Power distribution area carrying capacity evaluation method and system, electronic device and storage medium

Abstract

The application relates to a power distribution area carrying capacity evaluation method and system, an electronic device and a storage medium, and belongs to the field of power system state evaluation and intelligent analysis. The method solves the problems of inaccurate evaluation and high decision-making risk of the existing low-voltage power distribution area carrying capacity. The method comprises the following steps: collecting multi-dimensional time sequence data of each node of a low-voltage power distribution area and constructing features to obtain a time sequence state vector of each node; performing credibility scoring and anomaly detection of each node based on the time sequence state vector to obtain a node credibility score and a suspicious node set; based on the topology structure of the low-voltage power distribution area, the time sequence state vector of each node is fused into the corresponding credibility score to obtain a time graph sequence; features are extracted from the time graph sequence to obtain a space-time embedding vector of each node; based on the space-time embedding vector of each node, a carrying capacity estimation value, an anomaly score and an anomaly tracing path of each node are obtained; and the carrying capacity estimation value is corrected to obtain a final carrying capacity estimation value of each node.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system state evaluation and intelligent analysis, and particularly relates to a distribution area carrying capacity evaluation method and system, an electronic device and a storage medium. BACKGROUND

[0002] With the continuous advancement of new power system construction, low-voltage distribution areas are facing the challenges of increasingly complex source-load bilateral structure and highly dynamic operation state. Especially under the background of wide access of high proportion of distributed power sources, gradual popularization of electric vehicles, and highly differentiated load power consumption behaviors, the state parameters such as node voltage, current, active power and reactive power fluctuate significantly, which may cause overload operation of the system at a specific time period or a specific node, and increase the risk of power supply reliability.

[0003] The traditional low-voltage distribution area carrying capacity evaluation method mainly relies on steady-state power flow calculation and experience rule judgment, and is difficult to respond to the nonlinear change characteristics of node state in a timely manner, and lacks comprehensive description of real-time dynamic load capacity. Meanwhile, the measured data is affected by factors such as sensing accuracy, communication abnormalities or equipment aging, and there is a certain credibility fluctuation. If not identified and corrected, it may cause deviation of the carrying capacity evaluation result or even misjudgment.

[0004] In addition, under the current operating conditions, the traceability analysis capability of abnormal conditions is still weak, especially in terms of local node abnormal source positioning and propagation path judgment, there is a lack of systematic method, which is difficult to effectively support emergency handling and power grid control strategy optimization. SUMMARY

[0005] In view of the above analysis, the embodiments of the present application aim to provide a distribution area carrying capacity evaluation method fusing data credibility and abnormal source tracing, to solve the technical problems of inaccurate evaluation results and high decision risk caused by data credibility fluctuation and lack of abnormal source tracing capability in the existing low-voltage distribution area carrying capacity evaluation method.

[0006] The present application provides a distribution area carrying capacity evaluation method, comprising the following steps:

[0007] Collecting multi-dimensional time series data of each node in the low-voltage distribution area and constructing features to obtain the time series state vector of each node; based on the time series state vector, performing credibility scoring and abnormal detection of each node to obtain the node credibility score and the suspicious node set;

[0008] Based on the topology structure of the low-voltage distribution area, the time series state vector of each node is fused into the corresponding credibility score to obtain a time graph sequence; feature extraction is performed on the time graph sequence to obtain the space-time embedding vector of each node;

[0009] Based on the spatiotemporal embedding vectors of each node, the bearing capacity estimate, anomaly score, and anomaly tracing path of each node are obtained; the bearing capacity estimate is corrected to obtain the final bearing capacity estimate of each node.

[0010] Furthermore, the credibility score for each node is calculated, as shown below:

[0011] ;

[0012] in, For nodes Credibility score; This is a voltage over-limit indication parameter. If the node voltage amplitude meets the voltage amplitude over-limit rule, then... It is 1 if it is true, otherwise it is 0; This is a parameter indicating a sudden change in node load. If the active power meets the load change rule, then... It is 1 if it is true, otherwise it is 0; This is a parameter indicating missing node data; when data is missing, then... It is 1 if it is true, otherwise it is 0; As an indicator of excessive node volatility, if the standard deviation of active power satisfies the volatility rule, then... If it is 1; otherwise, it is 0. , , and They are respectively , , and The weight.

[0013] Furthermore, if a node's credibility score is less than the credibility threshold, the node is determined to be a suspicious node.

[0014] Furthermore, based on the topology of the low-voltage distribution transformer area, the confidence score is integrated into the time-series state vector to obtain a time-series diagram, including:

[0015] Based on the topology of the low-voltage distribution radio area, the adjacency matrix is ​​obtained. ;

[0016] Based on the adjacency matrix This yields the set of undirected edges between nodes. ;

[0017] The temporal state vector and credibility score of each node are combined. By splicing, fusion, and embedding, the node features of each node are obtained. ;

[0018] Based on nodes and the set of edges between nodes Node features The unit-time graph is obtained by combining the corresponding adjacency matrix. ;

[0019] Based on unit time plot Obtain the time-plotted series ;in, This represents the number of time steps.

[0020] Furthermore, a pre-trained spatiotemporal graph neural network model is used to extract features from the time graph sequence to obtain the spatiotemporal embedding vector of each node;

[0021] The spatiotemporal graph neural network model includes a spatial model and a temporal model in sequence;

[0022] The spatial model is constructed based on the graph attention network (GAT) and is used to obtain the spatial embedding features of each node based on the time graph sequence.

[0023] The time model is constructed based on the Long Short-Term Memory (LSTM) network and is used to capture the dynamic evolution of the spatial embedding features of each node, thereby obtaining the spatiotemporal embedding vector of each node.

[0024] Furthermore, based on the time-map sequence, the spatial embedding features of each node are obtained, including:

[0025] For the nodes in the time graph sequence Node features Perform a linear transformation to obtain the corresponding linear transformation embedding vector. ;

[0026] In the time-map sequence, compute nodes with neighboring nodes Attention score between ;

[0027] Scoring the attention Normalization is performed to obtain the nodes. Normalized attention weights ;

[0028] node Normalized attention weights The node is obtained by multiplying and accumulating the linear transformation embedding vectors of all its neighbors in the neighbor set, followed by ReLU activation. Spatial embedding features .

[0029] Furthermore, by capturing the dynamic evolution of the spatial embedding features of nodes, the spatiotemporal embedding vector of the nodes is obtained, including:

[0030] Will Time step node Spatial embedding features As input to the Long Short-Term Memory (LSTM) network, the hidden states are obtained as follows:

[0031] ;

[0032] in, This is the output gate vector of the Long Short-Term Memory (LSTM) network. This represents the cell state at the current time step. The product of Hadamard;

[0033] Will Hidden state of time As a node spatiotemporal embedding vector .

[0034] Furthermore, two parallel pre-trained MLP networks are used as downstream multitasking heads of the Long Short-Term Memory (LSTM) network; the two MLP networks include a capacity estimation MLP and an anomaly scoring MLP.

[0035] Based on nodes spatiotemporal embedding vector The bearing capacity estimation MLP outputs an estimated bearing capacity value;

[0036] Based on spatiotemporal embedding vectors The anomaly scoring MLP obtains anomaly probability values ​​in the interval [0,1] through the Sigmoid activation function in the output layer, which are used as anomaly scores.

[0037] Furthermore, the estimated bearing capacity is revised, including:

[0038] Inversely map the node anomaly score to the node credibility.

[0039] Nodes calculated based on bearing capacity estimates The global average bearing capacity;

[0040] Based on nodes The node reliability and global mean bearing capacity, and the estimated bearing capacity. Perform weighted corrections to obtain the nodes. Final bearing capacity estimate ,as follows:

[0041] ;

[0042] in, For node credibility, This represents the global average bearing capacity. For nodes The set of neighboring nodes.

[0043] Further, for the node , the normalized attention weight obtained based on the graph attention network GAT , the first maximum values in all the in-edge attention weights are selected;

[0044] The edges corresponding to the first maximum values and the nodes in the corresponding suspicious node set constitute an abnormal propagation path, which is defined as an abnormal tracing path.

[0045] Further, the pre-trained bearing capacity MLP and the abnormal score MLP are obtained through the following training process in parallel:

[0046] The bearing capacity MLP adopts the mean square error loss, and the abnormal score MLP adopts the binary cross entropy loss;

[0047] The total loss is constructed based on the MSE loss and the binary cross entropy loss ;

[0048] The bearing capacity MLP, the abnormal score MLP, the spatial model and the time model parameters are iteratively updated through back propagation and the optimizer Adam; until the total loss converges, the pre-trained bearing capacity MLP and the abnormal score MLP are obtained.

[0049] Further, the multi-dimensional time series data includes the voltage amplitude, the current, the active power, the reactive power, the active power average value of the historical time, the energy storage state, the energy storage charging and discharging power and the node attribute information of the node; wherein, the node attribute information includes the node function type and the partition number of the node in the low-voltage distribution area.

[0050] The node time series state vector is constructed, including:

[0051] The multi-dimensional time series data of each node is normalized to obtain the state vector sequence of each node; The state vector sequence of each node is constructed into a three-dimensional time series state vector

[0052] .

[0053] Further, the node function type includes the ordinary load, the distribution transformer, the distributed power access point, the energy storage access point and the bus node.

[0054] The application further discloses a power distribution area bearing capacity evaluation system, the system includes a data acquisition and credibility evaluation module M1, a space-time graph feature extraction module M2 and a bearing capacity and abnormality evaluation module M3.

[0055] ​​The data collection and credibility evaluation module M1 is used for collecting multi-dimensional time sequence data of each node of the low-voltage distribution area and performing feature construction to obtain a time sequence state vector of each node; credibility scores and abnormality detection of each node are performed based on the time sequence state vector to obtain node credibility scores and a suspicious node set;

[0056] The space-time graph feature extraction module M2 is used for obtaining a time graph sequence based on the topology structure of the low-voltage distribution area, the time sequence state vector of each node and the corresponding credibility score; feature extraction is performed on the time graph sequence to obtain a space-time embedding vector of each node;

[0057] The carrying capacity and abnormality evaluation module M3 is used for obtaining a carrying capacity estimation value, an abnormality score and an abnormality tracing path of each node based on the space-time embedding vector of each node; the carrying capacity estimation value is corrected to obtain a final carrying capacity estimation value of each node.

[0058] The application further discloses an electronic device for evaluating the carrying capacity of a distribution area, which comprises:

[0059] A memory is used for storing a computer program;

[0060] A processor is used for executing the computer program to realize the method for evaluating the carrying capacity of a distribution area.

[0061] The application further discloses a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize the steps of the method for evaluating the carrying capacity of a distribution area.

[0062] Compared with the prior art, the application can realize at least one of the following beneficial effects:

[0063] 1. The application introduces a data credibility quantification and correction mechanism, can identify and correct low credibility data affected by sensing errors and communication abnormalities, effectively identifies and reduces the interference of abnormal data and measurement errors on the carrying capacity evaluation result, improves the robustness and reliability of the output result in a real complex scene, and the evaluation result is more reliable;

[0064] 2. The application fuses the topology structure of the low-voltage distribution area to construct a space-time graph model, can dynamically capture the time sequence evolution characteristics of the node state, and can locate the abnormal source and propagation link through the abnormality tracing path, and makes up for the defects of the traditional method that only performs static analysis and has no abnormal explanation ability. By using the inherent characteristics of the graph attention mechanism, not only the abnormality score can be output, but also the abnormality tracing path can be automatically generated, the problem source is accurately located, a clear decision basis is provided for the operation and maintenance personnel, and the defect of poor explanation of the traditional "black box" model is overcome;

[0065] 3. This invention corrects the carrying capacity estimate based on the credibility score, and outputs the anomaly score and the final carrying capacity value. It provides complete support for flexible resource access planning and power grid operation and maintenance control, which is more practical for engineering than traditional single assessment results and can better adapt to the rapid fluctuation scenarios brought about by the high proportion of distributed energy access.

[0066] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0067] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0068] Figure 1 This is a flowchart of a method for assessing the load-bearing capacity of a distribution substation according to an embodiment of the present invention;

[0069] Figure 2 This is a flowchart illustrating the final bearing capacity estimate, anomaly score, and anomaly tracing path obtained in this embodiment of the invention.

[0070] Figure 3 This is a schematic diagram of a power distribution area bearing capacity assessment system module in an embodiment of the present invention. Detailed Implementation

[0071] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0072] Example 1:

[0073] To address the aforementioned technical issues, there is an urgent need for an evaluation method for low-voltage distribution transformer substation operation scenarios. This method should integrate the reliability assessment mechanism of low-voltage distribution transformer substation operation data with topology modeling to improve the ability to identify abnormal nodes and dynamically estimate the true carrying capacity of low-voltage distribution transformer substations in real time.

[0074] This invention aims to construct an evaluation method consisting of state credibility analysis, graph structure perception, carrying capacity prediction, and anomaly propagation path identification, so as to realize dynamic, efficient, and reliable evaluation of the carrying capacity of key nodes in low-voltage distribution transformer areas, and provide data support for edge sensing terminals and intelligent control decisions.

[0075] One specific embodiment of the present application discloses a power distribution area carrying capacity evaluation method, as shown in the formula (I), comprising the following steps: Figures 1-2

[0076] Step S1, collect multi-dimensional time sequence data of each node in the low-voltage power distribution area and construct features to obtain time sequence state vectors of each node; based on the time sequence state vectors, perform credibility scoring and anomaly detection of each node to obtain node credibility scores and a suspicious node set;

[0077] Step S2, based on the topology structure of the low-voltage power distribution area, the time sequence state vectors of each node are fused into the corresponding credibility scores to obtain a time graph sequence; features are extracted from the time graph sequence to obtain space-time embedding vectors of each node;

[0078] Step S3, based on the space-time embedding vectors of each node, obtain the carrying capacity estimation value, anomaly score and anomaly tracing path of each node; correct the carrying capacity estimation value to obtain the final carrying capacity estimation value of each node.

[0079] In the present application, the node refers to the load access point in the low-voltage power distribution area network, which is the basic analysis unit for carrying capacity evaluation and anomaly identification.

[0080] The key operating node refers to a functional node that has a significant impact on the overall voltage stability, load carrying capacity or abnormal conduction path in the low-voltage power distribution area operation structure. Such nodes have the following characteristics:

[0081] I. Located in a structure position with significant load concentration, topology convergence or reverse power flow;

[0082] II. The state parameters (voltage, current, active power, reactive power, etc.) fluctuation is sensitive to the influence of the surrounding nodes;

[0083] III. The carrying capacity evaluation result, credibility score or anomaly score has a high weight in the evaluation model and significantly affects the final access decision.

[0084] The present scheme collects multi-dimensional time sequence data of the key operating nodes in the low-voltage power distribution area to estimate the carrying capacity of the key operating nodes.

[0085] Step S1 includes steps S11-S13.

[0086] Step S11, collect multi-dimensional time sequence data of each node in the low-voltage power distribution area and construct features to obtain time sequence state vectors of each node.

[0087] For the key operating nodes among all nodes in the low-voltage power distribution area, collect their multi-dimensional time sequence operation data within a specified observation period.

[0088] ​Taking a typical low-voltage distribution area as an example, in order to evaluate the dynamic change trend of the short-term carrying capacity of the key node, the specified observation period is set to the past 14 days.

[0089] During this period, the voltage, current, active power, reactive power and other operation state time series data of the node are collected every 15 minutes;

[0090] A total of 14 days x 24 hours x 4 = 1344 time series data are collected; 3 times of abnormal load fluctuation events and 1 time of voltage limit early warning are recorded. These data are uniformly input into the model as the dynamic operation characteristics of the key operation node for node carrying capacity estimation and abnormality identification.

[0091] Exemplarily, the operation history data of the key operation node are collected through a multi-source heterogeneous data platform, mainly including the following several typical sources:

[0092] (1) Advanced Metering Infrastructure (AMI) system: real-time collection of voltage, current, active power, reactive power and other basic operation parameters of each user node;

[0093] (2) Transformer Terminal Unit (TTU) or Distribution Terminal Unit (DTU) equipment: used to collect load summary data, voltage drop, overcurrent and other abnormal event records of the low-voltage distribution area level, and synchronously record node voltage and current fluctuation information;

[0094] (3) Feeder Terminal Unit (FTU) and Distributed Fault Indicator (DTU): especially for key nodes containing distributed power access, collect characteristic data such as node power flow backfeed and voltage inversion through the feeder intelligent monitoring system.

[0095] (4) Edge Gateway: in the new type of distribution area, the operation state of photovoltaic, energy storage, electric vehicle and other equipment is accessed and fused through the edge gateway, and is uniformly uploaded to the main station system.

[0096] For example, in a certain distribution transformer area, the critical operating node numbered Node23 has an observation period of the past 14 days. It obtains voltage, current and power data every 15 minutes through the smart meter AMI system. At the same time, the low-voltage distribution transformer monitoring terminal (TTU) of its branch recorded 3 voltage drop abnormal events and 2 power rapid fluctuations, which are used as input features for subsequent load capacity estimation and anomaly scoring.

[0097] The multidimensional time-series data includes the node's voltage amplitude, current, active power, reactive power, historical average active power, energy storage status, energy storage charging and discharging power, and node attribute information; wherein, the node attribute information includes the node's functional type and the node's partition number in the low-voltage distribution substation area.

[0098] The construction node time-series state vector ,include:

[0099] Perform multi-dimensional time series data on each node After normalization, the state vector sequence of each node is obtained;

[0100] The state vector sequence of each node is used to construct a three-dimensional temporal state vector. .

[0101] The node function types include ordinary load, distribution transformer, distributed power source access point, energy storage access point, and bus node.

[0102] Historical runtime data of key running nodes are used to construct a multidimensional time-series state vector for each node. This serves as the input basis for subsequent credibility assessment and spatiotemporal neural network models.

[0103] Multidimensional temporal state vector of each node As shown below:

[0104] ;

[0105] in, For nodes At any moment The voltage amplitude; For nodes At any moment The current value; For nodes At any moment Active load; For nodes At any moment reactive load, The expected value is used as a baseline for normal node behavior. For nodes The state of charge of the energy storage system at time is a 0-1 variable; 0 means not running, 1 means running, is the function type of the node is the charge and discharge power of the energy storage device, positive value for discharge, negative value for charge; is the function type of the node, 0 represents a normal load node; 1 represents a transformer node; 2 represents a distributed power access point; 3 represents an energy storage access point; 4 represents a bus node; is the structural partition information of the node in the low-voltage distribution area, which is used to divide subgraphs and analyze local carrying capacity.

[0106] statistical expectation value is the function type of the node at time is the moving average value of the power load at time, which is used to represent the "average" behavior state in a period of time before the current time as a "normal" benchmark, as follows:

[0107] ;

[0108] wherein, represents the sliding period window length (i.e. how many past time points are taken); represents the interval length of the statistical period (for example, every 15 minutes, every 1 hour);

[0109] and are preset hyperparameters, which are automatically selected by analyzing the periodicity and volatility of the node power through historical data and . is the average value of the active load of the node at the historical time.

[0110] is the node partition number or identifier, which is used to divide the entire low-voltage distribution area network into several structural sub-areas (for example, according to branches, geographical location, voltage level, etc.). Each node partition is connected to a certain , as shown in Table 1. Each node is classified into a certain , which is used to:

[0111] (1) support local subgraph construction (such as dividing the distribution area into multiple connected sub-networks);

[0112] (2) assist in local operation analysis or abnormal positioning;

[0113] (3) improve model solving efficiency, such as local clustering, parallel analysis, regional scheduling, etc.

[0114]

[0115] To ensure data quality and model generalization, the original multi-dimensional time series data of each node needs to be normalized before constructing the tensor. For time-continuous sequence variables, such as 、 、 、 、 , normalization is performed to normalize the values to the interval [0, 1] as follows:

[0116] ;

[0117] wherein, is the value of 、 、 、 or after normalization; is the maximum value in the collected original data of 、 、 、 or ; is the minimum value in the collected original data of 、 、 、 or .

[0118] The state vector sequence of each node is constructed into a three-dimensional node state tensor as follows:

[0119] ;

[0120] wherein, is the total number of nodes, is the length of the time window, represents the feature dimension, i.e. the multi-dimensional time series state vector , represents the state tensor of the key operating nodes of the low-voltage distribution substation area network within time steps, which serves as the input for the subsequent model. represents the real number field, indicating that this is a three-dimensional tensor, and each element is a real number.

[0121] Step S12, based on the time series state vector, the credibility score and anomaly detection of each node are performed to obtain the node credibility score.

[0122] A weighted deduction mechanism is adopted to score the node credibility.

[0123] Based on the timing state vector and the preset rule base, the credibility score of each node is calculated as follows:

[0124] ;

[0125] wherein, is the credibility score of the node ; is the voltage out-of-limit indication parameter, if the voltage amplitude of the node meets the voltage amplitude out-of-limit rule, then is 1, otherwise 0; is the node load mutation indication parameter, if the active power meets the load mutation rule, then is 1, otherwise 0; is the node data missing indication parameter, when the data is missing, then is 1, otherwise 0; is the node fluctuation rate too high indication parameter, if the standard deviation of the active power meets the fluctuation rate rule, then is 1, otherwise 0; , , and are the weights of , , and respectively.

[0126] The higher the credibility score is, the more reliable it is, 0 means completely unreliable, and 1 means completely reliable.

[0127] For the weights , , and , it satisfies , .

[0128] Exemplarily, , , and each take the value of 0.2, 0.3, 0.4 and 0.1. In actual application, it can be changed according to specific needs.

[0129] The final output node credibility score vector and the suspicious node are:

[0130] ;

[0131] wherein, is the credibility score of the 1st, 2nd, …, the node.

[0132] The preset rule base includes a rule set for anomaly detection and credibility scoring, and the preset rule base includes a voltage amplitude out-of-limit rule, a load mutation rule, a missing data rule, and a volatility rule.

[0133] (1) The voltage amplitude out-of-limit rule is as follows:

[0134] ;

[0135] wherein, and are upper and lower voltage limits.

[0136] Exemplarily, and are 220V ± 10%, that is, is 198V, is 242V.

[0137] (2) The load mutation rule is as follows:

[0138] ;

[0139] wherein, is a preset mutation judgment threshold.

[0140] Exemplarily, The threshold is generally set based on engineering experience values or dispatching regulations; for low-voltage area conventional users, the power fluctuation is small, 0.5kW is taken; for nodes containing photovoltaic and energy storage, the power fluctuation is large, 2kW is taken.

[0141] (3) The missing data rule: whether there is a missing item; if there is a missing item, is 1.

[0142] Exemplarily, in a certain low-voltage distribution area, when any one of the voltage amplitude, current, active power, reactive power, historical time active power average, energy storage state, energy storage charge and discharge power, and node attribute information is missing in one of the time steps collected every 15 minutes, then is 1;

[0143] When all data are collected, is 0.

[0144] (4) The volatility rule: let the node be , and the load at the last time points be , then the standard deviation is:

[0145] ;

[0146] wherein, is the node standard deviation of load data, used to measure the volatility of load within the statistical time window; is the node active power value at time point ; is the time point index, the th time within the time window ; from 1 to ; is the node average active power within the time window , , represents the “normal” or “average” load level of the node within the time window, which is the baseline for calculating the fluctuation amplitude.

[0147] is the quantitative indicator of volatility, which measures the average deviation of the active power value at each time point from the average value ; the larger the value, the more intense the fluctuation of the power value around the average value, the worse the stability; the smaller the value, the more stable the power value, closely around the average value.

[0148] When , it means that the volatility of the load data of the node within the statistical time window exceeds the judgment threshold, and it is considered that there is an “high-frequency fluctuation” anomaly. is the judgment threshold of volatility, which is a pre-set threshold value.

[0149] Exemplarily:

[0150] For low-voltage residential areas, set = 0.15 kW;

[0151] For transformer incoming lines, set = 2.0 kW;

[0152] If the average power of a node is 5.0 kW, set = 0.5 kW, then if the standard deviation of the last time points is > 0.5, it is considered that the “volatility is too high”, and there is an anomaly.

[0153] Not only active load can be used for volatility rule, other physical quantities such as voltage, current, reactive load can also be applied to volatility rule, only need to replace with the corresponding variable.

[0154] Step S13, obtaining a suspicious node set.

[0155] If the trust score of a node is less than the trust threshold, the node is determined as a suspicious node.

[0156] ;

[0157] wherein, is the trust threshold; is the suspicious node set.

[0158] Exemplarily, the trust threshold is 0.9.

[0159] indicates that the node data trust score meets the standard. Based on the node trust score , it is divided into four levels, as shown in Table 2.

[0160] Based on the preset abnormality levels (mild, moderate and severe abnormality), the suspicious nodes are divided into multiple types. All types of suspicious nodes constitute the suspicious node set.

[0161]

[0162] Based on the preset abnormality levels, the suspicious nodes are divided into multiple types. All types of suspicious nodes constitute the suspicious node set. The abnormality levels include:

[0163] If , the node is determined as a mild abnormality node;

[0164] If , the node is determined as a moderate abnormality node;

[0165] If , the node is determined as a severe abnormality node;

[0166] wherein, , and are the first, second and third abnormality level thresholds, respectively, .

[0167] As shown in Table 2, the first abnormality level threshold is 0.9; the second abnormality level threshold is 0.7; and the third abnormality level threshold is 0.4.

[0168] The function of step S1 is to collect multi-dimensional time sequence data of key operation nodes in low-voltage distribution areas, construct normalized node time sequence state vectors, complete trust score calculation and abnormality detection of each node based on a preset rule library, output node trust scores and suspicious node sets, and provide high-quality data basis for subsequent load capacity prediction and abnormality scoring.

[0169] Step S2 includes steps S21-S22.

[0170] Step S21, based on the topology of the low-voltage power distribution area, the time sequence state vector of each node is fused into the corresponding credibility score to obtain a time graph sequence.

[0171] Based on the topology of the low-voltage power distribution area, the credibility score is fused into the time sequence state vector to obtain a time graph sequence, including:

[0172] Based on the topology of the low-voltage power distribution area, an adjacency matrix is obtained ;

[0173] Based on the adjacency matrix , a set of undirected edges between nodes is obtained ;

[0174] The time sequence state vector and the credibility score of each node are spliced and fused to obtain a node feature of each node ; ;

[0175] Based on the node, the set of edges between nodes , the node feature and the corresponding adjacency matrix, a unit time graph is obtained ;

[0176] Based on the unit time graph , a time graph sequence is obtained ; wherein, is the number of time steps.

[0177] In the state evaluation or optimization decision task of the low-voltage power distribution area, the electrical connection relationship between nodes, the state information of each node and the credibility of the observation data constitute important information sources that can be used for modeling.

[0178] By constructing a graph structure in a unit time scale, the topology relationship of the low-voltage power distribution area, the node state and the credibility score are fused into a unified graph representation, and input is provided for subsequent spatio-temporal graph neural network models.

[0179] Based on the topology of the low-voltage power distribution area, an adjacency matrix is obtained , defined as follows:

[0180] ;

[0181] wherein, is the number of nodes; represents the electrical connection relationship between nodes .

[0182] Exemplarily, the topology is obtained from a line diagram of the power supply system.

[0183] Two nodes exist electrical connection, not only physical cable link, but also conduction (that is, the physical link between the two nodes has the ability of power transmission, the current can flow smoothly in this physical link.).

[0184] Edge set Indicates the electrical connection relationship between nodes, and is constructed according to the actual topology of the low-voltage distribution area. The adjacency matrix Indicates the existence of edges. This definition ensures that the graph structure is consistent with the physical structure of the low-voltage distribution area, making the graph neural network learning process have engineering interpretability.

[0185] Node state feature With the credibility score of each node , the Node state feature of the corresponding node is spliced into , to get the node feature of each node , as follows:

[0186] ;

[0187] Wherein, The node feature vector after splicing the credibility score.

[0188] Node set Indicates all key measurement points or electrical nodes in the low-voltage distribution area, such as transformers, user access points, distributed power sources (such as photovoltaic), energy storage devices, etc. Each node Corresponds to a state feature vector ;

[0189] Unit time graph, as follows:

[0190] ;

[0191] Wherein, .

[0192] Based on the unit time graph, the final time graph sequence of each node is obtained, as follows:

[0193] ;

[0194] Step S22, feature extraction is performed on the time graph sequence to obtain the spatiotemporal embedding vector of each node.

[0195] The pre-trained spatiotemporal graph neural network model is used to extract features from the time graph sequence to obtain the spatiotemporal embedding vector of each node;

[0196] The spatiotemporal graph neural network model includes a spatial model and a temporal model in sequence;

[0197] The spatial model is constructed based on the graph attention network (GAT) and is used to obtain the spatial embedding features of each node based on the time graph sequence.

[0198] The time model is constructed based on the Long Short-Term Memory (LSTM) network and is used to capture the dynamic evolution of the spatial embedding features of each node, thereby obtaining the spatiotemporal embedding vector of each node.

[0199] Based on the time-map sequence, the spatial embedding features of each node are obtained, including:

[0200] For the nodes in the time graph sequence Node features Perform a linear transformation to obtain the corresponding linear transformation embedding vector. ;

[0201] In the time-map sequence, compute nodes with neighboring nodes Attention score between ;

[0202] Scoring the attention Normalization is performed to obtain the nodes. Normalized attention weights ;

[0203] node Normalized attention weights The node is obtained by multiplying and accumulating the linear transformation embedding vectors of all its neighbors in the neighbor set, followed by ReLU activation. Spatial embedding features .

[0204] The spatial model is built upon a graph attention mechanism. The Graph Attention Network (GAT) calculates the information fusion weights of neighboring nodes for each node to obtain a spatial structure representation. Feature extraction is performed on the temporal graph sequences of the nodes, and a shared linear transformation is applied.

[0205] ;

[0206] in, The features obtained by linearly transforming the time graph sequence of the node are used for subsequent operations such as attention weight calculation or feature aggregation. These are trainable hyperparameters.

[0207] For nodes with neighboring nodes Attention scores between nodes are scored as follows:

[0208]

[0209] where, is the transformed feature of node ; is the transpose of trainable parameter , denotes concatenation. The activation function is used to activate the role of the function, usually using a small linear correction with a negative slope to avoid the state of “complete non-update” of neurons.

[0210] Equation (16) is a formula for a single neighbor; all neighbor nodes of node are traversed and summarized through subsequent equation (18).

[0211] denotes the concatenation of the transformed feature vectors of the two nodes; the dimension of the concatenated is 2d. The dimension of

[0212] The softmax is used to obtain normalized attention weights:

[0213]

[0214] where, is the attention weight of node to its neighbor node , indicating the importance of neighbor when aggregating information of node ; is the neighbor node of ; the denominator represents the exponential sum of the scores of all neighbor nodes, which is used for Softmax normalization to ensure that the sum of all is 1; denotes the neighbor set of node .

[0215] The spatial embedding feature representation of the node is:

[0216]

[0217] where, is an activation function. Exemplarily, the activation function is a ReLU activation function.

[0218] The dynamic evolution of the spatial embedding feature of the node is captured to obtain the spatiotemporal embedding vector of the node, including:

[0219] The temporal step node ​​​Spatial embedding features As input to the Long Short-Term Memory (LSTM) network, the hidden states are obtained as follows:

[0220] ;

[0221] in, This is the output gate vector of the Long Short-Term Memory (LSTM) network. This represents the cell state at the current time step. The product of Hadamard;

[0222] Will Hidden state of time As a node spatiotemporal embedding vector .

[0223] Temporal modeling is based on an LSTM network to capture the dynamic evolution of node features. For each node... , its in Given the spatial representation sequence of time steps as input, given the input... The LSTM state is updated as follows:

[0224] ;

[0225] in, The forgetting gate vector determines how much of the past memory is retained; The weight matrix of the input features for the forget gate; Let be the weight matrix of the hidden states of the forget gate; For the bias term of the forget gate;

[0226] This is the output gate vector, which controls the final output. Let be the weight matrix of the input eigenvectors of the output gate. Let be the weight matrix of the hidden states of the output gate; This is the bias term for the input gate;

[0227] The input gate vector determines how much new input is accepted. Let be the weight matrix of the input feature vectors of the input gate. Let be the weight matrix of the hidden states of the input gate; This is the bias term for the output gate;

[0228] The cell state at the previous time step;

[0229] The weight matrix is ​​the input feature vector of the candidate state; The weight matrix for the hidden state of the candidate state; The bias term for the candidate state;

[0230] The Sigmoid activation function; The hyperbolic tangent function for controlling information nonlinear transformation;

[0231] The Hadamard product, corresponding elements are multiplied.

[0232] Finally, , The hidden state at the moment As a spatiotemporal embedding representation of the node.

[0233] The role of step S2 is to fuse the physical topology structure, node state and data credibility of the low-voltage power distribution area into a spatiotemporal graph structure, and extract node spatiotemporal features through the graph attention network GAT and the long short-term memory network LSTM, to provide deep feature representation for carrying capacity assessment and abnormality tracing.

[0234] Step S3 includes steps S31-S33.

[0235] Step S31 obtains the carrying capacity estimation value, abnormality score and abnormality tracing path of each node based on the spatiotemporal embedding vector of each node.

[0236] A pre-trained two-way MLP network is used as the downstream multi-task head of the long short-term memory network LSTM; the two-way MLP network includes a carrying capacity estimation MLP and an abnormality score MLP;

[0237] Based on the spatiotemporal embedding vector of the node , the carrying capacity estimation MLP outputs the carrying capacity estimation value.

[0238] Based on the spatiotemporal embedding vector , the abnormality score MLP outputs an abnormality probability value in the interval [0, 1] through the Sigmoid activation function in the output layer, as the abnormality score.

[0239] For the final embedding of each node , a two-way MLP network is used for carrying capacity prediction and abnormality scoring respectively:

[0240] The carrying capacity estimation is as follows:

[0241] ;

[0242] Wherein, is the node The predicted value of the carrying capacity of the node, in kilowatts, indicating the maximum power that the node can carry under the current operating state The predicted value of the maximum accessible power capacity under the current operating state; for example, if the predicted value is 5 kW, it means that the power of the newly connected distributed photovoltaic or energy storage equipment at the node should not exceed 5 kW to ensure the safe and stable operation of the low-voltage distribution area;

[0243] MLP for carrying capacity prediction, MLP dedicated to the carrying capacity task, independent of the MLP parameters for the anomaly score task, The spatio-temporal embedding vector of the node , which is the final result calculated by the graph attention network GAT and LSTM of the previous steps.

[0244] The physical meaning of formula (20) is to input the spatio-temporal embedding vector of the node into a trained MLP network to predict the corresponding node carrying capacity estimate.

[0245] Using the representation information extracted from the node under the spatial topology and temporal evolution, combined with the nonlinear mapping ability of the MLP network, the maximum carrying capacity of the node under the given operating condition is predicted.

[0246] Anomaly score, as follows:

[0247] ;

[0248] where, is the anomaly score prediction value of the node , a scalar, with a value between 0 and 1, used to measure the likelihood of the node being abnormal, the closer to 1, the higher the likelihood of the node being abnormal; the closer to 0, the higher the likelihood of the node being normal;

[0249] MLP dedicated to the anomaly score task, which receives the spatio-temporal embedding vector of the node as input, and outputs the anomaly probability of the node through multi-layer full connection and nonlinear transformation; MLP dedicated to the anomaly task, independent of the MLP parameters for the carrying capacity estimate task.

[0250] The physical meaning of formula (21) is to input the spatio-temporal embedding vector of the node into the anomaly detection MLP module to output the anomaly score of the node.

[0251] ​​Two MLP networks are used to predict carrying capacity and anomaly scoring, respectively. LSTM is a shared encoder, and MLP is a downstream multitasking head of the LSTM network.

[0252] The training of the bearing capacity MLP and the anomaly scoring MLP is as follows:

[0253] The parallel pre-trained bearing capacity MLP and anomaly scoring MLP are obtained through the following training process:

[0254] The load-bearing capacity MLP uses mean squared error loss; the anomaly scoring MLP uses binary cross-entropy loss.

[0255] Construct the total loss based on MSE loss and binary cross-entropy loss. ;

[0256] The parameters of the bearing capacity MLP, anomaly score MLP, spatial model, and temporal model are updated iteratively through backpropagation and the Adam optimizer until the total loss is reached. Convergence yields the pre-trained bearing capacity MLP and anomaly score MLP.

[0257] Sample data: Original multidimensional time-series data of nodes, and corresponding spatiotemporal embedding vectors;

[0258] Sample Label 1: The actual bearing capacity that the bearing capacity estimation MLP is expected to predict, corresponding to this sample data. ;

[0259] Sample Label 2: The actual anomaly score that the anomaly score MLP is expected to predict for this sample data.

[0260] Define two loss functions for supervised training:

[0261] Bearing capacity estimation MLP, using MSE loss As shown below:

[0262] ;

[0263] in, For the first The actual carrying capacity of each node.

[0264] Anomaly scoring MLP uses binary cross-entropy (BCE) loss. As shown below:

[0265] ;

[0266] in, For the first The actual value of the anomaly score for each node.

[0267] Total loss As follows:

[0268] ;

[0269] Wherein, Respectively And Is the task weight coefficient, and .

[0270] Step S32, get the abnormal source path.

[0271] For node , based on the normalized attention weight Obtained by graph attention network GAT Select the top Maximum value in all incoming edge attention weight;

[0272] The edge corresponding to the top Maximum value, and the node in the corresponding suspicious node set, constitute the abnormal propagation path, defined as the abnormal source path.

[0273] The attention weight Obtained in the space modeling stage is used to derive the abnormal path. For each target node , select the top-k maximum edge in all incoming edge attention to constitute the source path:

[0274] ;

[0275] Wherein, Is the abnormal path edge set of node ; Is the graph attention weight, indicating the information influence degree of neighbor node On target node ; Indicates that the top Maximum value (here, the edge with the largest weight) is selected from the candidate set.

[0276] The physical meaning of formula (22) is that for the target node , the top K maximum edge is selected from all incoming edge attention coefficients , to constitute the abnormal path explanation set . These edges are regarded as the path that has the greatest influence on the final prediction abnormal score Of the node in the graph neural network propagation stage, which can be understood as the "main source path of abnormality".

[0277] The path can explain the abnormal source of the node .

[0278] So far, the technical scheme obtains three values:

[0279] (1) Node bearing capacity estimation value , the first output head is used to fit the real node bearing capacity. After the model input historical measurement data, the LSTM encoder outputs ; after a linear transformation or MLP regression head, output ;

[0280] (2) Node anomaly score value , the second output head is used to judge whether the node is abnormal (such as load mutation, measurement drift, etc.), output ;

[0281] (3) Abnormal source path , which is obtained by spatial attention mechanism from suspicious nodes, represents the influence degree of other nodes on the current node abnormal state, so as to realize the positioning and path explanation of abnormal reason.

[0282] Step S33, correct the bearing capacity estimation value to obtain the final bearing capacity estimation value of each node.

[0283] The correction of the bearing capacity estimation value includes:

[0284] Reverse mapping the node anomaly score to the node credibility;

[0285] Calculate the global bearing capacity mean of the node based on the bearing capacity estimation value;

[0286] Based on the node credibility and the global bearing capacity mean of the node , the bearing capacity estimation value is weighted and corrected to obtain the final bearing capacity estimation value of the node , as follows:

[0287] ;

[0288] Wherein, is the node credibility, is the global bearing capacity mean, is the neighbor node set of the node .

[0289] In real-world low-voltage distribution transformer area operation scenarios, node capacity prediction is susceptible to interference from abnormal behavior or measurement errors. To improve the reliability of capacity estimation results under abnormal conditions, this step introduces a node credibility scoring mechanism to weight and correct the initial capacity prediction value, outputting a more credible recommended access capacity value, thus providing more robust decision support for the scheduling and planning of flexible resources.

[0290] Node credibility score Defined as a reflection of anomaly scoring, it indicates the reliability of node prediction results.

[0291] The following relationship is established between credibility and anomaly scoring:

[0292] ;

[0293] Introducing node credibility Preliminary prediction values ​​for node bearing capacity estimates After weighted adjustment, the final access capacity is obtained. .

[0294] The final result of this technical solution is the node. At any moment The final estimated bearing capacity Node anomaly score and abnormal source tracing path .

[0295] Step S3 is to output the bearing capacity estimate and anomaly score in parallel based on the spatiotemporal embedding vector, generate anomaly tracing path by combining graph attention mechanism, and correct the final bearing capacity estimate result by confidence weighting.

[0296] In practical applications, the historical operating information of the low-voltage distribution substation (including voltage, current, power, load, etc.) and the node number N17 to be evaluated are input into the spatiotemporal graph neural network model constructed in this invention. The model automatically integrates the node's state characteristics, network structure, and reliability information, outputting an initial load-bearing capacity estimate of 6.3kW for the node, and simultaneously providing an anomaly score for the node. This reflects a certain degree of deviation or disturbance in its data.

[0297] By reverse tracing the abnormal score, the model automatically identifies the upstream path that has the greatest impact on the abnormal score as N8→N12→N17 based on the attention weight graph, where node N12 is continuously assigned a high attention weight (maximum value of 0.84) in multiple rounds of training, indicating that it has the maximum weight in the abnormal score propagation and is suspected to be the key node that triggers the prediction uncertainty. Further analysis shows that N12 has mixed characteristics of power spikes and voltage drops in two consecutive time periods, which is most likely caused by load mutation or equipment jitter.

[0298] Combined with the abnormal score of this node, a weighted correction strategy is adopted:

[0299] ;

[0300] The final output of the reliability-enhanced carrying capacity is 6.06 kW, which is slightly lower than the initial prediction value, further avoiding the risk of excessive access caused by abnormal disturbances. This result provides stable and reliable support for the capacity planning of subsequent flexible resources (such as household energy storage systems and distributed photovoltaics), and also provides accurate abnormal positioning guidance for edge and end operation and maintenance personnel, effectively improving the risk perception and operation decision-making ability of low-voltage distribution areas in complex and uncertain scenarios.

[0301] This method introduces a node data reliability evaluation mechanism to weight and correct the node state affected by sensing errors or communication anomalies, significantly improving the accuracy of carrying capacity prediction in abnormal or incomplete data scenarios and reducing the risk of misjudgment.

[0302] This method combines graph structure modeling and state quantity sensing ability, can dynamically respond to changes in node state, and real-time depict the operation characteristics and carrying capacity of each node. Compared with static power flow analysis, it is more practical and timely.

[0303] By constructing an abnormal score mechanism and a spatial propagation path analysis model, abnormal nodes can be identified and their possible sources and propagation links can be located, improving the fine level of abnormal handling and grid regulation.

[0304] With the reliability-enhanced carrying capacity evaluation results, recommendations for subsequent flexible resource access such as energy storage devices and distributed power sources are provided, enhancing the robustness and response efficiency of intelligent edge control.

[0305] The scheme covers multiple aspects such as reliability analysis, graph modeling, capacity prediction, and abnormal source tracing, forming a unified and scalable carrying capacity evaluation system with strong engineering adaptability and application promotion value.

[0306] Example Two:

[0307] One specific embodiment of the present application discloses a power distribution area carrying capacity evaluation system, so as to realize the power distribution area carrying capacity evaluation method in embodiment one. The specific implementation of each module refers to the corresponding description in embodiment one.

[0308] As shown in Figure 3 A power distribution area carrying capacity evaluation system, the system comprises a data acquisition and credibility evaluation module M1, a space-time graph feature extraction module M2 and a carrying capacity and anomaly evaluation module M3;

[0309] The data acquisition and credibility evaluation module M1 is used for acquiring multi-dimensional time series data of each node of a low-voltage power distribution area and constructing features to obtain a time series state vector of each node; credibility scores and anomaly detection of each node are performed based on the time series state vector to obtain node credibility scores and a suspicious node set;

[0310] The space-time graph feature extraction module M2 is used for obtaining a time graph sequence based on the topology structure of the low-voltage power distribution area, the time series state vector of each node and the corresponding credibility score; feature extraction is performed on the time graph sequence to obtain a space-time embedding vector of each node;

[0311] The carrying capacity and anomaly evaluation module M3 is used for obtaining carrying capacity estimation values, anomaly scores and anomaly tracing paths of each node based on the space-time embedding vector of each node; the carrying capacity estimation values are corrected to obtain final carrying capacity estimation values of each node.

[0312] Since the system in this embodiment is related to the method in embodiment one, they can be mutually referred to, and here is repeated description, so here is not described again. Since the system embodiment and the above-mentioned method embodiment have the same principle, the system embodiment also has the corresponding technical effects of the above-mentioned method embodiment.

[0313] Embodiment three

[0314] Another specific embodiment of the present application discloses an electronic device for power distribution area carrying capacity evaluation, which comprises:

[0315] A memory for storing a computer program;

[0316] A processor for executing the computer program to realize the power distribution area carrying capacity evaluation method.

[0317] Embodiment four

[0318] Another specific embodiment of the present application discloses a computer readable storage medium, which stores a computer program, and the computer program realizes the steps of the power distribution area carrying capacity evaluation method when executed by a processor.

[0319] In summary, the power distribution area carrying capacity evaluation method, system, electronic device and storage medium provided by the embodiment of the application have the following beneficial effects.

[0320] 1、The application can identify and correct low-confidence data affected by sensing errors and communication abnormalities by introducing a data confidence quantification and correction mechanism, effectively identifying and reducing the interference of abnormal data and measurement errors on the carrying capacity evaluation results, improving the robustness and reliability of the output results in real complex scenarios, and the evaluation results are more reliable.

[0321] 2、The application fuses the topology structure of the low-voltage power distribution area to construct a space-time graph model, which can not only dynamically capture the time sequence evolution characteristics of the node state, but also can locate the abnormal source and propagation link through the abnormal source path, and make up for the defects of the traditional method that only analyzes statically and has no abnormal explanation ability. By using the inherent characteristics of the graph attention mechanism, not only the abnormal score can be output, but also the abnormal source path can be automatically generated to accurately locate the problem source, which provides clear decision basis for the operation and maintenance personnel and overcomes the poor explanation of the traditional "black box" model.

[0322] 3、The application corrects the carrying capacity estimation value based on the confidence score, simultaneously outputs the abnormal score and the final carrying capacity value, and provides complete support of "evaluation + explanation + decision" for flexible resource access planning and power grid operation and regulation, which is more practical than the traditional single evaluation result and can better adapt to the rapid fluctuation scenario caused by high proportion of distributed energy access.

[0323] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium, wherein the computer readable storage medium is a disk, an optical disk, a read-only memory or a random access memory, etc.

[0324] The above is only a preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A method for assessing the load-bearing capacity of a power distribution station area, characterized in that, include: Multidimensional time-series data of each node in the low-voltage distribution transformer area are collected and features are constructed to obtain the time-series state vector of each node. The credibility score and anomaly detection of each node are performed based on the temporal state vector to obtain the node credibility score and the set of suspicious nodes. Based on the topology of the low-voltage distribution transformer area, the temporal state vector of each node is integrated with the corresponding confidence score to obtain a time graph sequence; feature extraction is performed on the time graph sequence to obtain the spatiotemporal embedding vector of each node. Based on the spatiotemporal embedding vectors of each node, the estimated bearing capacity, anomaly score, and anomaly tracing path of each node are obtained. The bearing capacity estimates are corrected to obtain the final bearing capacity estimates for each node; Two parallel, pre-trained MLP networks are used as downstream multitasking heads of the Long Short-Term Memory (LSTM) network; the two MLP networks include a capacity estimation MLP and an anomaly scoring MLP. Based on nodes spatiotemporal embedding vector The bearing capacity estimation MLP outputs an estimated bearing capacity value; Based on spatiotemporal embedding vectors The anomaly scoring MLP obtains anomaly probability values ​​in the interval [0,1] through the Sigmoid activation function in the output layer, which are used as anomaly scores; The bearing capacity estimate is corrected, including: Inversely map the node anomaly score to the node credibility. Nodes calculated based on bearing capacity estimates The global average bearing capacity; Based on nodes The node reliability and global mean bearing capacity, and the estimated bearing capacity. Perform weighted corrections to obtain the nodes. Final bearing capacity estimate ,as follows: ； in, For node credibility, This represents the global average bearing capacity. For nodes The set of neighboring nodes; For nodes Normalized attention weights obtained based on the graph attention network GAT Select the first of all incoming edge attention weights One maximum value; forward The edges corresponding to the maximum values, and the nodes in the corresponding set of suspicious nodes, constitute the anomaly propagation path, which is defined as the anomaly tracing path.

2. The method for assessing the load-bearing capacity of a distribution substation according to claim 1, characterized in that, Based on the temporal state vector and a preset rule base, the credibility score of each node is calculated as follows: ； in, For nodes Credibility score; This is a voltage over-limit indication parameter. If the node voltage amplitude meets the voltage amplitude over-limit rule, then... It is 1 if it is true, otherwise it is 0; This is a parameter indicating a sudden change in node load. If the active power meets the load change rule, then... It is 1 if it is true, otherwise it is 0; This is a parameter indicating missing node data; when data is missing, then... It is 1 if it is true, otherwise it is 0; As an indicator of excessive node volatility, if the standard deviation of active power satisfies the volatility rule, then... If it is 1; otherwise, it is 0. , , and They are respectively , , and The weight.

3. The method for assessing the load-bearing capacity of a distribution substation according to claim 2, characterized in that, If a node's credibility score is less than the credibility threshold, the node is considered a suspicious node.

4. The method for assessing the load-bearing capacity of a distribution substation according to claim 1, characterized in that, Based on the topology of the low-voltage distribution transformer area, the confidence score is integrated into the time-series state vector to obtain a time-graph sequence, including: Based on the topology of the low-voltage distribution radio area, the adjacency matrix is ​​obtained. ; Based on the adjacency matrix This yields the set of undirected edges between nodes. ; The temporal state vector and credibility score of each node are combined. By splicing, fusion, and embedding, the node features of each node are obtained. ; Based on nodes and the set of edges between nodes Node features The unit-time graph is obtained by combining the corresponding adjacency matrix. ; Based on unit time plot Obtain the time-plotted series ;in, This represents the number of time steps.

5. The method for evaluating the bearing capacity of a distribution substation according to claim 4, characterized in that, The spatiotemporal graph neural network model is used to extract features from the time graph sequence to obtain the spatiotemporal embedding vector of each node; The spatiotemporal graph neural network model includes a spatial model and a temporal model in sequence; The spatial model is constructed based on the graph attention network (GAT) and is used to obtain the spatial embedding features of each node based on the time graph sequence. The time model is constructed based on the Long Short-Term Memory (LSTM) network and is used to capture the dynamic evolution of the spatial embedding features of each node, thereby obtaining the spatiotemporal embedding vector of each node.

6. The method for evaluating the bearing capacity of a distribution substation according to claim 5, characterized in that, Based on the time-map sequence, the spatial embedding features of each node are obtained, including: For the nodes in the time graph sequence Node features Perform a linear transformation to obtain the corresponding linear transformation embedding vector. ; In the time-map sequence, compute nodes with neighboring nodes Attention score between ; Scoring the attention Normalization is performed to obtain the nodes. Normalized attention weights ; node Normalized attention weights The node is obtained by multiplying and accumulating the linear transformation embedding vectors of all its neighbors in the neighbor set, followed by ReLU activation. Spatial embedding features .

7. The method for evaluating the bearing capacity of a distribution substation according to claim 6, characterized in that, By capturing the dynamic evolution of the spatial embedding features of nodes, the spatiotemporal embedding vector of the nodes is obtained, including: Will Time step node Spatial embedding features As input to the Long Short-Term Memory (LSTM) network, the hidden states are obtained as follows: ； in, This is the output gate vector of the Long Short-Term Memory (LSTM) network. This represents the cell state at the current time step. The product of Hadamard; Will Hidden state of time As a node spatiotemporal embedding vector .

8. The method for assessing the load-bearing capacity of a distribution substation according to claim 1, characterized in that, The parallel pre-trained bearing capacity MLP and anomaly scoring MLP are obtained through the following training process: The load-bearing capacity MLP uses mean squared error loss; the anomaly scoring MLP uses binary cross-entropy loss. Construct the total loss based on MSE loss and binary cross-entropy loss. ; The parameters of the bearing capacity MLP, anomaly score MLP, spatial model, and temporal model are updated iteratively through backpropagation and the Adam optimizer until the total loss is reached. Convergence yields the pre-trained bearing capacity MLP and anomaly score MLP.

9. The method for assessing the load-bearing capacity of a distribution substation according to claim 1, characterized in that, The multidimensional time-series data includes the node's voltage amplitude, current, active power, reactive power, historical average active power, energy storage status, energy storage charging and discharging power, and node attribute information; wherein, the node attribute information includes the node's functional type and the node's partition number in the low-voltage distribution substation area. Constructing node temporal state vectors ,include: Perform multi-dimensional time series data on each node After normalization, the state vector sequence of each node is obtained; The state vector sequences of each node are used to construct a three-dimensional node temporal state vector. .

10. The method for evaluating the bearing capacity of a distribution substation according to claim 9, characterized in that, The node function types include ordinary load, distribution transformer, distributed power source access point, energy storage access point, and bus node.

11. A power distribution area load-bearing capacity assessment system, characterized in that, The system includes a data acquisition and credibility assessment module M1, a spatiotemporal map feature extraction module M2, and a bearing capacity and anomaly assessment module M3. The data acquisition and credibility assessment module M1 is used to collect multi-dimensional time-series data of each node in the low-voltage distribution transformer area and construct features to obtain the time-series state vector of each node; based on the time-series state vector, the credibility score of each node and anomaly detection are performed to obtain the node credibility score and a set of suspicious nodes. The spatiotemporal graph feature extraction module M2 is used to integrate the temporal state vectors of each node into the corresponding confidence scores based on the topology of the low-voltage distribution transformer area to obtain a time graph sequence; and to extract features from the time graph sequence to obtain the spatiotemporal embedding vectors of each node. The bearing capacity and anomaly assessment module M3 is used to obtain the bearing capacity estimate, anomaly score and anomaly tracing path of each node based on the spatiotemporal embedding vector of each node. The bearing capacity estimates are corrected to obtain the final bearing capacity estimates for each node; Two parallel, pre-trained MLP networks are used as downstream multitasking heads of the Long Short-Term Memory (LSTM) network; the two MLP networks include a capacity estimation MLP and an anomaly scoring MLP. Based on nodes spatiotemporal embedding vector The bearing capacity estimation MLP outputs an estimated bearing capacity value; Based on spatiotemporal embedding vectors The anomaly scoring MLP obtains anomaly probability values ​​in the interval [0,1] through the Sigmoid activation function in the output layer, which are used as anomaly scores; The bearing capacity estimate is corrected, including: Inversely map the node anomaly score to the node credibility. Nodes calculated based on bearing capacity estimates The global average bearing capacity; Based on nodes The node reliability and global mean bearing capacity, and the estimated bearing capacity. Perform weighted corrections to obtain the nodes. Final bearing capacity estimate ,as follows: ； in, For node credibility, This represents the global average bearing capacity. For nodes The set of neighboring nodes; For nodes Normalized attention weights obtained based on the graph attention network GAT Select the first of all incoming edge attention weights One maximum value; forward The edges corresponding to the maximum values, and the nodes in the corresponding set of suspicious nodes, constitute the anomaly propagation path, which is defined as the anomaly tracing path.

12. An electronic device for assessing the load-bearing capacity of a distribution station area, characterized in that, The electronic device includes: Memory, used to store computer programs; A processor for executing the computer program to implement the distribution area carrying capacity assessment method as described in any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the distribution area carrying capacity assessment method as described in any one of claims 1 to 10.

Images