Power distribution network first-aid repair scheduling method and system based on fault prediction

Abstract

The invention relates to a power distribution network first-aid repair scheduling method and system based on fault prediction, and solves the problem of low efficiency of existing first-aid repair scheduling. Comprising the steps of collecting a power distribution network basic data time sequence in a historical period in a dispatching area and preprocessing to obtain multivariable time sequence data; dividing the scheduling area into a plurality of grids; predicting the user power failure quantity at the current moment based on the historical multivariable time sequence data in each grid and the corresponding user power failure quantity; on the basis of the multivariable time sequence data and the predicted user power failure number at the current moment, an allocation center site selection optimization model with the minimum allocation center first-aid repair efficiency and the comprehensive influence of the power failure loss as the target is constructed, and the optimal allocation center site selection position is obtained through solving; and on the basis of the optimal allocation center site selection position, the obtained current fault point list and the urgent repair team parameters, constructing an urgent repair team scheduling multi-target model by taking the minimum urgent repair path distance and repair time as targets, and solving to obtain an optimal urgent repair scheduling route of the power distribution network. And scientific first-aid repair scheduling is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution network security assurance, and in particular to a power distribution network emergency repair scheduling method and system based on fault prediction. Background Art

[0002] With the continuous expansion of power grids and the growth of electricity demand, the reliability and resilience of distribution networks have become core concerns. Frequent natural disasters (such as typhoons, floods, and earthquakes) can lead to power outages and losses for users. Efficiently completing emergency repairs and restoring power supply after a disaster has become a major challenge in power grid operations. Research on optimizing distribution network emergency repair and dispatch in natural disaster scenarios is not only crucial for the safe operation of power systems but also plays a crucial role in improving user satisfaction and reducing power outage losses.

[0003] Current research on distribution network emergency repair and dispatching focuses primarily on traditional optimization models and algorithms, such as optimizing dispatch center location and repair task allocation based on linear programming and genetic algorithms. These methods have limitations in practical applications: First, traditional optimization methods are insufficiently capable of solving complex multi-objective, multi-constraint problems, making it difficult to simultaneously balance repair efficiency and losses. Second, existing technologies often overlook the complex impact of disasters on distribution network operations and load fluctuations, lacking accurate predictions of key variables such as the number of user outages and load losses. Furthermore, existing methods are limited in their ability to handle missing data in dynamic disaster scenarios, which can easily lead to biased predictions and dispatch results. Summary of the Invention

[0004] In view of the above analysis, an embodiment of the present invention aims to provide a distribution network emergency repair scheduling method based on fault prediction, so as to solve the technical problem of low efficiency of distribution network fault emergency repair scheduling in existing methods.

[0005] The present invention provides a distribution network emergency repair scheduling method based on fault prediction, comprising the following steps:

[0006] Collect the basic data time series of the distribution network in a certain historical period in the dispatching area and preprocess it to obtain multivariate time series data;

[0007] Dividing the dispatch area into a plurality of grids; predicting the number of user power outages at the current moment based on the historical multivariate time series data and the corresponding number of user power outages in each grid;

[0008] Based on the multivariate time series data and the predicted number of users experiencing power outages at the current moment, a dispatch center location optimization model is constructed with the goal of minimizing the dispatch center's emergency repair efficiency and the combined impact of power outage losses, and the optimal dispatch center location is obtained.

[0009] Based on the optimal allocation center location, the current fault point list and repair team parameters obtained, a multi-objective model for repair team scheduling is constructed with the goal of minimizing the repair path distance and repair time, and the optimal repair scheduling route for the distribution network is obtained by solving the model.

[0010] Furthermore, the historical multivariate time series data in each grid is used as the feature variable, and the corresponding number of user power outages is used as the target variable to train a random forest to obtain the predicted power outage user data at the current moment; including:

[0011] For each grid's historical multivariate time series data set, multiple multivariate time series data are randomly extracted with replacement to form multiple subsample sets, each of which is used to train a decision tree.

[0012] For each decision tree, a portion of the features in the feature vector is randomly selected, and the Gini index of each feature vector is calculated at each node. The feature vector with the smallest Gini index is selected as the split feature, which is used as the basis for splitting the decision tree node;

[0013] Each decision tree splits its nodes until the decision tree stops growing.

[0014] The mean value of the target variable corresponding to all the feature vectors reaching the tree leaf nodes of the decision tree is used as the predicted value of the number of power outages for users in the grid at the current moment;

[0015] The conditions for stopping the growth of the decision tree include that the number of decision tree nodes is less than a specified value, the Gini index is less than a preset threshold, the depth of the decision tree reaches a specified value, or all multivariate time series data samples have been used up.

[0016] Furthermore, the allocation center location optimization model includes a location objective function and constraints;

[0017] The site selection objective function is as follows:

[0018]

[0019] Among them, f1 is the comprehensive impact of minimizing the emergency repair efficiency of the dispatch center and the power outage loss; F i is the predicted failure probability of line i; M L is the total number of faulty lines; N B (i) is the total number of nodes without load caused by fault line i; w ij is the level weight of load loss node j caused by line i failure; L ij is the load loss amount caused by node j during unit power outage time; t ijis the repair waiting time for the failure of load-loss node j caused by line i, that is, the time spent on the journey from the temporary dispatch center to the faulty line i, which is obtained based on the distance from the fault point to the dispatch center and the speed of the repair team; U ij,grid is the predicted number of power outages for users in the grid; grid is the priority weight of the grid; α1 and α2 are the weight coefficients of the dispatch center's emergency repair efficiency and power outage loss;

[0020] The site selection constraints include geographical location constraints, coverage constraints and load coverage rate constraints.

[0021] Furthermore, the improved grey wolf-shark algorithm is used to perform global and local searches to obtain the optimal location of the optimal allocation center, including:

[0022] Initialize the size of the gray wolf population and the white shark population, and set the maximum number of iterations;

[0023] Set the initial values α1 and α2 of the dispatch center's emergency repair efficiency and power outage loss weight coefficients;

[0024] Randomly generate a set of candidate locations for the allocation center in all grid spaces and ensure that all candidate locations for the allocation center meet the geographical location constraints;

[0025] Generate random numbers to initialize a group of gray wolf populations, each gray wolf represents a candidate location for the allocation center;

[0026] For each candidate location of the allocation center, the value of the site selection objective function is calculated as the fitness value of the gray wolf;

[0027] The three best candidate allocation center positions corresponding to the fitness values are selected and marked as α, β, and δ gray wolves as the best candidate allocation center positions in the current population; other gray wolves update their positions based on the positions of α, β, and δ gray wolves;

[0028] When the maximum number of iterations is reached, the three optimal candidate allocation center locations are searched locally using the white shark algorithm, introducing α, β, and δ white sharks. The white sharks move toward the optimal candidate allocation center location, and at the optimal candidate allocation center location, a decision is made based on the strategy selection threshold to determine whether to introduce random perturbations. The location of the gray wolf population is then updated using the positions of the three white sharks α, β, and δ, exploring the local optimum until the location selection objective function converges.

[0029] The location with the smallest site selection objective function value and satisfying the site selection constraints is selected as the optimal allocation center location.

[0030] Furthermore, the multi-objective model for dispatching the emergency repair team includes a dispatching objective function and constraints;

[0031] The scheduling objective function is as follows:

[0032]

[0033] Among them, M1 is the total repair path distance; M2 is the overall repair time of the region; p and q are damaged points, k is the power repair team; x pqk is a decision variable, and its value of 0 or 1 indicates that the power repair team k has not arrived at the damaged point p or has arrived at the damaged point q; c pq is the transportation distance from damaged point p to damaged point q; v is the speed of the repair team; R is the repair team set, which is the set of all available repair teams; V is the set of all damaged points; n is the number of damaged points; K is the number of repair teams; is the repair time of the damaged point p;

[0034] The dispatch constraints include repair team capacity constraints, task allocation constraints, path continuity constraints, distribution network radial radiation topology constraints, node voltage constraints and distribution flow constraints.

[0035] Furthermore, the improved Pareto-Zebra algorithm is used to solve the multi-objective model of repair team scheduling and obtain the optimal repair scheduling route, including:

[0036] Based on the optimal allocation center location and the list of fault points, a set of emergency repair routes is randomly generated, the zebra population size is set, and the zebra population is initialized;

[0037] Determine the weights of the two objectives: repair path distance and repair time;

[0038] Calculate the emergency repair path distance and repair time of each route, and calculate the objective function value by weighting the emergency repair path distance and repair time based on the weights of the two objectives as the fitness value of each zebra;

[0039] Perform a non-dominated sort on all routes in the zebra population and determine the Pareto rank of each route. If route A is not inferior to route B in terms of both repair path distance and repair time, and is superior to route B in at least one of its objectives, then route A dominates route B. All routes that are not dominated by other routes are assigned to the first non-dominated rank on the Pareto front.

[0040] After removing the first-level routes, repeat the non-dominated sorting on the remaining routes to determine the second non-dominated level, and so on;

[0041] Select the num routes with the lowest non-dominated level as the pioneer zebra routes; other zebra routes move closer to the pioneer zebra routes to simulate the foraging behavior of zebras;

[0042] Introduce random perturbations near each zebra's current route to explore new solutions; ensure that the new route satisfies the constraints;

[0043] Re-perform non-dominated sorting on the updated zebra population until the scheduling objective function converges;

[0044] Output the final set of routes on the Pareto front, where the routes in the set are optimal in terms of total repair path distance and overall regional repair time, thereby obtaining a set of Pareto optimal solutions;

[0045] Calculating the emergency repair path distance and the repair time weighted value F for each route in the Pareto optimal solution;

[0046] The optimal solution with the smallest F among the Pareto optimal solutions is selected as the final optimal emergency repair scheduling route.

[0047] Furthermore, the preprocessing includes deduplication, data correction, continuity testing, outlier processing and missing value filling;

[0048] The missing value filling includes:

[0049] Filling missing values in the time series of the basic data of the distribution network with zero values to obtain original data and initial time series data filled with zero values;

[0050] Generating a mask vector of the initial time series data;

[0051] Inputting the initial time series data and the corresponding mask vector into the trained BiLSTM-GAN model to obtain the multivariate time series data;

[0052] The BiLSTM-GAN model is constructed based on the Bi-LSTM and GAN networks, including a generator and a discriminator;

[0053] The generator uses a bidirectional LSTM network to capture the contextual dependency of the initial time series data from the forward and reverse directions to generate fitted values of missing values; retains the original data of the distribution network basic data time series sequence, and uses the fitted values of the missing values to replace the zero values of the missing positions to obtain the distribution network basic data time series sequence filled with missing values;

[0054] The discriminator forces the time series sequence of the basic data of the distribution network to be close to the real data, thereby obtaining multivariate time series data.

[0055] Furthermore, the generator uses a bidirectional LSTM network to capture the contextual dependencies of the initial time series data in the forward and reverse directions to generate fitting values of missing values, including:

[0056] The time series data X with missing data and its corresponding mask vector m are input into the forward LSTM to obtain the forward hidden state

[0057] Input the reverse sequence of X and its reverse mask vector into the reverse LSTM to obtain the backward hidden state

[0058] based on and Get the forward output respectively and backward output

[0059] based on and Calculate the fitted value x for the missing value uni,i ,as follows:

[0060]

[0061] in, are the forward and backward combination factors respectively.

[0062] Furthermore, the basic data of the distribution network includes distribution network data, distribution network fault data, distribution network load data, natural disaster data, user distribution data, user power outage quantity data, historical emergency repair data, geographical environment data and infrastructure data.

[0063] The present invention also provides a distribution network emergency repair and dispatching system based on fault prediction, comprising:

[0064] The basic data acquisition and preprocessing module is used to collect the basic data time series of the distribution network within a certain historical period in the dispatching area and preprocess it to obtain multivariate time series data;

[0065] A user power outage number prediction module is used to divide the scheduling area into multiple grids; based on the historical multivariate time series data and the corresponding user power outage number in each grid, the number of user power outages at the current moment is predicted;

[0066] An allocation center location selection module is configured to construct an allocation center location optimization model based on the multivariate time series data and the predicted number of users experiencing power outages at the current moment, with the goal of minimizing the allocation center's emergency repair efficiency and the combined impact of power outage losses, and to solve for the optimal allocation center location.

[0067] The emergency repair scheduling optimization module is used to build an emergency repair team scheduling multi-objective model based on the optimal allocation center location, the current fault point list and emergency repair team parameters obtained, with the goal of minimizing the emergency repair path distance and repair time, and solve it to obtain the optimal emergency repair scheduling route for the distribution network.

[0068] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0069] 1. The BiLSTM-GAN model constructed in the present invention integrates the generative adversarial network GAN and the bidirectional long short-term memory network Bi-LSTM. The generator of the GAN network uses bidirectional LSTM to capture the contextual dependencies of the initial time series data in both forward and reverse directions to generate fitting values for missing values. The discriminator optimizes the data quality so that the filled data is closer to the real data, effectively solving the problem that the existing methods are insufficient in processing missing data. In natural disaster scenarios, distribution network data often has large-scale missing data due to communication interruptions or equipment damage, resulting in the inability of existing methods to accurately reflect the actual operating status of the distribution network. The bidirectional LSTM network makes full use of the before and after information of the time series to improve the prediction accuracy of missing values of multivariate data such as load loss and the number of user power outages. The BiLSTM-GAN model can not only make up for the shortcomings of traditional methods in filling missing data, but also improve the accuracy of fault prediction in disaster scenarios, and provide a data basis for the subsequent allocation center site selection and repair team scheduling optimization;

[0070] 2. Existing methods for selecting the site of the allocation center have deficiencies in coping with disaster variability and multi-objective optimization, and it is difficult to take into account both emergency repair efficiency and power outage losses at the same time. The present invention optimizes the site selection of the allocation center based on the improved Gray Wolf-White Shark algorithm. The White Shark algorithm has a strong local search capability, and the Gray Wolf algorithm can enhance the global search performance. The combination of the two significantly improves the comprehensive solution capability of the algorithm, and effectively overcomes the problem that existing methods are prone to falling into local optimality and slow convergence speed. The present invention constructs an allocation center site selection optimization model with the goal of minimizing the comprehensive impact of the emergency repair time of the allocation center's response to the fault point and the power outage load loss. It can provide more reasonable and efficient emergency repair allocation center site selection in complex disaster scenarios, greatly improve emergency repair scheduling efficiency, and reduce power grid operation losses;

[0071] 3. In response to the problem that the existing methods have a single scheduling strategy for emergency repair teams and ignore task priorities, the present invention constructs a multi-objective model for emergency repair team scheduling based on multi-objective optimization, with the goal of minimizing the emergency repair path distance and repair time, and comprehensively considering constraints such as the number of teams, load capacity, and driving paths to achieve a reasonable allocation of emergency repair resources. The improved Pareto-Zebra algorithm is used to solve the scheduling problem. By combining the fast global search capability of the zebra optimization algorithm and the multi-objective balance processing capability of the Pareto optimization, the emergency repair team is given priority to serve high-priority areas with large power outage impacts. At the same time, the scheduling method makes full use of the path optimization characteristics of the improved Pareto-Zebra algorithm, significantly shortening the total driving distance of the emergency repair team from the allocation center to the damaged point, thereby improving the efficiency of emergency repairs;

[0072] 4. The existing allocation center site selection and repair team scheduling are often carried out independently, lacking linkage optimization, resulting in uneven resource allocation or low efficiency in the execution of repair tasks. The present invention combines the allocation center site selection location with the repair team scheduling through linkage optimization, and determines the optimal site selection method through the allocation center site selection optimization model, providing efficient resource support for the repair team scheduling; then, based on the optimal allocation center site selection location, as well as the current fault point list and repair team parameters obtained, a multi-objective model for repair team scheduling is constructed with the goal of minimizing the repair path distance and repair time, and the solution is performed to obtain the optimal repair scheduling route for the distribution network. This method dynamically balances the coupling relationship between the allocation center layout and the repair team allocation, avoids the global suboptimal problem caused by separate optimization, and provides an innovative solution for the efficient recovery of the distribution network under disasters.

[0073] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0075] Figure 1 This is a flow chart of a distribution network emergency repair and scheduling method based on fault prediction according to an embodiment of the present invention;

[0076] Figure 2 The figure is a schematic diagram of a distribution network emergency repair and dispatching system module based on fault prediction in an embodiment of the present invention. DETAILED DESCRIPTION

[0077] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0078] Example 1:

[0079] The present invention proposes a distribution network emergency repair scheduling method and system based on fault prediction. It combines GAN (Generative Adversarial Networks) and Bi-LSTM (Bi-directional Long Short-Term Memory) to make high-precision predictions of missing locations in the acquired distribution network basic data time series, obtains fitted data to fill the missing locations, and achieves accurate prediction of the number of faults and user power outages in disaster scenarios. It optimizes the location of the allocation center by improving the Gray Wolf-White Shark optimal algorithm, and constructs an allocation center location optimization model with the goal of minimizing the combined impact of the allocation center's emergency repair efficiency and power outage losses to obtain the optimal allocation center location. For emergency repair team scheduling, a multi-objective model for emergency repair team scheduling is constructed with the goal of minimizing the emergency repair path distance and repair time. The improved Pareto-Zebra algorithm is used to solve and obtain the optimal emergency repair scheduling route for the distribution network.

[0080] A specific embodiment of the present invention, as Figure 1 As shown, a distribution network emergency repair scheduling method based on fault prediction is disclosed, comprising the following steps:

[0081] Step S1: collecting the basic data time series of the distribution network within a certain historical period in the dispatching area and preprocessing it to obtain multivariate time series data;

[0082] Step S2: Divide the dispatch area into a plurality of grids; predict the number of user power outages at the current moment based on the historical multivariate time series data and the corresponding number of user power outages in each grid;

[0083] Step S3: Based on the multivariate time series data and the predicted number of users experiencing power outages at the current moment, construct an allocation center location optimization model with the goal of minimizing the allocation center's emergency repair efficiency and the combined impact of power outage losses, and solve the model to obtain the optimal allocation center location;

[0084] Step S4: Based on the optimal allocation center location, the current fault point list and repair team parameters obtained, a multi-objective model for repair team scheduling is constructed with the goal of minimizing the repair path distance and repair time, and the optimal repair scheduling route for the distribution network is obtained by solving the model.

[0085] Step S1 includes steps S11-S12.

[0086] Step S11: Collect the time series of basic data of the distribution network within a certain historical period in the dispatching area.

[0087] The basic data of the distribution network includes distribution network data, distribution network fault data, distribution network load data, natural disaster data, user distribution data, user power outage data, historical emergency repair data, geographical environment data and infrastructure data.

[0088] Exemplarily, three years of historical basic data within the scheduling area are collected.

[0089] (1) Distribution network data, including:

[0090] Equipment data: including equipment records and equipment technical parameters;

[0091] Line data: number of lines, total length, transmission capacity, transmission distance, insulation status, conductor condition;

[0092] Device geographic coordinates and geographic environment.

[0093] (2) Distribution network fault data, including:

[0094] Fault type: various fault types occurring in the distribution network, such as equipment failure, line power outage, switch failure, overload, etc.;

[0095] Fault occurrence time: the specific moment when the fault occurred;

[0096] Fault duration: the duration from the occurrence of a fault to its repair;

[0097] Fault location: the specific location where the fault occurred (transformer, switchgear, line, etc.);

[0098] Fault analysis: specific symptoms and manifestations of the fault;

[0099] Fault repair: dispatch of fault repair teams;

[0100] Failure rate: In the dispatching area, it represents the probability or frequency of failures per unit time. It reflects the ratio of the number of failures that occurred in a distribution network within a specific time period to the total operating time. The failure rate can be calculated based on the statistical number of failures and the failure time.

[0101] (3) Distribution network load data, including:

[0102] Distribution network load curve: records the load changes of the distribution network in each time period;

[0103] Load distribution: Record the load distribution in different areas of the distribution network.

[0104] Line length: The total length of each line.

[0105] (4) Natural disaster data, including:

[0106] Disaster occurrence time: the specific moment when the natural disaster occurred;

[0107] Disaster duration: the duration from the occurrence to the end of a natural disaster;

[0108] Disaster type: type of disaster, such as earthquake, typhoon, flood, lightning, etc.;

[0109] Disaster intensity: Disaster intensity data, such as wind speed, precipitation, etc.;

[0110] Time and location of disaster: the time and area where the disaster occurred;

[0111] Disaster prediction data: including weather forecasts before disasters occur, historical disaster patterns, disaster warning data, etc.

[0112] (5) User distribution data, including:

[0113] User location information: user's precise geographic location and detailed address information;

[0114] User types: Residential, commercial, industrial users, and users such as hospitals or schools that have high requirements for power supply reliability;

[0115] Number of users: number of regional users and user density;

[0116] User electricity usage behavior patterns: electricity usage time patterns and electricity consumption trends;

[0117] User power supply reliability requirements: reliability level, impact of power outages;

[0118] User access method: access point, access capacity.

[0119] (6) Data on the number of power outages for users, including:

[0120] Power outage time: the specific time when the user's power outage occurs;

[0121] Number of power outages for users: number of power outages for users due to faults or disasters;

[0122] Power outage load loss: power load loss caused by power outage;

[0123] Power outage area distribution: geographical distribution of users affected by the power outage;

[0124] Restoration time: the specific time when power supply is restored;

[0125] User feedback and complaints: complaint records, satisfaction survey records.

[0126] (7) Historical emergency repair data, including:

[0127] Emergency repair response time: the time from when the dispatch center receives the fault report to when the emergency repair team dispatches, the time the emergency repair team arrives at the fault site, the deployment of the emergency repair team, and the time from when the fault report is received to when the emergency repair begins;

[0128] Usage of emergency repair resources: equipment, personnel, materials and other resources used in emergency repairs;

[0129] Completion status of emergency repair tasks: whether the emergency repair tasks are completed on time and the quality of completion.

[0130] (8) Geographical environment data, including:

[0131] Topographic and geomorphic data;

[0132] Weather data: temperature, humidity, precipitation, maximum average wind speed, temperature difference and other meteorological factors.

[0133] (9) Infrastructure data, including:

[0134] Distribution transformer quantity: number of distribution transformers;

[0135] Power distribution equipment status: records the status, maintenance records, aging conditions, etc. of distribution network equipment.

[0136] The basic data collected from the distribution network is multiple time series data. Different data sets are aligned according to the unified collection time frequency (such as hourly), and missing time points are marked as to be filled.

[0137] Step S12: pre-process the distribution network basic data time series to obtain multivariate time series data.

[0138] The preprocessing includes deduplication, data correction, continuity test, outlier processing and missing value filling;

[0139] The missing value filling includes:

[0140] Filling missing values in the time series of the basic data of the distribution network with zero values to obtain original data and initial time series data filled with zero values;

[0141] Generating a mask vector of the initial time series data;

[0142] Inputting the initial time series data and the corresponding mask vector into the trained BiLSTM-GAN model to obtain the multivariate time series data;

[0143] The BiLSTM-GAN model is constructed based on the Bi-LSTM and GAN networks, including a generator and a discriminator;

[0144] The generator uses a bidirectional LSTM network to capture the contextual dependency of the initial time series data from the forward and reverse directions to generate fitted values of missing values; retains the original data of the distribution network basic data time series sequence, and uses the fitted values of the missing values to replace the zero values of the missing positions to obtain the distribution network basic data time series sequence filled with missing values;

[0145] The discriminator forces the time series sequence of the basic data of the distribution network to be close to the real data, thereby obtaining multivariate time series data.

[0146] Use MD5 to deduplicate data, and retain only one copy of the basic data of the distribution network with the same MD5 code; data correction includes format correction, spelling correction, and logic correction, which can be automated using data cleaning tools or scripts; check whether the timestamps of the time series are continuous; identify and process outliers in the data, including retention, deletion, and modification. If the outlier reflects the actual situation, retain it and mark it.

[0147] The BiLSTM-GAN model in the present invention is used to process missing values in the time series of basic data of the distribution network, especially for the prediction of highly intermittent and random data characteristics.

[0148] Fill the missing values in the distribution network basic data time series with zero values to obtain the original data and the initial time series data filled with zero values; for each initial time series data, generate the corresponding mask vector m, and the element m in m i =1 means data i exists, m i = 0 indicates that data i is missing. Through the mask vector, the BiLSTM-GAN model identifies missing values in the initial time series data and provides a basis for subsequent filling and prediction.

[0149] The generator uses a bidirectional LSTM network to capture the contextual dependencies of the initial time series data in the forward and reverse directions to generate fitting values for missing values, including:

[0150] The time series data X with missing data and its corresponding mask vector m are input into the forward LSTM to obtain the forward hidden state

[0151] Input the reverse sequence of X and its reverse mask vector into the reverse LSTM to obtain the backward hidden state

[0152] based on and Get the forward output respectively and backward output

[0153] based on and Calculate the fitted value x for the missing valueuni,i ,as follows:

[0154]

[0155] in, are the forward and backward combination factors respectively.

[0156] Bi-LSTM is an extension of the standard LSTM. When processing time series data, it can simultaneously utilize past and future information, thereby improving prediction accuracy. The forward LSTM predicts the current value based on historical data from the beginning of the time series data to the current state. The backward LSTM backtracks the current value based on future data from the end of the time series data to the current state. Combining the output values from both directions provides a more complete understanding of the current moment.

[0157] The GAN network consists of a generator and a discriminator. The generator generates missing values based on the input data and compares them with the real data. The generator's goal is to make it impossible for the discriminator to distinguish between generated values and real values. The generator is trained by minimizing the probability that the discriminator correctly identifies false data. The discriminator distinguishes between real and generated values. Training the discriminator strengthens the generator's performance, making the values it generates for missing values closer to real data.

[0158] The generator of the GAN network uses Bi-LSTM to capture the contextual dependencies of the initial time series data from both forward and backward directions. The fitting value of each missing position is generated from two directions: the forward output is generated from its initial time series data; the backward output is generated from the reverse initial time series data. The initial time series data X and the mask vector m are used as inputs of the generator to obtain the fitting data x of the missing position. uni .

[0159] After preprocessing the basic time series of the distribution network, the multivariate time series data is obtained as follows:

[0160]

[0161] in, is the preprocessed distribution network basic data time series sequence, which is multivariate time series data, where the missing positions are filled with the fitted values predicted by the generator; x is the initial time series data, including missing values; m is a mask vector used to identify which time points in the data are missing; ⊙ is the dot multiplication operation, x uni Fitted values for missing values generated by the generator based on Bi-LSTM;

[0162] x⊙m represents the non-missing part of the initial time series data; x uni ⊙(1-m) indicates that the missing values in the initial time series data are filled with the fitted values predicted by the generator.

[0163] Bi-LSTM is an improvement on the standard RNN. It improves the learning effect of time series data through bidirectional training, especially when processing complex time series data. Bidirectional LSTM considers both past and future information at the same time and is more effective in predicting data features with high intermittent and random characteristics. For a Bi-LSTM neural network unit, the hidden state h of the current period is i , which is expressed as follows:

[0164] h i =σ(W h h i-1 +U h x i +b n ) Formula (2)

[0165] Where σ is the sigmoid function; W h is the weight matrix between the hidden state of the previous period and the current hidden state; U h is the weight matrix between the input data and the current hidden state; b n is the bias term of the current hidden state; h i-1 is the hidden state of the previous period.

[0166] The output of the bidirectional recurrent layer of Bi-LSTM consists of two parts: the forward hidden state and the backward hidden state from the forward and backward LSTMs, respectively. and is the time decay factor added to the hidden state calculation, representing the confidence of the forward and backward prediction values,

[0167] Forward error correction δ i (f) and the backward error correction δ i (b) The higher the value, the greater the current error. and The smaller the value of . and confidence factor as follows:

[0168]

[0169] in, is the hidden state of the forward LSTM network at time i; is the weight matrix of the forward LSTM; is the hidden state at the previous moment; is the predicted value of the forward LSTM at time i; are the bias term and confidence factor of the forward LSTM respectively; and To calculate the forward confidence factor The weights and biases of the backward hidden state and confidence factor as follows:

[0170]

[0171] in, is the hidden state of the backward LSTM network at time i; is the weight matrix of backward LSTM; is the forward hidden state at the previous moment; is the predicted value of the backward LSTM at time i; is the bias term of the backward LSTM; is the confidence factor of the backward LSTM; and To calculate the backward confidence factor weights and biases.

[0172] Each period combines the results of the forward layer and the reverse layer as the output of the bidirectional LSTM unit, and the GAN generator generates the target data through the output of the recurrent layer. as follows:

[0173]

[0174] in, is the generated value of the forward LSTM at time i.

[0175] Input the reverse sequence and its reverse mask vector into the reverse LSTM to obtain the backward generated value as follows:

[0176]

[0177] in, is the generated value of the backward LSTM at time i.

[0178] Therefore, the generated data of the generator is calculated from the forward and backward generated values:

[0179]

[0180] in, are two combined factors,

[0181] The discriminator is trained to maximize the correct classification of actual real values and generated values, while the generator is trained to minimize the probability that the discriminator correctly identifies the fake component.

[0182] During the BiLSTM-GAN model training process, the following four losses are defined to evaluate the model filling and prediction effects.

[0183] During the generator training process, the MAE (Mean Absolute Error) metric is used to define the MAE value of the actual time series data and the corresponding generated time series data as the mask reconstruction loss LossR. Since the actual missing values are unknown during training, the mask reconstruction loss is calculated for the non-missing parts of the initial time series data. LossR is expressed as follows:

[0184]

[0185] Among them, x is the initial time series data, x uni,i is the generated time series data, μ is the number of missing data.

[0186] The prediction results of the bidirectional LSTM in both directions are preferably consistent, so the consistency loss LossC is defined to represent the difference between the forward and backward generation results, as shown in the following formula. To ensure that the magnitude of LossR and LossC are consistent, the MAE is also used to calculate LossC, as shown below:

[0187]

[0188] in, is the predicted data of missing values generated by the forward LSTM and backward LSTM.

[0189] During the discriminator training process, binary cross entropy loss LossD is used as follows:

[0190]

[0191] represents the classification probability that the true value is identified as true, Represents the probability that a generated value is identified as false.

[0192] The training goal of the generator is to make it difficult for the discriminator to identify fakes, that is, to minimize The value of , defines LossG to represent the classification loss of the generator as follows:

[0193]

[0194] During training, decreasing LossG is consistent with the discriminator being less likely to classify fake instances as fake. The loss function of the generator of the GAN model is as follows:

[0195] Loss GAN(G)=αLossR+βLossC+LossG formula (14)

[0196] α and β are custom weights used to avoid numerical imbalance among the three losses. α controls the weight of the mask reconstruction loss LossR in the total loss, while β controls the weight of the consistency loss LossC in the total loss. The generator classification loss LossG directly measures the overall difference between the generated data and the real data.

[0197] After obtaining the trained BiLSTM-GAN model, the generator fills the missing initial time series data based on the fitting data at the missing positions of the Bi-LSTM output to obtain multivariate time series data.

[0198] The function of step S1 is to collect the basic data time series of the distribution network within a certain historical period in the dispatching area, and generate complete multivariate time series data through preprocessing, providing a reliable data basis for subsequent analysis and modeling.

[0199] Step S2 includes steps S21-S22.

[0200] Step S21: Divide the scheduling area into multiple grids.

[0201] Based on geographical conditions, distribution network distribution, user distribution and the impact of natural disasters, a dispatching area is divided into multiple small grids.

[0202] (1) Quantification of geographical conditions

[0203] Elevation variation does not exceed 100 meters. Within the same grid, terrain elevation does not vary by more than 100 meters to ensure consistency of geographical conditions. The slope standard deviation of terrain complexity is less than 5%, indicating relatively small variations in terrain slope within the same grid. River density variation does not exceed 10%, indicating relatively uniform distribution of river density within the same grid, with no significant variations in river density.

[0204] (2) Quantification of distribution network

[0205] The distribution line density should not vary by more than 20%. Distribution lines within the same grid should be relatively evenly distributed, with no significant variations in line density. The number of transformers within each grid should be between 5 and 10. Ensure the appropriate number of transformers within each grid. Fault rate distribution should be within a 10% variance, ensuring a relatively even distribution within the same grid.

[0206] (3) Quantification of user distribution

[0207] The difference in user density does not exceed 15%; the total number of power outage events in each grid is between 100 and 500; and the difference in distribution load distribution is ±10%.

[0208] (4) Quantification of the impact of natural disasters

[0209] The wind speed difference does not exceed 5m / s; the precipitation variation range is within 10mm; the earthquake intensity distribution does not exceed one magnitude unit.

[0210] Meshing principles:

[0211] (1) No overlap: There is no overlapping area between each grid, ensuring that each location belongs to one grid;

[0212] (2) Full coverage: All grids together cover the entire dispatching area, leaving no areas unoccupied.

[0213] By dividing a dispatching area into multiple small grids and ensuring that the conditions within each grid are relatively consistent, the data within each grid can be analyzed more carefully, thereby achieving more accurate fault prediction and emergency repair dispatch.

[0214] Step S22: Based on the historical multivariate time series data in each grid and the corresponding number of user power outages, use random forest to predict the number of user power outages at the current moment.

[0215] The historical multivariate time series data in each grid is used as the feature variable, and the corresponding number of user power outages is used as the target variable. A random forest is trained to obtain the predicted power outage user data at the current moment; this includes:

[0216] For each grid's historical multivariate time series data set, multiple multivariate time series data are randomly extracted with replacement to form multiple subsample sets, each of which is used to train a decision tree.

[0217] For each decision tree, a portion of the features in the feature vector is randomly selected, and the Gini index of each feature vector is calculated at each node. The feature vector with the smallest Gini index is selected as the split feature, which is used as the basis for splitting the decision tree node;

[0218] Each decision tree splits its nodes until the decision tree stops growing.

[0219] The mean value of the target variable corresponding to all the feature vectors reaching the tree leaf nodes of the decision tree is used as the predicted value of the number of power outages for users in the grid at the current moment;

[0220] The conditions for stopping the growth of the decision tree include that the number of decision tree nodes is less than a specified value, the Gini index is less than a preset threshold, the depth of the decision tree reaches a specified value, or all multivariate time series data samples have been used up.

[0221] The conditions for stopping the growth of the decision tree include that the number of decision tree nodes is less than a specified value, the Gini index is less than a preset threshold, the depth of the decision tree reaches a specified value, or all multivariate time series data samples have been used up.

[0222] If the classification feature is "grid load", the condition for splitting the node may be "grid load > 100"; if the classification feature is "user distribution density", the condition for splitting the node may be "user distribution density ≤ 50".

[0223] After dividing the dispatching area into grids, random forest is used to predict the number of power outages for users in each grid at the current moment.

[0224] Random Forest model (RF) is composed of a set of decision trees {h(X,Θ k ),k=1,2,...,N}, X is the input multivariate time series data, N is the number of decision trees, and the parameter Θ k It is a random vector that is independent and identically distributed with the k-th decision tree and can represent the growth process of the k-th decision tree.

[0225] After a sample X to be classified is input into a trained RF model, it will enter all trained decision trees. Each decision tree will select and determine the type of data X based on its characteristics. After all decision trees have reached their classification results, the mean of the target variable Y corresponding to all feature vectors X that reach the tree leaf nodes is used as the predicted value, as shown below:

[0226]

[0227] Among them, H(X,Y) is the RF prediction value, h i (x) is the classification result of the i-th decision tree; N is the number of decision trees; X and Y are the feature vector and target vector respectively.

[0228] The training process of the RF model is as follows:

[0229] Construct a sample dataset. From each grid's historical multivariate time series dataset, randomly extract multiple multivariate time series data sets with replacement to form multiple subsample sets. Each subsample set is used to train a decision tree. The training set for each decision tree is different and may contain duplicate samples. The multiple multivariate time series data sets for each grid serve as feature vectors, and the corresponding number of user power outages serves as the target vector.

[0230] Split feature selection: The decision tree in the RF model selects only a subset of all features when splitting. The RF model first randomly selects a subset of all available features. Each time the decision tree splits, the optimal feature is selected from these randomly selected features. Each decision tree is grown as far as possible without pruning. The Gini index of the feature vector at each node in the decision tree is calculated, and the feature vector with the smallest Gini index is selected as the splitting feature, which is used as the basis for splitting the decision tree node. The Gini index is used to measure the impurity of samples in a dataset.

[0231] Gini index k Gini , as shown below:

[0232]

[0233] Among them, p(i|t) is the probability that the feature variable x belongs to class i; n is the number of subsample sets.

[0234] Select the Gini index k Gini The smallest eigenvector is used as the splitting feature.

[0235] Each decision tree splits a node. The decision tree cannot grow indefinitely. The condition for the decision tree to stop growing is: the amount of data at the node is less than the specified value; k Gini The index is less than the threshold; the depth of the decision tree reaches the specified value; all features have been used.

[0236] The mean of the target variable corresponding to all feature vectors reaching the tree leaf nodes of the decision tree is used as the predicted value of the number of power outages for users at the current moment.

[0237] The purpose of step S2 is to divide the dispatching area into multiple non-overlapping and fully covered grids. Using the random forest algorithm, the number of power outages for users in each grid at the current moment is predicted based on the historical multivariate time series data of each grid as the feature vector and the number of user power outages as the target vector.

[0238] Step S3 includes steps S31-S32.

[0239] Based on multivariate time series data and the predicted number of power outages at the current moment, an allocation center location optimization model is constructed with the goal of minimizing the emergency repair efficiency of the allocation center and the combined impact of power outage losses. The improved Grey Wolf-White Shark algorithm is used to solve the problem and the optimal allocation center location is obtained.

[0240] Step S31: Constructing an allocation center location optimization model, including a location objective function and constraints.

[0241] The allocation center location optimization model includes location objective function and constraint conditions;

[0242] The site selection objective function is as follows:

[0243]

[0244] Among them, f1 is the comprehensive impact of minimizing the emergency repair efficiency of the dispatch center and the power outage loss; F i is the predicted failure probability of line i; M L is the total number of faulty lines; N B (i) is the total number of nodes without load caused by fault line i; w ij is the level weight of load loss node j caused by line i failure; L ij is the load loss amount caused by node j during unit power outage time; t ij is the repair waiting time for the failure of load-loss node j caused by line i, that is, the time spent on the journey from the temporary dispatch center to the faulty line i, which is obtained based on the distance from the fault point to the dispatch center and the speed of the repair team; U ij,grid is the predicted number of power outages for users in the grid; grid is the priority weight of the grid; α1 and α2 are the weight coefficients of the dispatch center's emergency repair efficiency and power outage loss;

[0245] The site selection constraints include geographical location constraints, coverage constraints, and load coverage constraints. The site selection constraints are as follows:

[0246] (1) Geographical location constraints

[0247] The geographical location of the allocation center must be within the specified geographical range:

[0248] x min ≤x≤x max ,y min ≤y≤y max Formula (18)

[0249] x,y are the possible coordinates of the allocation center, which are longitude and latitude.

[0250] x min , x max ,y min ,y max The geographical boundaries of the area where the dispatch center is located are the boundaries of the dispatch area where the disaster occurs.

[0251] Geographical constraints limit the possible locations of the dispatch center, thus affecting the repair waiting time t in the objective function. ij Different geographical locations will result in different repair waiting times, which in turn affects the value of the objective function.

[0252] (2) Coverage constraints

[0253] The distance between the dispatch center and the task point (such as the fault point or damaged point) must not exceed the maximum coverage radius:

[0254] d ij ≤D max Formula (19)

[0255] The allocation center (x, y) and the task point i (x i ,y i ) is the Euclidean distance of x. i ,y i is the latitude and longitude coordinates of task point i; D max It is the maximum coverage radius of the allocation center, which is a preset value; for example, it is preset to 5 kilometers.

[0256] A grid contains one or more task points. The coverage constraint ensures that the dispatch center can cover all task points, avoiding long waiting time for repairs due to long distances. This directly affects the t in the objective function. ij , ensure that t ij Within a reasonable range.

[0257] (3) Load coverage constraint

[0258] The allocation center needs to cover a certain proportion of task point loads:

[0259]

[0260] Among them, SC is the set of task points covered by the allocation center; P i The load of task point i; S is the set of all task points; γ load coverage threshold, illustratively, is set to 0.9.

[0261] The load coverage constraint ensures that the dispatch center covers enough loads to avoid excessive load loss due to insufficient coverage. It directly affects the F in the location objective function. i and U ij,grid , ensuring that load losses and user losses are within a reasonable range.

[0262] Step S32: using the improved grey wolf-shark algorithm to solve the allocation center location optimization model and obtain the optimal allocation center location.

[0263] The improved grey wolf-shark algorithm is used to perform global and local searches to obtain the optimal location of the optimal allocation center. The process includes:

[0264] Initialize the size of the gray wolf population and the white shark population, and set the maximum number of iterations;

[0265] Set the initial values α1 and α2 of the dispatch center's emergency repair efficiency and power outage loss weight coefficients;

[0266] Randomly generate a set of candidate locations for the allocation center in all grid spaces and ensure that all candidate locations for the allocation center meet the geographical location constraints;

[0267] Generate random numbers to initialize a group of gray wolf populations, each gray wolf represents a candidate location for the allocation center;

[0268] For each candidate location of the allocation center, the value of the site selection objective function is calculated as the fitness value of the gray wolf;

[0269] The three best candidate allocation center positions corresponding to the fitness values are selected and marked as α, β, and δ gray wolves as the best candidate allocation center positions in the current population; other gray wolves update their positions based on the positions of α, β, and δ gray wolves;

[0270] When the maximum number of iterations is reached, the three optimal candidate allocation center locations are searched locally using the white shark algorithm, introducing α, β, and δ white sharks. The white sharks move toward the optimal candidate allocation center location, and at the optimal candidate allocation center location, a decision is made based on the strategy selection threshold to determine whether to introduce random perturbations. The location of the gray wolf population is then updated using the α, β, and δ white sharks, exploring the local optimum until the location selection objective function converges.

[0271] The location with the minimum site selection objective function value and that meets the site selection constraints is selected as the optimal allocation center location. The details are as follows:

[0272] Initialize the gray wolf population and white shark population, illustratively, to 30 each; set the maximum number of iterations to 100; set α1 = 0.6, α2 = 0.4;

[0273] In the Grey Wolf Optimizer (GWO), there are n wolves in a population of grey wolves. Each wolf has its own position, and the position of each wolf can be expressed as: X i =(x i1 ,x i2 ,...,x id ), where x id The d-th dimension value representing the position of the i-th wolf is the multivariate time series data of the candidate locations of the allocation center.

[0274] The first three wolves with the smallest fitness function values (i.e., the location selection objective function values) are called α, β, and δ wolves respectively, and the remaining wolves update their positions according to the positions of these three wolves.

[0275] D=|C·X p (t)-X(t)| Formula (21)

[0276] X(t+1)=X p(t)-A·D Formula (22)

[0277] Among them, X p (t) represents the position vector of the current prey (optimal solution) at time t. In the actual execution process algorithm, this value is replaced by α, β, and δ wolves to obtain D α ,D β ,D δ , X(t) is the current position vector of the gray wolf at time t; A is the coefficient vector used to control the step size and determine the distance the wolf moves; C is the coefficient vector used to calculate the distance between the current gray wolf X(t) and the prey.

[0278] α Wolf: Leader, representing the optimal solution for the current allocation center location;

[0279] β Wolf: auxiliary decision maker, representing the suboptimal solution for the current allocation center location;

[0280] δWolf: Executor (scout, guard, etc.), representing the third best solution for the current allocation center location;

[0281] Other wolves: followers, update their positions through α, β, and δ optimal wolves;

[0282] X(t+1) is the current position vector of the gray wolf at time t+1.

[0283] The update strategy of GWO first updates the positions of the three current optimal wolves to obtain D α ,D β ,D δ , these three values represent the distance between the current wolf’s X(t) and the three best wolves. X1, X2, and X3 are calculated using the following equations (23)-(25), and then the position of the current population is updated by X(t+1)=(X1+X2+X3) / 3; X(t+1) is the updated position vector, indicating the new position of the wolf at time t+1; A is the coefficient vector.

[0284] X1=X α -A1D α Formula (23)

[0285] X2=X β -A2D β Formula (24)

[0286] X3=X δ -A3D δ Formula (25)

[0287] Among them, X α 、X β and X δ are the three optimal wolf positions α, β, and δ respectively; D α 、Dβ and D δ They represent the distance between the current wolf's position X(t) and the three optimal wolves; A1, A2 and A3 correspond to the coefficient vectors of the three optimal wolves, which are used to control the step size; X1, X2 and X3 are the candidate positions of the current wolf calculated based on the positions and distances of the three optimal wolves.

[0288] The White Shark algorithm is optimized by drawing on the concept of the Grey Wolf Algorithm to optimize the convergence speed and global search capability of the White Shark algorithm, so as to enhance the global optimization capability of the White Shark algorithm. After optimization, the mathematical model of the White Shark movement is:

[0289]

[0290] in, are the position vector and velocity vector of the i-th white shark at the k+1 iteration respectively;

[0291] u and l are the boundaries of the variables in each dimension, respectively, which are used to limit the value range of the position;

[0292] m v Select a threshold for the preset strategy to decide whether to introduce random disturbances and control the probability of random disturbances occurring;

[0293] rand is a random number generated in the interval [0,1], used to introduce randomness; it is usually preset to a value between 0 and 1;

[0294] are the position vector and velocity vector of the i-th white shark at the k-th iteration respectively;

[0295] f represents the frequency parameter of the wave, which is a fixed constant;

[0296] μ contraction factor, used to control the effect of speed;

[0297] p1 and p2 represent control and right The impact ratio of the change;

[0298] c1, c2 are two random numbers generated uniformly between 0 and 1, where a larger c1 indicates a greater tendency to Plays a leading role, the larger the c2, the more inclined Plays a leading role. c1 and c2 are random values in the WSO algorithm and are mainly used to increase disturbance. represents the global optimal position found; represents the optimal position that the i-th shark has visited.

[0299] For example, m vis 0.3.

[0300] rand<m v Indicates that the current iteration meets the conditions for introducing random perturbations. When random perturbations are introduced, the position update is the product of a random number and a boundary range. At this point, the algorithm uses random perturbations to explore new solution spaces, enhancing global search capabilities.

[0301] rand≥m v This means that the current iteration does not meet the conditions for introducing random perturbations. In this case, the algorithm will adopt a more stable update strategy, where the position update is the weighted sum of the current position and velocity, and the existing information is used for local search.

[0302] The position of each white shark represents a candidate solution for the dispatch center. The position is a multidimensional vector, where each dimension represents a variable in a multivariate time series data set. The three sharks α, β, and δ update the position of the gray wolf population.

[0303] Calculate the location objective function values of all gray wolves in the gray wolf population (including the α, β, and δ gray wolves replaced by three white sharks), and take the location with the smallest location objective function value and satisfying the location constraints as the optimal allocation center location.

[0304] The Gray Wolf optimization algorithm is combined with the White Shark algorithm. The Gray Wolf algorithm has strong global search capabilities, while the White Shark algorithm excels at local optimization. To overcome the limitations of existing algorithms in finding local optimal solutions and slow convergence, combining the two algorithms leverages their strengths. The Gray Wolf algorithm conducts an extensive global search to ensure a suitable allocation center location within a larger area; while the White Shark algorithm, after finding a certain number of candidate solutions, uses a local search strategy to fine-tune the optimal solution.

[0305] By combining the optimization processes of the Gray Wolf and White Shark algorithms, the dispatch center locations are continuously explored and adjusted within the search space, and the comprehensive optimization results for each dispatch center location are evaluated. The White Shark algorithm provides local adjustments to potential solutions, while the Gray Wolf algorithm ensures the search for the global optimal solution. Ultimately, this hybrid optimization approach finds the optimal dispatch center location.

[0306] The function of step S3 is to construct an allocation center location optimization model based on multivariate time series data and the predicted number of user power outages, and use the improved gray wolf-white shark algorithm to solve it to obtain the optimal allocation center location to minimize the combined impact of emergency repair efficiency and power outage losses.

[0307] Step S4 includes steps S41-S42.

[0308] Based on the optimal allocation center location, the current fault point list and repair team parameters, a multi-objective model for repair team scheduling is constructed with the goal of minimizing the repair path distance and repair time. The improved Pareto-Zebra algorithm is used to solve the problem and obtain the optimal repair scheduling route for the distribution network.

[0309] Step S41: a multi-objective model for dispatching a repair team, including a site selection objective function and constraints.

[0310] (1) Objective function

[0311] The multi-objective model for dispatching the emergency repair team includes a dispatching objective function and constraints;

[0312] The scheduling objective function is as follows:

[0313]

[0314] Among them, M1 is the total repair path distance; M2 is the overall repair time of the region; p and q are damaged points, k is the power repair team; x pqk is a decision variable, and its value of 0 or 1 indicates that the power repair team k has not arrived at the damaged point p or has arrived at the damaged point q; c pq is the transportation distance from damaged point p to damaged point q; v is the speed of the repair team; R is the repair team set, which is the set of all available repair teams; V is the set of all damaged points; n is the number of damaged points; K is the number of repair teams; is the repair time of the damaged point p;

[0315] The dispatch constraints include repair team capacity constraints, task allocation constraints, path continuity constraints, distribution network radial radiation topology constraints, node voltage constraints and distribution flow constraints.

[0316] RThe collection of all available repair teams.

[0317] The scheduling constraints are as follows:

[0318] (1) Capacity constraints of the repair team:

[0319] The power repair team has limited vehicle capacity and limited energy, so each power repair team has limited repair hours, and the constraints are as follows:

[0320]

[0321] Among them, y pk is a decision variable, and its value equal to 1 indicates that the damaged point p is served by the power repair team k. zrepresents the working capacity limit of the zth repair team; Q is the maximum repair capacity of each team, which refers to the maximum amount of tasks or workload that the repair team can undertake.

[0322] (2) Task allocation constraints:

[0323] To ensure the independence of each repair team and prevent them from affecting each other, constraints are added to ensure that each damaged grid repair task is completed by one power repair team, namely:

[0324]

[0325] Among them, m is the number of repair teams; y ik is the decision variable, y pk =1, it means that team k is responsible for the repair task of damaged point p, y pk = 0, indicating that team k is not responsible for the repair task of damaged point p; V' is the set including all damaged points.

[0326] For each damaged point p, only one power repair team is responsible for its repair mission. The sum in the formula indicates that, out of all possible repair teams, only one will be assigned to that damaged point. This ensures that each damaged grid is repaired by an independent power repair team, preventing interference between teams.

[0327] (3) Path continuity constraints:

[0328] To ensure the continuity of the repair team's vehicle transportation route, constraints are added to ensure that only one power repair team arrives at and leaves a damaged grid.

[0329]

[0330] The transportation route from one damaged point to another should be continuous. That is, when the repair team starts from a damaged point, their path should be able to reach the next task point smoothly, avoiding path interruption or unreasonable driving route.

[0331] This continuity needs to be ensured through mathematical models or logical constraints. By adding constraints, we ensure that the task assignments for each repair team are reasonable. Ultimately, these conditions ensure that each repair team follows a reasonable, coherent route when performing their tasks, and that the repair team at each damaged site will not conflict with other teams within its mission scope. This makes repair scheduling more efficient and avoids unnecessary overlap or waste of resources.

[0332] (4) Radial radiation topology constraints of distribution network

[0333] g re ∈G MFormula (34)

[0334] Among them, g re Represents the distribution network topology after emergency repairs are completed and normal power supply is restored; G M Represents a set of structures that can ensure the radial topology of the distribution network.

[0335] (5) Node voltage constraints

[0336] U min ≤U p ≤U max Formula (35)

[0337] Among them, U p is the node voltage at the damaged point; U min is the lower limit of the node voltage; U max is the upper limit of the node voltage.

[0338] (6) Distribution network flow constraints

[0339] |S L |≤S L,max Formula (36)

[0340] Among them, S L is the flow of line L; S L,max is the maximum value of the line L power flow.

[0341] Step S42: Use the improved Pareto-Zebra algorithm to solve and obtain the optimal emergency repair dispatch route of the distribution network.

[0342] The improved Pareto-Zebra algorithm is used to solve the multi-objective model of repair team scheduling and obtain the optimal repair scheduling route, including:

[0343] Based on the optimal allocation center location and the list of fault points, a set of emergency repair routes is randomly generated, the zebra population size is set, and the zebra population is initialized;

[0344] Determine the weights of the two objectives: repair path distance and repair time;

[0345] Calculate the emergency repair path distance and repair time of each route, and calculate the objective function value by weighting the emergency repair path distance and repair time based on the weights of the two objectives as the fitness value of each zebra;

[0346] Perform a non-dominated sort on all routes in the zebra population and determine the Pareto rank of each route. If route A is not inferior to route B in terms of both repair path distance and repair time, and is superior to route B in at least one of its objectives, then route A dominates route B. All routes that are not dominated by other routes are assigned to the first non-dominated rank on the Pareto front.

[0347] After removing the first-level routes, repeat the non-dominated sorting on the remaining routes to determine the second non-dominated level, and so on;

[0348] Select the num routes with the lowest non-dominated level as the pioneer zebra routes; other zebra routes move closer to the pioneer zebra routes to simulate the foraging behavior of zebras;

[0349] Introduce random perturbations near each zebra's current route to explore new solutions; ensure that the new route satisfies the constraints;

[0350] Re-perform non-dominated sorting on the updated zebra population until the scheduling objective function converges;

[0351] Output the final set of routes on the Pareto front, where the routes in the set are optimal in terms of total repair path distance and overall regional repair time, thereby obtaining a set of Pareto optimal solutions;

[0352] Calculating the emergency repair path distance and the repair time weighted value F for each route in the Pareto optimal solution;

[0353] The optimal solution with the smallest F among the Pareto optimal solutions is selected as the final optimal emergency repair scheduling route.

[0354] The Zebra Optimization Algorithm (ZOA) simulates zebra behavior to perform multi-objective optimization. When determining the optimal path, ZOA balances multiple optimization objectives and uses non-dominated sorting to ensure solution diversity and find the optimal repair scheduling route.

[0355] Based on the optimal dispatch center location and the list of fault points as the optimization space, a zebra population is randomly initialized in the optimization space. Each zebra represents a potential solution for the optimal emergency repair dispatch route, while ensuring that the initial solution covers the solution space and satisfies the dispatch constraints.

[0356] Each zebra represents a possible repair route. The route starts and ends at the dispatch center and passes through the fault point in between.

[0357] Calculate the fitness value of each zebra;

[0358] The best member of the population is considered the pioneer zebra and leads other population members to its position in the search space. Constraint checking is performed to ensure that the updated solution satisfies all constraints.

[0359] ZOA indirectly handles multi-objective problems by maintaining the Pareto front and diversity of the solution set. Multi-objectives are converted into single objectives and expressed using a weighted method. For each route in the Pareto optimal solution, the weighted value F of the repair path distance and repair time is calculated as follows:

[0360] F=ω1M1+ω2M2 Formula (37)

[0361] Among them, ω1 and ω2 are the weights of targets M1 and M2 respectively; ω1≥0, ω2≥0, ω1+ω2=1.

[0362] Re-perform non-dominated sorting on the updated zebra population until the scheduling objective function converges;

[0363] The solution set on the current Pareto front is output as the optimal solution set for the multi-objective problem. These solutions represent the optimal trade-off between different total repair path distances and the overall repair time for the region under given constraints.

[0364] The optimal solution with the smallest F among the Pareto optimal solutions is selected as the final optimal emergency repair scheduling route.

[0365] The function of step S4 is to build a multi-objective optimization model for repair team scheduling based on the optimal allocation center location and fault point list, and use the improved Pareto-Zebra algorithm to solve it, so as to obtain the optimal distribution network repair scheduling route in terms of repair path distance and repair time.

[0366] Example 2:

[0367] Another embodiment of the present invention discloses a distribution network emergency repair and dispatching system based on fault prediction, thereby implementing the distribution network emergency repair and dispatching method based on fault prediction in embodiment 1. The specific implementation of each module refers to the corresponding description in embodiment 1.

[0368] like Figure 2 As shown, the system includes a basic data acquisition and preprocessing module M1, a grid division and power outage user prediction module M2, a positioning and allocation center site selection module M3, and an emergency repair scheduling optimization module M4.

[0369] The basic data acquisition and preprocessing module M1 is used to collect the basic data time series of the distribution network within a certain historical period in the dispatching area and preprocess it to obtain multivariate time series data;

[0370] The user power outage number prediction module M2 is used to divide the scheduling area into multiple grids; based on the historical multivariate time series data and the corresponding user power outage number in each grid, the user power outage number is predicted at the current moment;

[0371] The allocation center location selection module M3 is used to construct an allocation center location optimization model based on the multivariate time series data and the predicted number of users with power outages at the current moment, with the goal of minimizing the allocation center's emergency repair efficiency and the combined impact of power outage losses, and to solve for the optimal allocation center location;

[0372] The emergency repair scheduling optimization module M4 is used to build an emergency repair team scheduling multi-objective model based on the optimal allocation center location, the current fault point list and emergency repair team parameters obtained, with the goal of minimizing the emergency repair path distance and repair time, and solve it to obtain the optimal emergency repair scheduling route for the distribution network.

[0373] In summary, the method for detecting and automatically labeling multiple targets on a water surface based on threshold segmentation according to an embodiment of the present invention has the following beneficial effects:

[0374] 1. The BiLSTM-GAN model constructed in the present invention integrates the generative adversarial network GAN and the bidirectional long short-term memory network Bi-LSTM. The generator of the GAN network uses bidirectional LSTM to capture the contextual dependencies of the initial time series data in both forward and reverse directions to generate fitting values for missing values. The discriminator optimizes the data quality so that the filled data is closer to the real data, effectively solving the problem that the existing methods are insufficient in processing missing data. In natural disaster scenarios, distribution network data often has large-scale missing data due to communication interruptions or equipment damage, resulting in the inability of existing methods to accurately reflect the actual operating status of the distribution network. The bidirectional LSTM network makes full use of the before and after information of the time series to improve the prediction accuracy of missing values of multivariate data such as load loss and the number of user power outages. The BiLSTM-GAN model can not only make up for the shortcomings of traditional methods in filling missing data, but also improve the accuracy of fault prediction in disaster scenarios, and provide a data basis for the subsequent allocation center site selection and repair team scheduling optimization;

[0375] 2. Existing methods for selecting the site of the allocation center have deficiencies in coping with disaster variability and multi-objective optimization, and it is difficult to take into account both emergency repair efficiency and power outage losses at the same time. The present invention optimizes the site selection of the allocation center based on the improved Gray Wolf-White Shark algorithm. The White Shark algorithm has a strong local search capability, and the Gray Wolf algorithm can enhance the global search performance. The combination of the two significantly improves the comprehensive solution capability of the algorithm, and effectively overcomes the problem that existing methods are prone to falling into local optimality and slow convergence speed. The present invention constructs an allocation center site selection optimization model with the goal of minimizing the comprehensive impact of the emergency repair time of the allocation center's response to the fault point and the power outage load loss. It can provide more reasonable and efficient emergency repair allocation center site selection in complex disaster scenarios, greatly improve emergency repair scheduling efficiency, and reduce power grid operation losses;

[0376] 3. In response to the problem that the existing methods have a single scheduling strategy for emergency repair teams and ignore task priorities, the present invention constructs a multi-objective model for emergency repair team scheduling based on multi-objective optimization, with the goal of minimizing the emergency repair path distance and repair time, and comprehensively considering constraints such as the number of teams, load capacity, and driving paths to achieve a reasonable allocation of emergency repair resources. The improved Pareto-Zebra algorithm is used to solve the scheduling problem. By combining the fast global search capability of the zebra optimization algorithm and the multi-objective balance processing capability of the Pareto optimization, the emergency repair team is given priority to serve high-priority areas with large power outage impacts. At the same time, the scheduling method makes full use of the path optimization characteristics of the improved Pareto-Zebra algorithm, significantly shortening the total driving distance of the emergency repair team from the allocation center to the damaged point, thereby improving the efficiency of emergency repairs;

[0377] 4. The existing allocation center site selection and repair team scheduling are often carried out independently, lacking linkage optimization, resulting in uneven resource allocation or low efficiency in the execution of repair tasks. The present invention combines the allocation center site selection location with the repair team scheduling through linkage optimization, and determines the optimal site selection method through the allocation center site selection optimization model, providing efficient resource support for the repair team scheduling; then, based on the optimal allocation center site selection location, as well as the current fault point list and repair team parameters obtained, a multi-objective model for repair team scheduling is constructed with the goal of minimizing the repair path distance and repair time, and the solution is performed to obtain the optimal repair scheduling route for the distribution network. This method dynamically balances the coupling relationship between the allocation center layout and the repair team allocation, avoids the global suboptimal problem caused by separate optimization, and provides an innovative solution for the efficient recovery of the distribution network under disasters.

[0378] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0379] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A distribution network emergency repair scheduling method based on fault prediction, characterized in that: include: Collect the basic data time series of the distribution network in a certain historical period in the dispatching area and preprocess it to obtain multivariate time series data; Dividing the scheduling area into a plurality of grids; Predict the number of power outages for users at the current moment based on the historical multivariate time series data and the corresponding number of power outages for users in each grid; Based on the multivariate time series data and the predicted number of users experiencing power outages at the current moment, a dispatch center location optimization model is constructed with the goal of minimizing the dispatch center's emergency repair efficiency and the combined impact of power outage losses, and the optimal dispatch center location is obtained. Based on the optimal allocation center location, the current fault point list and repair team parameters obtained, a multi-objective model for repair team scheduling is constructed with the goal of minimizing the repair path distance and repair time, and the optimal repair scheduling route for the distribution network is obtained by solving the model.

2. The method according to claim 1, characterized in that The historical multivariate time series data in each grid is used as the feature variable, and the corresponding number of user power outages is used as the target variable. A random forest is trained to obtain the predicted power outage user data at the current moment; this includes: For each grid's historical multivariate time series data set, multiple multivariate time series data are randomly extracted with replacement to form multiple subsample sets, each of which is used to train a decision tree. For each decision tree, a portion of the features in the feature vector is randomly selected, and the Gini index of each feature vector is calculated at each node. The feature vector with the smallest Gini index is selected as the split feature, which is used as the basis for splitting the decision tree node; Each decision tree splits its nodes until the decision tree stops growing. The mean value of the target variable corresponding to all the feature vectors reaching the tree leaf nodes of the decision tree is used as the predicted value of the number of power outages for users in the grid at the current moment; The conditions for stopping the growth of the decision tree include that the number of decision tree nodes is less than a specified value, the Gini index is less than a preset threshold, the depth of the decision tree reaches a specified value, or all multivariate time series data samples have been used up.

3. The method according to claim 1, characterized in that The allocation center location optimization model includes location objective function and constraint conditions; The site selection objective function is as follows: Among them, f1 is the comprehensive impact of minimizing the emergency repair efficiency of the dispatch center and the power outage loss; F i is the predicted failure probability of line i; M L is the total number of faulty lines; N B (i) is the total number of nodes without load caused by fault line i; w ij is the level weight of load loss node j caused by line i failure; L ij is the load loss amount caused by node j during unit power outage time; t ij is the repair waiting time for the failure of load-loss node j caused by line i, that is, the time spent on the journey from the temporary dispatch center to the faulty line i, which is obtained based on the distance from the fault point to the dispatch center and the speed of the repair team; U ij,grid is the predicted number of power outages for users in the grid; grid is the priority weight of the grid; α1 and α2 are the weight coefficients of the dispatch center's emergency repair efficiency and power outage loss; The site selection constraints include geographical location constraints, coverage constraints and load coverage rate constraints.

4. The method according to claim 3, characterized in that The improved grey wolf-shark algorithm is used to perform global and local searches to obtain the optimal location of the optimal allocation center, including: Initialize the size of the gray wolf population and the white shark population, and set the maximum number of iterations; Set the initial values α1 and α2 of the dispatch center's emergency repair efficiency and power outage loss weight coefficients; Randomly generate a set of candidate locations for the allocation center in all grid spaces and ensure that all candidate locations for the allocation center meet the geographical location constraints; Generate random numbers to initialize a group of gray wolf populations, each gray wolf represents a candidate location for the allocation center; For each candidate location of the allocation center, the value of the site selection objective function is calculated as the fitness value of the gray wolf; The three best candidate allocation center positions corresponding to the fitness values are selected and marked as α, β, and δ gray wolves as the best candidate allocation center positions in the current population; other gray wolves update their positions based on the positions of α, β, and δ gray wolves; When the maximum number of iterations is reached, the three optimal candidate allocation center locations are searched locally using the White Shark algorithm, introducing α, β, and δ white sharks. The white sharks move toward the optimal candidate allocation center location, and at the optimal candidate allocation center location, a decision is made based on the strategy selection threshold to determine whether to introduce random perturbations. The location of the gray wolf population is then updated using the positions of the three white sharks α, β, and δ, exploring the local optimum until the location selection objective function converges. The location with the smallest site selection objective function value and satisfying the site selection constraints is selected as the optimal allocation center location.

5. The method according to claim 1, characterized in that The multi-objective model for dispatching the emergency repair team includes a dispatching objective function and constraints; The scheduling objective function is as follows: Among them, M1 is the total repair path distance; M2 is the overall repair time of the region; p and q are damaged points, k is the power repair team; x pqk is a decision variable, and its value of 0 or 1 indicates that the power repair team k has not arrived at the damaged point p or has arrived at the damaged point q; c pq is the transportation distance from damaged point p to damaged point q; v is the speed of the repair team; R is the repair team set, which is the set of all available repair teams; V is the set of all damaged points; n is the number of damaged points; K is the number of repair teams; is the repair time of the damaged point p; The dispatch constraints include repair team capacity constraints, task allocation constraints, path continuity constraints, distribution network radial radiation topology constraints, node voltage constraints and distribution flow constraints.

6. The method according to claim 5, characterized in that The improved Pareto-Zebra algorithm is used to solve the multi-objective model of repair team scheduling and obtain the optimal repair scheduling route, including: Based on the optimal allocation center location and the list of fault points, a set of emergency repair routes is randomly generated, the zebra population size is set, and the zebra population is initialized; Determine the weights of the two objectives: repair path distance and repair time; Calculate the emergency repair path distance and repair time of each route, and calculate the objective function value by weighting the emergency repair path distance and repair time based on the weights of the two objectives as the fitness value of each zebra; Perform a non-dominated sort on all routes in the zebra population and determine the Pareto rank of each route. If route A is not inferior to route B in terms of both repair path distance and repair time, and is superior to route B in at least one of its objectives, then route A dominates route B. All routes that are not dominated by other routes are assigned to the first non-dominated rank on the Pareto front. After removing the first-level routes, repeat the non-dominated sorting on the remaining routes to determine the second non-dominated level, and so on; Select the num routes with the lowest non-dominated level as the pioneer zebra routes; other zebra routes move closer to the pioneer zebra routes to simulate the foraging behavior of zebras; Introduce random perturbations near each zebra's current route to explore new solutions; ensure that the new route satisfies the constraints; Re-perform non-dominated sorting on the updated zebra population until the scheduling objective function converges; Output the final set of routes on the Pareto front, where the routes in the set are optimal in terms of total repair path distance and overall regional repair time, thereby obtaining a set of Pareto optimal solutions; Calculating the emergency repair path distance and the repair time weighted value F for each route in the Pareto optimal solution; The optimal solution with the smallest F among the Pareto optimal solutions is selected as the final optimal emergency repair scheduling route.

7. The method according to claim 1, characterized in that: The preprocessing includes deduplication, data correction, continuity test, outlier processing and missing value filling; The missing value filling includes: Filling missing values in the time series of the basic data of the distribution network with zero values to obtain original data and initial time series data filled with zero values; Generating a mask vector of the initial time series data; Inputting the initial time series data and the corresponding mask vector into the trained BiLSTM-GAN model to obtain the multivariate time series data; The BiLSTM-GAN model is constructed based on the Bi-LSTM and GAN networks, including a generator and a discriminator; The generator uses a bidirectional LSTM network to capture the contextual dependency of the initial time series data from the forward and reverse directions to generate fitted values of missing values; retains the original data of the distribution network basic data time series sequence, and uses the fitted values of the missing values to replace the zero values of the missing positions to obtain the distribution network basic data time series sequence filled with missing values; The discriminator forces the time series sequence of the basic data of the distribution network to be close to the real data, thereby obtaining multivariate time series data.

8. The method according to claim 7, characterized in that: The generator uses a bidirectional LSTM network to capture the contextual dependencies of the initial time series data in the forward and reverse directions to generate fitting values for missing values, including: The time series data X with missing data and its corresponding mask vector m are input into the forward LSTM to obtain the forward hidden state Input the reverse sequence of X and its reverse mask vector into the reverse LSTM to obtain the backward hidden state based on and Get the forward output respectively and backward output based on and Calculate the fitted value x for the missing value uni,i ,as follows: in, are the forward and backward combination factors respectively.

9. The method according to any one of claims 1 to 8, characterized in that The basic data of the distribution network includes distribution network data, distribution network fault data, distribution network load data, natural disaster data, user distribution data, user power outage data, historical emergency repair data, geographical environment data and infrastructure data.

10. A distribution network emergency repair and dispatching system based on fault prediction, characterized in that: include: The basic data acquisition and preprocessing module is used to collect the basic data time series of the distribution network within a certain historical period in the dispatching area and preprocess it to obtain multivariate time series data; A user power outage number prediction module is used to divide the scheduling area into multiple grids; based on the historical multivariate time series data and the corresponding user power outage number in each grid, the number of user power outages at the current moment is predicted; An allocation center location selection module is configured to construct an allocation center location optimization model based on the multivariate time series data and the predicted number of users experiencing power outages at the current moment, with the goal of minimizing the allocation center's emergency repair efficiency and the combined impact of power outage losses, and to solve for the optimal allocation center location. The emergency repair scheduling optimization module is used to build an emergency repair team scheduling multi-objective model based on the optimal allocation center location, the current fault point list and emergency repair team parameters obtained, with the goal of minimizing the emergency repair path distance and repair time, and solve it to obtain the optimal emergency repair scheduling route for the distribution network.

Images