CN-122000984-A - High-permeability distributed photovoltaic power distribution network reconstruction method and system

CN122000984ACN 122000984 ACN122000984 ACN 122000984ACN-122000984-A

Abstract

The invention discloses a reconstruction method and a reconstruction system of a high-permeability distributed photovoltaic power distribution network, which belong to the technical field of power system dispatching, wherein the method comprises the steps of designing a continuous state space based on system data of the distributed photovoltaic power distribution network to generate a plurality of attribute state matrixes; the method comprises the steps of generating a strategy sample by using an attribute state matrix based on a traditional distribution network topology optimization algorithm, screening high-quality strategy tracks to construct an expert strategy track library, training an intelligent body model by using the expert strategy track library based on a state simulation reinforcement learning algorithm, inputting current distribution network running state data into the intelligent body model to generate a real-time reconstruction strategy, evaluating the real-time reconstruction strategy based on a comprehensive index system, and dynamically updating the expert strategy track library adopted during intelligent body model training.

Inventors

GUO NING
LU XIAOXING
GUO ZIRAN
YUAN GUANGYU
CHEN WENJIA
LIU JIAN
YUAN YUBO
ZANG HAIXIANG
XUE ZHITONG
ZHAO YONGKAI
CHEN JINMING
XIAO XIAOLONG
SONG SHUANG

Assignees

国网江苏省电力有限公司电力科学研究院
国网江苏省电力有限公司
河海大学
江苏省电力试验研究院有限公司

Dates

Publication Date: 20260508
Application Date: 20251125

Claims (14)

1. The method for reconstructing the high-permeability distributed photovoltaic power distribution network is characterized by comprising the following steps of: inputting the acquired current distribution network running state data into a pre-trained agent model to generate a real-time reconstruction strategy; Evaluating the real-time reconstruction strategy based on a comprehensive index system, and dynamically updating an expert strategy track library adopted during training of the intelligent body model; The training method of the intelligent agent model comprises the following steps: based on the acquired system data of the distributed photovoltaic power distribution network, designing a continuous state space, and generating a plurality of attribute state matrixes; Based on a traditional distribution network topology optimization algorithm, generating a strategy sample by using the attribute state matrix, and screening high-quality strategy tracks to construct an expert strategy track library; training an agent model based on a state simulation reinforcement learning algorithm by using the expert strategy trajectory library; the intelligent agent model inputs the current attribute state matrix and the target attribute state matrix, and outputs a distribution network topology reconstruction policy vector.
2. The method of claim 1, wherein the system data comprises system grid topology data, line impedance data, node voltage data, node injection active data, and reactive data.
3. The method for reconstructing the high-permeability distributed photovoltaic power distribution network according to claim 1, wherein the method for acquiring the system data of the distributed photovoltaic power distribution network comprises the steps of acquiring parameters of reactive equipment, energy storage equipment and transformers configured by all nodes in the system, and voltage amplitude and voltage phase angle of all the nodes.
4. The method for reconstructing a high-permeability distributed photovoltaic power distribution network according to claim 3, wherein the designing the continuous state space comprises converting a topological connection relationship of nodes of the power distribution network system into two-dimensional coordinates by using a multidimensional scale analysis algorithm, wherein euclidean distances between the two-dimensional coordinates are as close as possible to a distance matrix calculated based on a line impedance module value, and the calculating process is as follows: ; wherein the matrix As a matrix of distances between nodes of the system, In the form of the square of the distance matrix, Is that A rank identity matrix having diagonal elements of 1, non-diagonal elements of 0, Is a vector of all 1 s and, For centering the matrix, for translating the data to the origin, Is an inner product matrix for reflecting the inner product relationship among the nodes, Is a matrix of eigenvalues, the eigenvalues of which are arranged in descending order, As a matrix of feature vectors, A diagonal matrix consisting of the first two largest eigenvalues, the eigenvalues being arranged in descending order, For the first two principal eigenvector matrices, Is a two-dimensional coordinate matrix.
5. The method of claim 4, further comprising fixing the node 0 coordinates at the origin and aligning the node 1 coordinates to the vertical axis to unify the coordinates as follows: ; Wherein, the As the original coordinates of the node 0, In order to translate the post-coordinate matrix, For the vertical axis coordinates and the horizontal axis coordinates of the node 1, In order for the angle of rotation to be a function of, As a four-quadrant arctangent function, And the system node coordinates after alignment.
6. The method for reconstructing a high-permeability distributed photovoltaic power distribution network according to claim 4, wherein the plurality of attribute state matrices are obtained by downsampling and interpolating within a fixed window range, wherein the attribute matrices include a voltage attribute matrix, an active power attribute matrix and a reactive power attribute matrix, and the attribute matrix coordinate calculation process is as follows: ; Wherein, the Is the first The two-dimensional abscissa of the node, Is the boundary range of transformed coordinates, i is the side length of a square window covering all coordinate points, d is a redundant length for avoiding coordinate out-of-limit under special conditions, Respectively representing the center coordinates and the boundary coordinates of the square window, Representing the number of sample points on a single side of the downsampling grid within a square window, Representing a coordinate sequence of grid points in the direction of the transverse and longitudinal axes, using Representing a global grid point coordinate matrix.
7. The method for reconstructing a high-permeability distributed photovoltaic power distribution network according to claim 4, wherein the downsampling interpolation uses a radial basis function interpolation method, and for sampling grid coordinate points represented by countless values, interpolation is performed according to attributes under original system node coordinates to generate a continuous attribute state matrix, and the radial basis function interpolation process is as follows: ; Wherein, the As a function of the interpolation of the values, Radial basis function weighting coefficients for the ith node, As a function of the radial basis function, Is a shape parameter, is used to control the decay rate of the function, For the euclidean distance of point (x, y) to the node, The term coefficients of the linear polynomials respectively, As a matrix element of the radial basis function, 、、 Respectively representing the polynomial condition matrix, the node attribute value vector and the interpolation result of the grid points.
8. The method for reconstructing the high-permeability distributed photovoltaic power distribution network according to claim 1, wherein the generating the policy sample by using the attribute state matrix based on the conventional distribution network topology optimization algorithm comprises solving an optimal reconstruction policy under the condition of meeting the power flow constraint, the line switch action frequency constraint and the radial topology constraint of the system to generate a binary decision vector as the policy sample.
9. The method for reconstructing the high-permeability distributed photovoltaic power distribution network according to claim 8, wherein the strategy samples are spliced to form a daily real-time reconstruction strategy track matrix in a set time scale, and the daily real-time reconstruction strategy track matrix is stored in an expert strategy track library, and each strategy track corresponds to three groups of attribute state matrixes.
10. The method for reconstructing a high-permeability photovoltaic power distribution network according to claim 9, wherein the state simulation reinforcement learning algorithm comprises two sets of neural network models, an action network model and a judgment network model; the action network model is input into three groups of attribute state matrixes which are complete at the current moment and a target voltage attribute state matrix from an expert strategy track library, and outputs a decision vector for transferring the current environment state to the target voltage state; And the evaluation network model is input into the decision vector and a complete attribute state matrix at the current moment, and is output into an estimated value of the system voltage attribute state matrix after the decision is executed.
11. The method for reconstructing a high-permeability distributed photovoltaic power distribution network according to claim 10, wherein the training of the intelligent body model uses a convolutional neural network as a pre-feature extraction layer, and uses a two-norm of the deviation of a system target voltage state matrix from an estimated voltage state as a loss function to guide parameter optimization, and the calculation process is as follows: ; Wherein, the As an internal network parameter of the model, For observation states in adjacent time steps, three types of attribute state matrices are broadly referred to herein, In order to predict the state transition action, For the fitting loss of the judgment model, the judgment model f is ensured to accurately predict the real action A transition of the state of the lower environment, For forward consistency loss, a two-norm measure is used to measure the degree of deviation of the transition state from the actual state, And (3) simulating decision actions in the expert strategy trajectory library through labeled training for standard action loss functions.
12. The method for reconstructing the high-permeability distributed photovoltaic power distribution network according to claim 1, wherein the step of inputting the acquired current power distribution network operation state data into a pre-trained agent model to generate a real-time reconstruction strategy comprises the steps of matching a state quantity most similar to a current state from an expert strategy track library, and using the current state matrix and a next-moment state matrix in the expert strategy track as agent model inputs to generate a reconstruction strategy vector.
13. The method for reconstructing the high-permeability distributed photovoltaic power distribution network according to claim 1, wherein dynamically updating the expert strategy trajectory library adopted during training of the intelligent body model comprises calculating economic benefit and safety benefit of the current system strategy trajectory by using a comprehensive index system, and when the value reaches a threshold value, determining the current strategy trajectory as a high-quality sample and updating the high-quality sample to the expert strategy trajectory library, wherein the calculation formula is as follows: ; Wherein, the Quantifying the goodness of the current decision with a comprehensive index system for the reward function, As the coefficient of the net loss, And the system network loss after the action is executed. For the economic coefficient of the operation of the switch, The number of switching actions.
14. A high-permeability distributed photovoltaic power distribution network reconstruction system for implementing the high-permeability distributed photovoltaic power distribution network reconstruction method according to any one of claims 1 to 13, comprising: The state space design module is used for designing a continuous state space based on the acquired system data of the distributed photovoltaic power distribution network and generating a plurality of attribute state matrixes; the expert policy track library module is used for generating a policy sample by using the attribute state matrix based on a traditional distribution network topology optimization algorithm, and screening high-quality policy tracks to construct an expert policy track library; The intelligent body training module is used for training an intelligent body model by using the expert strategy track library based on a state imitation reinforcement learning algorithm, wherein the intelligent body model inputs a current attribute state matrix and a target attribute state matrix and outputs a distribution network topology reconstruction strategy vector; The reconstruction strategy resolving module is used for inputting the acquired current distribution network running state data into a pre-trained agent model to generate a real-time reconstruction strategy; and the dynamic updating module is used for evaluating the real-time reconstruction strategy based on the comprehensive index system and dynamically updating an expert strategy track library adopted during training of the intelligent body model.

Description

High-permeability distributed photovoltaic power distribution network reconstruction method and system Technical Field The invention relates to a high-permeability distributed photovoltaic power distribution network reconstruction method and system, and belongs to the technical field of power system dispatching. Background With large-scale access of distributed photovoltaic power generation, the operation morphology of a power distribution network is deeply changed, the problems of voltage out-of-limit and network loss increase are increasingly outstanding, and the load flow distribution needs to be optimized through dynamic topology reconstruction so as to improve the bearing capacity of the system. The conventional reconstruction multi-reliance mathematical optimization method of the power distribution network solves the optimal topology scheme on the premise of meeting radial constraint, switch action frequency limitation and power flow feasibility. In recent years, reinforcement learning has been introduced into this field for its potential in complex decision problems, enabling autonomous decision-making by building environmental states and action spaces. In the prior art, a deep reinforcement learning model is adopted to generate a distribution network reconstruction strategy, and a neural network approximates a strategy function or a cost function to realize response to a real-time running state. However, the pure reinforcement learning method faces the problems of large fluctuation, slow convergence speed, low sample efficiency and the like in the practical application, and particularly, the decision stability and feasibility are difficult to ensure in a high-dimensional, nonlinear and safety-constrained power system. In addition, due to the lack of effective a priori knowledge guidance, the agent needs extensive exploration to obtain a superior strategy, resulting in high training costs and difficulty in rapid deployment to new scenarios. The existing method generally only depends on an online learning mechanism, history high-quality strategy tracks are not fully utilized as experience guidance, the generalization capability and real-time response performance of the model are limited, and the requirement of rapid reconstruction of frequent fluctuation in a hypertonic photovoltaic scene cannot be met. Disclosure of Invention The invention aims to overcome the defects in the prior art, and provides a high-permeability distributed photovoltaic power distribution network reconstruction method and system, which can rapidly apply a reinforcement learning model to a new scene by introducing an expert strategy track library, obtain a better decision effect, respond to distributed photovoltaic fluctuation in real time and rapidly and accurately adjust the network topology of the system. In order to achieve the above purpose, the invention is realized by adopting the following technical scheme: in a first aspect, the present invention provides a method for reconstructing a high-permeability distributed photovoltaic power distribution network, including: inputting the acquired current distribution network running state data into a pre-trained agent model to generate a real-time reconstruction strategy; Evaluating the real-time reconstruction strategy based on a comprehensive index system, and dynamically updating an expert strategy track library adopted during training of the intelligent body model; The training method of the intelligent agent model comprises the following steps: based on the acquired system data of the distributed photovoltaic power distribution network, designing a continuous state space, and generating a plurality of attribute state matrixes; Based on a traditional distribution network topology optimization algorithm, generating a strategy sample by using the attribute state matrix, and screening high-quality strategy tracks to construct an expert strategy track library; training an agent model based on a state simulation reinforcement learning algorithm by using the expert strategy trajectory library; the intelligent agent model inputs the current attribute state matrix and the target attribute state matrix, and outputs a distribution network topology reconstruction policy vector. Further, the system data includes system grid topology data, line impedance data, node voltage data, node injection active data and reactive data. Further, the system data acquisition method of the distributed photovoltaic power distribution network comprises the steps of acquiring parameters of reactive power equipment, energy storage equipment and transformers configured by all nodes in the system, and voltage amplitude values and voltage phase angles of all nodes. Further, the design continuous state space comprises the steps of converting a topological connection relation of nodes of the distribution network system into two-dimensional coordinates by using a multidimensional scale analysis algorithm, wherein