Search

CN-122026373-A - Distributed power distribution management method for power distribution cabinet

CN122026373ACN 122026373 ACN122026373 ACN 122026373ACN-122026373-A

Abstract

The invention discloses a distributed power distribution management method of a power distribution cabinet, and relates to the technical field of power distribution management. The method comprises the steps of constructing a standardized data set based on power distribution cabinet power operation data through a three-phase linear trend algorithm, training a graph neural network model based on a comparison learning mechanism and a generated countermeasure network through the standardized data set to obtain a first power distribution neural network model, constructing a federal learning architecture, performing federal learning on the first power distribution neural network model based on the federal learning architecture to obtain a second power distribution neural network model, inputting the standardized data set to obtain a power distribution prediction result and formulating a power dispatching basic strategy, optimizing a Di Jie St algorithm and an improved ant colony algorithm based on the power dispatching basic strategy, obtaining a power dispatching optimization strategy, executing the power dispatching optimization strategy, constructing an optimized data set by collecting feedback data in real time, updating the second power distribution neural network model, and finally obtaining a power dispatching final strategy.

Inventors

  • ZHOU XIANJUN
  • CHENG ZHONGYU
  • LIAN WEIZHENG
  • LIU FEI

Assignees

  • 国网湖北省电力有限公司洪湖市供电公司

Dates

Publication Date
20260512
Application Date
20251126

Claims (10)

  1. 1. A distributed power distribution management method of a power distribution cabinet is characterized by comprising the following steps: Step S1, collecting power operation data of a power distribution cabinet, processing the power operation data of the power distribution cabinet through a three-phase linear tide algorithm and constructing a standardized data set; Step S2, training a graph neural network model based on a contrast learning mechanism and a generated countermeasure network through the standardized data set to obtain a first power distribution neural network model; Step S3, constructing a federal learning architecture, performing federal learning on the first power distribution neural network model based on the federal learning architecture to obtain a second power distribution neural network model, inputting the standardized data set into the second power distribution neural network model to obtain a power distribution prediction result, and formulating a power dispatching basic strategy according to the power distribution prediction result; Step S4, based on a power dispatching basic strategy, adopting an improved Di Jie St-Lag algorithm and an improved ant colony algorithm to optimize to obtain a power dispatching optimization strategy, executing the power dispatching optimization strategy and collecting feedback data in real time; And S5, constructing an optimized data set according to the feedback data, updating the second power distribution neural network model based on the optimized data set, and adjusting the power dispatching optimization strategy according to an updating result to obtain a final power dispatching strategy so as to realize distributed power distribution management of the power distribution cabinet.
  2. 2. The method for distributed power distribution management of a power distribution cabinet according to claim 1, wherein the processing of the power operation data of the power distribution cabinet by a three-phase linear power flow algorithm and the construction of a standardized data set comprises the following specific steps: Collecting power running data of power distribution cabinet The electric cabinet power operation data comprises a basic power parameter subset And three-phase characteristic parameter subset ; Based on Calculating three-phase unbalance degree, quantifying the unbalance degree, and calculating the voltage unbalance degree according to the following formula: ; Wherein, the The reference voltage level of the power distribution cabinet is reflected for the average value of the three-phase voltage of the power distribution cabinet i at the time t, 、 And The effective values of A, B, C phase voltages of the power distribution cabinet i at the time t are respectively, Is the degree of voltage imbalance; the calculation formula of the current unbalance is as follows: ; Wherein, the For the three-phase current average value of the power distribution cabinet i at the time t, reflecting the reference load level of the power distribution cabinet, 、 And The effective values of A, B, C phases of currents of the power distribution cabinet i at the time t are respectively, Is the degree of current imbalance; the original voltage and current parameters are corrected, parameter deviation caused by three-phase unbalance is eliminated, and for a line ij connecting the power distribution cabinets i and j, the transmission power correction formula of each phase is as follows: ; ; Wherein, the And Respectively the line ij at the time t Phase corrected active and reactive power transfer, And For the power distribution cabinet i at time t Phase total active output power and reactive output power, For a contiguous set of power distribution cabinets directly connected to power distribution cabinet i, And For transmission of adjacent power distribution cabinets k to i The phase active power and the reactive power, And For the power distribution cabinet i at time t Phase load active and reactive power consumption; based on the corrected line transmission power, for the power distribution cabinet j The phase voltage is corrected, and the formula is: ; Wherein, the For the corrected power distribution cabinet j at time t Phase voltage, eliminating voltage deviation caused by line loss and three-phase unbalance, For the corrected power distribution cabinet i at time t The phase voltage is applied to the substrate at a voltage level, And Respectively the lines ij The phase resistance and the reactance of the phase, The average value of the three-phase voltages of the power distribution cabinet i at the time t is used for normalizing line loss items; defining the unbalance of the edge level and the line The calculation formula of the side level unbalance degree is as follows: ; ; Wherein, the And Respectively representing the voltage unbalance degree of the node v and the node u at the time t, For the three impedance differences of the line, The value of the impedance difference influence coefficient is 0.05, Is a circuit Is added to the degree of side-level voltage imbalance, Is a circuit Is a side level current imbalance; Integrating the corrected parameters with the unbalance index to construct a standardized data set The dataset contains the following dimensions: Node characteristic subset of corrected three-phase voltage of each power distribution cabinet i Three-phase current after correction Critical component temperature , Is a load power factor, and the side characteristic subset is the corrected three-phase transmission power of each line ij 、 Line resistance Reactance of A subset of unbalance degree, namely voltage unbalance degree of each power distribution cabinet i Degree of unbalance of current Side voltage imbalance, side current imbalance.
  3. 3. The method for distributed power distribution management of a power distribution cabinet according to claim 2, wherein the training of the graph neural network model based on a contrast learning mechanism and a generated countermeasure network by the standardized data set to obtain a first power distribution neural network model comprises the following steps: Construction of initial graph neural network Based on standardized data sets Construction of Node feature matrix and edge feature matrix of (a) and defining a network layer structure: Node characteristic matrix : Wherein For node feature dimensions, matrix elements Representing the k-th dimension eigenvalue of node v, edge eigenvalue matrix The method comprises the steps of line resistance, line reactance, average transmission efficiency, average apparent power, side current unbalance, side voltage unbalance and load priority; The activation function adopts ReLU, the weight initialization adopts He normal distribution, and the learning rate initial value =0.001; Constructing a set of pairs of samples Wherein the sample is anchored Embedding features for node v at time t, positive samples At the moment of time for node v Node embedded features of (1); negative example Embedding a feature for a node u non-contiguous to node v at time t; Definition of contrast loss function The feature similarity of the anchoring sample and the positive and negative samples is measured, and the calculation formula of the contrast loss function is as follows: ; Wherein, the To compare the loss values, a smaller loss represents a higher similarity between the anchor sample and the positive sample, a lower similarity between the anchor sample and the negative sample, a stronger feature discrimination, For the number of sets of pairs of samples, Is a cosine similarity function and is used for measuring the directional consistency of embedded features of two nodes, The value of the temperature parameter is 0.08, Embedding features for nodes of the kth negative sample; the contrast loss is fed back to the weight adjustment, and the weight calculation formula after adjustment is as follows: ; Wherein, the For the line-contrast loss value, In order to adjust the coefficient, the value is 0.15, To balance the correction factor, take a value of 1.2, Is the adjusted weight; the calculation formula of the prediction loss of (2) is as follows: ; Wherein, the As a real feature of the device, it is, In order to predict the characteristics of the feature, Is that A predictive loss of (2); Will be And (3) with Predictive loss of (a) Weighting and fusing to obtain total loss The calculation formula of (2) is as follows: ; Wherein, the In order to predict the weight of the loss, =0.7, In order to compare the loss value of the sample, Is that Is used for the prediction of the loss of (a), Is the total loss; based on after optimization Executing a forward propagation process by combining the corrected power parameters, simulating the transmission and influence relation of power among the distributed power distribution cabinets, and generating a data propagation trend; Introducing and generating an antagonism network optimization weight matrix W to finally obtain a trained first graph neural network model 。
  4. 4. The method for distributed power distribution management of a power distribution cabinet according to claim 3, wherein said constructing an initial graph neural network comprises the steps of: mapping a power distribution system into an initial graph neural network based on physical connection relation and power transmission paths of a distributed power distribution cabinet V is the number of nodes and E is the number of lines, where each component is defined as follows: The node set V is that each power distribution cabinet corresponds to one node V, the physical meaning of the node V is a power distribution unit with independent power metering and control functions, and the initial attribute of the node set V comprises corrected parameters of the node feature subset in the step S1; edge set E, wherein each section of power transmission line for connecting node v with node u corresponds to one edge The physical meaning of the edge is a cable circuit for carrying power transmission among nodes, and the initial attribute of the cable circuit comprises the corrected parameters of the edge feature subset in the step S1; Based on edge feature matrix The initial edge weight is calculated by 7 dimensions, and the calculation formula of the initial edge weight is as follows: ; Wherein, the Is a normalized value for the kth edge feature dimension, As the dimension weight coefficient of the object to be measured, Is the initial edge weight.
  5. 5. The method for distributed power distribution management of a power distribution cabinet according to claim 4, wherein said introducing generates an antagonism network optimization weight matrix, comprising the steps of: Generator G random noise vector input as node v And 4-dimensional key indexes are output as simulated node characteristics Adopting a 3-layer full-connection layer+1-layer batch normalization layer structure, wherein the input is a real edge characteristic or a simulation characteristic, and the output is a real degree score of the characteristic; The GAN training adopts an alternate optimization strategy, optimizes D and then optimizes G, and the calculation formulas of the loss of the discriminator and the loss of the generator are as follows: ; ; Wherein, the For the discriminator to true characteristics Is used for the degree of authenticity score of (c) in the database, Generator input noise And degree of unbalance 、 The analog characteristics of the post-output, For the arbiter to score the authenticity of the generated feature, In order for the arbiter to be lost, Generator loss; initializing parameters of G and D, adopting Xavier normal distribution of weights, and learning rate Iterative training 150 rounds, when And (3) with When stable, GAN reaches Nash equilibrium; When GAN reaches Nash equilibrium, the weight is optimized by using 10 groups of simulation features, and the formula is modified as follows: ; Wherein, the In order to optimize the edge weights after the optimization, For the k-th set of analog features, Is a balancing constraint factor.
  6. 6. The method for distributed power distribution management of a power distribution cabinet according to claim 5, wherein said constructing a federal learning architecture comprises the steps of: the method comprises the steps of constructing a local node-federation server two-cascade bang learning architecture, wherein functions and symbols of each component are defined as follows: Local node Dividing according to distribution areas, wherein each distribution cabinet corresponds to 1 local node The core function is local model training and feature comparison checking, the step S1 of standardizing a data set is not uploaded in the step S2 of the original power data, and the step S2 of the first power distribution neural network model The node characteristics and the edge weight parameters of the network node, and RSA encryption and differential privacy are adopted when the local node uploads the parameters; The edge gateway multiplexes the LoRaWAN communication group of S1 to realize the encryption and forwarding of model parameters without additional deployment of communication modules; The federation server S is deployed in a regional power distribution room and executes aggregation and strategy generation; Local model Each of the following Based on Is used for constructing a local model by a network structure of the system, and initial parameter inheritance A kind of electronic device Extracting layer parameters from node characteristics; The federation server S calculates a dynamic aggregation period based on the load timing data of step S1 The calculation formula is as follows: ; Wherein, the For the dynamic aggregation period at time t, For the polymerization period at time t-1, For the period smoothing weight, the value is 0.6, Taking the value of 10min as the reference polymerization period, As the load fluctuation coefficient at the time t, The average side level imbalance at time t.
  7. 7. The distributed power distribution management method of the power distribution cabinet of claim 6, wherein the federal learning architecture-based federal learning is used for performing federal learning on the first power distribution neural network model to obtain a second power distribution neural network model, and the method comprises the following steps: Local model The training of the model adopts three factors of a total loss function of prediction loss, characteristic alignment loss and edge level unbalance degree fitting loss, and the total loss function The calculation formula of (2) is as follows: ; Wherein, the Is a local node Is used to determine the total loss value of (a), To predict the loss weight, take a value of 0.7, The loss weight is aligned for the feature, the value is 0.3, In order to predict the loss locally, In order to account for the loss of alignment of features, Fitting the loss for the edge level imbalance, Fitting loss weight for the unbalance degree of the side level, and taking a value of 0.2; Introducing a contrast learning checking mechanism, and receiving a local model by the federal server S The two-dimensional comparison and verification is carried out, and the specific flow is as follows: Firstly, checking the feature similarity, and calculating each local model Output characteristics of (a) And (3) with Average cosine similarity of (2) : , wherein, Is of a local model Is characterized by an average degree of similarity of features of (3), For the number of features of the global reference feature library, Is that Is provided with a single reference feature of the pattern, Is of a local model Setting a similarity threshold for the feature extraction result of the node m =0.85; Verifying the prediction error of the local model on the side unbalance, judging that the feature similarity verification and the unbalance fitting accuracy verification reach standards to be effective models and incorporating aggregation, judging that any model does not reach standards to be abnormal models, feeding back the local node to adjust the learning rate, and retraining; Local node After training and passing the comparison and verification, the method The core parameters of the (E) are uploaded to a federation server S, the federation server S monitors the number of the uploaded effective models in real time, and aggregation is triggered when the following conditions are met, namely the number of the effective models reaches 80% of the total local node number or the time from last aggregation reaches a dynamic aggregation period ; The federal server S performs weighted aggregation on the received effective local models to generate a second power distribution neural network model The aggregation formula is: ; Wherein, the For the second distribution neural network model Is used for the weight matrix of the (c), To be an index set of the active local model, Is of a local model Is a weight of aggregation of (1); Local model The calculation formula of the aggregate weight of (2) is as follows: ; Wherein, the For the local data quality score, For the purpose of obtaining the precision score of the model, The value of the data quality weight coefficient is 0.4, Is of a local model Is used to determine the aggregate weight of the (c) for the (c), The model precision weight coefficient takes a value of 0.3 as a load priority weight coefficient and takes a value of 0.3, Scoring load priority; finally obtaining a second power distribution neural network model 。
  8. 8. The method for distributed power distribution management of a power distribution cabinet according to claim 7, wherein the step of inputting the standardized data set into a second power distribution neural network model to obtain a power distribution prediction result comprises the following specific steps: inputting the standardized data set into a second power distribution neural network model, and obtaining a power distribution prediction result after each round of aggregation by a federal server S Based on the multidimensional power distribution prediction result, a structured power dispatching basic strategy is generated, and the specific steps are as follows: Calculating the load rate of each node Deviation from equilibrium : ; ; Wherein, the For the load factor of node m at time t, For the rated load power of node m, For the average load rate of all nodes at time t, Load balancing deviation for node m; the high load node is > + The low-load node is < - For high load nodes, based on Selecting an optimal power transfer path Path selection rule, preference selection and Directly connected and <10% Low load node Path weight , Selection for optimization weights of each edge in Path Maximum path, calculate transfer power: ; Wherein, the In order to transfer the power, Indicating that the high load node may transfer power, Indicating that the low load node can accept power; Sequencing according to 'critical load guarantee' serious unbalance management 'common load scheduling', and outputting a power scheduling basic strategy Is a structured instruction of (a).
  9. 9. The method for distributed power distribution management of a power distribution cabinet according to claim 8, wherein the modified Di Jie Style algorithm is as follows: Path cost function The definition is as follows: ; Wherein, the For the integrated cost of the path p, 、 And As a result of the cost-weight coefficient, =0.3、 =0.25、 =0.25, =0.2, For the transmission loss cost of the path p, For the cost of the response time of path p, For the three-phase imbalance cost of path p, Cost for real-time edge weighting; Based on Is used for executing the improved Dijkstra algorithm and outputting an optimal power supply path The flow is as follows: first, initializing, and setting a start point set s= Endpoint set t= (Low load node), path cost matrix = Optimal path precursor node matrix prev (n) = For each starting point, it directly adjoins the initial cost of the line ; Selecting a node with the minimum cost from nodes with undetermined optimal paths , For node set of determined optimal path, pair Updating the path cost of its neighboring nodes: If (1) < Then update = And records the precursor node prev (u) = Will (i) be Joining an optimal path set Repeating the above steps until all the end points T are included ; For each high-load node, backtracking the precursor node prev (u) from the corresponding low-load node to obtain an optimal path 。
  10. 10. The distributed power distribution management method of a power distribution cabinet according to claim 9, wherein the improved ant colony algorithm is as follows: Fitness function The calculation formula of (2) is as follows: ; Wherein, the The q-th level load is allocated with the adaptability, the higher the adaptability is, the more preferentially the power is allocated, The value of the priority weight coefficient is 0.7, As a load priority weight, Is a low-load node Is used to receive the power available from the power source, For the total required power of the q-th stage load, Is a low-load node Is used to determine the current three-phase imbalance of (a), Is a three-phase imbalance limit; Based on improved Dijkstra output Executing an improved ant colony algorithm, dynamically distributing loads of all nodes and outputting a load distribution scheme The specific flow is as follows: Each ant starts from a high-load node and follows an optimal path Moving to a low load node, distributing load based on a roulette selection strategy: ; Wherein, the For the pheromone concentration of the q-th stage load, The value of the weight coefficient of the pheromone is 1.0, The q-th stage load is assigned a fitness, Distributing load for the q-th stage; after ants distribute loads, ensuring that q=1-2 level key loads are 100 percent reserved, reasonably distributing q=3-5 level loads according to available power, after distribution, node three-phase unbalance is less than or equal to 2 percent, and load rate is not more than rated value, and if the load rate is not more than rated value, reselecting, and after iteration is finished, updating pheromone Wherein For the average fitness of the q-th level load, after 30 rounds of iteration, selecting a load distribution scheme with maximum sum of fitness and up to the standard of balance 。

Description

Distributed power distribution management method for power distribution cabinet Technical Field The invention relates to the technical field of power distribution cabinets, in particular to a distributed power distribution management method of a power distribution cabinet. Background With the increase of the access proportion of the distributed power supply and the increase of the electric automobile and the controllable load, the distributed power distribution cabinet needs to deal with complex working conditions such as three-phase unbalance, frequent load fluctuation and the like, and management of the distributed power distribution cabinet needs to rely on multi-dimensional power data cooperation and accurate scheduling so as to meet the requirements of power supply reliability and power quality, and the traditional power distribution management mode is difficult to adapt to the dynamic operation requirement under the access of high-proportion renewable energy sources. In distributed power distribution management, a federal learning method is often adopted in a traditional method, namely, a local node-federal server architecture is constructed, the local node uploads model parameters to a server according to a fixed period based on a power data training model, the server weights and aggregates the model parameters and generates a global model, and a power distribution scheduling strategy is formulated according to the global model, so that model cooperation is realized while data privacy is ensured. However, the conventional federal learning adopts a fixed aggregation period, cannot be dynamically adjusted according to the load fluctuation characteristic of the power distribution cabinet, and when the industrial load suddenly increases and decreases, the fixed period can lead to aggregation hysteresis, so that the cooperative efficiency of distributed power data is limited, the real-time operation change of a power distribution system is difficult to quickly respond, the timeliness and the suitability of a subsequent scheduling strategy are finally affected, and the management requirement of a dynamic power distribution scene cannot be met. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a distributed power distribution management method of a power distribution cabinet, which solves the problems existing in the background art. In order to achieve the purpose, the distributed power distribution management method of the power distribution cabinet comprises the following steps of: Step S1, collecting power operation data of a power distribution cabinet, processing the power operation data of the power distribution cabinet through a three-phase linear tide algorithm and constructing a standardized data set; Step S2, training a graph neural network model based on a contrast learning mechanism and a generated countermeasure network through the standardized data set to obtain a first power distribution neural network model; Step S3, constructing a federal learning architecture, performing federal learning on the first power distribution neural network model based on the federal learning architecture to obtain a second power distribution neural network model, inputting the standardized data set into the second power distribution neural network model to obtain a power distribution prediction result, and formulating a power dispatching basic strategy according to the power distribution prediction result; Step S4, based on a power dispatching basic strategy, adopting an improved Di Jie St-Lag algorithm and an improved ant colony algorithm to optimize to obtain a power dispatching optimization strategy, executing the power dispatching optimization strategy and collecting feedback data in real time; And S5, constructing an optimized data set according to the feedback data, updating the second power distribution neural network model based on the optimized data set, and adjusting the power dispatching optimization strategy according to an updating result to obtain a final power dispatching strategy so as to realize distributed power distribution management of the power distribution cabinet. Preferably, the processing the power operation data of the power distribution cabinet by using a three-phase linear power flow algorithm and constructing a standardized data set comprises the following specific steps: Collecting power running data of power distribution cabinet The electric cabinet power operation data comprises a basic power parameter subsetAnd three-phase characteristic parameter subset; Based onCalculating three-phase unbalance degree, quantifying the unbalance degree, and calculating the voltage unbalance degree according to the following formula: Wherein, the The reference voltage level of the power distribution cabinet is reflected for the average value of the three-phase voltage of the power distribution cabinet i at the time t,、AndThe effective values of A, B, C phase