in For the set of power supply path branches, For the branch reference resistor, The path reference resistor; Based on the equivalent temperature state, thermal inertia retention coefficient, equivalent ambient temperature, heating mapping coefficient, and equivalent current from the previous moment, the formula for calculating the equivalent temperature state is as follows: ; in This represents the equivalent temperature state at the previous moment. The coefficient of thermal inertia retention. Equivalent ambient temperature The heat mapping coefficient is... This is the equivalent temperature state; The formula for calculating the equivalent path resistance is as follows, based on the path reference resistance, temperature sensitivity coefficient, equivalent temperature state, and reference temperature: ; in For temperature sensitivity coefficient, As the reference temperature, Equivalent path resistance; The formula for calculating the electrothermal response density based on the equivalent path resistance is as follows: ; in The equivalent path resistance at the previous moment. Electrothermal response density; Based on the discrete time and the electrothermal response density, the formula for calculating the electrothermal response frontier is as follows: ; in It is the leading edge quantity of the electrothermal response; Based on the electrothermal response front and the electrical disturbance front, the formula for calculating the thermoelectric misalignment is as follows: ; in This represents the leading edge quantity of the electrical disturbance. This is the thermoelectric misalignment.

[0020] It should be noted that active power is the actual active power output by a node to or absorbed from the network at the current discrete moment, which can be collected through smart meters, distribution transformer monitoring devices, feeder terminal devices, inverter interfaces, or master station telemetry databases. Reactive power is the reactive power exchanged between a node and the network at the current discrete moment, which can be collected through smart meters, feeder terminal devices, distribution transformer monitoring devices, or protection and control devices. Voltage amplitude is the effective voltage amplitude of a node at the current discrete moment, which can be collected through feeder terminal devices, power quality monitoring devices, smart meters, or synchronous measurement devices. Apparent power is the combined power amplitude characterized by active and reactive power, used to represent the overall magnitude of power injected by the node. Equivalent current is the heat-driving proxy quantity calculated using apparent power and voltage amplitude, used to describe the electrothermal load intensity corresponding to the node when real-time full-path current measurement is lacking. The power supply path branch set is the set of all series branches obtained from the power source point to the target node along the static asset map, used to determine the path resistance and electrothermal propagation path of the node. The branch reference resistance is the nominal resistance parameter of a branch at a reference temperature. It serves as the basis for calculating the path reference resistance and equivalent path resistance. It can be obtained from line parameter ledgers, cable nameplate parameters, design drawings, equipment files, or on-site primary parameter measurements. The path reference resistance is the cumulative sum of the reference resistances of all branches within the power supply path branch set, used to characterize the reference resistive path length from the source point to the target node. The thermal inertia retention factor is a recursive coefficient that controls the proportion of the equivalent temperature state retained at the current moment from the previous moment. A preferred value is 0.70 to 0.995, with 0.70 to 0.90 preferred for overhead lines and 0.90 to 0.995 preferred for underground cables. In other words, the stronger the thermal inertia and the slower the heat dissipation, the larger the thermal inertia retention factor should be.

[0021] It should be noted that the equivalent ambient temperature is a characterization of the ambient temperature along the node path, calculated by applying the air temperature around overhead lines, the soil or pipe gallery temperature around underground cables, and necessary wind cooling corrections. It is used to describe the external thermal boundary conditions. The heat mapping coefficient is a proportionality coefficient that maps the square of the equivalent current to the equivalent temperature rise driving force. A preferred value is 0.001 to 0.05 when modeling in per-unit format, meaning it should be calibrated from historical load and temperature rise response, and consistent with the line material, cross-section, and unit system. The inertia retention component is the temperature component retained from the previous moment's equivalent temperature state multiplied by the thermal inertia retention coefficient, used to characterize the continuation of historical thermal states. The heat driving component is the current heat driving quantity formed by superimposing the equivalent ambient temperature and the equivalent current square heat term, used to characterize the heat input caused by the current external environment and load. The equivalent temperature state is a characterization of the node thermal state derived recursively from the previous moment's thermal state, current environmental conditions, and current heat driving quantity, used to calculate the equivalent path resistance. The temperature sensitivity coefficient is a material temperature-sensitive parameter that maps temperature difference to the relative change in resistance. A preferred value is 0.003 to 0.0045, typically close to 0.0039 for copper conductors and close to 0.0040 for aluminum conductors, and can be fine-tuned according to the circuit material and historical calibration results. The reference temperature is used to define the branch reference resistance and temperature difference, preferably 20 or 25 degrees Celsius, meaning it should be consistent with the standard resistance temperature used in the circuit parameter ledger. The temperature difference is the deviation of the equivalent temperature state from the reference temperature, used to characterize the degree of resistance correction caused by changes in thermal state. The equivalent path resistance is the path reference resistance corrected by the temperature sensitivity coefficient under the current equivalent temperature state, used to characterize the real-time resistive change under electrothermal action. The electrothermal response density is a time series formed by the absolute values ​​of the equivalent path resistance changes at adjacent moments, used to describe the response intensity of electrothermal impedance changes on the time axis. The electrothermal response frontier is the central moment corresponding to the time centroid of the electrothermal response density, used to characterize the moment of the main change in electrothermal impedance. Thermoelectric misalignment is the time difference between the leading edge of the electrothermal response and the leading edge of the electrical disturbance, used to characterize the degree of misalignment between changes in electrical quantity and changes in electrothermal quantity.

[0022] It should be noted that the method for determining the power supply path branch set is as follows: First, a radial search tree is established with the power supply point in the static asset diagram as the root node. Then, the shortest level path search or parent node backtracking is performed on the target node to extract all continuous branches between the power supply point and the target node as the power supply path branch set. When there are connecting branches or backup power supplies, the actual power supply path should be selected based on the current effective switch status before backtracking. The equivalent ambient temperature is constructed as follows: for overhead branches, the air temperature is taken and superimposed with wind speed heat transfer correction. For underground cable branches, the soil temperature, pipe gallery temperature, or trench temperature is taken. Then, the equivalent ambient temperature is weighted and summed according to the resistance ratio or length ratio of each branch in the path to form the node-level equivalent ambient temperature. When there is only a single-point ambient quantity on site, correction coefficients can be assigned according to the line type before summing. The values ​​for thermal inertia retention coefficient, heat generation mapping coefficient, temperature sensitivity coefficient, and reference temperature are determined as follows: initial values ​​are first extracted from line design parameters, material types, and historical operating data, and then offline calibration is performed using historical measurement windows; overhead lines can preferentially set a lower thermal inertia retention coefficient, while underground cables can preferentially set a higher thermal inertia retention coefficient; the temperature sensitivity coefficient can be selected around 0.0039 for copper conductors and around 0.0040 for aluminum conductors; and the reference temperature is consistent with the standard resistance value temperature in the ledger.

[0023] It should be noted that the initialization method for the equivalent temperature state at the previous moment is as follows: when the system runs for the first time, the equivalent temperature state at the previous moment can be set to the equivalent ambient temperature or the reference temperature. After entering stable operation, the equivalent temperature state at the end of the previous identification window is directly inherited. When the device is offline for a long time and then comes back online, the initialization should be re-executed to prevent the thermal state error from continuing. The handling method for the first moment of electrothermal response density and the zero denominator scenario is as follows: the electrothermal response density at the first discrete moment can be set to 0 or the equivalent path resistance at the end of the previous window can be used as a historical reference. When calculating the electrothermal response front quantity, if the sum of the electrothermal response density within the entire window is less than the set small threshold, the electrothermal response front quantity is set to the window center moment or the previous valid value, and the current window is simultaneously marked as a window with no significant electrothermal response. The path reference resistance is updated during topology switching as follows: once a change in switch state is detected, the actual power supply path should be updated according to the new switch state first, and then the power supply path branch set and path reference resistance should be recalculated. For topology switching that occurs in the middle of a window, the window can be split into two sub-windows according to the switching time and recursively calculated to prevent the mixing of old and new path parameters.

[0024] Specifically, because it is usually impossible to continuously deploy high-precision current sensors on every conductor segment along the power supply path on-site, reconstructing the equivalent current using active power, reactive power, and voltage amplitude is to construct a proxy quantity consistent with the heating intensity under the condition that node measurements are available but path measurements are insufficient. For example, when the active power and reactive power at a certain node increase simultaneously while the voltage amplitude decreases slightly, the equivalent current will increase synchronously, thus reflecting the trend of intensified heating. Since the thermal state of lines and cables has a significant memory effect, the current temperature is not only determined by the current load but also by the residual heat from the previous moment. Therefore, recursively constructing the equivalent temperature state along the static asset diagram requires incorporating the equivalent temperature state from the previous moment, the current environmental term, and the current heating term. Furthermore, this invention needs to distinguish between actual structural anomalies and apparent drift caused by thermal memory. Therefore, it is not enough to simply consider whether electrical disturbances occur; the chronological relationship between the main moments of the electrothermal response and the main moments of the electrical disturbance must also be compared. The thermoelectric misalignment quantity can be obtained by the difference between the leading edge quantity of the electrothermal response and the leading edge quantity of the electrical disturbance, so as to quantitatively express the phenomenon that electricity changes first and then heat changes.

[0025] In one embodiment of the present invention, the second-order difference generation of nodal shear curvature is calculated based on the thermoelectric misalignment amount at the branch electrical distance in the static asset map, and the segment shear curvature is generated by weighting the nodal shear curvature using nodal perturbation energy, including: The formula for calculating the electrical distance of a branch, based on its resistance and reactance, is as follows: ; in For branch resistance, For branch reactor, For branch electrical distance; Based on the thermoelectric misalignment and the branch electrical distance, the formula for calculating the nodal shear curvature is as follows: ; in This refers to the thermoelectric misalignment at the nodes. As the first parent node, This represents the thermoelectric misalignment of the first parent node. It is the second parent node. This represents the thermoelectric misalignment of the second parent node. The branch electrical distance between the node and its first parent node. This represents the branch electrical distance between the first parent node and the second parent node. To prevent zero positive numbers, For node shear curvature; Based on the nodal perturbation energy and nodal shear curvature, the formula for calculating the segment shear curvature is as follows: ; in For section numbering, For the set of nodes within the physical segment, The node perturbation energy. The energy of perturbation when traversing nodes within the set. The segment shear curvature.

[0026] It should be noted that branch resistance is a resistive parameter of the branch between nodes, used to characterize the active power loss characteristics when electrical energy is transmitted along the branch. It can be obtained from line parameter ledgers, cable nameplates, equipment design drawings, line resistance test results, or asset archives. Branch reactance is an inductive reactance parameter of the branch between nodes, used to characterize the reactive power voltage drop characteristics when electrical energy is transmitted along the branch. It can be obtained from line parameter ledgers, design drawings, impedance model calculation results, or equipment archives. Branch electrical distance is the combined electrical length formed by branch resistance and branch reactance, used to standardize the spatial scale between branches with different line types and impedance levels. The first parent node is the directly adjacent node upstream of the target node along the power supply direction, used to establish the first-order spatial differential relationship of thermoelectric misalignment. The second parent node is the directly adjacent node upstream of the first parent node along the power supply direction, used to establish the second-order spatial differential relationship of thermoelectric misalignment. The first spatial gradient is the normalized result of the difference between the thermoelectric misalignment of a node and the thermoelectric misalignment of the first parent node, expressed as the electrical distance of the corresponding branch. It is used to characterize the rate of change of thermoelectric misalignment between the current node and its upstream node. The second spatial gradient is the normalized result of the difference between the thermoelectric misalignment of the first parent node and the thermoelectric misalignment of the second parent node, expressed as the electrical distance of the corresponding branch. It is used to characterize the rate of change of thermoelectric misalignment in a further upstream path. The node shear curvature is the absolute value of the difference between the first and second spatial gradients, expressed as the second-order spatial bending intensity of the thermoelectric misalignment along the power supply path. The node set within a physical segment is a set of all nodes within the same physical spatial segment, used to aggregate node-level features within a switch segment or line segment. The disturbance normalization weight is a normalized weight obtained by using the node disturbance energy of a single node relative to the sum of the node disturbance energy within the physical segment. It is used to ensure that nodes with strong disturbances account for a higher proportion when aggregating segments. The segment shear curvature is a segment-level structural feature obtained by weighting and accumulating the nodal shear curvatures of each node within a physical segment according to the perturbation normalization weight. It is used to characterize the overall thermoelectric double-front bending strength of the segment.

[0027] It should be noted that the selection method for the first and second parent nodes is as follows: a power supply tree is established with the power source point as the root. For any target node, its direct upstream node in the power supply tree is taken as the first parent node, and then the direct upstream node of the first parent node is taken as the second parent node. When the target node is the first node of a first-level branch, only the first parent node exists, and the second parent node does not exist. The physical segment division method is as follows: the segment boundaries are defined by sectionalizing switches, tie switches, line type change points, cable and overhead line conversion points, and important equipment access points. Lines and nodes that are continuous and have similar electrical properties between the boundaries are divided into the same physical segment. In the transformer substation scenario, it can also be divided according to the continuous power supply segment within the range of a single public transformer outgoing line. The node shear curvature calculation method for root nodes and boundary nodes is as follows: For power source root nodes and first-level nodes with only one upstream node, the node shear curvature can be set to 0, or a one-sided difference can be used instead of a second-order difference; for switching boundaries and the first and last nodes of a segment, the node shear curvature can be calculated only when the effective upstream chain is complete, otherwise it is left empty and ignored during segment aggregation. The handling method for the minimum branch electrical distance scenario is as follows: When the branch resistance and branch reactance cause the branch electrical distance to be less than the preset small threshold, the denominator should be replaced with the sum of the branch electrical distance and the zero-prevention positive number, or the minimum distinguishable impedance value in the line parameter ledger should be directly used to prevent the spatial gradient from being abnormally amplified. In addition, when the sum of the node disturbance energy of all nodes in the physical segment is 0 or less than the set threshold, the disturbance normalization weight of all nodes can be reset to an equal weight value, or the segment shear curvature can be directly recorded as 0 and the segment can be marked as a segment without significant disturbance.

[0028] Specifically, because overhead lines and underground cables differ in thermal inertia and impedance temperature sensitivity, the same disturbance propagating along the feeder will produce leading-edge slip and gradient torsion in the transition section. Therefore, the node shear curvature can characterize the spatial bending of the thermoelectric double leading edge, rather than just the time delay of a single node. For example, a smooth increase in a single load may raise the overall thermoelectric misalignment, but it will not necessarily form a significant peak in the node shear curvature. In addition, distribution feeders are usually radial structures that unfold in layers from the power supply point downstream. The propagation of thermoelectric misalignment along the path naturally has a chain relationship from upstream to downstream. Therefore, the combination of the parent node chain recursion and the radial power supply structure is not a replacement of the arbitrary adjacency relationship in a general graph structure, but rather uses the power supply direction to form an ordered spatial differential chain, which will not be elaborated here.

[0029] In one embodiment of the present invention, the alignment measurement tensor is input into a convolutional layer to extract temporal features, which are then incorporated into the node shear curvature to construct attention weights that aggregate temporal features. Physical constraints are then applied to output node embedding vectors and node pseudo-state amplitudes, including: Based on the alignment measurement tensor and the scale weight proportion component, the formula for calculating the temporal features is as follows: ; in These are the operating parameters for multi-scale temporal convolutional layers. To normalize the input tensor matrix, For scale weight proportions, It is a time-series feature; Based on global temporal characteristics, the formula for calculating the scale weight proportion component is as follows: ; in For global time features, as well as This is a global mapping parameter vector; Based on temporal characteristics, node shear curvature, and endpoint discrete state variables, the formula for calculating graph attention weights is as follows: ; in The weight matrix is ​​a linear parameter mapping matrix. For the temporal characteristics of nodes, as well as The temporal characteristics of neighboring nodes Let be the node shear curvature. as well as Shear curvature for neighboring nodes. , as well as For discrete state quantities at the endpoints, For splicing symbols, For attention score mapping vectors, To correct the nonlinear activation function, Let be the set of the neighborhood range of a node. For graph attention weights; Based on the graph attention weights and temporal features, the formula for calculating the node embedding vector is as follows: ; in It is a non-linear activation mapping. Embed vectors for nodes; Based on the node embedding vector, the formulas for calculating the estimated magnitude of the node voltage and the magnitude of the node pseudo-state are as follows: ; ; in For smoothing positive nonlinear network activation functions, and For voltage-connected network mapping parameters, and For pseudo-state fully connected network mapping parameters, Estimate the magnitude of the node voltage. The amplitude of the node pseudo-state; Based on the branch reference resistance and the node pseudo-state amplitude, the formula for calculating the equivalent adjusted branch resistance is as follows: ; in For the branch reference resistor, The pseudo-state amplitude of adjacent nodes. To adjust the branch resistance equivalently; Based on the estimated magnitude of the node voltage, the equivalent adjusted branch resistance, the average active power flow, and the average reactive power flow, the formula for calculating the physical constraint loss is as follows: ; ; ; in For average active power flow, For average reactive power flow, For the set of downstream terminal nodes of the branch road, Active power Data on photovoltaic power output Reactive power For the set of physical network branches, The total number of branches in the physical network branch set. Estimate the voltage amplitude of the adjacent node on the right end of the branch. Estimate the amplitude of the initial node voltage at the left end of the branch. For branch reactor, This represents the loss due to physical constraints.

[0030] It should be noted that the convolution scale set is the set of discrete-time receptive fields used by multi-scale temporal convolutional layers to simultaneously capture fast-changing perturbations, medium-speed fluctuations, and slow hot tails. Preferred values ​​are 1, 3, and 5, or 3, 5, and 7, meaning it should cover three typical temporal scales: rapid changes, continuous changes, and slow tails within the current event's time step. The preliminary feature matrix is ​​a set of feature maps formed after the alignment measurement tensor is initially extracted at different convolution scales, used to carry the local temporal texture at each scale. The global temporal feature is a global representation obtained by compressing the temporal information of the entire recognition window, used to determine the proportion of different convolutional scales at the current node. The scale weight proportion component is the weight of each convolutional scale generated based on the global temporal feature, used for adaptive fusion of multi-scale convolutional outputs. The temporal feature is the node temporal representation obtained by weighting and summing the preliminary feature matrices at different scales according to the scale weight proportion component, used to participate in subsequent graph attention propagation. The global mapping parameter vector is a set of trainable parameters that maps global temporal features to scale weight components, used to learn the preference of different temporal patterns for convolution scale. The linear parameter mapping weight matrix is ​​a set of trainable matrix parameters that maps node temporal features and neighborhood information to the attention scoring space, used to form the basic linear transformation for graph attention. Endpoint discrete state quantities are discrete information such as node corresponding switch positions, protection action states, equipment activation / deactivation states, or operating modes, used to introduce equipment state constraints in graph attention propagation. These can be obtained through switch quantity acquisition modules, protection and control devices, feeder terminal devices, inverter status words, or the master station remote signaling database.

[0031] It should be noted that the information interaction vector is a comprehensive input vector formed by concatenating the node's temporal features, neighbor node temporal features, node shear curvature, and endpoint discrete state variables, used to calculate the graph attention score. The attention score mapping vector is a trainable parameter that compresses the information interaction vector into a single attention score, used to reflect the strength of the influence of different neighbors on the target node. The graph attention score is the unnormalized association strength between the node and its neighbors under the current temporal and structural conditions, used to calculate the graph attention weight. The graph attention weight is a neighborhood weighting coefficient obtained after normalizing the graph attention score, used to control the contribution ratio of each neighbor node's temporal features during aggregation. The node's neighborhood range set is the set of adjacent nodes directly connected to the target node on the static asset graph or connected according to a set number of hops, used to limit the scope of graph attention propagation. The node embedding vector is the implicit state representation of the node after comprehensively considering temporal features, neighborhood relationships, and node shear curvature, used to output the node voltage estimate amplitude and node pseudo-state amplitude. The voltage fully connected network mapping parameters are a set of trainable parameters that map node embedding vectors to estimated node voltage magnitudes, used to form the voltage output header. The pseudo-state fully connected network mapping parameters are a set of trainable parameters that map node embedding vectors to estimated node pseudo-state magnitudes, used to form the pseudo-state output header.

[0032] It should be noted that the estimated node voltage amplitude is the network's estimation of the node voltage amplitude, used for physical voltage drop verification and operational status identification. The node pseudo-state amplitude is the network's estimation of the apparent drift intensity of node thermal memory, used to correct the equivalent adjustment branch resistance and participate in the section state intensity calculation. The equivalent adjustment branch resistance is the branch resistance obtained by correcting the branch reference resistance using the pseudo-state amplitudes of the nodes at both ends of the branch, used to reflect the impact of pseudo-states on the voltage drop relationship. The downstream terminal node set of a branch is the set of all terminal load nodes covered by a downstream terminal node of a certain branch, used to statistically analyze the average power flow borne by that branch. Photovoltaic output data is the actual output sequence of the distributed photovoltaic system corresponding to the node at each discrete moment, used to deduct the local photovoltaic injection effect from the average active power flow, and can be obtained through inverter communication interfaces, photovoltaic grid-connected meters, distribution area acquisition terminals, or distributed energy access platforms. The average active power flow is the time average of the net active power demand of the downstream terminal nodes of the branch within the identification window, used to construct the physical voltage drop verification item. Average reactive power flow is the time average of the reactive power demand of the downstream terminal nodes within the identification window, used to construct the physical voltage drop verification term. The total number of branches in the connection set is a statistical value of the number of branches in the physical network connection set, used to average the physical constraint loss. The physical constraint loss is a voltage drop error term constructed using estimated node voltage amplitude, equivalent adjusted branch resistance, average active power flow, average reactive power flow, and branch reactance, used to monitor whether the network output meets the physical laws of the distribution network.

[0033] It should be noted that the convolution scale set and convolutional layer structure are set as follows: the aligned measurement tensor of each node is input into at least three sets of one-dimensional temporal convolutional layers. The kernel length can be 1, 3, 5 or 3, 5, 7, with a stride of 1. The boundaries are zero-padding or mirror padding. The number of convolutional output channels can be set layer by layer in the order of 16, 32, 64. When the sampling period is slow, a larger kernel length can be selected according to the equivalent time scale. The global temporal feature is generated as follows: first, average pooling or max pooling is performed on the time axis for the output of each convolution scale, then the multi-scale pooling results are concatenated and compressed into a fixed-length vector through one or two fully connected layers as the global temporal feature of the current node. In engineering implementation, the length can be set to 16-dimensional, 32-dimensional, or 64-dimensional. The encoding method for endpoint discrete state quantities is as follows: discrete states such as switch on / off positions, protection action positions, equipment on / off positions, and inverter operation positions are first converted into 0 and 1 codes, and then concatenated into a discrete state vector according to the endpoint order. One-hot encoding can be used for multi-class operation modes, and an additional invalid flag bit is added when a missing state position appears. The method for determining the neighborhood range set of a node is as follows: based on the static asset graph, the neighborhood range set of a node is formed by default from one-hop directly adjacent nodes. When the feeder structure is sparse or it is necessary to strengthen the upstream and downstream coupling, it can be expanded to a two-hop neighborhood, but the number of neighborhood layers should be fixed before training to prevent inconsistencies in the graph structure scale between different samples.

[0034] It should be noted that the downstream end node set of the branch is determined as follows: First, a depth-first search is performed on the static asset map with the downstream end of the branch as the root to extract all leaf nodes and intermediate load nodes below that end. Then, nodes with actual power injection or load consumption are included in the downstream end node set. If there are normally open connecting branches in the middle, nodes not covered by this branch should be removed according to the current effective power supply direction. The calculation window method for average active power flow and average reactive power flow is as follows: consistent with the current identification window, the downstream net active and reactive power at each discrete time within the window are summed and then averaged over time. Short-term missing values ​​can be imputed by adjacent valid points, while long-term missing values ​​are not included in the mean calculation, and the number of valid samples is recorded simultaneously. The joint optimization method of physical constraint loss and supervised training is as follows: the physical constraint loss, the supervised loss for node voltage estimation amplitude, and the state classification loss are weighted and summed to form the total loss function. In the initial stage, the weight of the supervised loss can be increased to stabilize convergence, and in subsequent stages, the weight of the physical constraint loss can be gradually increased to enhance voltage drop consistency. Common weight ratios can be set to 1:0.5:1 or 1:1:1. The method for setting the boundary of node pseudo-state amplitude constraints is as follows: non-negative activation output is used for node pseudo-state amplitudes, and upper limit pruning or regularization constraints are set during training to avoid distortion of the equivalent adjustment branch resistance caused by abnormally large values. In engineering, the node pseudo-state amplitude can be limited to between 0 and 1, or limited to a range that ensures the equivalent adjustment branch resistance does not exceed twice the reference resistance. The training samples and labels are constructed as follows: historical operating data is divided into samples by a fixed window. The input includes the alignment measurement tensor, static asset graph and structural features calculated from steps two to four. The label includes reference values ​​of node voltage estimation amplitude and labels of the entire network status category. The status category labels can be jointly labeled by scheduling logs, switch operation records, fault handling records and on-site manual verification results, and the sample set is established according to four categories: normal steady state, new energy disturbance-dominated state, thermal memory pseudo-state drift state, and abnormal entity structure state.

[0035] Specifically, because power disturbances, thermal tails, and equipment status changes have different time scales, it is difficult for a single convolutional receptive field to simultaneously detect rapid abrupt changes and slow lags. Therefore, multi-scale temporal convolution to extract temporal features is not a direct replacement for ordinary single-scale convolution. Instead, the short scale is responsible for capturing switching and impact loads, the medium scale is responsible for capturing photovoltaic fluctuations, and the long scale is responsible for capturing thermal memory tails.

[0036] In one embodiment of the present invention, the segment state intensity is obtained by multiplying the node perturbation energy, the node pseudo-state amplitude, and the node shear curvature; the network-wide state probability is obtained by inputting the node embedding vector and the segment shear curvature into the classification layer. The formula for calculating the segment state intensity is as follows, based on the nodal disturbance energy, nodal pseudo-state amplitude, and nodal shear curvature: ; in For a set of nodes in the same physical space segment, For positive bias, The intensity of the segment state; The formula for calculating the global embedding state vector based on the node embedding vector is as follows: ; in For the combination of nodes across the entire power distribution network, This is the global embedded state vector; The formula for calculating the overall network state probability is as follows, based on the global embedded state vector, segment shear curvature, and segment state strength: ; in The system structural feature tensor is constructed by splicing the shear curvature of all segments. The system state amplitude feature tensor is constructed by splicing the state intensity of all segments. and For the classification layer network mapping parameters, This is a normalized numerical scaling exponential function mapping operation. This represents the probability of the entire network state.

[0037] It should be noted that the section comprehensive feedback factor is a comprehensive quantity obtained by continuously multiplying and summing the node disturbance energy, node pseudo-state amplitude, and node shear curvature of each node within the same physical space section, and is used to characterize the overall feedback intensity of the section. The section normalization denominator is the sum of the node disturbance energy of all nodes within the same physical space section, used to eliminate differences in the number of nodes and the total disturbance scale. The section state intensity is the section-level state quantity obtained by dividing the section comprehensive feedback factor by the section normalization denominator, used to represent the degree of dominance of the section on the overall network operation state. The global embedded state vector is the global implicit state representation obtained by averaging the node embedding vectors of all nodes in the entire network, used to describe the overall operation characteristics of the entire distribution network. The system structural feature tensor is a sequence of network-wide structural features formed by splicing the shear curvature of all sections in a fixed section order, used to characterize the spatial bending distribution of the entire network. The system state amplitude feature tensor is a sequence of network-wide state amplitudes formed by splicing the state intensities of all sections in a fixed section order, used to characterize the section-level intensity distribution of the entire network. The classification layer network mapping parameters are a set of trainable parameters that map the global embedded state vector, system structure feature tensor, and system state amplitude feature tensor to the network-wide state probabilities, used to form the network-wide classification output head. The network-wide state probabilities are the class probability distributions obtained after the classification layer normalizes the global operating states, used to output the most likely operating state category of the current network.

[0038] It should be noted that the definition of the entire network state category set is as follows: based on scheduling and operation and maintenance business needs, the state categories are divided into at least four categories: normal steady state, new energy disturbance-dominated state, hot memory pseudo-state drift state, and entity structure abnormal state. Clear criteria and sample label sources are established for each category. If it is necessary to expand to include fault precursor states or control switching states on-site, new categories should be added within the same classification system, and the model should be retrained synchronously. The sorting method for system structure feature tensors and system state amplitude feature tensors is: fixed sorting according to physical segment codes from smallest to largest, or sorting according to the power supply path order from the power source point to the end point. Regardless of the sorting method used, it must remain completely consistent in the training set, validation set, and online inference phases to prevent the same tensor position from corresponding to different segments. The classification layer network structure is configured as follows: the global embedded state vector, system structure feature tensor, and system state amplitude feature tensor are concatenated and input into one or two fully connected classification layers. The hidden layer dimension can be 32, 64, or 128, and the output layer dimension is consistent with the number of state categories. The hidden layers use non-linear activation and appropriate regularization, and the output layer uses normalized probability mapping. Imbalanced class training is handled by using class-weighted loss, oversampling, or event window augmentation to address hot-memory pseudo-state drift states and abnormal entity structures with insufficient sample size, ensuring that a minority of classes receive sufficient gradient contributions during training. During online deployment, the results are still output according to the true prior probability, without changing the category definition during the inference stage. Outliers in the normalized segment state intensity are handled by setting a minimum threshold for windows with excessively small segment normalized denominators. Upper and lower limit pruning or moving average suppression can be added to the segment state intensity to prevent extremely small denominators from causing a single node to amplify the entire segment's state intensity.

[0039] Specifically, because the dominant state at the segment level depends not only on the strength of the disturbance but also on whether the disturbance is accompanied by significant pseudo-state amplitudes and spatial curvature, the segment state strength is obtained by coupling node disturbance energy, node pseudo-state amplitudes, and node shear curvature. This allows for the differentiation between scenarios with strong disturbances but smooth structures and scenarios with strong disturbances and twisted structures. Furthermore, since the overall network state category is influenced by both local segment anomalies and the overall network operating background, using only local segment features or only global node average features is insufficient for stable classification. It is necessary to jointly classify the globally embedded state vector with segment-level features to simultaneously preserve information about the overall background and the dominant local segments.

[0040] In one embodiment of the present invention, the item with the highest probability of state across the entire network is taken as the state category code, and the item with the highest state strength of a segment is taken as the dominant segment code. The segment state strength corresponding to the dominant segment code is then output synchronously, including: Based on the network-wide state probabilities, the formula for calculating the state category code is as follows: ; in For category classification index number, For the network state probability independent components, Positioning operations for comparing the absolute size of numerical values. For state category codes; The formula for calculating the dominant segment code is as follows, based on the segment status strength: ; in This represents the total number of paragraphs in the system. The dominant section code; Based on the segment state strength and the dominant segment code, the formula for calculating the pseudo-state strength is as follows: ; in To extract the true state intensity magnitude across sequences for the dominant segment code, The strength of the pseudo-state; Based on the state category code, dominant segment code, and pseudo-state strength, the final joint identification matrix is ​​calculated using the following formula: ; in The final output is a joint information matrix containing all analyzed components.

[0041] It should be noted that the state category code is the output code obtained from the category index with the largest value in the overall network state probability, used to indicate the category to which the current overall network operating state belongs. The category classification index number is the position number of each state category in the overall network state probability vector, used to complete the location of the highest probability category. The independent component of the overall network state probability is the single probability value corresponding to a single state category in the overall network state probability vector. The total number of system segments is the total number of physical segments obtained from the current distribution network division. The dominant segment code is the output code obtained from the segment index with the largest value in the segment state intensity, used to indicate the physical segment where the current dominant state is located. The extraction of the true state intensity amplitude across sequences of the dominant segment code is the extraction of the corresponding segment state intensity value in the segment state intensity sequence according to the dominant segment code, used to form the pseudo state intensity. The pseudo state intensity is the segment state intensity result corresponding to the dominant segment code, used to quantify the pseudo state level of the current dominant segment. The final output, a joint information matrix containing all analysis components, is a result structure composed of the state category code, the dominant segment code, and the pseudo state intensity, used to synchronously display the identification results to the main station interface, edge terminals, or operation and maintenance system.

[0042] It should be noted that the mapping method between state category codes and business states is as follows: a correspondence table between category indexes and business states is pre-established. For example, 1 corresponds to the normal steady state, 2 corresponds to the new energy disturbance-dominated state, 3 corresponds to the hot memory pseudo-state drift state, and 4 corresponds to the entity structure abnormal state. This correspondence table should be fixed in the model configuration file and the main station display rules, and maintained synchronously when the model is updated. The encoding method for the dominant segment code is as follows: according to the physical segment division results, the codes are sequentially numbered from the power supply side to the end side, or a composite code is formed by adding the feeder number to the segment sequence number. The code within the same feeder must be unique. When deployed across feeders, a feeder prefix should be added to prevent different segments from having the same code. The decision method when multiple segment state intensities are tied for the highest is as follows: first, compare the segment shear curvatures corresponding to the tied segments, and retain the segment with the larger segment shear curvature; if they are still tied, select the segment closer to the power supply side or the segment that has maintained a high value for a longer period in the historical continuous window to ensure that the dominant segment code is unique. The final output, which includes a joint information matrix of all analyzed components, is displayed as follows: the status category code, dominant segment code, and pseudo-status intensity are simultaneously output on both the edge terminal and the main station interface, along with time labels, feeder identifiers, and optional status description text; the main station side can display the results with a single line of results records and a segment highlighting information, while the edge side can display them with event alarm cards.

[0043] Specifically, the actual deployment of this invention can be divided into three parts: the field access layer, the edge computing layer, and the main station application layer. The field access layer deploys intelligent converged terminals, feeder terminal devices, smart meters, distribution transformer monitoring devices, inverter communication interfaces, and environmental data acquisition devices at the feeder heads, sectionalizing switches, ring main units, public transformer areas, and distributed power supply access points. Various devices upload voltage, current, active power, reactive power, switching quantities, photovoltaic output, air temperature, soil temperature, or pipe gallery temperature according to existing protocols. The main station simultaneously imports line parameter ledgers, primary wiring diagrams, and geographic information ledgers, forming a set of physical network nodes and physical network branches for the distribution network. To ensure the availability of accurate timestamps, field terminals can use satellite time synchronization or unified time synchronization from the main station.

[0044] The runtime state identification program of this invention is deployed at the edge computing layer. The program first reads the multi-source raw data within the current window, performs Gaussian smoothing alignment according to a unified discrete event time and time smoothing scale, and generates an aligned measurement tensor and a static asset graph. Subsequently, it sequentially calculates the charge perturbation density, charge perturbation front, node perturbation energy, equivalent current, equivalent temperature state, electrothermal response density, electrothermal response front, thermoelectric misalignment, node shear curvature, and segment shear curvature. Then, the aligned measurement tensor is input into a convolutional and graph attention network, outputting node embedding vectors, node voltage estimation amplitudes, and node pseudo-state amplitudes. These are combined with average active power flow, average reactive power flow, and equivalent adjusted branch resistance to complete physical constraint verification. This process can be performed locally on the intelligent fusion terminal using lightweight inference or centrally on the main station edge server.

[0045] It should be noted that the final output results are the state category code, the dominant segment code, and the pseudo-state strength. For example, if a 10kV overhead line and cable hybrid feeder experiences cloud shadowing and industrial load startup at 14:12, after the system calculates the most recent 300-second window, the state category code is 3, the dominant segment code is 5, and the pseudo-state strength is 0.72. If we pre-define 1 as the normal steady state, 2 as the new energy disturbance dominant state, 3 as the thermal memory pseudo-state drift state, and 4 as the physical structure anomaly state, then the above results indicate that the current event mainly belongs to the thermal memory pseudo-state drift state, mainly falling in the 5th physical segment, and the apparent drift degree of this segment is relatively high. The main station interface can simultaneously display the line name, the equipment at both ends, and the suggested inspection content for this segment, such as prioritizing the inspection of cable current-carrying tail and distributed photovoltaic power output fluctuations in this segment, rather than directly determining it as a branch fault.

[0046] Specifically, the application value of this invention is reflected in three aspects: First, it can distinguish between apparent state drift caused by the difference in thermal inertia of hybrid feeders and actual structural anomalies, reducing the possibility of misjudging thermal tailing as topology changes or equipment failures. Second, it can simultaneously provide the overall network status assessment and the location of the dominant section, facilitating rapid organization of verification and handling by the master station. Third, in scenarios with widespread integration of distributed photovoltaic, energy storage, and fluctuating loads, it can provide unified quantitative results for distribution network operation monitoring, transformer substation maintenance, condition-based maintenance, and event review, especially suitable for distribution networks with a large amount of multi-source asynchronous data and where traditional single-point criteria are difficult to operate stably; details will not be elaborated here.

[0047] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0048] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for identifying the operating status of a distribution network based on multi-source data fusion using intelligent fusion terminals, characterized in that, Includes the following steps: Step S1: Smoothly align the multi-source raw data to the event time axis according to the collection interval to generate the aligned measurement tensor and static asset graph; Step S2: Calculate the change amplitude of the alignment measurement tensor to obtain the charge perturbation density, take the time centroid of the charge perturbation density as the charge perturbation frontier quantity, and calculate the nodal perturbation energy. Step S3: Calculate the equivalent current based on the alignment measurement tensor, and recursively calculate the electrothermal response density along the static asset diagram. Take the time centroid of the electrothermal response density as the electrothermal response frontier quantity, and subtract the charge disturbance frontier quantity from the electrothermal response frontier quantity to obtain the thermoelectric misalignment quantity. Step S4: Calculate the second-order difference generation node shear curvature of thermoelectric misalignment on the branch electrical distance of the static asset map, and use the node perturbation energy to weight the node shear curvature to generate the segment shear curvature. Step S5: Input the alignment measurement tensor into the convolutional layer to extract temporal features, incorporate node shear curvature to construct attention weights to aggregate temporal features, and apply physical constraints to output node embedding vectors and node pseudo-state amplitudes. Step S6: Multiply the node perturbation energy, the node pseudo-state amplitude, and the node shear curvature to obtain the segment state intensity; input the node embedding vector and the segment shear curvature into the classification layer to obtain the overall network state probability. Step S7: Take the item with the highest probability of state across the entire network as the state category code, take the item with the highest state intensity of a segment as the dominant segment code, and output the segment state intensity corresponding to the dominant segment code simultaneously.

2. The method for identifying the operating status of a distribution network based on multi-source data fusion using an intelligent fusion terminal according to claim 1, characterized in that, The time difference generated by subtracting the real timestamps of the multi-source original data from the time of the unified discrete event is calculated. The time difference is divided by the time smoothing scale and the square value is calculated. The negative half of the square value is extracted and used to generate Gaussian smoothing weights by exponential function operation. The Gaussian smoothing weights are multiplied by the original data from multiple sources and the sequence numbers are accumulated. Then, the result is divided by the sum of the accumulated Gaussian smoothing weights to generate physical measurement data. Calculate the difference between the physical measurement data and the channel time mean, divide the difference by the sum of the channel time standard deviation and the zero-prevention positive number to generate normalized measurement data; By combining physical measurement data of all source data types with normalized measurement data, an aligned measurement tensor is constructed, and the relationships between physical network nodes and the connections between physical network branches are extracted to generate a static asset graph.

3. The method for identifying the operating status of a distribution network based on multi-source data fusion using an intelligent fusion terminal according to claim 1, characterized in that, Calculate the absolute value of the discrete difference of physical measurement data in adjacent time intervals for the alignment measurement tensor, and generate the magnitude of change; The total change of each power channel is obtained by summing the changes over time. The total change of each independent channel is then divided by the sum of the total changes of all power channels in the set to generate an energy adaptive weight. The energy adaptive weights are multiplied by their corresponding change magnitudes and accumulated within the set to extract the energy disturbance density; The weighted total disturbance is obtained by multiplying the discrete time by the electrical disturbance density and summing the results on the time axis. The nodal perturbation energy is obtained by simply summing the charge perturbation densities over time. Divide the total weighted disturbance by the node disturbance energy to extract the time centroid, and determine it as the leading edge of the electrical disturbance.

4. The method for identifying the operating status of a distribution network based on multi-source data fusion using an intelligent fusion terminal according to claim 1, characterized in that, Extract the aligned measurement tensor to obtain the arithmetic square root of the sum of squares of active and reactive power, extract the apparent power, and divide the apparent power by the voltage amplitude with zero-prevention positive number to generate the equivalent current. The reference resistances of the branches included in the power supply path from the power source point to the target node are extracted along the static asset map and accumulated to obtain the path reference resistance. The equivalent temperature state at the previous moment is multiplied by the thermal inertia retention coefficient to generate the inertia retention component. The product of the square of the equivalent current and the heating mapping coefficient is calculated and added to the equivalent ambient temperature to generate the heating driving component. The heating driving component is multiplied by the difference between the thermal inertia retention coefficient and the inertia retention component, and then added to the inertia retention component to recursively obtain the equivalent temperature state.

5. The method for identifying the operating status of a distribution network based on multi-source data fusion using an intelligent fusion terminal according to claim 4, characterized in that, The equivalent temperature state is subtracted from the reference temperature to extract the temperature difference. The temperature difference is multiplied by the temperature sensitivity coefficient plus one and then multiplied by the path reference resistance to generate the equivalent path resistance. Calculate the absolute value of the difference in equivalent path resistance between adjacent time points to obtain the electrothermal response density; The time centroid is obtained by summing the product of discrete time and electrothermal response density, and then dividing by the sum of time-based electrothermal response density. This is then extracted as the electrothermal response frontier quantity. Subtract the electrical disturbance front from the electrothermal response front to generate the thermoelectric misalignment.

6. The method for identifying the operating status of a distribution network based on multi-source data fusion using an intelligent fusion terminal according to claim 1, characterized in that, The electrical distance of a branch is calculated by taking the square root of the sum of the squares of the branch resistance and the squares of the branch reactance. The difference between the thermoelectric misalignment of the node and the thermoelectric misalignment of the first parent node is calculated and divided by the branch electrical distance to generate the first spatial gradient. Calculate the difference between the thermoelectric misalignment of the first parent node and the thermoelectric misalignment of the second parent node, and divide it by the branch electrical distance to generate the second spatial gradient; The absolute value of the difference between the first spatial gradient and the second spatial gradient is taken, and the second-order difference is performed to obtain the nodal shear curvature. Extract the node perturbation energy of a single node within the physical segment, and divide it by the sum of the node perturbation energy of all nodes within the physical segment to generate the perturbation normalization weight. The segment shear curvature is extracted by multiplying and summing the perturbation normalized weights and the node shear curvature.

7. The method for identifying the operating status of a distribution network based on multi-source data fusion using an intelligent fusion terminal according to claim 1, characterized in that, The alignment measurement tensor is input into a multi-scale temporal convolutional layer to obtain a preliminary feature matrix, and scale weight proportion components are generated based on the global temporal feature mapping. The temporal features are obtained by multiplying the scale weight proportions with the preliminary feature matrix and summing the results. The temporal features of the spliced ​​nodes, the temporal features of the neighboring nodes, the node shear curvature, the neighboring node shear curvature, and the discrete state variables of the endpoints are used to construct an information interaction vector. The information interaction vector is then substituted into a linear parameter mapping, and the graph attention score is calculated based on the leakage correction nonlinear activation function. Perform a normalized exponential operation on the graph attention score to extract the graph attention weights.

8. The method for identifying the operating status of a distribution network based on multi-source data fusion using an intelligent fusion terminal according to claim 7, characterized in that, The temporal features of neighboring nodes within the neighborhood are weighted and combined using graph attention weights, and the node embedding vector is output using nonlinear activation mapping. The node embedding vector is multiplied and added using an independent fully connected mapping layer, and the output is separated by a smooth positive nonlinear network to obtain the estimated magnitude of the node voltage and the magnitude of the node pseudo-state. Extract the pseudo-state amplitudes of the nodes connecting adjacent nodes at both ends of the branch, sum them, divide by two, and use this sum with the branch reference resistance to update and generate the equivalent adjusted branch resistance. The average active power flow is obtained by extracting the difference between the active power contained in the downstream end of the branch and the photovoltaic output data, and by calculating the average time value. The average reactive power flow is obtained by extracting the reactive power and calculating the average time value. Network supervision training is conducted by constructing a physical voltage drop verification item for the distribution network based on the estimated amplitude of node voltage, equivalent adjusted branch resistance, average active power flow, average reactive power flow, and branch reactance, which serves as the physical constraint loss.

9. The method for identifying the operating status of a distribution network based on multi-source data fusion using an intelligent fusion terminal according to claim 1, characterized in that, The node perturbation energy, node pseudo-state amplitude, and node shear curvature within the same physical space segment are extracted, multiplied continuously, and combined. The data are then summed over the entire range of nodes in the physical space segment to obtain the segment's comprehensive feedback factor. The normalized denominator term of the segment is obtained by summing the node perturbation energy across all nodes in the physical space segment. The section's comprehensive feedback factor is divided by the section's normalized denominator to calculate the section's state strength. Extract the node embedding vectors of all nodes in the entire power distribution network, add them together to calculate the average value, and construct a global embedding state vector; The shear curvature of all sections of the distribution network is spliced ​​together to construct the system structural feature tensor, and the state intensity of all sections of the distribution network is spliced ​​together to construct the system state amplitude feature tensor. The global embedded state vector, system structure feature tensor, and system state amplitude feature tensor are concatenated by vector length, input into the classification layer network topology to perform data dimensionality reduction mapping operation, and output the whole network state probability based on the normalized numerical proportional exponential function mapping operation.

10. The method for identifying the operating status of a distribution network based on multi-source data fusion using an intelligent fusion terminal according to claim 1, characterized in that, Within the overall network state probability, numerical comparison and retrieval are performed to capture the component with the highest absolute value, and the index number of the highest component is extracted and converted into a state category code. The system performs numerical comparison and retrieval within the state intensity of all segments to capture the parameter with the highest absolute value. The highest amplitude parameter is then associated with the corresponding segment number and converted into the dominant segment code. Based on the dominant segment code of the location, the magnitude of the real state intensity is extracted across sequences and directly defined as the pseudo state intensity; The execution result is synchronously displayed by constructing a joint information matrix consisting of state category code, dominant segment code, and pseudo-state strength.

Images