CN116523148B - Distribution network distribution transformer overload early warning method, device and equipment - Google Patents

Abstract

The invention discloses a distribution transformer overload early warning method, device and equipment for a distribution transformer of a distribution network. The method comprises the steps of generating a predicted load rate through a preset target load rate prediction model, calculating a heavy overload risk value according to the predicted load rate, and further determining an early warning result of heavy overload of the distribution transformer. The preset target load rate prediction model comprises a graph sampling aggregation layer, a long-short-period memory layer, a first full-connection layer and a second full-connection layer, wherein the graph sampling aggregation layer integrates operation time sequence data into high-dimensional time sequence data, and combines the learning advantage of the long-short-period memory layer on the time sequence data, so that the accuracy of distribution change heavy overload early warning under the operation condition of high-dimensional characteristic variables and small sample data is improved; and the heavy overload risk value is calculated according to the predicted load rate, and then the early warning result is determined, so that the conservation of dividing the early warning level directly according to the predicted load rate interval is avoided, and the accuracy of the heavy overload early warning of the distribution transformer is further improved.

Description

Distribution network distribution transformer overload early warning method, device and equipment

Technical Field

The invention relates to the technical field of power distribution network risk early warning, in particular to a power distribution network distribution transformer overload early warning method, device and equipment.

Background

With popularization and promotion of roof photovoltaic power generation, a large number of medium-voltage distribution network areas implement photovoltaic grid connection. The full generation of the photovoltaic transformer area internet users can cause the reverse heavy overload of a distribution transformer (hereinafter referred to as distribution transformer) in the period of power generation peak, and if the distribution transformer is not processed in time, serious consequences such as distribution transformer damage, the photovoltaic grid-connected inverter quit operation, electric equipment burnout and the like can be caused. Therefore, the problem of heavy overload of the distribution transformer is effectively pre-warned in time, so that the power supply quality is improved and the equipment damage is avoided.

Because the advancing time of the photovoltaic on the roof of the county is relatively late, the training samples of part of the photovoltaic transformer areas at the current stage are insufficient, the running situation of high-dimensional characteristic variables and small sample data is difficult to comprehensively consider in the existing distribution transformer overload early warning method, the early warning result is low in accuracy, and the effect of eliminating the heavy overload hidden danger cannot be achieved in the transformer areas accessed by the photovoltaic; when the distribution transformer actually operates, short-time heavy overload can be performed under certain conditions without influencing the service life of the distribution transformer, but the pre-warning level of the current distribution transformer heavy overload pre-warning method based on load rate prediction is divided directly according to a predicted load rate interval, so that the accuracy of the distribution transformer heavy overload prediction is further reduced due to high conservation over-strong error.

Disclosure of Invention

The invention provides a distribution transformer overload early warning method, a distribution transformer overload early warning device and distribution transformer overload early warning equipment, and solves the technical problem that the existing distribution transformer overload early warning method is low in accuracy.

The invention provides a distribution transformer overload early warning method for a distribution network, which comprises the following steps:

when a request of distribution network distribution transformer overload early warning is received, acquiring actual operation time sequence data corresponding to the request, performing first preprocessing by adopting the actual operation time sequence data, and constructing a target association diagram;

inputting the target association graph into a preset target load rate prediction model corresponding to the request, and outputting a predicted load rate; the preset target load rate prediction model comprises a graph sampling aggregation layer, a long-short-period memory layer, a first full-connection layer and a second full-connection layer;

and calculating a heavy overload risk value according to the predicted load rate, and determining an early warning result corresponding to the request.

Optionally, the performing the first preprocessing with the actual operation time sequence data and constructing a target association graph includes:

performing first preprocessing by adopting the actual operation time sequence data to generate first time sequence data; wherein the first time sequence data comprises load rate data, characteristic variable data and influence factor data;

The characteristic variable data is taken as a node, the load rate data and the influence factor data are taken as node characteristics, a node connection relation is determined according to the distance between the nodes, and an initial association diagram is constructed;

calculating residual vectors among the nodes according to the node characteristics;

and updating the node connection relation in the initial association graph by adopting the residual vector and combining with preset node connection conditions to generate a target association graph.

Optionally, inputting the target association graph to a preset target load rate prediction model, and outputting a predicted load rate, including:

inputting the target association graph to the graph sampling aggregation layer for sampling, aggregating, transferring updating and sequencing, and outputting an allocation data sequence;

encoding elements in the data sequence to generate an encoded data sequence;

inputting the encoded data sequence into the long-period and short-period memory layer for decoding, and outputting a first intermediate predicted value;

inputting the first intermediate predicted quantity to the first full-connection layer and outputting a second intermediate predicted quantity;

inputting the second intermediate pre-measurement to the second full connection layer and outputting a third intermediate pre-measurement;

And performing inverse first preprocessing on the third intermediate pre-measurement to generate a predicted load rate.

Optionally, the calculating the heavy overload risk value according to the predicted load factor, and determining the early warning result corresponding to the request, includes:

determining a distribution state corresponding to the predicted load rate by adopting the predicted load rate and a preset load rate interval;

calculating a heavy overload risk value corresponding to the request according to the configuration state, the predicted load rate and the historical operation time sequence data corresponding to the request;

and determining the early warning grade and the early warning color corresponding to the request by adopting the heavy overload risk value and combining a preset risk value early warning interval.

Optionally, the calculation formula of the heavy overload risk value is:

；

wherein, for the heavy overload risk value, +.>For the predicted load factor,/->Time ratio of the preset load rate interval where the predicted load rate is located in the historical operation time sequence data is +.>For the state of the distribution transformer corresponding to the predicted load factor,>is->The duty cycle in the historical operating schedule data,for the load rate of->Cause->Is a severity of (c).

Optionally, the severity is calculated by the formula:

；

Wherein, for a preset heavy overload rate threshold, < ->Is the preset firstA parameter of->Is natural constant (18)>Is a preset second parameter.

Optionally, the step of obtaining the preset target load rate prediction model includes:

acquiring historical operation time sequence data of all the areas of the power distribution network, and performing second preprocessing to generate training time sequence data;

respectively constructing training association diagrams corresponding to different platform types by adopting the training time sequence data and combining the platform types corresponding to each platform;

and respectively training preset initial load rate prediction models by adopting the training association diagrams corresponding to different zone types, and generating preset target load rate prediction models corresponding to different zone types.

Optionally, the method for determining the zone type includes:

respectively constructing load photovoltaic power curves corresponding to the areas according to the historical operation time sequence data;

calculating the distance between all the load photovoltaic power curves, and constructing a distance distribution matrix;

the distance distribution matrix is adopted, the local density of all the load photovoltaic power curves is calculated respectively, and the load photovoltaic power curves with the number of preset clusters and the maximum local density are selected as initial cluster centers;

Respectively calculating the distance between each load photovoltaic power curve and each initial clustering center, classifying the initial clustering centers with the smallest distance from the load photovoltaic power curve into one type, and generating a cluster of a preset clustering number;

calculating a middle cluster center of each cluster of the transformer areas, and judging whether the middle cluster center meets a preset convergence condition or not;

if yes, confirming the type of the area corresponding to each area according to the area cluster;

if the distance between each load photovoltaic power curve and each initial clustering center is calculated, the initial clustering centers with the smallest distance are classified, and the cluster area clusters with the preset clustering number are generated.

The second aspect of the invention provides a distribution network distribution transformer overload early warning device, which comprises:

the data processing module is used for acquiring actual operation time sequence data corresponding to a request when receiving the request of distribution transformer overload early warning of the power distribution network, and carrying out first preprocessing and constructing a target association diagram by adopting the actual operation time sequence data;

the model prediction module is used for inputting the target association graph into a preset target load rate prediction model corresponding to the request and outputting a predicted load rate; the preset target load rate prediction model comprises a graph sampling aggregation layer, a long-short-period memory layer, a first full-connection layer and a second full-connection layer;

And the risk early warning module is used for calculating a heavy overload risk value according to the predicted load rate and determining an early warning result corresponding to the request.

A third aspect of the present invention provides an electronic device, comprising: a memory and a processor;

the memory is used for storing program codes and transmitting the program codes to the processor;

the processor is configured to execute the distribution transformer overload pre-warning method according to any one of the first aspects of the present invention according to the instructions in the program code.

From the above technical scheme, the invention has the following advantages:

according to the method, the predicted load rate is generated through the preset target load rate prediction model, the heavy overload risk value is calculated according to the predicted load rate, and then the pre-warning result of the heavy overload of the distribution transformer is determined. The preset target load rate prediction model comprises a graph sampling aggregation layer, a long-short-period memory layer, a first full-connection layer and a second full-connection layer, wherein the graph sampling aggregation layer integrates operation time sequence data into high-dimensional time sequence data, and combines the learning advantage of the long-short-period memory layer on the time sequence data, so that the accuracy of distribution change heavy overload early warning under the operation condition of high-dimensional characteristic variables and small sample data is improved; and the heavy overload risk value is calculated according to the predicted load rate, and then the early warning result is determined, so that the conservation of dividing the early warning level directly according to the predicted load rate interval is avoided, and the accuracy of the heavy overload early warning of the distribution transformer is further improved.

Drawings

In order to more clearly illustrate the embodiments of the invention or the technical solutions of the prior art, the drawings which are used in the description of the embodiments or the prior art will be briefly described, it being obvious that the drawings in the description below are only some embodiments of the invention, and that other drawings can be obtained from these drawings without inventive faculty for a person skilled in the art.

Fig. 1 is a flowchart of steps of a distribution transformer overload early warning method for a distribution network according to a first embodiment of the present invention;

fig. 2 is a flowchart of steps of a distribution transformer overload early warning method for a distribution network according to a second embodiment of the present invention;

FIG. 3 is a diagram illustrating a long-short-term memory layer according to a second embodiment of the present invention;

FIG. 4 is a schematic diagram of average absolute errors of prediction models of different network structures according to a second embodiment of the present invention;

FIG. 5 is a schematic diagram of root mean square error of prediction models of different network structures according to a second embodiment of the present invention;

FIG. 6 is a graph of clustering results of various load photovoltaic power curves provided in a second embodiment of the present invention;

fig. 7 is a schematic diagram of precision and recall of a distribution transformer overload early warning method and a conventional method for a distribution transformer in a second embodiment of the present invention;

FIG. 8 is a diagram showing a change in a loss function of a training set and a verification set according to a third embodiment of the present invention;

FIG. 9 is a graph comparing the predicted load rate and the actual load rate of the test set according to the third embodiment of the present invention;

fig. 10 is a structural block diagram of a distribution transformer overload early warning device provided in a fourth embodiment of the present invention.

Detailed Description

The embodiment of the invention provides a distribution transformer overload early warning method, a distribution transformer overload early warning device and distribution transformer overload early warning equipment, which are used for solving the technical problem that the existing distribution transformer overload early warning method is low in accuracy.

In order to make the objects, features and advantages of the present invention more comprehensible, the technical solutions in the embodiments of the present invention are described in detail below with reference to the accompanying drawings, and it is apparent that the embodiments described below are only some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Referring to fig. 1, fig. 1 is a flowchart illustrating steps of a distribution transformer overload early warning method according to a first embodiment of the present invention.

The first embodiment of the invention provides a distribution transformer overload early warning method for a distribution network, which comprises the following steps:

and step 101, when a request of distribution network distribution transformer heavy overload early warning is received, acquiring actual operation time sequence data corresponding to the request, performing first preprocessing by adopting the actual operation time sequence data, and constructing a target association diagram.

It can be understood that the actual operation time sequence data is the actual operation parameters of the distribution transformer of the distribution network including the load rate, and can be obtained from a production and management system related to distribution network marketing, distribution and scheduling or obtained by calculation through the available data; the first preprocessing may be any data preprocessing method of data normalization; the target association diagram is constructed according to the actual operation time sequence data after the first preprocessing and can be used as the input of a preset target load rate prediction model in the next step.

102, inputting a target association diagram into a preset target load rate prediction model corresponding to a request, and outputting a predicted load rate; the preset target load rate prediction model comprises a graph sampling aggregation layer, a long-short-period memory layer, a first full-connection layer and a second full-connection layer.

It should be noted that, the Graph SAmple aggregation layer, that is, graph SAmple aggregation network (Graph SAmple and aggreGatE, graphSAGE), is a Graph neural network (Graph Neural Network, GNN), in this embodiment, the Graph SAmple aggregation layer can learn the hidden information of Graph data by using the topology information and the feature embedding information of the vertex of the target association Graph, so as to SAmple the local area of each node in the target association Graph and aggregate the features to generate an embedding vector, and can integrate the similar operation scene data around the early warning platform region distribution transformer and mine the hidden relationship between the actual operation time sequence data, thereby playing the effects of feature enhancement and data expansion. The Long-Term Memory layer, namely a Long-Term Memory network (LSTM), can capture Long-distance dependency in actual operation time sequence data, and decodes and outputs a load rate prediction result.

And 103, calculating a heavy overload risk value according to the predicted load rate, and determining an early warning result corresponding to the request.

It will be appreciated that the heavy overload risk value refers to a comprehensive evaluation index of the severity and likelihood of heavy overload, and can be expressed by the product of the probability and severity of occurrence of a heavy overload event. Under the condition of limited available resources, the distribution transformer can be overloaded for a short time under certain conditions without influencing the service life of the distribution transformer. Therefore, different from the traditional method that the early warning grades are directly divided according to the predicted load rate interval, the embodiment calculates the heavy overload risk value by the load rate and then determines the early warning result, so that the conservation of heavy overload evaluation in the traditional method can be reduced, and meanwhile, the sensitivity of the distribution transformer reverse heavy overload early warning is improved, and the accuracy of the distribution transformer heavy overload early warning is improved.

According to the first embodiment of the invention, the predicted load rate is generated through the preset target load rate prediction model, the heavy overload risk value is calculated according to the predicted load rate, and then the pre-warning result of the heavy overload of the distribution transformer is determined. The preset target load rate prediction model comprises a graph sampling aggregation layer, a long-short-period memory layer, a first full-connection layer and a second full-connection layer, wherein the graph sampling aggregation layer integrates operation time sequence data into high-dimensional time sequence data, and combines the learning advantage of the long-short-period memory layer on the time sequence data, so that the accuracy of distribution change heavy overload early warning under the operation condition of high-dimensional characteristic variables and small sample data is improved; and the heavy overload risk value is calculated according to the predicted load rate, and then the early warning result is determined, so that the conservation of dividing the early warning level directly according to the predicted load rate interval is avoided, and the accuracy of the heavy overload early warning of the distribution transformer is further improved.

Referring to fig. 2, fig. 2 is a flowchart illustrating steps of a distribution transformer overload early warning method according to a second embodiment of the present invention.

The second embodiment of the invention provides a distribution transformer overload early warning method for a distribution network, which comprises the following steps:

step 201, when a request of distribution network distribution transformer heavy overload early warning is received, acquiring actual operation time sequence data corresponding to the request, performing first preprocessing by adopting the actual operation time sequence data, and constructing a target association diagram.

It should be noted that, the request for pre-warning of distribution transformer overload of the power distribution network includes information of a target area to be pre-warned, and the actual operation time sequence data corresponding to the acquisition request is specifically the actual operation time sequence data of the target area and other areas of the same area type as the target area.

The actual operation time sequence data comprises non-preprocessed load rate data, influence factor data and characteristic variable data corresponding to a plurality of characteristic variables, wherein the characteristic variables can comprise distribution transformer nameplate capacity, distribution transformer type, week, holiday, solar radiation intensity, temperature, humidity, wind speed, wind direction, single-phase active power of a measuring point and other distribution transformer operation parameters, the characteristic variable data corresponding to each characteristic variable comprises characteristic components of a plurality of sampling moments of a plurality of sampling days, one sampling day comprises a plurality of sampling moments, and the time interval of the sampling moments is a preset sampling time threshold value, and is preferably 15min in the embodiment; the influence factor data can comprise date factors, weather factors, standing account data factors and the like, the representation mode of the influence factor data can be referred to in a table 1, and the table 1 is a distribution transformer load factor influence factor type table; the characteristic variable data and the influence factor data can be directly obtained from production and management systems related to distribution network marketing, distribution and scheduling, including a scheduling management system (DMS), a Production Management System (PMS), an electricity consumption information acquisition system, a metering system, a distribution network scheduling intelligent management and control platform (SMD), an meteorological system and the like; the load rate data is calculated after the distribution transformer account data and the metering data corresponding to the same sampling time point are acquired:

；

In the method, in the process of the invention,for the load factor of the distribution transformer at the moment, +.>To change in->Power factor angle of moment +.>Single-phase active power for measuring point of transformer in transformer area,/->And reporting capacity for the transformer in the transformer area.

In this embodiment, the degree of correlation between all preselected feature variables and the load factor may be calculated by pearson correlation coefficient analysis (Pearson correlation coefficient, PCC), and feature variable data corresponding to feature variables whose pearson correlation coefficients are within a preset R value interval may be selected to be obtained. The value of the pearson correlation coefficient is between-1 and 1, and a positive value or a negative value represents that two variables belong to positive correlation or negative correlation, and the larger the absolute value is, the stronger the correlation is, and the preset R value interval can be set according to the actual application scenario, and in this embodiment, it is preferable to (- ++0.5)/(0.5, ++). The calculation formula of the pearson correlation coefficient analysis method is as follows:

；

in the method, in the process of the invention,is the pearson correlation coefficient between the characteristic variable and the load factor, ++>For the load rate sequence, +.>For characteristic variable sequences, ++>For the dimension of the sequence of characteristic variables +.>For the average value of all elements in the load-factor sequence, +.>Is the average of all elements in the sequence of feature variables.

In a preferred embodiment, step 201 comprises the following sub-steps S11 to S14:

S11, performing first preprocessing by adopting actual operation time sequence data to generate first time sequence data.

The first time sequence data includes load rate data, characteristic variable data and influence factor data after the first preprocessing. The first preprocessing may be any data preprocessing method with normalized data, preferably normalization processing, where the formula of normalization processing is:

；

in the method, in the process of the invention,representing data to be processed before normalization processing, +.>Represents the data after the normalization process,for the minimum value in the characteristic variable data corresponding to the data to be processed, < >>And the maximum value in the characteristic variable data corresponding to the data to be processed.

And S12, taking the characteristic variable data as nodes, taking the load rate data and the influence factor data as node characteristics, determining a node connection relation according to the distance between the nodes, and constructing an initial association graph.

It should be noted that, the feature variable data, the load factor data and the influencing factor data in step S12 all belong to the first time sequence data after the first preprocessing.

The initial association diagram may be expressed asWherein->Representing the construction of data from characteristic variablesmThe number of nodes in the network is,mnumber of actual stations =number of sampling days ×number of sampling times in one day, the number of actual stations is the number of target stations and other stations of the same type as the target station, +. >Represent the firstiPersonal node->Represent the firstiThe first node included in the individual nodenThe feature components;representing initial associationsConnection relation of nodes in the graph +.>Representation ofmNode characteristics of individual nodes->Comprising nodesiLoad rate data of (a)t-1 influencing factor data.

The connection relation of the nodes in the initial association graph is determined by the distance between the nodes:if nodejBelonging to and nodeiFront with minimum Euclidean distance between themkThe individual nodes are assigned->Otherwise add->，kCan be set according to the actual application scene.

S13, calculating residual vectors among the nodes according to the node characteristics.

NodeiSum nodejThe residual vector of the node features of (a) can be expressed asWherein->Is a nodeiNode feature vector, ">Is a nodejIs described.

S14, updating the node connection relation in the initial association graph by adopting the residual vector and combining with preset node connection conditions to generate a target association graph.

The residual threshold vector can be preset according to experience knowledgeAnd node connection conditions. If nodeiSum nodejIs>And residual threshold vector->The relation of (a) satisfies the node connection condition, and the node can be determinediSum nodejIf there is an association relationship, add->If node iSum nodejIs>And residual threshold vector->If the relation of (2) does not satisfy the preset node connection condition, the method is to assign +.>Update the initial association diagram with this +.>Node connection relation->：

；

In a more preferred embodiment, the acquired influence factor data types comprise solar radiation, weather type, holiday, distribution transformer type, humidity, temperature, wind speed and annual energy growth rate of the platform, then the nodeiNode characteristics of (2)Comprising nodesiLoad factor of->Solar radiation->Weather type->Holiday->Type of distribution transformation->Moisture->Temperature->Wind speed->Power consumption increase rate comparable to that of the district>NodeiSum nodejIs>Thereby setting the residual threshold vector +.>Corresponding node connection condition->The method comprises the following steps:

；

it can be understood that steps S11 to S14 analyze the correlation between the time series data of each distribution transformer by using euclidean distance and experience knowledge, and construct a correlation diagram of the distribution transformer of each distribution transformer, so as to ensure the fusion between samples and the learning diversity.

Step 202, inputting a target association diagram into a preset target load rate prediction model corresponding to a request, and outputting a predicted load rate; the preset target load rate prediction model comprises a graph sampling aggregation layer, a long-short-period memory layer, a first full-connection layer and a second full-connection layer.

In a preferred embodiment, step 202 specifically comprises the following substeps S21 to S26:

s21, inputting the target association graph into a graph sampling aggregation layer for sampling, aggregating, transferring updating and sequencing, and outputting an allocation data sequence.

After the target association graph is input into a preset target load rate prediction model, the graph sampling aggregation layer firstly samples first-order neighbor nodes of nodes in the target association graph according to a preset rule, and respectively generates neighbor node sets of the first-order neighbor nodes including preset sampling numbers corresponding to each node. The preset rule can be set to random sampling, and the preset sampling number can be set according to the actual application scene. If the number of the first-order neighbor nodes of the node is smaller than the preset sampling number, adopting a random sampling rule with the substitution, otherwise adopting a random sampling rule without the substitution.

Then adopt the neighbor node setAggregating all nodes in the target association graph to obtain the aggregated embedded features:

；

；

in the method, in the process of the invention,representing nodesiIs at the first order neighbor node of (1)kEmbedding vector of layer, ">Representing data nodesjIs at (1)k-embedding vector of layer-1,>representing data nodesiIs at (1)kThe embedding vector of a layer, i.e. the target layer embedding vector, ，Sampling the number of aggregate layers for a graph, +.>Representing the connection operation, the last embedded vector of the connection node and the embedded vector of its neighbor set, +.>Representing nodesiNeighbor node set,/->Weight matrix representing a linear transformation, +.>Representing a function that characterizes the aggregated neighbor nodes.

The data sequence of the configuration data generated after the sampling, aggregation, transfer update and sequencing of the image sampling aggregation layer。

In a more preferred embodiment of the present invention,an aggregation function that takes into account the mechanism of attention may be employed, comprising in particular the steps of: applying the shared attention structure and the shared weights on the nodes, calculating correlation coefficients between the data nodes:

；

in the method, in the process of the invention,is a nodeiSum nodejCorrelation coefficient between->For sharing attention structure, add>To share weights +.>Is a nodeiNode feature vector, ">Is a nodejIs defined by the node feature vector of (a);

all correlation coefficients were normalized and linearly varied using a normalized exponential function (softmax function), resulting in an attention coefficient matrix:

；

in the method, in the process of the invention,is a nodeiSum nodejAttention coefficient matrix between +_>To activate the function +.>For the connection operation +.>Representation ofeA kind of electronic devicexA power of the second;

using matrices of attention coefficients and nodes iIs collected at a neighbor node of (1)k-layer 1 embedding vector calculation to get aggregated embedding features:

；

s22, coding elements in the data sequence to generate a coded data sequence.

The encoding process specifically includes extracting a complex data sequenceFront of (2)mAnd carrying out matrixing segmentation on the item data, namely the operation time sequence data corresponding to the target station area, according to the time step, obtaining a plurality of one-dimensional vectors with the length of the time step, and carrying out coding in sequence. Wherein, mnumber of sampling times in a day.

S23, inputting the encoded data sequence into the long-short-period memory layer for decoding, and outputting a first intermediate predicted value.

Referring to fig. 3, fig. 3 is a block diagram of a long-short-period memory layer according to a second embodiment of the invention. The dimension of the output characteristic of the long-short-period memory layer is 64, and the calculation formula is as follows:

；

；

；

；

；

；

in the method, in the process of the invention,、and->For weight parameter, ++>、And->For bias parameter +.>Outputting a hidden state vector for the present time step, +.>For the last time step memory hidden state vector, < >>For the memory hidden state vector, ">For the candidate memory hidden state vector, +.>For the entrance door->Is a forgetful door, is a->For the output door->Is thattInput vector of time long-short-term memory layer, +. >Is the last time step memory cell.

S24, inputting the first intermediate predicted quantity into the first full-connection layer, and outputting the second intermediate predicted quantity.

The output feature dimension of the first fully connected layer is 32.

S25, inputting the second intermediate predicted quantity into the second full-connection layer, and outputting a third intermediate predicted quantity.

The output feature dimension of the second fully connected layer is 1.

The activation functions of the first fully-connected layer and the second fully-connected layer both adopt linear rectification functions (Linear rectification function, reLU functions):

；

in the method, in the process of the invention,output for full connection layer, +.>Output weight for full connection layer, +.>Output bias for full connection layer,)>Is a linear rectification function.

Because the first preprocessing for normalizing the data is performed on the actual operation time sequence data before the target association graph is constructed, after the prediction result is output by the preset target load rate prediction model, the output prediction data is restored to the original data scale by performing inverse first preprocessing. In a more preferred embodiment the first pre-processing is set to normalization processing, the inverse first pre-processing may be an inverse normalization processing.

In a preferred embodiment, the step of obtaining the preset target load factor prediction model includes the following steps S31 to S33:

S31, acquiring historical operation time sequence data of all the areas of the power distribution network, and performing second preprocessing to generate training time sequence data.

It will be appreciated that the second preprocessing may include data preprocessing operations for data cleansing and data normalization.

S32, training time sequence data are adopted, and the types of the areas corresponding to the areas are combined to respectively construct training association diagrams corresponding to different types of the areas.

The training time sequence data is classified according to the types of the areas, the classified data is used for training a preset initial load rate prediction model, and a preset target load rate prediction model aiming at different types of the areas can be generated. The process of constructing the training association diagram may refer to the corresponding process of constructing the target association diagram, which is not described herein.

S33, respectively training preset initial load rate prediction models by adopting training association diagrams corresponding to different zone types, and generating preset target load rate prediction models corresponding to different zone types.

It may be understood that the initial load rate prediction model is a preset target load rate prediction model before training, and the algorithm steps after the training association chart is input to the initial load rate prediction model may refer to the algorithm steps of the preset target load rate prediction model, which are not described herein.

The training association diagram is input into the initial load rate prediction model, then the training prediction load rate is output, the weight parameters and bias parameters of each neuron in the initial load rate prediction model can be iteratively updated by combining the momentum of the self-adaptive motion estimation algorithm (Adam algorithm) and the self-adaptive learning rate, so that the output value of the loss function is optimal, and the optimal parameters of the model are saved. The mathematical expression of the loss function is:；

wherein:for the loss function of the predictive model, +.>Representing the number of training samples>Represents the actual load factor at time t, +.>Respectively representing training prediction load rates at the time t;

the method can also adopt training prediction load rate, calculate the error of the current prediction model by combining accuracy (Precision), recall (Recall), root mean square error (root mean square error, RMSE) and average absolute error (mean absolute Percentage error, MAPE), so as to adjust the network layer number, iteration times, learning rate and other super parameters of the initial load rate prediction model:

；

；

；

；

in the method, in the process of the invention,for the accuracy of->For recall->Is root mean square error>Mean absolute error, +.>Time points which indicate that the distribution transformer states corresponding to the actual load rate and the training predicted load rate are reverse heavy overload states, and +. >Sample number representing that the distribution state corresponding to the actual load rate is normal and the distribution state corresponding to the training predicted load rate is reverse heavy overload state, +.>The number of samples for training and predicting that the distribution state corresponding to the load rate is normal when the distribution state corresponding to the actual load rate is reverse heavy overload state is represented by +.>Is thattActual load factor of time of day +.>Is thattTime training predicts load rate, +.>The number of data for the example test.

From the practical engineering perspective, the loss caused by judging the reverse heavy overload situation as normal is larger than the loss caused by misjudging the normal situation as the reverse heavy overload accident, so that the adoption of the recall ratio index to adjust the super-parameters of the prediction model is more scientific. The accuracy rate, recall rate, root mean square error and average absolute error can also be used for evaluating the prediction performance of a preset target load rate prediction model, and for the distribution transformer weight overload early warning of a distribution network containing photovoltaic access, the higher the recall rate score is, the lower the root mean square error and average absolute error is, the better the model prediction effect is. In a more preferred embodiment, referring to fig. 4 and 5, fig. 4 is a schematic average absolute error diagram of prediction models of different network structures, fig. 5 is a schematic root mean square error diagram of prediction models of different network structures, where model 1 is a prediction model including a Back Propagation (BP) neural network, model 2 is a prediction model including a long and short term memory network, model 3 is a prediction model including a convolutional neural network (Convolutional Neural Networks, CNN) and a long and short term memory network, and model 4 is a preset target load rate prediction model including a graph sampling aggregation layer, a long and short term memory layer, a first fully connected layer and a second fully connected layer, provided by the present invention, and the root mean square error and average absolute error of the prediction model 4 are smaller when the number of samples is small.

In a more preferred embodiment, the method of determining the zone type includes the following steps S41 to S47:

s41, respectively constructing load photovoltaic power curves corresponding to each area according to the historical operation time sequence data.

The load photovoltaic power curves (P-L curves) of all the areas of the distribution network can be expressed as a setWherein->For the number of all the areas of the power distribution network, the firstiLoad photovoltaic power curve of each platform area，For the number of sampling instants in the day, +.>Is at the ith station areaNPhotovoltaic output value of each sampling point in time, < >>Is at the ith station areaNLoad values for each sampling instant.

S42, calculating the distance between all the load photovoltaic power curves, and constructing a distance distribution matrix.

The distance distribution matrix may be expressed as:

；

in the method, in the process of the invention,is distance distribution matrix>And->For the collection->A load light Fu Gonglv curve in the middle,，is->And->Is a euclidean distance of (c).

S43, calculating the local density of all the load photovoltaic power curves by adopting a distance distribution matrix, and selecting the load photovoltaic power curves with the number of preset clusters and the maximum local density as initial cluster centers.

In a more preferred embodiment, the preset number of clusters can be determined by Elbow method (Elbow method) and CH index (Calinski-Harabasz index) calculation. First, an initial cluster number is set K=2, 3, ⋯,8, each initial cluster number was calculated separately using the elbow ruleKError square sum of cluster center and cluster sample (Sum of Squared Error, SSE) is used to make SSE-K relation graph. When (when)KAfter the threshold value of the cluster number is reached,Kthe increase of (1) can lead the return of the aggregation degree to be rapidly reduced, namely, the optimal candidate cluster number corresponds to the point with the largest slope change of the segmentation broken line in the SSE-K, and the critical point can be considered as the point with better clustering performance; the essence of the CH index is the ratio of the inter-cluster distance to the intra-cluster distance, divided intoThe compactness of each curve in the cluster and the similar center curve and the separation of each center curve and the data center curve are respectively reflected. The larger the CH index is, the better the clustering effect is, and the initial clustering number with the largest CH index can be determined as the preset clustering number. The calculation formula of the CH index is as follows:

；

in the method, in the process of the invention,CHis a CH index of the number of the n-type,is the covariance matrix of the data between classes,is covariance matrix of data in class, +.>Trace representing matrix, +.>Is of the classiCenter curve of>For the collection->Center curve of>Representation classiNumber of curves of>Representation classiIs a set of curves for the model (a).

S44, respectively calculating the distance between each load photovoltaic power curve and each initial clustering center, classifying the initial clustering centers with the smallest distance from the load photovoltaic power curve into one type, and generating the cluster of the areas with the preset clustering number.

S45, calculating a middle cluster center of each cluster of the areas, and judging whether the middle cluster center meets a preset convergence condition.

It is understood that the preset convergence condition may be set such that the cluster center curves of the various types are not changed.

And S46, if the information is satisfied, confirming the type of the area corresponding to each area according to the area cluster.

The areas corresponding to the load photovoltaic power curves belonging to the same area cluster can be determined to be the same area type, so that associated areas with the same load and photovoltaic output time sequence characteristics can be mined, historical operation time sequence data can be classified according to the area types, and the classified data are respectively used for training a preset initial load rate prediction model to generate preset target load rate prediction models aiming at different area types.

And S47, if the distance between each load photovoltaic power curve and each initial clustering center is not met, performing skip execution, respectively calculating the distance between each load photovoltaic power curve and each initial clustering center, classifying the initial clustering centers with the smallest load photovoltaic power curve and the smallest distance into one type, and generating a cluster of the preset clustering number of the clusters.

Steps S41 to S47 utilize a density optimization-based modified k-means clustering algorithm (k-means clustering algorithm) to perform cluster analysis on the load photovoltaic power curves of all the zones. Because the load photovoltaic power curves corresponding to different zone types have different trend directions, the zone clusters corresponding to the different zone types can be obtained through the cluster analysis of the steps S41 to S47. Referring to fig. 6, fig. 6 is a graph of clustering results of various load photovoltaic power curves according to a more preferred embodiment, wherein the sampling time number of the load photovoltaic power curve is 24, and the load photovoltaic power curve of the industrial user is bimodal and is represented by day 10:00-18: 00. evening 19:00-22:00, and because the industrial factory building has a larger solar photovoltaic panel installation area, the photovoltaic output part of the load photovoltaic power curve has steeper peak values; the load photovoltaic power curve of the resident user is three peaks, and except for the single peak of the output of the distributed roof photovoltaic, the load photovoltaic power curve is expressed as peak electricity consumption in the rest time of the house in the middle of noon and evening; the load photovoltaic power curve of the commercial user is four-peak, and the peak value of the photovoltaic output is the lowest because the roof photovoltaic resource of the commercial area is relatively less, and the commercial area generally comprises the catering industry, so that the load part of the curve is shown to have peak phenomena in the early, middle and late periods, and the peak values are slightly different.

And 203, determining a configuration state corresponding to the predicted load rate by combining the predicted load rate with a preset load rate interval.

The distribution state corresponding to the load rate interval and the load rate interval can be set according to experience knowledge. In a preferred embodiment, the load factor interval and the configuration state are set to the correspondence relationship as shown in table 2.

And 204, calculating a heavy overload risk value corresponding to the request according to the configuration transformer state, the predicted load rate and the historical operation time sequence data corresponding to the request.

It should be noted that, the historical operation time sequence data may be historical operation time sequence data of the target station area needing to be pre-warned in a preset time period, where the preset time period may be set according to actual requirements, for example, set as historical operation time sequence data of the last month.

The calculation formula of the heavy overload risk value is as follows:

；

in the method, in the process of the invention,for overload risk value->For predicting load factor, +.>Time ratio of preset load rate interval in which predicted load rate is located in historical operation time sequence data>To predict the corresponding distribution state of the load rate, including normal,E ₃ 、E ₂ 、E ₁ AndE ₀ ，E ₀ indicating a heavy load such as a heavy load,E ₁ indicating an overload of the device and the system,E ₂ indicating a reverse heavy load such as a heavy load,E ₃ indicating reverse overload->Is->Duty ratio in historical operation time sequence data, +. >For the load rate of->The time causesIs a severity of (c). The preset load factor interval may be set according to the actual distribution of the load factor data, and the interval is preferably 0.1, such as [0.5,0.6 ], [0.6, 0.7), etc.

It can be understood that in the distribution transformer overload early warning of the photovoltaic access platform area, the reverse overload and overload have more serious consequences, which can greatly reduce the distribution transformer operation efficiency and service life and bring huge losses to power grid companies and users. Further, as the reverse heavy overload is represented by that the load rate is greatly deviated from the normal value, the severity is comprehensively considered, and therefore, the severity of the heavy overload of the distribution transformer is quantified by adopting the risk-type utility function, and the calculation formula of the severity is as follows:

；

in the method, in the process of the invention,for a preset heavy overload rate threshold, < ->For a first predetermined parameter, +.>Is natural constant (18)>Is a preset second parameter.

Heavy overload load rate thresholdMay be set based on empirical knowledge. In a more preferred embodiment, the configuration state is set to be normal when the predicted load factor is within the [ -0.8,0.8) interval, and +.>. First parameter->And second parameter->All are positive parameters, and can be set according to factors such as the type of the power load of the station, the capacity increase and decrease of the business expansion, the allowable load coefficient for short-term overload, the allowable continuous operation time and the like.

Under the condition of limited available resources, the distribution transformer can be overloaded for a short time under certain conditions without influencing the service life of the distribution transformer. The risk-type utility function enables the severity of the heavy overload of the distribution transformer to be exponentially increased, and when the value of the predicted load rate is higher, the higher risk value can be shown no matter the proportion of the interval where the predicted load rate is located in the historical operation time sequence data and the proportion of the distribution transformer state in the historical operation time sequence data; when the predicted load rate is lower in value, the corresponding severity is lower, and the duty ratio of the interval where the predicted load rate is located in the historical operation time sequence data and the duty ratio of the distribution transformer state in the historical operation time sequence data are comprehensively considered, so that the distribution transformer is allowed to be subjected to heavy overload to a certain degree or for a certain period of time, the actual working and operation conditions of the distribution transformer are more met, and the accuracy of distribution transformer heavy overload early warning is improved. Referring to fig. 7, fig. 7 is a schematic diagram of an early warning method for distribution transformer overload in a power distribution network according to a second preferred embodiment of the present invention and a conventional method for dividing early warning levels directly according to a predicted load rate interval, so that the early warning method for distribution transformer overload in the power distribution network provided by the present invention has a higher recall rate and a higher early warning accuracy than the conventional method.

And 205, determining an early warning level and an early warning color corresponding to the request by combining the heavy overload risk value with a preset risk value early warning interval.

The early warning grade and the early warning color corresponding to the risk value early warning interval can be set according to experience knowledge. In a preferred embodiment, the risk value early warning interval, the early warning level and the early warning color are set to be corresponding to those shown in table 3.

According to the embodiment of the invention, the predicted load rate is generated through the preset target load rate prediction model, the heavy overload risk value is calculated according to the predicted load rate, and then the pre-warning result of the heavy overload of the distribution transformer is determined. The preset target load rate prediction model comprises a graph sampling aggregation layer, a long-short-period memory layer, a first full-connection layer and a second full-connection layer, wherein the graph sampling aggregation layer integrates operation time sequence data into high-dimensional time sequence data, and combines the learning advantage of the long-short-period memory layer on the time sequence data, so that the accuracy of distribution change heavy overload early warning under the operation condition of high-dimensional characteristic variables and small sample data is improved; according to the predicted load rate and the risk type utility function, the heavy overload risk value is calculated, the duty ratio of the interval where the predicted load rate is located in the historical operation time sequence data and the duty ratio of the distribution transformer state in the historical operation time sequence data are comprehensively considered, the actual working and operation conditions of the distribution transformer are more met, and the accuracy of distribution transformer heavy overload early warning is further improved.

In order to more specifically describe the embodiments of the present application in detail, a third embodiment of a distribution network distribution transformer overload early warning method is described below by taking a typical medium voltage distribution feeder line in Guangdong province as an example:

taking a typical medium-voltage distribution feeder line in Guangdong province as an example, the upper-level main transformer capacity of the area is 5MVA; obtaining 18 distribution transformers of a residential area, 4 distribution transformers of a commercial area and 15 distribution transformers of an industrial area through cluster analysis; four commercial areas with rated capacity of distribution transformer of 160kVA, 250kVA, 100kVA and 80kVA are selected, one of the commercial areas is set as a target area, and the maximum conveying capacity of each branch is not considered; selecting distribution and transformation metering data of four areas 2019, 11 months, 1 month, 11 months and 30 months, wherein each 15 minutes in the metering data is a sampling point, 96 data points are contained in one day, and corresponding load rates are calculated according to the metering data to generate historical operation time sequence dataWherein 4 is the number of areas, 30 is the number of sampling days, 96 is the number of sampling moments in a day, and 11 is the number of characteristic variables; the historical operation time sequence data is divided into a training set, a verification set and a test set, and the proportion is 26:3:1, respectively constructing corresponding training association diagrams; the method comprises the steps of calculating a loss function through a training association diagram corresponding to a training set, then carrying out back propagation, continuously adjusting weight parameters and bias parameters of a prediction model, calculating error indexes and the loss function through a training association diagram corresponding to a verification set to adjust super parameters of the prediction model, and finally calculating the prediction performance of the error index evaluation model through a training association diagram corresponding to a test set; in the online application stage, the actual operation time sequence data is used for predicting the distribution transformer load rate at the next moment of the early warning area, the distribution transformer heavy overload degree is further judged through a preset heavy overload risk value interval, and the early warning color and the early warning grade are issued by combining with the visualization technologies such as SVG rendering. Referring to fig. 8, fig. 9 and table 4, fig. 8 is a schematic diagram of a change of a loss function of a training set and a verification set provided in the third embodiment of the present application, fig. 9 is a comparative diagram of a predicted load rate and an actual load rate of a test set provided in the third embodiment of the present application, and table 4 is a table of early warning results of the test set provided in the third embodiment of the present application.

Referring to fig. 10, fig. 10 is a block diagram illustrating a distribution transformer overload early warning device for a distribution network according to a fourth embodiment of the present invention.

The fourth embodiment of the invention provides a distribution network distribution transformer overload early warning device, which comprises:

the data processing module 1001 is configured to, when receiving a request for pre-warning of a distribution transformer overload of the power distribution network, obtain actual operation time sequence data corresponding to the request, perform first preprocessing by using the actual operation time sequence data, and construct a target association graph;

the model prediction module 1002 is configured to input the target association graph to a preset target load rate prediction model corresponding to the request, and output a predicted load rate; the preset target load rate prediction model comprises a graph sampling aggregation layer, a long-short-period memory layer, a first full-connection layer and a second full-connection layer;

and the risk early warning module 1003 is configured to calculate a heavy overload risk value according to the predicted load factor, and determine an early warning result corresponding to the request.

The embodiment of the invention also provides electronic equipment, which comprises: a memory and a processor; a memory for storing program code and transmitting the program code to the processor; and the processor is used for executing the distribution network distribution transformer overload early warning method according to the instructions in the program codes.

It will be clearly understood by those skilled in the art that, for convenience and brevity of description, specific working procedures of the apparatus, module and device described above may refer to corresponding procedures in the foregoing method embodiments, which are not repeated herein.

In the several embodiments provided by the present application, it should be understood that the disclosed apparatus, device and method may be implemented in other manners. For example, the apparatus embodiments described above are merely illustrative, e.g., the division of the units is merely a logical function division, and there may be additional divisions when actually implemented, e.g., multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, which may be in electrical, mechanical or other form.

The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in the embodiments of the present invention may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units may be implemented in hardware or in software functional units.

The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied essentially or in part or all of the technical solution or in part in the form of a software product stored in a storage medium, including instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

The above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. The utility model provides a distribution transformer heavy overload early warning method which is characterized by comprising the following steps:

when a request of distribution network distribution transformer overload early warning is received, acquiring actual operation time sequence data corresponding to the request, performing first preprocessing by adopting the actual operation time sequence data, and constructing a target association diagram;

inputting the target association graph into a preset target load rate prediction model corresponding to the request, and outputting a predicted load rate; the preset target load rate prediction model comprises a graph sampling aggregation layer, a long-short-period memory layer, a first full-connection layer and a second full-connection layer;

calculating a heavy overload risk value according to the predicted load rate, and determining an early warning result corresponding to the request;

The step of inputting the target association graph to a preset target load rate prediction model and outputting a predicted load rate comprises the following steps:

inputting the target association graph to the graph sampling aggregation layer for sampling, aggregating, transferring updating and sequencing, and outputting an allocation data sequence;

encoding elements in the data sequence to generate an encoded data sequence;

inputting the encoded data sequence into the long-period and short-period memory layer for decoding, and outputting a first intermediate predicted value;

inputting the first intermediate predicted quantity to the first full-connection layer and outputting a second intermediate predicted quantity;

inputting the second intermediate pre-measurement to the second full connection layer and outputting a third intermediate pre-measurement;

performing inverse first preprocessing on the third intermediate pre-measurement to generate a predicted load rate;

the step of calculating a heavy overload risk value according to the predicted load rate and determining an early warning result corresponding to the request comprises the following steps:

determining a distribution state corresponding to the predicted load rate by adopting the predicted load rate and a preset load rate interval;

calculating a heavy overload risk value corresponding to the request according to the configuration state, the predicted load rate and the historical operation time sequence data corresponding to the request;

And determining the early warning grade and the early warning color corresponding to the request by adopting the heavy overload risk value and combining a preset risk value early warning interval.

2. The distribution network distribution transformer overload pre-warning method according to claim 1, wherein the performing a first pre-process and constructing a target association graph by using the actual operation time sequence data comprises:

performing first preprocessing by adopting the actual operation time sequence data to generate first time sequence data; wherein the first time sequence data comprises load rate data, characteristic variable data and influence factor data;

the characteristic variable data is taken as a node, the load rate data and the influence factor data are taken as node characteristics, a node connection relation is determined according to the distance between the nodes, and an initial association diagram is constructed;

calculating residual vectors among the nodes according to the node characteristics;

and updating the node connection relation in the initial association graph by adopting the residual vector and combining with preset node connection conditions to generate a target association graph.

3. The distribution transformer overload pre-warning method of claim 1, wherein the calculation formula of the overload risk value is as follows:

；

Wherein, for the heavy overload risk value, +.>For the predicted load factor,/->Time ratio of the preset load rate interval where the predicted load rate is located in the historical operation time sequence data is +.>For the state of the distribution transformer corresponding to the predicted load factor,>is->The duty cycle in the historical operating schedule data,for the load rate of->Cause->Is a severity of (c).

4. The distribution transformer overload pre-warning method of claim 3, wherein the severity is calculated by the following formula:

；

wherein, for a preset heavy overload rate threshold, < ->For a first predetermined parameter, +.>Is natural constant (18)>Is a preset second parameter.

5. The distribution transformer overload pre-warning method of claim 1, wherein the step of obtaining the preset target load rate prediction model comprises the following steps:

acquiring historical operation time sequence data of all the areas of the power distribution network, and performing second preprocessing to generate training time sequence data;

respectively constructing training association diagrams corresponding to different platform types by adopting the training time sequence data and combining the platform types corresponding to each platform;

and respectively training preset initial load rate prediction models by adopting the training association diagrams corresponding to different zone types, and generating preset target load rate prediction models corresponding to different zone types.

6. The distribution network distribution transformer overload pre-warning method according to claim 5, wherein the method for determining the type of the station area comprises the following steps:

respectively constructing load photovoltaic power curves corresponding to the areas according to the historical operation time sequence data;

calculating the distance between all the load photovoltaic power curves, and constructing a distance distribution matrix;

the distance distribution matrix is adopted, the local density of all the load photovoltaic power curves is calculated respectively, and the load photovoltaic power curves with the number of preset clusters and the maximum local density are selected as initial cluster centers;

respectively calculating the distance between each load photovoltaic power curve and each initial clustering center, classifying the initial clustering centers with the smallest distance from the load photovoltaic power curve into one type, and generating a cluster of a preset clustering number;

calculating a middle cluster center of each cluster of the transformer areas, and judging whether the middle cluster center meets a preset convergence condition or not;

if yes, confirming the type of the area corresponding to each area according to the area cluster;

if the distance between each load photovoltaic power curve and each initial clustering center is calculated, the initial clustering centers with the smallest distance are classified, and the cluster area clusters with the preset clustering number are generated.

7. The utility model provides a distribution network joins in marriage and becomes heavy overload early warning device which characterized in that includes:

the data processing module is used for acquiring actual operation time sequence data corresponding to a request when receiving the request of distribution transformer overload early warning of the power distribution network, and carrying out first preprocessing and constructing a target association diagram by adopting the actual operation time sequence data;

the model prediction module is used for inputting the target association graph into a preset target load rate prediction model corresponding to the request and outputting a predicted load rate; the preset target load rate prediction model comprises a graph sampling aggregation layer, a long-short-period memory layer, a first full-connection layer and a second full-connection layer;

the risk early warning module is used for calculating a heavy overload risk value according to the predicted load rate and determining an early warning result corresponding to the request;

the model prediction module is specifically used for:

inputting the target association graph to the graph sampling aggregation layer for sampling, aggregating, transferring updating and sequencing, and outputting an allocation data sequence;

encoding elements in the data sequence to generate an encoded data sequence;

inputting the encoded data sequence into the long-period and short-period memory layer for decoding, and outputting a first intermediate predicted value;

Inputting the first intermediate predicted quantity to the first full-connection layer and outputting a second intermediate predicted quantity;

inputting the second intermediate pre-measurement to the second full connection layer and outputting a third intermediate pre-measurement;

performing inverse first preprocessing on the third intermediate pre-measurement to generate a predicted load rate;

the risk early warning module is specifically used for:

determining a distribution state corresponding to the predicted load rate by adopting the predicted load rate and a preset load rate interval;

calculating a heavy overload risk value corresponding to the request according to the configuration state, the predicted load rate and the historical operation time sequence data corresponding to the request;

and determining the early warning grade and the early warning color corresponding to the request by adopting the heavy overload risk value and combining a preset risk value early warning interval.

8. An electronic device, comprising: a memory and a processor;

the memory is used for storing program codes and transmitting the program codes to the processor;

the processor is configured to execute the distribution network distribution transformer overload early warning method according to any one of claims 1 to 6 according to the instructions in the program code.