Power distribution network reconstruction optimization method for improving distributed new energy bearing capacity

Abstract

The invention discloses a power distribution network reconstruction optimization method for improving distributed new energy bearing capacity, and relates to the technical field of intelligent power distribution network optimization. The method comprises the following steps: constructing an evolution state path diagram for describing a system operation state space-time evolution rule based on historical and real-time operation data of a target power distribution network; identifying a plurality of potential weak moments when the bearing capacity margin of the system is lower than a dynamic safety threshold value in a future prediction time domain based on the prediction operation data of time-space association; analyzing relevance and time sequence proximity of the running states corresponding to the plurality of potential weak moments in the evolution state path diagram, and clustering to generate one or more demand time periods with internal state consistency; and for each demand period, determining a unique optimal network topology and a corresponding switch operation instruction, and integrating the instructions of each period to generate a multi-period reconstruction scheduling scheme. According to the invention, unnecessary switching operation times are reduced while the new energy consumption capability is ensured, and the reconstruction optimization effect of the power distribution network is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent power distribution network optimization, in particular to a power distribution network reconstruction optimization method for improving the carrying capacity of distributed new energy. BACKGROUND

[0002] With high proportion of wind power, photovoltaic and other distributed new energy connected to the power distribution network, the output of new energy and the demand of load present a dramatic fluctuation characteristic with time. The existing technology mainly adjusts the network topology through power distribution network reconstruction to optimize the power flow distribution, relieve branch overload and voltage out-of-limit, so as to improve the carrying capacity of the system to new energy.

[0003] However, the existing reconstruction method is mainly based on the optimization idea of static or fixed time interval, assuming that the topology remains unchanged in a long period of time, which is difficult to adapt to the rapid time-varying characteristics of new energy and load. If the reconstruction is frequent to track fluctuations, it will lead to too many switch actions, affecting the service life of the equipment. At the same time, the traditional piecewise optimization method ignores the space-time correlation of system state evolution, and only optimizes independently according to time slices, which is easy to cause switch operation redundancy and overall optimization effect decline. SUMMARY

[0004] The present application provides a power distribution network reconstruction optimization method for improving the carrying capacity of distributed new energy, aiming to solve the technical problems of switch operation redundancy and poor optimization effect in the existing power distribution network reconstruction.

[0005] In view of the above problems, the present application provides a power distribution network reconstruction optimization method for improving the carrying capacity of distributed new energy, comprising: Based on the historical and real-time operation data of the target power distribution network, an evolution state path graph is constructed to describe the space-time evolution law of system operation state; Based on the predicted operation data of the space-time correlation, a plurality of potential weak moments in the future prediction time domain are identified, in which the system carrying capacity margin is lower than the dynamic safety threshold; The correlation and time sequence proximity of the operation state corresponding to the plurality of potential weak moments in the evolution state path graph are analyzed, and one or more demand periods with internal state consistency are clustered; For each demand period, a unique optimal network topology and corresponding switch operation instruction are determined, and the instructions of each period are integrated to generate a multi-period reconstruction scheduling scheme.

[0006] The one or more technical solutions provided in the present application have at least the following technical effects or advantages: This invention provides a distribution network reconfiguration optimization method to enhance the carrying capacity of distributed renewable energy. First, it constructs an evolutionary state path diagram based on historical and real-time operational data to accurately depict the spatiotemporal evolution of the distribution network's operating state, laying a data and model foundation for subsequent optimization. Then, it identifies potential future weak points based on spatiotemporal correlation prediction data, enabling early prediction and precise location of system carrying capacity risks. Next, it generates demand periods with internal consistency by analyzing state correlations and temporal proximity clustering, effectively avoiding the drawbacks of traditional segmented optimization that ignores temporal correlations. Finally, it determines the optimal network topology and switch operation instructions for each demand period and integrates them into a multi-period scheduling scheme. This scheme adapts to the rapid time-varying characteristics of renewable energy and loads, reduces redundant switch operations and equipment losses, enhances the carrying capacity of distributed renewable energy in the distribution network, effectively avoids safety hazards such as branch overload and voltage exceeding limits, and ensures the safe, efficient, and economical operation of the distribution network. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a flowchart illustrating the distribution network reconfiguration optimization method for enhancing the carrying capacity of distributed new energy sources, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the process for constructing an evolutionary state path diagram in the distribution network reconfiguration optimization method for improving the carrying capacity of distributed new energy sources provided in an embodiment of the present invention. Detailed Implementation

[0009] This invention provides a distribution network reconfiguration optimization method for improving the carrying capacity of distributed new energy sources, which addresses the technical problems of redundant switching operations and poor optimization effects in existing distribution network reconfiguration.

[0010] Examples, such as Figure 1 As shown, this invention provides a distribution network reconfiguration optimization method for improving the carrying capacity of distributed renewable energy sources, the method comprising: S100: Based on the historical and real-time operation data of the target distribution network, construct an evolutionary state path diagram that describes the spatiotemporal evolution of the system's operating state.

[0011] In this embodiment of the invention, an evolutionary state path diagram describing the spatiotemporal evolution of the system's operating state is constructed based on historical and real-time operational data of the target distribution network. In target distribution networks containing a high proportion of wind and solar power, the output of new energy sources and load demand exhibit significant time-varying characteristics, and the operating states at different times show spatiotemporal correlations. Traditional distribution network reconfiguration and optimization methods lack precise characterization of spatiotemporal evolution patterns, making it difficult to effectively predict the future operating state of the system, resulting in poor adaptability of reconfiguration schemes and unsatisfactory optimization effects. Therefore, it is necessary to construct an evolutionary state path diagram to transform scattered historical and real-time operational data into a visualized and structured model that can characterize state correlations and temporal evolution, providing data support and a theoretical foundation for subsequent steps.

[0012] like Figure 2 As shown, step S100 in the method provided in this embodiment of the invention includes: Extract the node voltage amplitude and branch current amplitude at each sampling time from the historical operation data, and map the electrical quantity set at each time moment to an operation status node; Connect the running state nodes at adjacent sampling times in chronological order to form directed edges that represent state transitions; The frequency of each directed edge in the historical running data is counted, and the frequency is normalized to a transition probability, which is used as the weight of the corresponding directed edge. Map the current node voltage amplitude and branch current amplitude to a new operating state node, and update the directed edges and weights between the current operating state node and existing state nodes based on historical statistical patterns. Integrate all running state nodes and their corresponding directed edges and weights to generate the evolutionary state path graph.

[0013] First, the node voltage amplitude and branch current amplitude at each sampling time are extracted from the historical operational data, mapping the electrical quantity set at each time point to an operational state node. The electrical quantity set includes the node voltage amplitude set (consisting of the voltage amplitudes of all nodes) and the branch current amplitude set (consisting of the current amplitudes of all branches). Here, node voltage amplitude refers to the voltage magnitude of each power-consuming node in the distribution network, and branch current amplitude refers to the current magnitude of each power-supplying branch in the distribution network. An operational state node is an abstract representation of the electrical quantity set at a specific sampling time. Each node uniquely corresponds to the distribution network operational state at that sampling time and is the fundamental building block of the evolutionary state path diagram.

[0014] First, complete operational data for each sampling time under a preset sampling period is extracted from the historical operational database of the target distribution network, and the voltage amplitude data of all nodes and the current amplitude data of all branches are filtered out. Second, the voltage amplitudes of all nodes at the same sampling time are organized into a node voltage amplitude set, and the current amplitudes of all branches are organized into a branch current amplitude set. The two are combined to form the electrical quantity set at that time. Finally, using a data abstraction mapping method such as vector mapping, the electrical quantity set is transformed into an operational state node in the evolution state path diagram, ensuring that each sampling time corresponds to a unique operational state node.

[0015] For example, taking a 10kV target distribution network as an example, assume that the distribution network contains 5 power consumption nodes and 3 power supply branches, with power consumption nodes numbered N1-N5 and power supply branches numbered L1-L3, and a preset sampling period of 15 minutes. Extract historical sampling times. The operating data is as follows: the node voltage amplitude set is [10.2kV, 10.1kV, 10.3kV, 10.0kV, 10.2kV], corresponding to N1-N5 respectively; the branch current amplitude set is [280A, 320A, 250A], corresponding to L1-L3 respectively; the combination of the two forms... The set of electrical quantities at time [10.2, 10.1, 10.3, 10.0, 10.2, 280, 320, 250] is used; this set of electrical quantities is mapped to running state nodes in the evolution state path diagram using a vector mapping method. .

[0016] Secondly, the operating state nodes at adjacent sampling times are connected in chronological order to form directed edges representing state transitions. A directed edge is a connecting line segment with a direction attribute, used to connect operating state nodes at different times in the evolutionary state path diagram, with the direction pointing from the node at the previous sampling time to the node at the next sampling time. State transition refers to the change in the operating state of the distribution network from one sampling time to the next adjacent sampling time; the direction of the directed edge represents the temporal direction of the state transition. Based on the chronological order of the sampling times, all mapped operating state nodes are sorted. For two adjacent operating state nodes after sorting, a directed edge is drawn from the previous node to the next node, pointing from the starting point to the ending point. Each directed edge uniquely corresponds to one distribution network operating state transition at adjacent times, ensuring that all operating state nodes form a continuous state transition link in chronological order.

[0017] For example, extracting adjacent sampling times The set of electrical quantities is mapped to the running status node. In chronological order , Draw a line from point to directed edges The directed edge represents the distribution network from The running status at any time The process of state transition at any given moment; similarly, extracting... runtime status node Drawing from point to directed edges This process continues, forming a continuous state transition link.

[0018] Next, the frequency of each directed edge appearing in the historical operating data is statistically analyzed, and the frequency is normalized to a transition probability, which is then used as the weight of the corresponding directed edge. The directed edge frequency refers to the number of times a directed edge appears in all extracted historical operating data; a higher frequency indicates a more frequent state transition process. The transition probability is a value obtained after normalizing the directed edge frequency, ranging from 0 to 1, representing the probability that the distribution network will transition from the operating state corresponding to the starting node to the operating state corresponding to the ending node. The directed edge weight is used to quantify the importance of the state transition process represented by the directed edge. In this embodiment, the transition probability is used as the weight; a higher weight indicates a greater probability of state transition and a stronger correlation.

[0019] First, iterate through all directed edges generated in the historical execution data and count the frequency of each directed edge, such as... Number of times it appears The process involves several steps: first, calculating the frequency of each directed edge relative to the total frequency of all directed edges, and then normalizing the result. This ratio represents the transition probability of the corresponding directed edge. Finally, the transition probability is assigned to the corresponding directed edge as its weight, thus quantifying the correlation between state transitions.

[0020] For example, suppose we extract one year's worth of historical operating data from this 10kV distribution network, assuming a sampling frequency of 15 minutes / sample, 365 days × 24 hours × 4 samples / hour, resulting in 35,040 sampling times and corresponding to 35,039 directed edges; statistical analysis reveals that the directed edges... It appeared 280 times in historical data, and the total frequency of all directed edges was 35039 times; the transition probability = 280 / 35039 ≈ 0.008, and this value is taken as... The weight of the directed edge; It occurred 420 times, so the transition probability = 420 / 35039 ≈ 0.012. This value is taken as... The weight.

[0021] Then, the node voltage amplitude and branch current amplitude at the current moment are mapped to new operating state nodes, and the directed edges and weights between the current operating state node and existing state nodes are updated based on historical statistical patterns.

[0022] Specifically, the node voltage amplitude and branch current amplitude at the current moment are mapped to new operating state nodes, and the directed edges and weights between the current operating state node and existing state nodes are updated based on historical statistical patterns, including: Using the node voltage amplitude and branch current amplitude as characteristics, calculate the Euclidean distance between the new operating state node and each existing state node in the evolution state path diagram; The K existing state nodes with the smallest Euclidean distance are selected as the set of similar state nodes, where K is a positive integer and its value is positively correlated with the total number of nodes; In the evolutionary state path graph, directed edges are established from each similar state node to the new running state node; Based on the update rules of state transition probability in historical statistical patterns, and combined with the Euclidean distance and time decay factor, the initial weights of the newly created directed edges are calculated and assigned.

[0023] First, using the node voltage amplitude and branch current amplitude as features, the Euclidean distance between the new operating state node and each existing state node in the evolution state path graph is calculated. Euclidean distance is a commonly used metric to measure the distance between two n-dimensional vectors. It quantifies the similarity between two vectors by calculating the square root of the sum of the squares of the differences between corresponding elements; the smaller the distance, the higher the vector similarity. First, the set of node voltage amplitudes and the set of branch current amplitudes at the current moment are combined to form the feature vector at the current moment, denoted as […]. =[ , ,..., , , ,..., ],in to The node voltage amplitude, to The first step is to determine the branch current amplitude. The second step is to extract the feature vectors corresponding to all existing state nodes in the evolutionary state path graph, denoted as... =[ , ,..., , , ,..., ], where k=1,2,...,N, and N is the total number of existing state nodes; finally, the Euclidean distance between the current feature vector and the feature vector of each existing state node is calculated according to the Euclidean distance formula: Complete the calculation of all distances.

[0024] For example, the feature vector at the current time. =[10.2,10.1,10.2,10.1,10.3,290,310,260]; Existing state nodes. eigenvectors =[10.2,10.1,10.3,10.0,10.2,280,320,250], eigenvectors =[10.1,10.2,10.1,10.3,10.2,300,290,270]; Substituting these values ​​into the Euclidean distance formula yields d( , )=10.5,d( , =12.3.

[0025] Secondly, the K existing state nodes with the smallest Euclidean distance are selected as the set of similar state nodes, where K is a positive integer and its value is positively correlated with the total number of nodes. The set of similar state nodes refers to the set of existing state nodes with the highest similarity to the feature vector of the current running state node, and the running state of the nodes in the set is closest to the current state. The total number of nodes refers to the total number of existing running state nodes in the evolutionary state path graph, denoted as N. Adaptive value selection means that the value of K is dynamically adjusted according to the total number of existing state nodes N, without manual preset, ensuring that the value of K matches the total number of nodes and avoiding deviations in the selection of similar nodes due to changes in the number of nodes.

[0026] First, sort all the calculated Euclidean distances in ascending order. Second, determine the value of K based on the total number of existing state nodes N. The value of K is positively correlated with N, i.e., K = α × N, where α is an adaptive coefficient with a value ranging from 0.1 to 0.5, which can be adjusted according to actual needs. Finally, select the existing state nodes corresponding to the first K Euclidean distances after sorting to form a set of similar state nodes.

[0027] For example, if the total number of state nodes N=3 and the adaptive coefficient α=0.3, then K=3×0.3=0.9, rounded down to K=1; the node with the smallest Euclidean distance after sorting is d( , =10.5, corresponding to the existing state node Selected as a set of similar state nodes.

[0028] Next, directed edges are established in the evolutionary state path graph, pointing from each similar state node to the new operating state node. The new operating state node refers to the operating state node generated based on the mapping between the node voltage magnitude and the branch current magnitude at the current moment, denoted as... First, define the new running status node. Let be the current running state node; secondly, for each node in the set of similar state nodes, let be denoted as . Establish a path from the evolutionary state path diagram point to The directed edges are then defined; finally, it is ensured that the direction of each directed edge is consistent with the temporal direction of the state transition, that is, from the historical similar state node to the current state node.

[0029] For example, the set of similar state nodes is { The new running status node is Establish a path from the evolutionary state diagram point to directed edges The directed edge represents the distribution network from Corresponding historical state The corresponding transition process of the current state.

[0030] Furthermore, based on the update rules of state transition probabilities in historical statistical patterns, and combined with the Euclidean distance and time decay factor, the initial weights of the newly created directed edges are calculated and assigned.

[0031] Specifically, based on the update rule of state transition probability in historical statistical patterns, and combined with the Euclidean distance and time decay factor, the initial weights of newly created directed edges are calculated and assigned, including: For each newly created directed edge, calculate the reciprocal of the corresponding Euclidean distance as the basic association strength; Calculate the time decay factor, wherein the time decay factor is expressed as the result of an exponential function calculated with a preset decay constant as the base and the absolute value of the time difference between the current time and the historical sampling time corresponding to the similar state node as the exponent; Multiplying the basic correlation strength by the time decay factor yields the time-corrected correlation strength; The total association strength is obtained by summing the time-corrected association strengths of all newly created directed edges. Divide the time-corrected association strength of each newly created directed edge by the total association strength to obtain the normalized initial weight, and assign it to the corresponding directed edge.

[0032] First, for each newly created directed edge, the reciprocal of the corresponding Euclidean distance is calculated as the basic association strength. Basic association strength is a quantitative indicator characterizing the degree of state similarity between similar state nodes connected by the newly created directed edge and the currently running state node; the smaller the Euclidean distance, the greater the basic association strength. First, determine the newly created directed edges to be calculated, with each newly created directed edge corresponding to a connection between a similar state node and the currently running state node; second, extract the Euclidean distance value d corresponding to the newly created directed edge; finally, calculate the reciprocal of the Euclidean distance as the basic association strength, using the formula: =1 / d, where Based on the strength of the basic association. For example, creating a new directed edge. The corresponding Euclidean distance d = 10.5, then the basic correlation strength is... =1 / 10.5≈0.0952.

[0033] Next, the time decay factor is calculated. This time decay factor is expressed as the result of an exponential function calculated with a preset decay constant as the base and the absolute value of the time difference between the current moment and the historical sampling moment corresponding to a similar state node as the exponent. The time decay factor is a parameter used to correct the temporal correlation between the historical state and the current state. The larger the time difference, the smaller the decay factor, indicating a lower degree of influence of the historical state on the current state. The preset decay constant is determined based on a combination of statistical analysis of historical data and engineering experience, typically taking a positive number close to 1 but less than 1, used to control the rate of time decay. It can be adjusted according to the actual operating characteristics of the distribution network. The absolute value of the time difference refers to the absolute value of the time interval between the current moment and the historical sampling moment corresponding to a similar state node, reflecting the temporal difference between the historical state and the current state.

[0034] First, determine the preset attenuation constant β, such as β=0.9; second, calculate the current time. Historical sampling times corresponding to similar state nodes Absolute value of time difference: Δt = | - |; Finally, according to the formula Calculate the time decay factor λ, where λ is the time decay factor and its value ranges from 0 to 1.

[0035] For example, the preset attenuation constant β = 0.9, at the current time =January 1, 2026, 12:00, similar state node N t1 Corresponding historical sampling time = January 1, 2026, 00:00, the absolute value of the time difference Δt = 12 hours, then the time decay factor. .

[0036] Next, the basic association strength is multiplied by the time decay factor to obtain the time-corrected association strength. The time-corrected association strength is a weighted index that comprehensively considers state similarity and temporal correlation, reflecting both the similarity between the current state and historical similar states, and the influence of historical states on the current state. First, the basic association strength is obtained. With time decay factor λ; secondly, according to the formula = ×λ calculates the time-corrected association strength, where Adjust the association strength for time. For example, the basic association strength. =0.0952, time decay factor λ= Then the time-corrected correlation strength =0.0952× ≈0.0269.

[0037] Then, the time-corrected association strengths of all newly created directed edges are summed to obtain the total association strength. The total association strength is the sum of the time-corrected association strengths of all newly created directed edges, used to normalize the time-corrected association strengths of each newly created directed edge, ensuring that the sum of the weights is 1. First, the number of all newly created directed edges is counted, denoted as M; second, the time-corrected association strengths of all newly created directed edges are iterated. , ,... Finally, according to the formula Calculate the total association strength, where This represents the total association strength. For example, if only one new directed edge exists, its time-corrected association strength is... =0.0269, then the total association strength is =0.0269.

[0038] Subsequently, the time-corrected association strength of each newly created directed edge is divided by the total association strength to obtain a normalized initial weight, which is then assigned to the corresponding directed edge. The normalized initial weight refers to converting the time-corrected association strength of each newly created directed edge into a weight value ranging from 0 to 1, ensuring that the sum of the weights of all newly created directed edges is 1, facilitating subsequent visualization and application in the evolutionary state path graph. First, the time-corrected association strength of each newly created directed edge is obtained. With overall correlation strength Secondly, according to the formula Calculate the normalized initial weight of each newly created directed edge, where The initial weights are normalized; finally, the calculated normalized initial weights are assigned to the corresponding newly created directed edges, completing the weight assignment. For example, the time-corrected association strength of newly created directed edges. =0.0269, total association strength =0.0269, then the normalized initial weights =1, assign weight 1 to the directed edge. .

[0039] Finally, all operating state nodes and their corresponding directed edges and weights are integrated to generate the evolutionary state path graph. The evolutionary state path graph is a structured graphical model with operating state nodes as basic constituent units, directed edges as state transition links, and normalized weights as quantitative indicators of the correlation strength of state transitions. It is an overall map that can intuitively and quantitatively reflect the spatiotemporal evolution of the distribution network's operating state, including the correlation and temporal transition characteristics of all historical and current operating states of the distribution network.

[0040] First, all components of the evolutionary state path graph are summarized, including all running state nodes mapped from historical sampling times, new running state nodes mapped from the current time, and all generated directed edges and their corresponding normalized weights. Second, the basic layout rules of the graph are determined. Using the time axis as the horizontal axis, all running state nodes are arranged sequentially from left to right along the horizontal axis according to the order of their corresponding sampling times, with the new node at the current time placed on the far right of the horizontal axis to ensure the temporal order of the node layout. Finally, all directed edges are drawn according to the connection relationships of the nodes, strictly following the direction of the start and end points of the directed edges. The corresponding normalized weight value is clearly marked next to each directed edge. All the fully marked nodes and directed edges are integrated to form a clear and complete evolutionary state path graph.

[0041] For example, assume that historical running status nodes have been generated. , , The new running state node at the current moment History has a direction (Weight 0.008) (Weight 0.012) Create a new directed edge (Weight 1). Summarize all 4 running state nodes, 3 directed edges, and their corresponding weights; then, using time as the horizontal axis, [the following is a list of weights]. , , Arranged from left to right according to the sampling time, with the current node... Place it on the far right of the horizontal axis; finally draw it according to the connection relationship. , , Three directed edges are labeled with weight values ​​of 0.008, 0.012, and 1 respectively. After integrating and labeling all elements, an evolutionary state path diagram is generated that clearly shows the time-series transition of the historical state of the distribution network and the correlation strength between the current state and similar historical states.

[0042] In this embodiment of the invention, by constructing an evolutionary state path diagram, the dispersed historical and real-time operational data of the distribution network are transformed into a structured and visualized state correlation model, achieving an accurate characterization of the spatiotemporal evolution of the system's operational state. This solves the problem that traditional methods struggle to quantify state correlation and temporal evolution characteristics. Through state node mapping and directed edge weight quantification, the correlation between the current operational state and historically similar states can be accurately captured, providing reliable model support and data foundation for subsequent identification of potential weak points and analysis of state correlation clustering demand periods based on spatiotemporal correlation prediction data. Simultaneously, the generated evolutionary state path diagram can be dynamically updated to adapt to the time-varying nature of the distribution network's operational state, laying the foundation for the dynamic adjustment of subsequent reconfiguration and optimization schemes.

[0043] S200: Based on spatiotemporal correlation of predictive operational data, identify multiple potential weak points in the future prediction time domain where the system's capacity margin is lower than the dynamic safety threshold.

[0044] In this embodiment of the invention, based on spatiotemporally correlated predicted operational data, multiple potential weak points in the future predicted time domain where the system's carrying capacity margin is lower than the dynamic safety threshold are identified. In target distribution networks containing a high proportion of wind and solar power, the output of new energy sources is affected by the natural environment, and load demand is affected by electricity consumption behavior. Both exhibit significant spatiotemporally coupled fluctuation characteristics, and the operating states of different nodes and time periods are strongly correlated. In traditional distribution network weak point identification methods, the prediction model training lacks standardized samples and precise error constraints, the prediction data is not fused and aligned in the spatiotemporal dimensions, and the carrying capacity margin calculation does not combine real-time topology and continuous power flow methods. This easily leads to distorted prediction data and large deviations in margin calculation. At the same time, the static safety threshold cannot adapt to the dynamic changes in the operating state, ultimately resulting in missed or misjudged potential weak points, failing to provide accurate time node support for subsequent distribution network reconfiguration and optimization. Therefore, it is necessary to achieve accurate identification of potential weak points through prediction model training, prediction data integration, margin calculation, and threshold comparison.

[0045] Step S200 in the method provided in this embodiment of the invention includes: Based on the historical and real-time operation data of the target distribution network, a spatiotemporal correlation prediction model is trained, and the output load prediction data and distributed new energy prediction data are obtained. The load forecast data includes the predicted active power and reactive power values ​​for each time point and node within the future forecast time domain. The distributed renewable energy prediction data includes the predicted active power output and reactive power output of distributed renewable energy at each time and at each access node within the future prediction time domain. The training steps of the spatiotemporal correlation prediction model include: Historical operation data of each node of the distribution network and each distributed renewable energy access point are collected within a historical time range. After data cleaning and standardization, historical sample data containing multiple sets of node active load, node reactive load, renewable energy active output and renewable energy reactive output are obtained. Each set of samples is labeled with the corresponding future short-term time series real load and real output labels. Construct the network architecture of the spatiotemporal correlation prediction model, wherein the network architecture is configured as a time series prediction structure with multi-head output; The spatiotemporal correlation prediction model is trained under supervision using the historical sample data and the corresponding real load and real output labels until the average absolute percentage error between the node load prediction sequence and the real load label output by the model is lower than the first error threshold, and the root mean square error between the distributed new energy output prediction sequence and the real output label is lower than the second error threshold, thus completing the model training. The load forecast data and distributed renewable energy forecast data are spatiotemporally aligned and integrated to form spatiotemporally correlated forecast operation data. Based on the predicted operational data and current network topology associated with the spatiotemporal correlation, the continuous power flow method is used to calculate the system carrying capacity margin at each moment in the future prediction time domain. The system bearing capacity margin calculated at each time point is compared with the dynamic safety threshold at the corresponding time point. All times when the system bearing capacity margin is lower than the dynamic safety threshold are selected and recorded as the multiple potential weak points.

[0046] First, based on historical and real-time operational data of the target distribution network, a spatiotemporal correlation prediction model is trained to output load forecast data and distributed renewable energy forecast data. The spatiotemporal correlation prediction model is a time-series prediction model that simultaneously captures the spatial and temporal correlations of distribution network operational data, used to output load and renewable energy output data for the future forecast time domain.

[0047] The load forecast data includes the predicted active power and reactive power values ​​for each node at each time point in the future forecast time domain; the distributed renewable energy forecast data includes the predicted active power output and reactive power output values ​​for each access node at each time point in the future forecast time domain.

[0048] The training steps of the spatiotemporal correlation prediction model include: First, historical operating data of each node of the distribution network and each distributed renewable energy access point are collected within a historical time range. After data cleaning and standardization, historical sample data containing multiple sets of node active load, node reactive load, renewable energy active output and renewable energy reactive output are obtained. Each set of samples is labeled with the corresponding future short-term time series real load and real output labels.

[0049] Data cleaning involves removing outliers and adding missing values ​​to the collected data to eliminate the impact of monitoring errors. Data standardization transforms raw data of different dimensions and magnitudes into values ​​within a unified range, eliminating dimensional interference and ensuring the convergence of model training. Historical sample data refers to the set of input data that, after cleaning and standardization, can be directly used for model training; each set of samples corresponds to a continuous historical operational data sequence. Future short-term time series labels refer to the actual load and actual renewable energy output data matched to each set of historical sample data for a specific future period. These serve as the verification basis for supervised model training and include actual active / reactive load at real nodes and actual active / reactive energy output from real renewable sources.

[0050] First, collect all operational data from the target distribution network within its historical timeframe, with the sampling period matching the actual monitoring period of the distribution network, covering active and reactive power data from all power-consuming nodes and all distributed renewable energy access points. Second, use the 3σ criterion to remove outliers from the data and supplement missing values ​​using linear interpolation to complete data cleaning. Subsequently, use the min-max normalization method to transform the cleaned data into the [0,1] interval to complete data standardization. Finally, using standardized historical data of a continuously preset length as a set of input samples, label each set of samples with the actual load and actual renewable energy output data of the future short-term time series of a preset length, forming complete historical sample data.

[0051] For example, assuming a 10kV distribution network contains 5 power consumption nodes N1~N5 and 2 distributed renewable energy access points N3 (PV) and N5 (Wind), historical operating data of the distribution network over the past year is collected, covering the active / reactive load of N1~N5 and the active / reactive output of N3 PV and N5 Wind. The 3σ criterion is used to remove outliers in the output of N3 PV, and data from 8 missing sampling times are supplemented by linear interpolation. All data are normalized to the [0,1] interval. For example, the original range of active load of node N1 is 0~500kW, which corresponds to 0~1 after normalization. The standardized data of 24 consecutive sampling times are used as a set of input samples. The actual node load and actual renewable energy output data of the subsequent 12 sampling times are labeled for each set of samples, and finally 10,000 sets of historical sample data are generated.

[0052] Secondly, the network architecture of the spatiotemporal correlation prediction model is constructed, wherein the network architecture is configured as a time series prediction structure with multiple outputs. A time series prediction structure with multiple outputs refers to a model output layer containing multiple parallel output branches, each branch corresponding to a type of prediction target, which can simultaneously output multi-dimensional prediction results, and the output data of each branch maintains temporal synchronization.

[0053] First, the basic architecture of the model is determined. Convolutional Neural Networks (CNNs) are used to extract spatial features of the data, capturing the power coupling relationship between different nodes and different distributed renewable energy access points. Long Short-Term Memory Networks (LSTMs) are combined to extract temporal features of the data, capturing the temporal variation pattern of power, thus constructing a CNN-LSTM hybrid feature extraction architecture. Second, the model input layer is designed, with the input dimension matching the length and number of features of historical sample data. The number of features is the number of nodes × 2 + the number of distributed renewable energy access points × 2. Subsequently, a feature fusion layer is built to integrate the spatiotemporal features extracted by CNNs and LSTMs. Finally, a multi-head output layer is constructed, with two output branches corresponding to load forecasting and distributed renewable energy forecasting, respectively. Each branch is further subdivided into active power and reactive power forecasting sub-branches, forming a time series forecasting structure with multi-head outputs, thus completing the construction of the entire network architecture.

[0054] For example, a CNN-LSTM hybrid network architecture is constructed: the input layer has a dimension of 24 (time step) × 14 (feature number = 5 nodes × 2 + 2 access points × 2); the feature extraction layer consists of 1 CNN convolutional layer (3 × 3 kernels, stride 1) and 2 LSTM layers (64 hidden units) to extract spatial and temporal features respectively; the feature fusion layer integrates spatiotemporal features through a fully connected layer; the output layer is configured with dual-core multi-head output branches, the first branch is the load prediction branch, which is subdivided into active and reactive power prediction sub-branches for nodes N1-N5, and the second branch is the distributed new energy prediction branch, which is subdivided into active and reactive power prediction sub-branches for photovoltaic and wind power N3 and N5 respectively. All branches output the predicted values ​​at 12 sampling times, thus completing the construction of a time series prediction structure with multi-head output.

[0055] Furthermore, using the historical sample data and corresponding actual load and actual output labels, the spatiotemporal correlation prediction model is subjected to supervised training until the mean absolute percentage error (MAPE) between the model's output node load prediction sequence and the actual load label is lower than a first error threshold, and the root mean square error (RMSE) between the distributed renewable energy output prediction sequence and the actual output label is lower than a second error threshold, at which point model training is complete. The mean absolute percentage error (MAPE) is an indicator used to evaluate the accuracy of node load prediction, reflecting the relative deviation between the predicted and actual values; a smaller value indicates higher prediction accuracy. The root mean square error (RMSE) is an indicator used to evaluate the accuracy of distributed renewable energy output prediction, reflecting the absolute deviation between the predicted and actual values; a smaller value indicates higher prediction accuracy. The first error threshold and the second error threshold are preset criteria for terminating model training, representing the maximum allowable values ​​for MAPE and RMSE, respectively. For example, the first error threshold is 5%, and the second error threshold is 0.02. When the model prediction error is lower than the corresponding threshold, model training is considered complete.

[0056] First, historical sample data is divided into training, validation, and test sets in a 7:2:1 ratio. Second, model training parameters are set, and the Adam optimizer is used to iteratively train the model with mean squared error as the loss function. Then, after each iteration, the model performance is evaluated using the validation set, and the MAPE of the node load prediction sequence and the RMSE of the distributed renewable energy output prediction sequence and the real label are calculated. Finally, training is continuously iterated until the model's MAPE is lower than the first error threshold and the RMSE is lower than the second error threshold, at which point training is stopped and the optimal model parameters are saved. The current and recent real-time operation data of the target distribution network are preprocessed and input into the trained model, the future prediction time domain is set, and the load prediction data and distributed renewable energy prediction data are output.

[0057] For example, 10,000 sets of historical sample data are divided into 7,000 training sets, 2,000 validation sets, and 1,000 test sets. The learning rate is set to 0.001, the number of iterations to 100, the batch size to 32, and the loss function to mean squared error. The Adam optimizer is used to train the model. At the 72nd iteration, the MAPE of the node load prediction on the validation set is 4.5%, lower than the first error threshold of 5%; the RMSE of the renewable energy output prediction is 0.018, lower than the second error threshold of 0.02. Training is then stopped and the parameters are saved. The preprocessed real-time operating data of the distribution network at the current moment and the previous 23 sampling moments are input into the trained spatiotemporal correlation prediction model. The future prediction time domain is set to 24 hours. The model outputs corresponding load prediction data and distributed renewable energy prediction data. The load prediction data includes active / reactive power prediction values ​​for times N1 to N5, and the distributed renewable energy prediction data includes active / reactive power prediction values ​​for times N3 and N5.

[0058] Based on this, the load forecast data and distributed renewable energy forecast data are spatiotemporally aligned and integrated to form spatiotemporally correlated forecast operation data. Spatiotemporal alignment refers to matching and calibrating the load forecast data and distributed renewable energy forecast data in the time and spatial dimensions to ensure that the two types of forecast data at the same time and the same node / access point are time-series synchronized and spatially corresponding, without misalignment or deviation. Spatiotemporally correlated forecast operation data refers to a unified data set that integrates load and distributed renewable energy forecast information after spatiotemporal alignment and integration. This data simultaneously contains power information in both the time and spatial dimensions, and the data in each dimension maintains a strong correlation.

[0059] First, based on the time dimension, the load forecast data and distributed renewable energy forecast data are sorted according to the sampling time within the future forecast time domain to ensure that the two types of data correspond one-to-one at the same time, thus completing the time dimension alignment. Second, based on the spatial dimension, the forecast data of each distributed renewable energy access point is matched to its respective distribution network node to ensure that the load and renewable energy output data of the same node correspond spatially, thus completing the spatial dimension alignment. Finally, according to the preset format: forecast time-node / access point-active power-reactive power, the two types of forecast data after time and space alignment are integrated into a whole to form time-space related forecast operation data.

[0060] For example, for the load forecast data and distributed renewable energy forecast data output by the spatiotemporal correlation forecast model, the two types of data are first sorted according to the sampling time to ensure... The load forecast data at each time point corresponds one-to-one with the renewable energy forecast data, completing time alignment; secondly, the forecast data for N3 photovoltaic and N5 wind power are matched to their respective nodes N3 and N5, completing spatial alignment; finally, they are integrated according to a preset format, such as... -N1-220kW-80kvar, t0+1-N3-150kW-20kvar, ultimately forming the spatiotemporal correlation prediction operation data of this distribution network for the next 24 hours.

[0061] Furthermore, based on the spatiotemporally correlated predicted operating data and the current network topology, the continuous power flow method is used to calculate the system carrying capacity margin at each moment in the future prediction time domain.

[0062] Specifically, based on the spatiotemporally correlated predicted operational data and the current network topology, the continuous power flow method is used to calculate the system carrying capacity margin at each time step within the future prediction time domain, including: The load forecast data and distributed renewable energy forecast data at the corresponding time in the spatiotemporally correlated forecast operation data are used as the benchmark operation point; Keeping the current network topology and the load forecast data unchanged, the output of distributed new energy sources will be gradually increased in a preset direction and step size. The continuous power flow method is used to track system state changes until the power flow calculation detects that any branch current has reached the thermal stability limit, or any node voltage has exceeded the allowable deviation range. Record the final percentage increase that the output of distributed renewable energy can achieve at the baseline operating point, as the system carrying capacity margin at the corresponding time. By iterating through all moments within the future prediction time domain, the processes of setting the baseline operating point, increasing the output of distributed renewable energy sources, calculating continuous power flow, and recording the proportion are repeatedly executed to obtain the system carrying capacity margin at each moment within the future prediction time domain.

[0063] First, the load forecast data and distributed renewable energy forecast data at the corresponding time point in the spatiotemporally correlated forecast operation data are used as the benchmark operating point. The benchmark operating point refers to the initial operating state point of the distribution network at a certain forecast time, determined based on the spatiotemporally correlated forecast operation data. It is the sole reference benchmark for calculating the system carrying capacity margin at that time, and includes the load power of all nodes and the output power of all renewable energy access points at that time. First, the full data of a certain time point to be calculated within the future forecast time domain is extracted from the spatiotemporally correlated forecast operation data. Second, the active / reactive load forecast data of all power consumption nodes and the active / reactive output forecast data of all distributed renewable energy access points at that time are separated. Finally, the two types of data are integrated to determine the initial operating parameters of the distribution network at that time, which serve as the benchmark operating point for calculating the system carrying capacity margin at that time.

[0064] For example, extract the time to be calculated. Forecast data (12:45 PM): Active power load [250kW, 200kW, 220kW, 180kW, 280kW] and reactive power load [90kvar, 70kvar, 80kvar, 60kvar, 100kvar] for N1~N5; Active power output of N3 photovoltaic power is 350kW and reactive power output is 50kvar; Active power output of N5 wind power power is 150kW and reactive power output is 20kvar. The two types of data will be integrated as follows: The reference operating point at any given time.

[0065] Secondly, keeping the current network topology and the aforementioned load forecast data unchanged, the output of distributed renewable energy will be gradually increased according to a preset direction and step size. The current network topology refers to the current switch connection relationships and network structure determined by the remote terminal units (RTUs) on all sectional switches and tie switches of the distribution network, which monitor and report the open / closed status of each switch in real time. This forms the physical basis for power transmission. The preset direction is determined by the changing trend of the distributed renewable energy forecast data in the spatiotemporal correlation forecast data, i.e., the direction in which renewable energy output is most likely to increase, such as the natural growth direction of photovoltaic power at midday or the rising direction of wind power output at night. The preset step size refers to the percentage increase in renewable energy output each time, determined based on engineering experience or previous trial calculations, such as 1% or 0.5% of the baseline output, balancing calculation accuracy and efficiency.

[0066] First, the real-time open / closed status of all sectional switches and tie switches is obtained through RTU, the current network topology is locked, and the switch status remains unchanged during the calculation process. Second, the load forecast data at the time to be calculated is fixed, and the active / reactive load of all nodes is maintained at the baseline operating point value. Finally, according to the preset direction, the active power output is gradually increased according to the preset step size, with the renewable energy output at the baseline operating point as the initial value, and the reactive power output is adjusted synchronously according to the preset power factor of the distribution network. New renewable energy output data is generated after each increment.

[0067] For example, the current network topology is monitored by the RTU: segment switches S1~S5 are closed, and tie switches S6~S8 are closed, locking the topology unchanged; fixed. Load data for all nodes at any given time; because The preset direction is the natural growth direction of photovoltaic power output, and the preset step size is 1% of the baseline output. The initial renewable energy output is 350kW for N3 photovoltaic and 150kW for N5 wind power. After the first increment, it becomes 353.5kW for N3 photovoltaic and 151.5kW for N5 wind power. Each subsequent increment is 1%, and the reactive power output increases simultaneously at a power factor of 0.98.

[0068] Secondly, the continuous power flow method is used to track system state changes until the power flow calculation detects that any branch current has reached the thermal stability limit, or any node voltage exceeds the allowable deviation range. The continuous power flow method is a precise calculation method for distribution network operation limit analysis. Its core logic is to accurately locate the boundary state from safe operation to critical limit exceedance through gradual power adjustment and continuous state tracking. The thermal stability limit refers to the maximum long-term operating current value of the distribution network branch conductors, determined by their material, cross-sectional area, and laying method. Exceeding this value will cause the conductors to overheat and be damaged. Allowable voltage deviation range: According to industry standards, the allowable voltage deviation of a 10kV distribution network node is ±7%, i.e., 9.3kV~10.7kV. Exceeding this range will affect the normal operation of electrical equipment.

[0069] The implementation logic of the continuous power flow method is as follows: First, starting from a certain initial operating state of the distribution network (such as the reference operating point), the power injection amount (such as the output of distributed renewable energy) is slowly adjusted according to a preset step size; after each adjustment, electrical parameters such as node voltage and branch current are calculated synchronously to form a continuous state change trajectory; during the process, numerical iteration technology is used to avoid the non-convergence problem of conventional power flow calculation when approaching the limit state, and the tracking continues until critical conditions such as the branch current reaching the thermal stability limit and the node voltage exceeding the allowable deviation range are detected, and finally the power carrying capacity limit and critical operating state of the system are accurately locked.

[0070] First, the new energy output data after each increment is combined with the fixed load data and input into the continuous power flow calculation model. Second, the calculation is executed, and the voltage amplitude of all nodes and the current amplitude of all branches under this state are output. Then, the calculation results are checked: if all branch currents are ≤ thermal stability limit and all node voltages are within the allowable deviation range, the increment is continued by step size. If any index exceeds the limit, the increment and calculation are stopped immediately, and the current state is recorded as the operating limit state.

[0071] For example, for The output was continuously calculated after each 1% increment: During the first 20 increments, the branch currents L1: 320A~345A, L2: 330A~358A, both ≤360A thermal stability limit, and the node voltage 9.5kV~10.6kV within ±7% range; After the 21st increment, the output of the new energy source was 1.21 times the baseline, and the calculation results showed that the branch current L2 was 362A, exceeding the thermal stability limit. The increment was immediately stopped, and the 21st increment was determined to be the limit state, while the first 20 increments were the safe state.

[0072] Subsequently, the final percentage increase in distributed renewable energy output at the baseline operating point is recorded as the system carrying capacity margin at that corresponding moment. The system carrying capacity margin characterizes the renewable energy absorption capacity margin of the distribution network at that moment and under the current topology, and is calculated using the following formula: ,in For the system's load-bearing capacity margin, This represents the ratio of renewable energy output after the (n-1)th increment to the baseline output, where n is the number of increments that first exceed the limit. First, determine the number of increments that first exceed the limit, n. Second, extract the total renewable energy output after the (n-1)th increment and calculate its ratio to the baseline output. Finally, substitute into the formula This yields the system's carrying capacity margin at that moment. For example, the first exceedance occurs during the 21st increment (n=21), and after the 20th increment, the output of new energy sources is 1.2 times the baseline. =1.2; Substituting into the formula: λ=1.2-1=0.2, that is... The system capacity margin at any given time is 20%.

[0073] Then, by iterating through all moments within the predicted future time domain, the processes of setting the baseline operating point, increasing the output of distributed renewable energy sources, and calculating and recording continuous power flow ratios are repeated to obtain the system carrying capacity margin at each moment within the predicted future time domain. The full-time-domain system carrying capacity margin set is a dataset containing margin values ​​for all predicted moments, serving as the basis for subsequent comparison with dynamic safety thresholds and identification of potential vulnerable moments. First, all sampling moments within the predicted future time domain are arranged chronologically; second, for each moment, the above steps are repeated sequentially to complete the margin calculation for that moment; finally, the margin values ​​for all moments are integrated chronologically to form the full-time-domain margin set. For example, with a sampling frequency of 15 minutes per sampling, there are 96 moments in the 24-hour predicted future time domain. to In chronological order , Repeat the above steps at each moment, and finally integrate the margin values ​​of 20%, 19%, 18%...15% at 96 moments to form a set of system carrying capacity margins across the entire time domain.

[0074] The calculation steps for the dynamic security threshold include: Obtain the historical forecast error statistics of load forecast data and distributed renewable energy forecast data at each time point in the future forecast time domain; The historical forecast error statistical characteristics include the historical forecast average absolute percentage error of the load forecast data and the historical forecast root mean square error of the distributed renewable energy forecast data. Based on the aforementioned historical prediction error statistical characteristics, the uncertainty index of the prediction data at each time point is calculated. Obtain the existing state node in the evolution state path diagram that is most similar to the running state at each predicted time, and extract the standard deviation of the corresponding node voltage amplitude and branch current amplitude within the historical time window as historical fluctuation features. By integrating the uncertainty index with the historical fluctuation characteristics, dynamic adjustment coefficients are generated for each time point; The preset static bearing capacity safety threshold is multiplied by the dynamic adjustment coefficient to obtain the dynamic safety threshold at each time point in the future prediction time domain.

[0075] First, the historical forecast error statistical characteristics of load forecast data and distributed renewable energy forecast data at each time point within the future forecast time domain are obtained. These historical forecast error statistical characteristics include the historical forecast mean absolute percentage error of the load forecast data and the historical forecast root mean square error of the distributed renewable energy forecast data.

[0076] Among them, the historical forecast error statistics are indicators reflecting the pattern of forecast deviation, based on the forecast and actual data of the distribution network within a preset period in the past. These include the mean absolute percentage error (MAPE) of load forecasting and the root mean square error (RMSE) of distributed renewable energy forecasting. The historical mean absolute percentage error of load forecasting is the average percentage of the absolute deviation between the forecast and the actual load value in each past period, reflecting the relative deviation level of load forecasting. The historical root mean square error of distributed renewable energy forecasting is the square root of the average of the squares of the deviations between the forecast and actual renewable energy output in each past period, reflecting the absolute deviation level of renewable energy forecasting.

[0077] First, load forecast data, distributed renewable energy forecast data, and corresponding real-time operation data for the distribution network within a preset period are collected. Second, the MAPE (Modal Activity Prediction) for load forecasting and the RMSE (Renewable Energy Prediction Result) for renewable energy forecasting are calculated for each time period. Finally, the MAPE and RMSE for the same historical period at each time point within the future forecast time domain are extracted as historical forecast error statistical characteristics for that time point. For example, by collecting historical data from the past year, the following calculations are performed: The historical same-period load forecast MAPE at that time was 4.2%, and the renewable energy forecast RMSE was 0.02kW; this set of data was extracted as... Statistical characteristics of historical prediction errors at any given time.

[0078] Secondly, based on the historical forecast error statistical characteristics, the uncertainty index of the forecast data at each time point is calculated. The uncertainty index is a comprehensive indicator that quantifies the risk of deviation between future forecast data and actual operating data, integrating the error characteristics of load and renewable energy forecasts. The larger the value, the higher the forecast uncertainty. First, the min-max normalization method is used to normalize the MAPE and RMSE at each time point to the [0,1] interval. Second, load error weights and renewable energy error weights are set to ensure that the sum of the weights is 1. Finally, the uncertainty index U at each time point is calculated according to the formula U = load error weight × MAPE + renewable energy error weight × RMSE.

[0079] For example, the historical MAPE range is 2%~6%. After time-normalization, MAPE = 0.35, and historical RMSE ranges from 0.01 to 0.03 kW. After time-normalization, RMSE = 0.4; with load error weight set at 0.4 and new energy error weight at 0.6, the uncertainty index U = 0.4 × 0.35 + 0.6 × 0.4 = 0.38. Similarly, the uncertainty indexes for the remaining time periods are obtained.

[0080] Furthermore, the existing state nodes in the evolutionary state path diagram that are most similar to the operating state at each predicted time are obtained, and the standard deviations of the corresponding node voltage amplitude and branch current amplitude within the historical time window are extracted as historical fluctuation characteristics. The most similar existing state node is the historical state node in the evolutionary state path diagram that has the smallest Euclidean distance to the operating state node at the current predicted time, and its historical operating characteristics are closest to those at the predicted time. The historical time window refers to a preset fixed-length time interval, such as 24 hours or one week, used to extract historical operating data of similar nodes. The length is determined based on engineering experience. The historical fluctuation characteristics are represented by the standard deviations of the node voltage amplitude and branch current amplitude of similar nodes within the historical time window, reflecting the degree of temporal fluctuation of this type of operating state. The larger the standard deviation, the more severe the fluctuation.

[0081] First, calculate the Euclidean distance between the running state node at a certain moment in the future prediction time domain and all existing state nodes in the evolution state path diagram, and select the node with the smallest distance as the most similar state node; second, set a historical time window, and extract the voltage amplitude and branch current amplitude data of all historical nodes within the window for the similar node; finally, calculate the standard deviation of the two types of data respectively, take the average value as the historical fluctuation feature F at that moment, and use the min-max normalization method to process the original historical fluctuation feature F so that F is in the interval [0,1].

[0082] For example, In the graph of running state nodes and historical nodes in the evolution state path, the current state nodes are constantly being updated. The Euclidean distance is minimized, thus determining... Find the most similar node; set the historical time window to 24 hours, and extract... The standard deviation of the node voltage amplitude within the corresponding 24 hours is 0.03kV, and the standard deviation of the branch current amplitude is 5A; the historical fluctuation characteristic F = (0.03 + 5) / 2 ≈ 2.515. The range of F for the historical fluctuation characteristic is 0.5~5, and after normalization, F ≈ 0.448.

[0083] Furthermore, the uncertainty index and the historical fluctuation characteristics are integrated to generate dynamic adjustment coefficients for each time point. Uncertainty weights and volatility weights are assigned to the uncertainty index U and the historical fluctuation characteristics F, based on the degree of influence of the two types of characteristics on the safety threshold. For example, the uncertainty weight is 0.5, the volatility weight is 0.5, and the sum of the weights is 1. The dynamic adjustment coefficient is a coefficient calculated based on the preliminary fusion value, with a value range of (0,1]. The larger the fusion value, the smaller the adjustment coefficient, and the lower the dynamic safety threshold; conversely, the larger the adjustment coefficient, the higher the threshold. First, the uncertainty weight a and the volatility weight b are set; second, the preliminary fusion value S is calculated according to the formula S=a×U+b×F; finally, the formula... Calculate the dynamic adjustment coefficient K, where e is the natural constant, approximately 2.718. For example, setting a=0.5, b=0.5; the initial fusion value S=0.5×0.38+0.5×0.448=0.414; the dynamic adjustment coefficient... ≈0.661.

[0084] Then, the preset static bearing capacity safety threshold is multiplied by the dynamic adjustment coefficient to obtain the dynamic safety threshold at each moment in the future prediction time domain. The static bearing capacity safety threshold is a fixed threshold preset based on distribution network design standards and operating experience, such as 10% or 15%, which is the benchmark value of the dynamic safety threshold and reflects the basic safety bearing requirements of the distribution network. The dynamic safety threshold is the product of the static threshold and the dynamic adjustment coefficient, which changes dynamically with the prediction uncertainty and operational volatility at each moment, and is the ultimate basis for judging whether the system's bearing capacity margin is safe. For example, the preset static bearing capacity safety threshold is 15%; The dynamic adjustment coefficient K ≈ 0.661; therefore, the dynamic safety threshold = 0.15 × 0.661 = 0.09915 ≈ 9.9%, which is... The dynamic safety threshold at any given time is 9.9%.

[0085] Finally, the system carrying capacity margin calculated at each time point is compared with the corresponding dynamic safety threshold. All times when the system carrying capacity margin is lower than the dynamic safety threshold are selected and recorded as potential weak points. Potential weak points refer to times in the future prediction time domain where the system carrying capacity margin is lower than the corresponding dynamic safety threshold. At these times, the distribution network's renewable energy absorption capacity is insufficient, and safety risks such as branch overload and voltage exceeding limits are likely to occur, requiring reconfiguration and optimization to address these issues.

[0086] First, the calculation results for all moments within the future prediction time domain are compiled to form a three-dimensional corresponding dataset: moment - system carrying capacity margin - dynamic safety threshold, ensuring that the margin and threshold at each moment match one by one; second, the values ​​are compared moment by moment in chronological order to determine whether the system carrying capacity margin is less than the corresponding dynamic safety threshold; finally, all moments that satisfy the condition of system carrying capacity margin < dynamic safety threshold are recorded and arranged in chronological order to form a set of multiple potential weak moments.

[0087] For example, organize the data for a portion of the time period: The system's capacity margin is 20% and the dynamic safety threshold is 9.9%. The margin is greater than the threshold, indicating that the system is not at a vulnerable point. A time margin of 8% and a threshold of 10% indicate a weak time. If the margin is less than the threshold, the time is considered a weak time. A time margin of 7% and a threshold of 8.5% indicate a weak time. If the margin is less than the threshold, the time is considered a weak time. With a time margin of 13% and a threshold of 11%, the time margin exceeds the threshold, indicating non-weak moments; records are then finalized. , , Multiple times that meet the conditions are used to form a set of potential vulnerable moments for the power distribution network in the next 24 hours.

[0088] In this embodiment of the invention, standardized sample construction, targeted network architecture design, and supervised training with dual error threshold constraints ensure the prediction accuracy of the spatiotemporal correlation prediction model. The output load and renewable energy prediction data can accurately characterize the power change characteristics of the future distribution network, solving the problem of data distortion in traditional prediction methods. Spatiotemporal alignment and integration of the prediction data form unified prediction operation data with strong spatiotemporal correlation, ensuring the temporal synchronization and spatial correspondence of data in subsequent capacity margin calculations and improving the reliability of the calculation basis data. Combining the current network topology with the continuous power flow method to calculate the system capacity margin time-by-time, it can accurately capture the renewable energy absorption capacity of the distribution network at each time. Compared with traditional power flow calculation methods, it is more in line with the actual operating characteristics of the distribution network and improves the accuracy of margin calculation. By comparing the capacity margin at each time with the dynamic safety threshold to screen weak moments, the drawback of static thresholds being unable to adapt to dynamic changes in operating status is avoided. This achieves accurate and complete identification of potential weak moments, providing accurate time node support for subsequent demand period clustering and distribution network reconfiguration optimization, and effectively avoiding the operational risks of branch overload and voltage exceeding limits.

[0089] S300: Analyze the correlation and temporal proximity of the operating states corresponding to the multiple potential weak moments in the evolutionary state path diagram, and cluster them to generate one or more demand periods with internal state consistency.

[0090] In this embodiment of the invention, the correlation and temporal proximity of the operating states corresponding to the multiple potential weak points in the evolutionary state path diagram are analyzed, and clustering is used to generate one or more demand periods with internal state consistency. The multiple potential weak points are scattered independent time points. Directly refactoring the distribution network for each time point would lead to excessively frequent switching operations, exacerbating equipment losses. Simultaneously, the operating states of some weak points are highly similar and temporally adjacent, possessing the conditions to share the same optimal topology. Therefore, it is necessary to quantify the state correlation and temporal proximity to cluster the scattered weak points, generating demand periods with consistent internal states. This lays the foundation for subsequent unified topology optimization, balancing the demand for renewable energy consumption with equipment lifespan protection.

[0091] Step S300 in the method provided in this embodiment of the invention includes: Extract the running state nodes corresponding to the multiple potential weak moments in the evolutionary state path diagram, and use them as a set of states to be analyzed; Calculate the shortest path distance between any two running state nodes in the set of states to be analyzed in the evolutionary state path graph, and use it as a measure of state correlation. Calculate the absolute value of the time difference between any two running state nodes in the set of states to be analyzed, and use it as a measure of temporal proximity. Based on the state correlation metric and the temporal proximity metric, a comprehensive distance matrix is ​​constructed for all running state nodes in the set of states to be analyzed. The comprehensive distance matrix is ​​processed using a hierarchical clustering algorithm, and the running status nodes that meet the preset comprehensive distance threshold are aggregated into the same cluster; Connect the time points corresponding to the running status nodes within the same cluster in chronological order to form the required time period with internal state consistency.

[0092] First, the operational state nodes corresponding to the multiple potential weak moments in the evolutionary state path diagram are extracted as the set of states to be analyzed. The set of states to be analyzed consists of all operational state nodes corresponding to the potential weak moments. Each node uniquely corresponds to the distribution network operational state at a weak moment and is the core object of subsequent correlation and temporal proximity analysis. Operational state node mapping refers to converting the load and renewable energy forecast data at the potential weak moments into nodes in the evolutionary state path diagram according to the S100 mapping rule, ensuring a one-to-one correspondence between moments and nodes.

[0093] First, all identified potential vulnerability moments are compiled. Second, for each vulnerability moment, the set of electrical quantities for that moment is extracted from the spatiotemporally correlated predicted operational data. Finally, using the vector mapping method of S100, the corresponding operational state node is found in the evolutionary state path diagram, and all nodes are summarized to form the set of states to be analyzed. For example, the identified potential vulnerability moments are... , , , Extract the electrical quantity sets at each moment, find the corresponding operating state nodes N10, N15, N25, and N30 in the evolution state path diagram, and summarize them to form the state set to be analyzed {N10, N15, N25, N30}.

[0094] Secondly, the shortest path distance between any two running state nodes in the set of states to be analyzed is calculated in the evolutionary state path graph, serving as a measure of state correlation. The shortest path distance refers to the length of the connected path with the minimum weight between two running state nodes in the evolutionary state path graph; that is, the minimum sum of the weights of all directed edges on the path. The smaller the distance, the stronger the correlation and the higher the similarity between the running states corresponding to the two nodes. The state correlation measure quantifies the similarity between two weak moments of running states, based on the topological connections and edge weight calculations of the evolutionary state path graph, reflecting the inherent correlation between states.

[0095] First, the connection relationships and directed edge weights of each node in the evolutionary state path graph are defined. Second, Dijkstra's algorithm is used to calculate the shortest path distance between any two nodes in the set of states to be analyzed. Finally, the calculation results are used as the state correlation metric for the corresponding two nodes and recorded to form a state correlation matrix.

[0096] For example, the set of states to be analyzed is {N10, N15, N25, N30}. The weights of the directed edges between each node in the evolutionary state path graph are known: N10-N15 weight 0.8, N10-N25 weight 1.5, N15-N25 weight 0.6, N15-N30 weight 1.2, and N25-N30 weight 0.7. Using Dijkstra's algorithm, the shortest path distance between N10 and N15 is calculated to be 0.8, between N10 and N25 to be 1.4, between N15 and N25 to be 0.6, and between N25 and N30 to be 0.7, forming a state association matrix.

[0097] Next, the absolute value of the time difference between any two corresponding moments in the set of states to be analyzed is calculated as a temporal proximity metric. Temporal proximity is an indicator that quantifies the closeness of two vulnerability moments in the time dimension, expressed as the absolute time difference between moments; the smaller the time difference, the higher the temporal proximity. The absolute value of the time difference refers to the time interval between two potential vulnerability moments, with the unit consistent with the sampling period. First, the potential vulnerability moments corresponding to each node in the set of states to be analyzed are extracted; second, the absolute value of the time difference between any two corresponding moments is calculated; finally, this absolute value is used as the temporal proximity metric for the two nodes, and recorded to form a temporal proximity matrix.

[0098] For example, the corresponding times for each node are: N10 (14:00), N15 (15:15), N25 (17:45), and N30 (19:00); the absolute values ​​of the time differences are calculated: N10 and N15 are 75 minutes, N10 and N25 are 225 minutes, N15 and N25 are 150 minutes, and N25 and N30 are 75 minutes, forming a temporal proximity matrix.

[0099] Furthermore, based on the state correlation metric and the temporal proximity metric, a comprehensive distance matrix is ​​constructed for all running state nodes in the set of states to be analyzed.

[0100] Specifically, based on the state correlation metric and the temporal proximity metric, a comprehensive distance matrix is ​​constructed for all running state nodes in the set of states to be analyzed, including: The state correlation metric and the temporal proximity metric are normalized, and correlation weights and temporal weights are configured respectively. For any two running state nodes in the set of states to be analyzed, the comprehensive distance value between the two running state nodes is calculated by weighted fusion. Traverse all running state node pairs in the set of states to be analyzed, calculate the comprehensive distance value for each node pair, and fill all comprehensive distance values ​​into a symmetric matrix, setting the diagonal elements of the matrix to 0, to form the comprehensive distance matrix.

[0101] First, the state correlation metric and the temporal proximity metric are normalized, and correlation weights and temporal weights are configured respectively. The correlation weight reflects the importance of operational state similarity to clustering, and the temporal weight reflects the importance of temporal proximity to clustering, ensuring that the sum of the weights is 1. First, the state correlation matrix of the set of states to be analyzed is extracted, i.e., the shortest path distance of all node pairs, and the temporal proximity matrix is ​​extracted, i.e., the absolute value of the time difference of all node pairs. Second, min-max normalization is performed on both matrices respectively, transforming all metric values ​​to the [0,1] interval. Finally, the correlation weight and temporal weight are configured based on engineering experience, ensuring that their sum is 1.

[0102] For example, the set of states to be analyzed is {N10, N15, N25, N30}: In the original state correlation metric, the shortest path distance between N10 and N15 is 0.8, which is 0.44 after normalization; the shortest path distance between N25 and N30 is 0.7, which is 0.39 after normalization, and the same applies to other node pairs; In the original temporal proximity metric, the time difference between N10 and N15 is 75 minutes, which is 0.25 after normalization; the time difference between N25 and N30 is 75 minutes, which is 0.25 after normalization, and the same applies to other node pairs; The correlation weight is configured as 0.6 and the temporal weight as 0.4, and the sum of the two is 1.

[0103] Secondly, for any two running state nodes in the set of states to be analyzed, a weighted fusion calculation is performed to determine the comprehensive distance between the two running state nodes. The comprehensive distance value = normalized state correlation × correlation weight + normalized temporal proximity × temporal weight; a larger value indicates a greater comprehensive difference. First, the normalized state correlation value and normalized temporal proximity value corresponding to any two nodes are extracted; second, the two values ​​are multiplied by their corresponding weights respectively; finally, the two products are added together to obtain the comprehensive distance between the nodes.

[0104] For example, to calculate the combined distance between N10 and N15: extract their normalized state correlation value of 0.44 and normalized temporal proximity value of 0.25, and the combined distance value = 0.44 × 0.6 + 0.25 × 0.4 = 0.364; the same logic applies to other node pairs, such as the combined distance value between N25 and N30 being 0.334, and between N10 and N25 being 0.768, and the remaining node pairs can be calculated using the same logic.

[0105] Then, all running state node pairs in the set of states to be analyzed are traversed, the comprehensive distance value of each node pair is calculated, and all comprehensive distance values ​​are filled into a symmetric matrix, with the diagonal elements of the matrix set to 0, to form the comprehensive distance matrix. A node pair refers to any combination of two different nodes in the set of states to be analyzed. Because the comprehensive distance value is symmetric, it only needs to be calculated once. The comprehensive distance matrix is ​​a symmetric matrix with the number of rows / columns equal to the number of nodes to be analyzed. The elements are the comprehensive distance values ​​of the node pairs. Nodes are indistinguishable from themselves, and the diagonal elements are 0, fully representing the comprehensive differences of all node pairs. First, an empty symmetric matrix is ​​initialized, with all elements initialized to 0. Second, all unordered node pairs are traversed, and after calculating the comprehensive distance value, it is simultaneously filled into the two corresponding symmetric positions in the matrix. Finally, the diagonal elements are kept at 0 to complete the matrix construction.

[0106] For example, if there are 4 nodes to be analyzed, initialize a 4×4 empty matrix: fill the matrix with the combined distance value of N10 and N15, 0.364, in the first row and second column and the second row and first column of the matrix; fill the matrix with the combined distance value of N25 and N30, 0.334, in the third row and fourth column and the fourth row and third column of the matrix; fill the matrix with the combined distance values ​​of other node pairs in the same way, and keep the four diagonal elements of the matrix at 0, thus forming a complete combined distance matrix.

[0107] Subsequently, a hierarchical clustering algorithm is used to process the comprehensive distance matrix, aggregating running nodes that meet a preset comprehensive distance threshold into the same cluster. Hierarchical clustering is an unsupervised clustering algorithm that continuously merges the closest clusters, where each node initially forms a cluster, gradually building a clustering tree until the distance between clusters exceeds the preset threshold. It is suitable for node clustering based on distance matrices. The preset comprehensive distance threshold is a critical value for determining whether two clusters should be merged, determined by engineering experience or trial calculations. Clusters are merged if the distance is less than the threshold, ensuring that nodes within the same cluster have small comprehensive differences and strong internal consistency.

[0108] First, initialize the clustering state: each running node is an independent cluster; second, calculate the comprehensive distance between all clusters using the average distance between clusters; third, merge the two clusters with the smallest comprehensive distance that is less than a preset threshold and update the cluster set; finally, repeat the inter-cluster distance calculation and merging steps until the comprehensive distance between all clusters is greater than the preset threshold, stop clustering, and obtain the final cluster set.

[0109] For example, the set of states to be analyzed is {N10, N15, N25, N30}, the preset comprehensive distance threshold is 0.37, and the comprehensive distance values ​​of each node pair are: N10-N15=0.364, N10-N25=0.768, N10-N30=1.0, N15-N25=0.398, N15-N30=0.732, N25-N30=0.334. Initialize four independent clusters: {N10}, {N15}, {N25}, and {N30}. The first calculation of inter-cluster distances: the combined distance between {N25} and {N30} is the smallest (0.334) and less than 0.37, so they are merged into a new cluster {N25, N30}. The second calculation of inter-cluster distances: the combined distance between {N10} and {N15} is the smallest (0.364) and less than 0.37, so they are merged into a new cluster {N10, N15}. The third calculation of inter-cluster distances: the average inter-cluster distance between the new clusters {N10, N15} and {N25, N30} is (0.768 + 0.732 + 0.398 + 0.334) / 4 ≈ 0.558, which is greater than 0.37, so they cannot be merged. Finally, two clusters are obtained: {N10, N15} and {N25, N30}.

[0110] Finally, the moments corresponding to the operating states of nodes within the same cluster are connected in chronological order to form a demand period with internal state consistency. A demand period refers to a continuous or semi-continuous time period formed by connecting all potential weak moments within the same cluster in chronological order. The operating states of all weak moments within this period are highly consistent and can share the same optimal topology. Internal state consistency means that the operating states of all weak moments within the demand period are strongly correlated, temporally adjacent, and have small differences in parameters such as node voltage and branch current, thus adapting to the operating requirements of the same topology.

[0111] First, for each cluster, extract the potential weak moments corresponding to all running status nodes within the cluster; second, sort all moments in chronological order; finally, form continuous demand periods with the earliest moment as the starting point and the latest moment as the ending point. If there are small intervals between moments within a cluster, they are still merged into one demand period to ensure the integrity of the time period.

[0112] For example, the nodes of the two clusters correspond to the following times: In cluster {N10, N15}, N10 corresponds to time 14:00 and N15 corresponds to time 15:15, which are sorted by time to form the demand period [14:00, 15:15]; In cluster {N25, N30}, N25 corresponds to time 17:45 and N30 corresponds to time 19:00, which are sorted by time to form the demand period [17:45, 19:00]; The weak moments in both periods have strong internal state consistency, which ultimately generates two demand periods.

[0113] In this embodiment of the invention, by quantifying state correlation and temporal proximity, scientific clustering of dispersed potential weak moments is achieved, avoiding the blindness of traditional segmented optimization and ensuring strong consistency of operating states within the demand period. By employing a hierarchical clustering algorithm combined with a comprehensive distance matrix, the time period range adapted to the same topology can be accurately divided, providing a clear target for subsequent targeted topology optimization and reducing unnecessary switching operations. The generated demand period not only covers all potential weak moments but also reduces the number of reconstructions through internal consistency screening, balancing the demand for new energy consumption and equipment life protection, and laying an efficient foundation for the formulation of multi-time period reconstruction scheduling schemes.

[0114] S400: For each demand period, determine the unique optimal network topology and the corresponding switching operation instructions, and integrate the instructions from each period to generate a multi-period reconfiguration scheduling scheme.

[0115] In this embodiment of the invention, a unique optimal network topology and corresponding switching operation instructions are determined for each demand period, and the instructions from each period are integrated to generate a multi-period reconfiguration scheduling scheme. Potentially weak points in different demand periods have internal state consistency, but the operating characteristics of each period differ, requiring targeted matching of the optimal network topology. Simultaneously, frequent switching operations accelerate equipment aging, necessitating a minimum number of operations to balance the safety of renewable energy consumption with equipment lifespan protection. Therefore, a scientific multi-period reconfiguration scheduling scheme is formed by defining decision variables, clarifying optimization objectives and constraints, using algorithms to search for the optimal solution, and generating integrated instructions, ensuring the safe and efficient operation of the distribution network throughout the entire prediction time domain.

[0116] Step S400 in the method provided in this embodiment of the invention includes: The decision variable is defined as the set of target states to be adjusted for all controllable switches in the target distribution network during the demand period. The optimization objective is set as minimizing the total change of the target state set relative to the current state, which is used to characterize the number of switching operations; Construct composite constraints, wherein the composite constraints require that, based on the network topology formed by the target state set, at each potential weak point in the demand period, the calculated result of the system carrying capacity margin is not lower than the dynamic safety threshold at the corresponding point, and all branch currents and node voltages meet the preset safe operation limits. A hybrid particle swarm optimization algorithm is used to search under the composite constraints to obtain the optimal set of target switching states for each demand period. The optimal set of target switch states for each demand period is compared with the current switch state to generate a switch operation instruction containing the switch identifier to be operated and the operation type. Based on the chronological order of each demand period on the time axis, the switching operation instructions corresponding to all demand periods are sorted and integrated to form a multi-period reconstruction scheduling scheme covering the entire prediction time domain.

[0117] First, the decision variable is defined as the set of target states to be adjusted for all controllable switches in the target distribution network during the demand period. The decision variable is the core variable characterizing the distribution network topology adjustment, focusing on the target operating states of all controllable switches, and is the core object of the optimization solution. A controllable switch refers to a switching device in the distribution network whose open / closed state can be changed through remote control or manual operation; it is the key carrier for adjusting the network topology. The target state set refers to the set of expected operating states of all controllable switches during the demand period. States are represented in binary, such as 1 representing closed and 0 representing open, denoted as . , where n is the number of controllable switches.

[0118] First, compile a list of all controllable switches in the target distribution network, clarifying the switch identifiers and current operating status. Second, define the target state variable for each controllable switch, using 1 to represent a closed target state and 0 to represent an open target state. Finally, integrate the target state variables of all switches to form a decision variable X, which uniquely corresponds to a network topology.

[0119] To illustrate with an example continuing the 10kV distribution network example, the list of controllable switches consists of sectionalizing switches S1~S5 and tie switches S6~S8, with the current state being {S1=1,S2=1,S3=1,S4=1,S5=1,S6=1,S7=1,S8=1}. A decision variable is defined: if X={1,1,1,0,1,1,0,1}, it indicates that the target state for S4 and S7 is open, while the rest are closed, corresponding to a specific network topology.

[0120] Secondly, the optimization objective is set as minimizing the total change of the target state set relative to the current state, which characterizes the number of switching operations. Minimizing the number of switching operations balances equipment losses and topology optimization effects. The total change is the sum of the absolute values ​​of the differences between the elements of the target state set and the current state set; each non-zero difference corresponds to one switching operation. First, the current state set of all controllable switches is extracted. Secondly, define the optimization objective function: Finally, the function value f(X) represents the number of switching operations, and the solution process aims to minimize this value. For example, in the current state... ={1,1,1,1,1,1,1,1}, a candidate target state Xa={1,1,1,0,1,1,0,1} has a total change of 2, which corresponds to 2 switching operations, S4 is open and S7 is open; another candidate state Xb={1,0,1,0,1,0,0,1} has a total change of 3. Since 2<3, Xa is closer to the optimization target.

[0121] Secondly, a composite constraint condition is constructed. This composite constraint condition requires that, based on the target state set, the network topology formed during the demand period, at each potential weak point, has a calculated system carrying capacity margin that is not lower than the corresponding dynamic safety threshold, and all branch currents and node voltages meet preset safe operation limits. The composite constraint condition is a set of multiple constraints ensuring the safety and feasibility of the reconstructed network topology, including carrying capacity margin constraints, branch current constraints, and node voltage constraints; none can be omitted. The carrying capacity margin constraint means that at all potential weak points within the demand period, the system carrying capacity margin of the reconstructed topology must be ≥ the corresponding dynamic safety threshold, ensuring that the renewable energy absorption capacity meets standards. The safe operation limits refer to the preset electrical safety boundaries of the distribution network. The branch current limit is the conductor thermal stability limit, such as 360A, and the node voltage limit is ±7% of the national standard, corresponding to 9.3kV~10.7kV for a 10kV distribution network.

[0122] Specifically, three types of constraints are constructed to form a composite constraint set: Bearing capacity margin constraint: For each potential weak point t within the demand period, the system bearing capacity margin λ(t) under the reconstructed topology is calculated, requiring λ(t) ≥ λ. dynamic (t), λ dynamic (t) represents the dynamic safety threshold at time t; Branch current constraint: For all branches k under the reconstructed topology, the current I at any time during the demand period. k (t)≤I lim (k), I lim (k) represents the thermal stability limit of branch k; Node voltage constraint: For all nodes m in the reconstructed topology, the voltage U at any time during the demand period. m (t)∈[U min U max ].

[0123] For example, the demand period [14:00, 15:15] contains two potentially weak moments t1=14:00 (λ dynamic =9.9%), t2=15:15 (λ) dynamic =10.5%); the reconstructed topology must satisfy the following conditions: λ(t1)≥9.9% at time t1, λ(t2)≥10.5% at time t2; all branch currents≤360A; all node voltages are between 9.3kV and 10.7kV, otherwise the topology is an infeasible solution.

[0124] Furthermore, a hybrid particle swarm optimization algorithm is employed to search under the aforementioned composite constraints, obtaining the optimal set of target switching states for each demand period. The hybrid particle swarm optimization algorithm is an optimization algorithm that integrates particle swarm optimization (PSO) with a loop search strategy. It retains the global search capability of PSO while improving local optimization efficiency through loop search, thus adapting to the high-dimensional nonlinear solution requirements of distribution network reconfiguration.

[0125] First, initialize the particle swarm: each particle corresponds to a set of decision variables, namely the set of target states for switching, and set parameters such as the number of particles, the maximum number of iterations, and the learning factor. Second, calculate the fitness value of each particle; the fitness of a feasible solution is the number of switching operations, and an infeasible solution is given a penalty value. Third, iterate to find a better feasible solution through particle position updates and loop search optimization. Finally, when the maximum number of iterations is reached or the fitness value converges, select the set of states corresponding to the particle with the best fitness as the optimal set of target states for switching during this demand period.

[0126] For example, for the demand period [14:00, 15:15], the initial particle swarm contains 50 particles, corresponding to 50 sets of on / off states, with a maximum of 100 iterations; after iterative calculation, the target state set X corresponding to a certain particle is calculated. opt1 ={1,1,1,0,1,1,0,1}, which satisfies all composite constraints and has the fewest operations among all feasible solutions, is determined as the optimal target state set for the current time period.

[0127] Then, the optimal target switch state set corresponding to each demand period is compared with the current switch state to generate a switch operation instruction containing the identifier of the switch to be operated and the operation type. The switch operation instruction is a specific instruction that guides the actual switch operation, containing the identifier of the switch to be operated and the operation type, and serves as the basis for topology reconfiguration. Operation type determination: If the current switch state is 1 and the target state is 0, the operation type is open; if the current state is 0 and the target state is 1, the operation type is closed; if the states are the same, no operation is performed.

[0128] First, extract the optimal set of target switching states X for the demand period. opt With the current state set X curr Secondly, the target state of each switch is compared with its current state one by one, and switches with inconsistent states are filtered out. Finally, an operation type is assigned to each switch to be operated, and a switch operation instruction is generated in the format of switch identifier-operation type.

[0129] For example, the current state X curr ={1,1,1,1,1,1,1,1}, the optimal state X for the demand period [14:00, 15:15]. opt1={1,1,1,0,1,1,0,1}; After comparison, the states of S4 and S7 are inconsistent, generating switch operation commands: S4-Open, S7-Open; The optimal state X for another demand period [17:45, 19:00] is... opt2 ={1,1,1,0,1,0,0,1}, after comparing with the current state, generates the instruction: S6 - Disconnect. S4 and S7 remain disconnected and do not need to be repeated.

[0130] Finally, according to the chronological order of each demand period on the timeline, the switching operation instructions corresponding to all demand periods are sorted and integrated to form a multi-period reconstruction scheduling scheme covering the entire forecast time domain. This multi-period reconstruction scheduling scheme is a unified scheduling scheme covering the entire future forecast time domain, integrating the switching operation instructions of all demand periods, clearly defining the time sequence, switch identifier, and operation type, facilitating actual execution. First, the chronological order of all demand periods within the future forecast time domain is analyzed; second, the switching operation instructions for each period are extracted according to the time sequence, eliminating duplicate operations; finally, all instructions are integrated into a unified multi-period reconstruction scheduling scheme in the format of operation time-switch identifier-operation type, ensuring the scheme is clear and executable.

[0131] For example, the time sequence of two demand periods within the future forecast time domain is [14:00, 15:15] and [17:45, 19:00]. The integrated instructions are: 14:00 execute [S4-disconnect, S7-disconnect], 17:45 execute [S6-disconnect]. If a third demand period [21:00, 22:30] subsequently exists, its optimal instruction is [S7-close], which is then added to the end of the scheme, ultimately forming a scheduling scheme covering the entire time domain and clearly defining the operational content at each time point. Distribution network dispatchers can directly carry out switching operations and network topology adjustments based on this multi-time-period reconfiguration scheduling scheme, ensuring the safe and efficient operation of the distribution network throughout the entire forecast time domain, while maximizing the absorption of distributed renewable energy.

[0132] In this embodiment of the invention, minimizing the number of switching operations is the optimization objective, effectively reducing the frequency of equipment operation, lowering the risk of equipment aging, and balancing the needs of topology optimization and equipment protection. The composite constraints comprehensively cover the requirements of bearing capacity margin, branch current, and node voltage safety, ensuring that the reconstructed topology operates safely and stably during the demand period and meets the needs of new energy consumption. The hybrid particle swarm optimization algorithm improves the search efficiency and accuracy of the optimal topology, adapts to the high-dimensional nonlinear solution difficulties of distribution network reconstruction, and ensures that a feasible optimal solution is found within a reasonable time. The operation instructions of multiple time periods are integrated in chronological order to form a unified scheduling scheme, avoiding operation conflicts between time periods, improving the operability of the scheme, and providing clear and reliable execution basis for distribution network dispatchers.

[0133] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing distribution network reconfiguration to enhance the carrying capacity of distributed renewable energy sources, characterized in that, The method includes: Based on the historical and real-time operation data of the target distribution network, an evolutionary state path diagram describing the spatiotemporal evolution of the system's operating state is constructed. Based on spatiotemporal correlation of predictive operational data, identify multiple potential weak points in the future prediction time domain where the system's carrying capacity margin is lower than the dynamic safety threshold. The correlation and temporal proximity of the operating states corresponding to the multiple potential weak moments in the evolutionary state path diagram are analyzed, and clustering is used to generate one or more demand periods with internal state consistency. For each demand period, a unique optimal network topology and corresponding switching operation instructions are determined, and instructions from each period are integrated to generate a multi-period reconfiguration scheduling scheme.

2. The distribution network reconfiguration optimization method for improving the carrying capacity of distributed new energy sources according to claim 1, characterized in that, Based on historical and real-time operational data of the target distribution network, an evolutionary state path diagram describing the spatiotemporal evolution of the system's operational state is constructed, including: Extract the node voltage amplitude and branch current amplitude at each sampling time from the historical operation data, and map the electrical quantity set at each time moment to an operation status node; Connect the running state nodes at adjacent sampling times in chronological order to form directed edges that represent state transitions; The frequency of each directed edge in the historical running data is counted, and the frequency is normalized to a transition probability, which is used as the weight of the corresponding directed edge. Map the current node voltage amplitude and branch current amplitude to a new operating state node, and update the directed edges and weights between the current operating state node and existing state nodes based on historical statistical patterns. Integrate all running state nodes and their corresponding directed edges and weights to generate the evolutionary state path graph.

3. The distribution network reconfiguration optimization method for improving the carrying capacity of distributed new energy sources according to claim 2, characterized in that, Map the current node voltage magnitude and branch current magnitude to a new operating state node, and update the directed edges and weights between the current operating state node and existing state nodes based on historical statistical patterns, including: Using the node voltage amplitude and branch current amplitude as characteristics, calculate the Euclidean distance between the new operating state node and each existing state node in the evolution state path diagram; The K existing state nodes with the smallest Euclidean distance are selected as the set of similar state nodes, where K is a positive integer and its value is positively correlated with the total number of nodes; In the evolutionary state path graph, directed edges are established from each similar state node to the new running state node; Based on the update rules of state transition probability in historical statistical patterns, and combined with the Euclidean distance and time decay factor, the initial weights of the newly created directed edges are calculated and assigned.

4. The distribution network reconfiguration optimization method for improving the carrying capacity of distributed new energy sources according to claim 3, characterized in that, Based on the update rule of state transition probability in historical statistical patterns, and combined with the Euclidean distance and time decay factor, the initial weights of newly created directed edges are calculated and assigned, including: For each newly created directed edge, calculate the reciprocal of the corresponding Euclidean distance as the basic association strength; Calculate the time decay factor, wherein the time decay factor is expressed as the result of an exponential function calculated with a preset decay constant as the base and the absolute value of the time difference between the current time and the historical sampling time corresponding to the similar state node as the exponent; Multiplying the basic correlation strength by the time decay factor yields the time-corrected correlation strength; The total association strength is obtained by summing the time-corrected association strengths of all newly created directed edges. Divide the time-corrected association strength of each newly created directed edge by the total association strength to obtain the normalized initial weight, and assign it to the corresponding directed edge.

5. The distribution network reconfiguration optimization method for improving the carrying capacity of distributed new energy sources according to claim 1, characterized in that, Based on spatiotemporal correlation of predictive operational data, multiple potential vulnerability moments in the future prediction time domain where the system's capacity margin falls below the dynamic safety threshold are identified, including: Based on the historical and real-time operation data of the target distribution network, a spatiotemporal correlation prediction model is trained, and the output load prediction data and distributed new energy prediction data are obtained. The load forecast data includes the predicted active power and reactive power values ​​for each time point and node within the future forecast time domain. The distributed renewable energy prediction data includes the predicted active power output and reactive power output of distributed renewable energy at each time and at each access node within the future prediction time domain. The training steps of the spatiotemporal correlation prediction model include: Historical operation data of each node of the distribution network and each distributed renewable energy access point are collected within a historical time range. After data cleaning and standardization, historical sample data containing multiple sets of node active load, node reactive load, renewable energy active output and renewable energy reactive output are obtained. Each set of samples is labeled with the corresponding future short-term time series real load and real output labels. Construct the network architecture of the spatiotemporal correlation prediction model, wherein the network architecture is configured as a time series prediction structure with multi-head output; The spatiotemporal correlation prediction model is trained under supervision using the historical sample data and the corresponding real load and real output labels until the average absolute percentage error between the node load prediction sequence and the real load label output by the model is lower than the first error threshold, and the root mean square error between the distributed new energy output prediction sequence and the real output label is lower than the second error threshold, thus completing the model training. The load forecast data and distributed renewable energy forecast data are spatiotemporally aligned and integrated to form spatiotemporally correlated forecast operation data. Based on the predicted operational data and current network topology associated with the spatiotemporal correlation, the continuous power flow method is used to calculate the system carrying capacity margin at each moment in the future prediction time domain. The system bearing capacity margin calculated at each time point is compared with the dynamic safety threshold at the corresponding time point. All times when the system bearing capacity margin is lower than the dynamic safety threshold are selected and recorded as the multiple potential weak points.

6. The distribution network reconfiguration optimization method for improving the carrying capacity of distributed new energy sources according to claim 5, characterized in that, Based on the predicted operational data and current network topology associated with the aforementioned spatiotemporal correlation, the system carrying capacity margin at each time step within the future prediction time domain is calculated using the continuous power flow method, including: The load forecast data and distributed renewable energy forecast data at the corresponding time in the spatiotemporally correlated forecast operation data are used as the benchmark operation point; Keeping the current network topology and the load forecast data unchanged, the output of distributed new energy sources will be gradually increased in a preset direction and step size. The continuous power flow method is used to track system state changes until the power flow calculation detects that any branch current has reached the thermal stability limit, or any node voltage has exceeded the allowable deviation range. Record the final percentage increase that the output of distributed renewable energy can achieve at the baseline operating point, as the system carrying capacity margin at the corresponding time. By iterating through all moments within the future prediction time domain, the processes of setting the baseline operating point, increasing the output of distributed renewable energy sources, calculating continuous power flow, and recording the proportion are repeatedly executed to obtain the system carrying capacity margin at each moment within the future prediction time domain.

7. The distribution network reconfiguration optimization method for improving the carrying capacity of distributed new energy sources according to claim 5, characterized in that, The calculation steps for the dynamic security threshold include: Obtain the historical forecast error statistics of load forecast data and distributed renewable energy forecast data at each time point in the future forecast time domain; The historical forecast error statistical characteristics include the historical forecast average absolute percentage error of the load forecast data and the historical forecast root mean square error of the distributed renewable energy forecast data. Based on the aforementioned historical prediction error statistical characteristics, the uncertainty index of the prediction data at each time point is calculated. Obtain the existing state node in the evolution state path diagram that is most similar to the running state at each predicted time, and extract the standard deviation of the corresponding node voltage amplitude and branch current amplitude within the historical time window as historical fluctuation features. By integrating the uncertainty index with the historical fluctuation characteristics, dynamic adjustment coefficients are generated for each time point; The preset static bearing capacity safety threshold is multiplied by the dynamic adjustment coefficient to obtain the dynamic safety threshold at each time point in the future prediction time domain.

8. The distribution network reconfiguration optimization method for improving the carrying capacity of distributed new energy sources according to claim 1, characterized in that, Analyze the correlation and temporal proximity of the operating states corresponding to the multiple potential weak points in the evolutionary state path diagram, and cluster them to generate one or more demand periods with internal state consistency, including: Extract the running state nodes corresponding to the multiple potential weak moments in the evolutionary state path diagram, and use them as a set of states to be analyzed; Calculate the shortest path distance between any two running state nodes in the set of states to be analyzed in the evolutionary state path graph, and use it as a measure of state correlation. Calculate the absolute value of the time difference between any two running state nodes in the set of states to be analyzed, and use it as a measure of temporal proximity. Based on the state correlation metric and the temporal proximity metric, a comprehensive distance matrix is ​​constructed for all running state nodes in the set of states to be analyzed. The comprehensive distance matrix is ​​processed using a hierarchical clustering algorithm, and the running status nodes that meet the preset comprehensive distance threshold are aggregated into the same cluster; Connect the time points corresponding to the running status nodes within the same cluster in chronological order to form the required time period with internal state consistency.

9. The distribution network reconfiguration optimization method for improving the carrying capacity of distributed new energy sources according to claim 8, characterized in that, Based on the state correlation metric and the temporal proximity metric, a comprehensive distance matrix is ​​constructed for all running state nodes in the set of states to be analyzed, including: The state correlation metric and the temporal proximity metric are normalized, and correlation weights and temporal weights are configured respectively. For any two running state nodes in the set of states to be analyzed, the comprehensive distance value between the two running state nodes is calculated by weighted fusion. Traverse all running state node pairs in the set of states to be analyzed, calculate the comprehensive distance value for each node pair, and fill all comprehensive distance values ​​into a symmetric matrix, setting the diagonal elements of the matrix to 0, to form the comprehensive distance matrix.

10. The distribution network reconfiguration optimization method for improving the carrying capacity of distributed new energy sources according to claim 1, characterized in that, For each demand period, a unique optimal network topology and corresponding switching operation instructions are determined, and instructions from each period are integrated to generate a multi-period reconfiguration scheduling scheme, including: The decision variable is defined as the set of target states to be adjusted for all controllable switches in the target distribution network during the demand period. The optimization objective is set as minimizing the total change of the target state set relative to the current state, which is used to characterize the number of switching operations; Construct composite constraints, wherein the composite constraints require that, based on the network topology formed by the target state set, at each potential weak point in the demand period, the calculated result of the system carrying capacity margin is not lower than the dynamic safety threshold at the corresponding point, and all branch currents and node voltages meet the preset safe operation limits. A hybrid particle swarm optimization algorithm is used to search under the composite constraints to obtain the optimal set of target switching states for each demand period. The optimal set of target switch states for each demand period is compared with the current switch state to generate a switch operation instruction containing the switch identifier to be operated and the operation type. Based on the chronological order of each demand period on the time axis, the switching operation instructions corresponding to all demand periods are sorted and integrated to form a multi-period reconstruction scheduling scheme covering the entire prediction time domain.

Images