Graph-model fusion technology-based high-reliability power distribution network visual planning method

Abstract

The invention discloses a high-reliability power distribution network visual planning method based on a graph-model fusion technology, and relates to the technical field of power distribution network planning. The method comprises the steps of multi-source data acquisition and preprocessing, topological graph construction, unified data model establishment, graph-model fusion, prediction optimization and visual platform development, acquisition of real-time / static data of scheduling, metering and other systems, and processing of unstructured data through OCR and NLP. A unified data pool is formed through cleaning, conversion and other purification steps; constructing a topological graph containing equipment nodes and edges and verifying the topological graph; establishing a unified model sharing database for quantifying static and dynamic data; graph-model fusion is realized through hash index and synchronous verification; lSTM is used to predict the load, and a Weibull model is used to predict the fault; genetic and JADE algorithms are adopted to optimize reliability, economical efficiency and energy efficiency; according to the method, graph-model interaction, scheme comparison and simulation verification are supported by developing a visual platform, and the reliability and efficiency of power distribution network planning are improved.

Description

Technical Field

[0001] This invention relates to the field of power distribution network planning technology, and in particular to a highly reliable visual planning method for power distribution networks based on graph-model fusion technology. Background Technology

[0002] As a crucial terminal link connecting users in the power system, the scientific and reliable planning of the distribution network directly determines the quality of electricity consumption and the efficiency of system operation. However, existing distribution network planning technologies still have many pain points: First, the ability to integrate multi-source data is insufficient. Real-time data (dispatch automation, SCADA, etc.), static data (production management, GIS, etc.), and unstructured data (maintenance reports, meteorological texts) are scattered and heterogeneous in format. The lack of standardized data purification processes leads to poor data quality, high integration barriers, and difficulty in supporting accurate analysis. Second, the topology map and data model are separated, and an efficient two-way ID mapping and state synchronization mechanism has not been established. This results in inaccurate quantification of topology differences and faults. The calculation efficiency of load transfer paths is low, making it impossible to achieve "graphics-data" linkage; third, the prediction and optimization accuracy is limited. The traditional load forecasting model MAPE is difficult to control within 8%, equipment failure prediction lacks reliable Weibull proportional risk model support, and planning scheme optimization mostly adopts a single algorithm, making it difficult to balance multiple objectives such as reliability (SAIDI / SAIFI), economy (investment cost), and energy efficiency (line loss rate); fourth, the visualization planning capability is weak. The existing platform lacks multi-dimensional data dashboards, graphic model fusion interaction, fault simulation, 3D scheme preview, and scene simulation functions, resulting in poor interactivity and compatibility, and failing to intuitively present the planning effect and support decision-making.

[0003] Patent CN112287423B discloses a method and terminal for visual planning of power distribution networks based on integrated diagrams and models. This invention acquires power grid graphic data and equipment ledger data, and generates a current power grid diagram based on the correlation and matching relationship between the power grid graphic data and equipment ledger data. The current power grid diagram is displayed on a graphical page. Based on the planning information entered by the user on the graphical page, the current power grid diagram is redesigned and the topological relationships between equipment are automatically maintained to generate a planned power grid diagram. The planned power grid diagram is then displayed on the graphical page. This invention acquires and correlates power grid graphic data, power grid topology data, and equipment ledger data, solving the problem of insufficient data integration. It generates a current power grid diagram based on the information and a planned power grid diagram after the planning information is entered, solving the problem of insufficient visualization methods. Furthermore, power grid planning can be performed on the graphical page, providing efficient visual interaction. However, the above invention is limited to the correlation and matching of power grid graphic data and equipment ledger data, and does not involve the integration and purification of complex data. Its data processing capabilities are weak and cannot support accurate planning. Therefore, those skilled in the art urgently need to solve the above technical problems. Summary of the Invention

[0004] One objective of this invention is to propose a highly reliable visual planning method for distribution networks based on graph-model fusion technology. This method breaks down information barriers by integrating and refining real-time, static, and unstructured multi-source data; it utilizes bidirectional hash indexing and synchronous verification to achieve bidirectional interaction and topology consistency between graphs and data, ensuring real-time synchronization; it combines LSTM load forecasting, Weibull fault prediction models, and genetic + JADE secondary optimization algorithms to achieve multi-objective planning prioritizing reliability; and it presents the planning results intuitively through a multi-dimensional visualization platform, supporting mobile operation and three-level access control, significantly improving the scientific nature, reliability, and decision-making efficiency of distribution network planning.

[0005] A highly reliable distribution network visualization planning method based on graph-model fusion technology according to an embodiment of the present invention includes the following steps:

[0006] Step S1: Multi-source data acquisition, obtaining real-time / static / unstructured data from multiple system sources;

[0007] Step S2: Perform data preprocessing using data cleansing methods;

[0008] Step S3: Construct the distribution network topology;

[0009] Step S4: Construct a unified data model for the power distribution network;

[0010] Step S5: Achieve topology mapping and fusion of the unified data model of the distribution network based on fusion verification and topology analysis;

[0011] Step S6: Achieve high-reliability prediction of distribution network conditions based on long short-term memory networks and the Weibull proportional hazards model;

[0012] Step S7: Utilize a genetic algorithm to achieve high reliability optimization of the power distribution network conditions;

[0013] Step S8: Generate a visual planning platform.

[0014] By adopting the above technical solutions, the overall method for high-reliability distribution network visualization planning based on graph-model fusion technology systematically solves the problems of data dispersion, graph-model separation, insufficient prediction accuracy, single optimization methods, and low visualization in traditional distribution network planning by integrating multi-source data acquisition, preprocessing, topology map and unified model construction, graph-model fusion, prediction optimization, and visualization platform development. It forms a closed-loop system from data integration to planning implementation, significantly improving the scientific nature, reliability, and decision-making efficiency of distribution network planning, and providing comprehensive technical support for the high-reliability operation of distribution networks.

[0015] Furthermore, in step S1, multi-source data acquisition is performed to obtain real-time / static / unstructured data from multiple system sources. Specifically, the operation is as follows:

[0016] Step S1-1: Obtain real-time data from external dispatch automation system, metering automation system, monitoring and data acquisition system, and achieve second-level real-time transmission using IEC61850 standard interface;

[0017] Step S1-2: Obtain static data from the production management information system and geographic information system, and use ETL tools to achieve periodic synchronization on a daily or weekly basis;

[0018] Steps S1-3: Obtain unstructured data from maintenance reports and meteorological text information, and convert it into structured data through OCR recognition and natural language processing to ensure that the data format is consistent for access;

[0019] OCR recognition stage: For scanned copies of paper maintenance reports or electronic image-formatted text, an adaptive binarization algorithm is first used to process the image, and then the text region is located using the EAST text detection model;

[0020] The EAST text detection model uses PVANet as its backbone network and outputs the coordinates and confidence of the text region after fusing feature maps of different scales.

[0021] Define the text region set as ,in This represents the top-left corner of the k-th text region. and bottom right corner coordinate;

[0022] Subsequently, a CRNN recognition model is used to extract the text content of each region. The CRNN recognition model consists of a convolutional layer for extracting image features, a bidirectional LSTM layer for capturing sequence dependencies, and a transcription layer for mapping features to characters, resulting in an unstructured text sequence. ,in For the region The corresponding recognized text;

[0023] NLP Structured Transformation Stage: A BIO annotation system is constructed based on a power distribution network domain dictionary. Entity annotation is performed on the text sequence T, and a BiLSTM-CRF model is used to identify key entities. The BiLSTM-CRF model includes an embedding layer that converts text into vectors, a bidirectional LSTM layer that captures contextual semantics, and a CRF layer that optimizes the rationality of the annotation sequence. The entity recognition probability calculation formula is as follows:

[0024] ;

[0025] in For input text features, For labeled sequences, Here is the state transition matrix. Here is the emission probability matrix. Normalization factor;

[0026] Key entities identified include device ID, fault type, maintenance time, and meteorological factors;

[0027] Finally, structured data is generated through entity-attribute mapping rules, in the format {Device ID: [Fault Type, Maintenance Time, Meteorological Parameter 1, Meteorological Parameter 2, ...]}.

[0028] By adopting the above technical solutions, the specific operations of multi-source data acquisition are optimized. Real-time data is obtained through standardized interfaces, static data is periodically synchronized, and unstructured data is transformed into structured data. This effectively breaks down data barriers between different systems, realizes unified access and format standardization of multi-source data, and provides a comprehensive and high-quality data foundation for subsequent data purification, model building, and graph-model fusion. It also solves the problem of biased analysis results caused by scattered data sources and inconsistent formats in traditional planning.

[0029] Furthermore, in step S2, a data cleansing method is used to perform data preprocessing, specifically as follows:

[0030] Step S2-1. Data cleaning; For missing values ​​in the data, continuous data is filled with the mean or median and corrected by combining it with historical data from the same period of the device; discrete data is filled with a pattern and matched with the default parameters of the same type of device.

[0031] Step S2-2. Data Conversion; The collected data is standardized in terms of format and dimensions: the time format is unified to "YYYY-MM-DDHH:MM:SS", the geographic coordinates are unified to the WGS84 coordinate system, and the units of equipment parameters are unified.

[0032] Dimensional normalization: The min-max normalization method is used to normalize data of different magnitudes to the [0,1] interval. The normalization formula is as follows: ,in The original data, For normalized data, The minimum value of this type of data. The maximum value of this type of data;

[0033] Step S2-3. Data correction and rejection: Compare the equipment nameplate parameters with the collected data. If the deviation exceeds 5%, use the nameplate data as the benchmark for calibration.

[0034] The route length is corrected by combining GIS terrain data. First, the starting point of the route is extracted through the GIS system. ,end The WGS84 coordinates, after being converted to planar coordinates using the Gauss-Kruger projection, are substituted into the formula. ; Calculate horizontal distance;

[0035] Then, the terrain type of the area traversed by the line is obtained by GIS, and the correction coefficient k for each terrain type is determined according to the power distribution network engineering regulations, ultimately determining the actual length of the line. ;

[0036] Delete duplicate data, data outside the reasonable range, and irrelevant data;

[0037] Step S2-4. Data Integration: Using the "Unique Device ID" as the key field, link data from multiple systems to form a unified data pool that links "device, geography, user, and load".

[0038] By adopting the above technical solutions and performing refined preprocessing operations such as data cleaning, transformation, correction, elimination, and integration, the data quality has been significantly improved, the data format and dimensions have been unified, erroneous data has been corrected, and data from multiple systems has been linked to form a unified data pool. This ensures the accuracy of subsequent topology construction, model analysis, and prediction optimization, avoids unreasonable planning schemes due to data quality issues, and provides reliable data support for power distribution network planning.

[0039] Furthermore, the construction of the distribution network topology in step S3 is specifically carried out as follows:

[0040] Step S3-1: Identify the core elements of the distribution network and construct a topology diagram from the data using the following formula:

[0041] ;

[0042] in, This represents a set of equipment nodes, including physical equipment such as transformers, lines, and switches. Represents a set of topological edges, characterizing the electrical connections between devices. Represents a set of timestamps. This represents a set of operational scenarios, including normal operation, emergency repair, and maintenance plans.

[0043] Step S3-2. Import each element in the topology graph into the Neo4j graph database, construct the topology network in the form of "node-relationship-node", and support topology traversal query;

[0044] Step S3-3. Topology Consistency Verification: Calculate the model topology signature, count the number of connected edges for each node, and determine the node degree. ,in Represent and sort the i-th device node in the topology graph G;

[0045] Obtain the connected region number to which each node belongs, where the electrical island... ,in Mapping functions for numbering connected components;

[0046] Substitute into the formula Generate model signature, where The model topological signature representing the topological graph G. This is an ascending sorting function. The set consisting of the degrees of all nodes. A set consisting of the numbered connected regions of all nodes; "+" indicates that the sets are concatenated.

[0047] Generate a comparison graph topology signature. If the Hamming distance between the two graphs is greater than 2, indicating two or more topological differences, then automatically perform the correction: delete isolated nodes in the graph that do not have a corresponding model ID, add nodes or edges that exist in the model but are missing from the graph, and generate a correction audit log.

[0048] By adopting the above technical solutions, a distribution network topology diagram containing equipment nodes, connection relationships, timestamps, and operating scenarios was constructed. Through graph database storage and topology consistency verification, the electrical connection relationships of the distribution network were clearly presented, supporting fast topology queries and timely correction of topology differences. This ensured the accuracy and consistency of the topology data, providing a solid topology foundation for subsequent graph-model fusion, fault analysis, and planning optimization.

[0049] Furthermore, in step S4, a unified data model for the distribution network is constructed, specifically through the following operations:

[0050] Step S4-1. Use static attributes Dynamic state For device nodes in the topology graph Data quantification; based on the device's unique ID. Collect device types one by one Rated capacity GIS coordinates Annual failure rate ,pass Implement structured storage;

[0051] Real-time acquisition of equipment active power through SCADA / metering automation system reactive power Bus voltage Switch status Equipment temperature Use the formula with a 1-minute step size. Record dynamic changes;

[0052] Topological edges in a topological graph Includes sampling resistor Reactance susceptance ,distance Rated current carrying capacity Through formula Integration;

[0053] Step S4-2. Abnormal Data Processing: Calculate Real-time Active Power Second-order difference: ,in, This is the current timestamp. Timestamps are adjacent 1-minute intervals;

[0054] Calculate the standard deviation of the device's 72-hour power data. Substitute into the formula If the condition is met, the data is marked as abnormal, triggering the SCADA system to issue a real-time re-sampling command for that node.

[0055] Step S4-3. Set the node's static attributes Edge parameters Store the device ID as the primary key in a PostgreSQL relational database and create an index to support batch updates;

[0056] Dynamic state of nodes Stored in the InfluxDB time-series database by partitioning by "Device ID and Timestamp".

[0057] By adopting the above technical solutions, a unified data model for the power distribution network was constructed, separating the static attributes and dynamic states of storage devices. A re-sampling mechanism was implemented for abnormal data, optimizing data storage and query efficiency and ensuring the real-time performance and accuracy of the data. This provided structured and standardized data model support for graph-model fusion, predictive analysis, and planning optimization, solving the problems of chaotic data storage and untimely dynamic updates in traditional models.

[0058] Furthermore, in step S5, the unified data model of the distribution network is realized based on fusion verification and topology analysis to achieve topology mapping and fusion. The specific operation is as follows:

[0059] Step S5-1. Implement bidirectional ID indexing based on hashing: Substitute the model-side ID and the front-end SVG element ID of each device ID into the formula. In the formula, The output hash value is represented by id, which represents the model-side device ID or the front-end SVG element ID of the input graphics ID. mod represents the modulo operation, where 65521 is less than 1. Find the largest prime number and calculate its hash value.

[0060] Create a two-way index in the Redis database using hash value, model ID, graph ID, and last synchronization timestamp to enable mapping queries. Clicking on a graph element will allow you to quickly retrieve the corresponding model data using the hash value.

[0061] Step S5-2. Fusion verification to achieve front-end and back-end state synchronization: When the device state changes, the back-end pushes a message to the front-end via the MQTT protocol, and the message carries a state change timestamp. ;

[0062] After receiving the message, the frontend updates the state of the graphic element and records the graphic refresh timestamp. Substitute into the formula If the threshold is exceeded, it is determined to be a synchronization error, and the front end calls the MAPI interface to fetch the latest model data again and perform graphics redrawing;

[0063] Step S5-3. Topological Difference Quantization: Calculate the topological signature of each model. and graph topology signature ;

[0064] Substitute into the formula ; Calculate the topological dissimilarity, where This indicates the degree of difference between the model and the graph topology. Represents the L2 norm;

[0065] like The correction is performed according to the following priority: first delete redundant nodes in the graph, then add missing edges to the model, and after correction, update the graph signature and synchronize it to the model side.

[0066] Step S5-4. When a line is disconnected or a transformer is out of service in the distribution network, for areas where the N-1 safety check is not met, Dijkstra's algorithm is used to traverse all possible load transfer paths based on the fused graph data.

[0067] For each path, substitute into the formula Calculate the total impedance and select... The path with the shortest path is taken as the optimal load transfer path, in the formula. This represents the total impedance of the current load transfer path. Indicates the first in the path The resistance value of the branch circuit, Indicates the first in the path The reactance value of the branch circuit, This indicates that all branches in the current path will be traversed.

[0068] By adopting the above technical solutions, fast mapping of graph models is achieved through bidirectional hash indexing, fusion verification ensures synchronization of front-end and back-end states, topology difference quantification and correction improve graph model consistency, and the optimal load transfer path is quickly calculated in case of a fault. This effectively solves the problems of graph model separation, state asynchrony, difficulty in correcting topology differences, and low fault handling efficiency in traditional distribution networks. It realizes bidirectional interaction and real-time synchronization of graph models, significantly improving the operation and management efficiency and fault response capability of distribution networks.

[0069] Furthermore, in step S6, high-reliability prediction of distribution network conditions is achieved based on long short-term memory networks and the Weibull proportional hazards model. The specific operation is as follows:

[0070] Step S6-1. Construct an LSTM load forecasting model: Input features include load data of the past 168 hours, average daily temperature, and holiday labels, where 1 represents a holiday and 0 represents a workday. Output the annual maximum load and peak-valley difference for the next 1-5 years.

[0071] The LSTM load forecasting model uses a three-layer LSTM hidden layer, where the first layer has 64 neurons, dropout is set to 0.2, and L2 regularization is 0.001 to capture intraday load fluctuations.

[0072] The second layer consists of 32 neurons, with parameters basically the same as the first layer, extracting cross-sky patterns;

[0073] The third layer consists of 16 neurons, focusing on the core temporal pattern. Each LSTM layer is followed by a batch normalization layer connected to an intermediate fully connected layer with ReLU activation. The L2 regularization parameter is set to 0.001 and the output layer of the Linear activation function is used. The first 5 dimensions represent the maximum load and the last 5 dimensions represent the peak-to-valley difference.

[0074] The loss function was set to mean squared error, the optimizer was Adam, and the learning rate was set to 0.001.

[0075] Historical load and related data of the power distribution network over the past 5 years were collected and divided into training set, validation set and test set in a ratio of 7:2:1.

[0076] Verify model accuracy using a test set: Calculate the true power for N test samples. With predicted power Substitute the deviation into the formula: Ensure that the model prediction accuracy meets the requirement of MAPE less than or equal to 8%. If it does not meet the requirement, readjust the model input features or hyperparameters until the target is met.

[0077] Step S6-2. Equipment failure prediction and reliability index calculation based on Weibull proportional hazards model;

[0078] Collect operational data from the past 5 years for transformers, lines, and switches in the distribution network, including the equipment's service life. Real-time load rate Ambient temperature and historical fault records;

[0079] The parameters of the Weibull proportional hazards model are fitted using maximum likelihood estimation. The model formula is as follows:

[0080] ;

[0081] in, As the baseline failure rate, Characteristic lifetime, Shape parameters This indicates that the failure rate increases over time. , , These are the weighting coefficients for each influencing factor;

[0082] Reliability is assessed using the SAIDI / SAIFI metrics.

[0083] ;

[0084] ;

[0085] Where t represents the duration of the power outage, n represents the number of affected users, m represents the number of power outages, and N represents the total number of users.

[0086] By adopting the above technical solutions, the LSTM model is used to accurately predict future loads, and the Weibull proportional hazards model is used to predict equipment failures and calculate reliability indicators. This provides accurate load and failure data support for distribution network planning, helps to identify load growth trends and equipment failure risks in advance, effectively improves the foresight and reliability of distribution network planning, and avoids planning defects caused by inaccurate load forecasting or insufficient failure prediction.

[0087] Furthermore, in step S7, a genetic algorithm is used to optimize the high reliability of the distribution network. Specifically, the operation is as follows:

[0088] Step S7-1. Construction of the multi-objective optimization objective function: Set three main optimization objectives:

[0089] Reliability objective: Minimize the average system outage duration ;

[0090] Economic objective: Minimize the investment cost of power distribution network planning. ;

[0091] Energy efficiency target: Minimize distribution network line loss rate ;

[0092] Where x represents the distribution network planning scheme to be optimized, including line routes, transformer capacity, distributed power source access parameters, etc.

[0093] The multi-objective optimization is transformed into a single-objective optimization using the weighted summation method. The fitness function formula is as follows:

[0094] ;

[0095] in, , , These are the current SAIDI benchmark value, investment cost benchmark value, and line loss rate benchmark value of the distribution network, respectively, with weighting coefficients satisfying the following conditions: , , ;

[0096] Step S7-2. Preliminary optimization based on genetic algorithm: Generate the Pareto optimal solution set. First, convert the planning scheme into chromosome codes that the algorithm can process, denoted as... ,in ; represents the binary encoding layer, where m is the total number of devices or lines to be modified. 1 indicates that new construction or expansion is needed, and 0 indicates that the current status should be maintained;

[0097] ; represents the real-number encoding layer, where p is the total number of modification parameters. , These are the upper and lower limits of the parameter;

[0098] The initial population was generated using random sampling. N is the population size; then, iterative optimization is performed using three genetic operators, with the selection operator randomly selecting 3 individuals from the population each time. Calculate its fitness value Select the individual with the highest fitness to enter the offspring population, and repeat this process N times to generate the offspring selection population. ;

[0099] Crossover operator pairs Individuals are randomly paired with a probability of 0.8.

[0100] The binary layer uses single-point intersection, with the intersection point randomly selected. ,implement Swap bit values;

[0101] The real number layer uses arithmetic crossover, according to... ; Generate new values, where, , The values ​​are the real values ​​of the parent individuals, obtained after crossover. ;

[0102] Mutation operator pairs Individuals mutate with a probability of 0.05, executed at the binary layer. Bit flip;

[0103] Gaussian mutation is used for real number layers, according to Generate new values, where, To vary the asynchronous length, take 5%-10% of the parameter range. Following a standard normal distribution, the offspring population obtained after mutation... ;

[0104] After 50 to 100 iterations, a local optimum is obtained, i.e., the Pareto optimum set;

[0105] Step S7-3. Secondary optimization based on JADE differential evolution algorithm to refine the optimal solution: The Pareto optimal solution set obtained by the genetic algorithm is used as the initial solution; the individual encoding completely follows the two-layer structure of "binary and real number" in step S7-2;

[0106] An elite mutation strategy is employed to modify individuals in the g-th generation population. Two different ordinary individuals were randomly selected. , and one elite individual Mutant individuals are randomly selected from the top 10% of the population and generated using the following formula: ;

[0107] In the formula, This is a scaling factor, initially set to 0.5-1.0, used to control the magnitude of variation;

[0108] Subsequently, crossover operations were used to generate experimental individuals. The mutated individuals were randomly paired with an 80% probability. Crossover was performed separately for the two-layer coding. One crossover point was randomly selected by identifying the modified object through the binary coding layer, and the binary bits after the crossover point of the paired individuals were swapped.

[0109] The modification parameters are recorded through a real-number encoding layer and then arithmetically cross-referenced. The parameters of paired individuals are then weighted and fused using a random coefficient α of 0.2-0.8. The formula is as follows: ;

[0110] Experimental individuals with fused parental traits were generated through cross-generation. ;

[0111] Selection of high-quality individuals is achieved through a selection process: The greedy selection criterion is used to calculate the number of experimental individuals. With the original individual fitness value;

[0112] like fitness ≥ The fitness level is retained. As the next generation of individuals Otherwise, retain the original individual. To ensure that the overall performance of the population does not degrade;

[0113] To optimize the dynamic matching process, after each iteration, only the effective scaling factor F and crossover probability CR that "can generate better individuals" are collected, forming sets S_F and S_CR respectively.

[0114] Update the global baseline parameters using the following formula: , ;

[0115] The next generation of individuals' F from Cauchy ( ,0.1) distribution sampling, CR from Normal ( The parameters are sampled in a 0,1) distribution and truncated to the intervals [0,1.2] and [0,1] respectively to achieve adaptive parameter adjustment;

[0116] Calculate the maximum fitness of the population in each generation. with minimum fitness When 5 consecutive generations meet the requirements This indicates that the solution has stabilized, so stop iterating and output the final optimal planning solution;

[0117] Step S7-4. Reliability Verification of the Optimized Scheme: Substitute the optimal planning scheme output in Step S7-3 into the LSTM load forecasting model in Step S6-1 and the Weibull fault prediction model and reliability index calculation formula in Step S6-2, and recalculate the final reliability index corresponding to the scheme. and ;

[0118] The verification process ends when the results meet the requirements.

[0119] minutes / household / year and per household per year

[0120] If the above criteria are not met, adjust the optimization algorithm parameters as follows: multiply the number of iterations of the genetic algorithm by 1.5, and adjust the reliability weight. Increase by 0.1, and re-execute steps S7-1 to S7-3 to optimize the process until the verification criteria are met.

[0121] By adopting the above technical solutions, multi-objective optimization is achieved through genetic algorithms and JADE differential evolution algorithms, taking into account reliability, economy and energy efficiency. The optimal planning scheme is generated and its reliability is verified to ensure that the scheme meets high standards. This solves the problems of single-objective optimization and insufficient feasibility of schemes in traditional planning, significantly improves the scientificity and feasibility of the planning scheme, and provides the optimal planning strategy for the high-reliability operation of the distribution network.

[0122] Furthermore, the visualization planning platform in step S8 includes a data visualization module, a graph-model fusion visualization module, a planning scheme visualization module, a simulation module, and a platform interaction and compatibility module, constructed as follows:

[0123] The data visualization module uses a dashboard as its core, integrating real-time load heatmaps, equipment failure rate statistics pie charts, and key indicator components such as SAIDI / SAIFI trend line charts; it also supports multi-condition filtering by region and time dimension.

[0124] The graphic model fusion visualization module integrates GIS geographic data with power distribution network topology models to achieve linked display of geographic wiring diagrams and topology diagrams; clicking on graphic elements allows you to view detailed model data of the corresponding equipment, and modifying model parameters will synchronously update the graphic status; the built-in fault simulation function allows the system to automatically calculate and visualize the optimal load transfer path and power outage range after setting equipment faults.

[0125] The planning scheme visualization module provides route adjustment maps, transformer capacity configuration comparison tables, and reliability index improvement bar charts. In addition, it realizes the 3D preview function of the scheme based on GIS terrain data, which shows the spatial layout relationship of lines and transformers from a three-dimensional perspective, and helps to evaluate the engineering feasibility of the planning scheme.

[0126] The simulation module supports the simulation of distribution network operation status in multiple scenarios, including extreme weather and typical load growth scenarios; it uses visualization to present the voltage drop range and line loss hotspots in the scenarios, helping users to verify the adaptability and stability of the planning scheme in extreme or future scenarios;

[0127] The platform's interaction and compatibility modules support drag-and-drop scheme adjustments; provide multi-format export functionality, generating CAD drawings, Excel parameter tables, and PDF reports; are compatible with tablets and mobile devices; and feature a three-tiered permission system: administrator, planner, and approver. Administrators are responsible for system configuration and data management, planners can edit schemes, and approvers are responsible for approval.

[0128] By adopting the above technical solutions, the visualization planning platform integrates multi-dimensional visualization modules and interactive compatibility functions. Through data dashboards, graphical model fusion display, planning scheme comparison, multi-scenario simulation, and convenient interaction, it intuitively presents the operation status and planning effect of the power distribution network. It supports collaborative work and mobile access for different roles, significantly improving the intuitiveness, efficiency, and decision support capabilities of planning, lowering the planning threshold, and facilitating the use and management of different roles such as planners and reviewers.

[0129] The beneficial effects of this invention are:

[0130] 1. This invention integrates real-time, static, and unstructured data from multiple systems, and through refined preprocessing such as cleaning, transformation, correction, elimination, and integration, it breaks down data barriers between different systems, unifies data formats and dimensions, ensures data quality, and provides a comprehensive and reliable data foundation for power distribution network planning.

[0131] 2. This invention achieves fast mapping between model and graph through hash bidirectional indexing, ensures front-end and back-end state synchronization through MQTT protocol, improves graph-model consistency through topology difference quantification correction, and automatically calculates the optimal load transfer path in case of fault. It effectively solves the problems of graph-model separation and state asynchrony in traditional distribution networks, realizes bidirectional graph-model interaction and real-time synchronization, and significantly improves the efficiency of distribution network operation and management and fault response capabilities.

[0132] 3. This invention accurately predicts the load and peak-valley difference for the next 1-5 years using an LSTM model, combines the Weibull proportional hazards model to predict equipment failures and calculate reliability indicators, and then uses genetic algorithms and JADE differential evolution algorithms to achieve multi-objective optimization of reliability, economy and energy efficiency, generating and verifying the optimal planning scheme that meets high standards, effectively improving the foresight, scientificity and feasibility of distribution network planning.

[0133] 4. This invention constructs a visual planning platform that integrates data visualization, graphical model fusion display, planning scheme comparison, multi-scenario simulation, and convenient interaction. It intuitively presents the operation status and planning effect of the power distribution network, supports collaborative work by different roles and mobile access, significantly improves the intuitiveness of planning, decision support capabilities and work efficiency, and lowers the planning threshold. Attached Figure Description

[0134] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0135] Figure 1 This is a flowchart of the method of the present invention;

[0136] Figure 2This is a comparison chart of the topology difference degree as a function of the number of distribution network devices, between the present invention and traditional methods;

[0137] Figure 3 This is a comparison chart showing the variation of the mean absolute percentage error of different methods of the present invention with the prediction time;

[0138] Figure 4 This is a comparison chart of the convergence speed of the optimized algorithm of this invention. Detailed Implementation

[0139] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0140] refer to Figures 1-4 A highly reliable distribution network visualization planning method based on graph-model fusion technology includes the following steps:

[0141] Step S1: Multi-source data acquisition, obtaining real-time / static / unstructured data from multiple system sources;

[0142] Step S2: Perform data preprocessing using data cleansing methods;

[0143] Step S3: Construct the distribution network topology;

[0144] Step S4: Construct a unified data model for the power distribution network;

[0145] Step S5: Achieve topology mapping and fusion of the unified data model of the distribution network based on fusion verification and topology analysis;

[0146] Step S6: Achieve high-reliability prediction of distribution network conditions based on long short-term memory networks and the Weibull proportional hazards model;

[0147] Step S7: Utilize a genetic algorithm to achieve high reliability optimization of the power distribution network conditions;

[0148] Step S8: Generate a visual planning platform.

[0149] In this embodiment, during data collection:

[0150] In step S1, multi-source data acquisition is performed to obtain real-time / static / unstructured data from multiple system sources. The specific operation is as follows:

[0151] Step S1-1: Obtain real-time data from external dispatch automation system, metering automation system, monitoring and data acquisition system, and achieve second-level real-time transmission using IEC61850 standard interface;

[0152] Step S1-2: Obtain static data from the production management information system and geographic information system, and use ETL tools to achieve periodic synchronization on a daily or weekly basis;

[0153] Steps S1-3: Obtain unstructured data from maintenance reports and meteorological text information, and convert it into structured data through OCR recognition and natural language processing to ensure that the data format is consistent for access;

[0154] OCR recognition stage: For scanned copies of paper maintenance reports or electronic image-formatted text, an adaptive binarization algorithm is first used to process the image, and then the text region is located using the EAST text detection model;

[0155] The EAST text detection model uses PVANet as its backbone network and outputs the coordinates and confidence of the text region after fusing feature maps of different scales.

[0156] Define the text region set as ,in This represents the top-left corner of the k-th text region. and bottom right corner coordinate;

[0157] Subsequently, a CRNN recognition model is used to extract the text content of each region. The CRNN recognition model consists of a convolutional layer for extracting image features, a bidirectional LSTM layer for capturing sequence dependencies, and a transcription layer for mapping features to characters, resulting in an unstructured text sequence. ,in For the region The corresponding recognized text;

[0158] NLP Structured Transformation Stage: A BIO annotation system is constructed based on a power distribution network domain dictionary. Entity annotation is performed on the text sequence T, and a BiLSTM-CRF model is used to identify key entities. The BiLSTM-CRF model includes an embedding layer that converts text into vectors, a bidirectional LSTM layer that captures contextual semantics, and a CRF layer that optimizes the rationality of the annotation sequence. The entity recognition probability calculation formula is as follows:

[0159] ;

[0160] in For input text features, For labeled sequences, Here is the state transition matrix. Here is the emission probability matrix. Normalization factor;

[0161] Key entities identified include device ID, fault type, maintenance time, and meteorological factors;

[0162] Finally, structured data is generated through entity-attribute mapping rules, in the format {Device ID: [Fault Type, Maintenance Time, Meteorological Parameter 1, Meteorological Parameter 2, ...]}.

[0163] First, for real-time data, dynamic information such as equipment operating status is obtained from external dispatch automation systems, metering automation systems, and Supervisory Control and Data Acquisition (SCADA) systems. The IEC 61850 standard interface is used to achieve second-level real-time transmission, ensuring data timeliness. Second, for static data, fixed information such as basic equipment parameters and geographic coordinates is extracted from Production Management Information Systems (PMS) and Geographic Information Systems (GIS). ETL tools are used to synchronize this information on a daily or weekly basis, maintaining the periodic updates of static data. Finally, unstructured data is processed. For scanned maintenance reports, meteorological texts, etc., an OCR recognition process is first performed using adaptive... Binarization algorithms optimize image quality, and then the EAST text detection model, based on PVANet, locates text regions (outputting coordinates and confidence scores). Next, a CRNN recognition model (including convolutional layers, bidirectional LSTM layers, and transcription layers) extracts text sequences. Following this, the system enters the NLP structured transformation stage. A BIO annotation system is constructed based on a power distribution network domain dictionary, and a BiLSTM-CRF model (embedding layer, bidirectional LSTM layer, and CRF layer) is used to identify key entities such as equipment ID, fault type, maintenance time, and meteorological elements. Finally, structured data formats are generated through entity-attribute mapping rules, enabling unified access to all data. This approach covers real-time, static, and unstructured data while ensuring data format consistency through professional technical means, providing a solid data foundation for subsequent power distribution network planning.

[0164] Next, the collected data is preprocessed, specifically as follows:

[0165] In step S2, data preprocessing is performed using a data cleansing method. The specific operation is as follows:

[0166] Step S2-1. Data cleaning; For missing values ​​in the data, continuous data is filled with the mean or median and corrected by combining it with historical data from the same period of the device; discrete data is filled with a pattern and matched with the default parameters of the same type of device.

[0167] Step S2-2. Data Conversion; The collected data is standardized in terms of format and dimensions: the time format is unified to "YYYY-MM-DDHH:MM:SS", the geographic coordinates are unified to the WGS84 coordinate system, and the units of equipment parameters are unified.

[0168] Dimensional normalization: The min-max normalization method is used to normalize data of different magnitudes to the [0,1] interval. The normalization formula is as follows: ;in The original data, For normalized data, The minimum value of this type of data. The maximum value of this type of data;

[0169] Step S2-3. Data correction and rejection: Compare the equipment nameplate parameters with the collected data. If the deviation exceeds 5%, use the nameplate data as the benchmark for calibration.

[0170] The route length is corrected by combining GIS terrain data. First, the starting point of the route is extracted through the GIS system. ,end The WGS84 coordinates, after being converted to planar coordinates using the Gauss-Kruger projection, are substituted into the formula. Calculate the horizontal distance;

[0171] Then, the terrain type of the area traversed by the line is obtained by GIS, and the correction coefficient k for each terrain type is determined according to the power distribution network engineering regulations, ultimately determining the actual length of the line. ;

[0172] Delete duplicate data, data outside the reasonable range, and irrelevant data;

[0173] Step S2-4. Data Integration: Using the "Unique Device ID" as the key field, link data from multiple systems to form a unified data pool that links "device, geography, user, and load".

[0174] This patent employs a four-step refined data processing workflow to ensure the high quality and consistency of power distribution network planning data. The first step is data cleaning, which uses a classification strategy to address missing values ​​in the collected data: continuous data, such as load values ​​and equipment temperatures, are first filled with the mean or median, and then corrected by combining historical operating data from the same period to avoid bias from a single filling method; discrete data, such as equipment types and fault codes, are filled with patterns, i.e., the most frequently occurring values, and matched with the default parameters of similar equipment to ensure the rationality of discrete attributes.

[0175] Secondly, data transformation was performed to achieve format and dimension unification: For format standardization, time was standardized to "YYYY-MM-DDHH:MM:SS" format, geographic coordinates were converted to the WGS84 coordinate system, and equipment parameter units (such as power and length) were standardized according to industry standards; dimension normalization adopted the min-max method, mapping data of different magnitudes (such as voltage, current, and load) to the [0,1] interval, using the following formula: To eliminate the impact of differences in data volume on subsequent model calculations.

[0176] The third step is data correction and elimination to improve data accuracy: The equipment nameplate parameters are compared with the collected data; if the deviation exceeds 5%, the nameplate data is used as the benchmark for calibration. The line length correction process is more detailed. First, the WGS84 coordinates of the line's starting and ending points are extracted using the GIS system, converted to plane coordinates via Gauss-Kruger projection, and the horizontal distance L0 is calculated. Then, based on the terrain type obtained from the GIS (e.g., plains or mountains), the correction coefficient k is determined according to the power distribution network engineering regulations, ultimately yielding the actual length L = L0 × k. Simultaneously, duplicate data, data exceeding reasonable limits (e.g., negative load values), and irrelevant data (e.g., meteorological data unrelated to the power distribution network) are deleted, purifying the data pool. Finally, data integration is performed, using the "unique equipment ID" as the key field to link data from multiple systems such as dispatching, metering, GIS, and PMS, forming a unified, linked data pool containing equipment attributes, geographic information, user information, and load data. This provides a structured, standardized, high-quality data foundation for subsequent topology map construction, model analysis, and planning optimization. The entire process is interconnected, from filling in missing information and standardizing the format, to correcting errors and integrating data, to comprehensively improve data quality and avoid deviations in planning schemes due to data issues.

[0177] When constructing the distribution network topology diagram, the specific steps are as follows:

[0178] The specific steps for constructing the distribution network topology in step S3 are as follows:

[0179] Step S3-1: Identify the core elements of the distribution network and construct a topology diagram from the data using the following formula:

[0180] ;

[0181] in, This represents a set of equipment nodes, including physical equipment such as transformers, lines, and switches. Represents a set of topological edges, characterizing the electrical connections between devices. Represents a set of timestamps. This represents a set of operational scenarios, including normal operation, emergency repair, and maintenance plans.

[0182] Step S3-2. Import each element in the topology graph into the Neo4j graph database, construct the topology network in the form of "node-relationship-node", and support topology traversal query;

[0183] Step S3-3. Topology Consistency Verification: Calculate the model topology signature, count the number of connected edges for each node, and determine the node degree. ,in Represent and sort the i-th device node in the topology graph G;

[0184] Obtain the connected region number to which each node belongs, where the electrical island... ,in Mapping functions for numbering connected components;

[0185] Substitute into the formula ; Generate model signature, where The model topological signature representing the topological graph G. This is an ascending sorting function. The set consisting of the degrees of all nodes. A set consisting of the numbered connected regions of all nodes; "+" indicates that the sets are concatenated.

[0186] Generate a comparison graph topology signature. If the Hamming distance between the two graphs is greater than 2, indicating two or more topological differences, then automatically perform the correction: delete isolated nodes in the graph that do not have a corresponding model ID, add nodes or edges that exist in the model but are missing from the graph, and generate a correction audit log.

[0187] This method constructs a distribution network topology map by first identifying the core elements of the distribution network, treating physical equipment such as transformers, lines, and switches as nodes, and electrical connections between equipment as edges. It also records timestamps and operational scenarios such as normal operation, fault repair, and maintenance plans to form a complete topology map structure. Next, the nodes, edges, timestamps, and scenarios in the topology map are imported into the Neo4j graph database, constructing a topology network in a "node-relationship-node" format to support subsequent topology traversal queries. Finally, topology consistency is verified. First, the model's topology signature is calculated—the number of connected edges for each device node is counted and sorted, and then the connected region number to which each node belongs is obtained. These two pieces of information are combined into a model signature, and then a topology signature for comparison with the graph is generated. If the two differ in more than two places, automatic correction is performed: deleting isolated nodes in the graph that do not correspond to the model ID, supplementing nodes or edges that exist in the model but are missing from the graph, and generating a correction audit log to record the operation process.

[0188] Next, a unified data model for the power distribution network is constructed, specifically as follows:

[0189] In step S4, the unified data model of the distribution network is constructed, and the specific operation is as follows:

[0190] Step S4-1. Use static attributes Dynamic state For device nodes in the topology graph Data quantification; based on the device's unique ID. Collect device types one by one Rated capacity GIS coordinates Annual failure rate ,pass Implement structured storage;

[0191] Real-time acquisition of equipment active power through SCADA / metering automation system reactive power Bus voltage Switch status Equipment temperature Use the formula with a 1-minute step size. Record dynamic changes;

[0192] Topological edges in a topological graph Includes sampling resistor Reactance susceptance ,distance Rated current carrying capacity Through formula Integration;

[0193] Step S4-2. Abnormal Data Processing: Calculate Real-time Active Power Second-order difference: ;in, This is the current timestamp. Timestamps are adjacent 1-minute intervals;

[0194] Calculate the standard deviation of the device's 72-hour power data. Substitute into the formula If the condition is met, the data is marked as abnormal, triggering the SCADA system to issue a real-time re-sampling command for that node.

[0195] Step S4-3. Set the node's static attributes Edge parameters Store the device ID as the primary key in a PostgreSQL relational database and create an index to support batch updates;

[0196] Dynamic state of nodes Stored in the InfluxDB time-series database by partitioning by "Device ID and Timestamp".

[0197] This method constructs a unified data model for the distribution network by first quantifying the data of equipment nodes and edges in the topology graph. For equipment nodes, static attributes and dynamic states are collected separately according to a unique ID. Static attributes include fixed information such as equipment type, rated capacity, GIS coordinates, and annual failure rate, which are stored in a structured manner. Dynamic states are collected in real time through SCADA and metering automation systems, including active power, reactive power, bus voltage, switch status, and equipment temperature, with changes recorded every minute. For topology edges, parameters such as resistance, reactance, susceptance, distance, and rated current carrying capacity are collected and integrated into edge parameters. Next, abnormal data is processed: the second-order difference of real-time active power is calculated, and the standard deviation of the equipment's 72-hour power data is first calculated. If the absolute value of the second-order difference exceeds three times the standard deviation, it is marked as abnormal data, and the SCADA system is triggered to re-collect the node's data. Finally, the data is stored in categories: the node's static attributes and edge parameters are stored in a PostgreSQL relational database with the equipment ID as the primary key and an index is created to support batch updates. The node's dynamic states are partitioned by equipment ID and timestamp and stored in an InfluxDB time-series database.

[0198] To optimize and train the model, specifically:

[0199] In step S5, the unified data model of the distribution network is realized based on fusion verification and topology analysis to achieve topology mapping and fusion. The specific operation is as follows:

[0200] Step S5-1. Implement bidirectional ID indexing based on hashing: Substitute the model-side ID and the front-end SVG element ID of each device ID into the formula. In the formula, The output hash value is represented by id, which represents the model-side device ID or the front-end SVG element ID of the input graphics ID. mod represents the modulo operation, where 65521 is less than 1. Find the largest prime number and calculate its hash value.

[0201] Create a two-way index in the Redis database using hash value, model ID, graph ID, and last synchronization timestamp to enable mapping queries. Clicking on a graph element will allow you to quickly retrieve the corresponding model data using the hash value.

[0202] Step S5-2. Fusion verification to achieve front-end and back-end state synchronization: When the device state changes, the back-end pushes a message to the front-end via the MQTT protocol, and the message carries a state change timestamp. ;

[0203] After receiving the message, the frontend updates the state of the graphic element and records the graphic refresh timestamp. Substitute into the formula If the threshold is exceeded, it is judged as a synchronization abnormality, and the front end calls the MAPI interface to fetch the latest model data again and perform graphics redraw.

[0204] Step S5-3. Topological Difference Quantization: Calculate the topological signature of each model. and graph topology signature ;

[0205] Substitute into the formula ; Calculate the topological dissimilarity, where This indicates the degree of difference between the model and the graph topology. Represents the L2 norm;

[0206] like The correction is performed according to the following priority: first delete redundant nodes in the graph, then add missing edges to the model, and after correction, update the graph signature and synchronize it to the model side.

[0207] Step S5-4. When a line is disconnected or a transformer is out of service in the distribution network, for areas where the N-1 safety check is not met, Dijkstra's algorithm is used to traverse all possible load transfer paths based on the fused graph data.

[0208] For each path, substitute into the formula Calculate the total impedance and select... The path with the shortest path is taken as the optimal load transfer path, in the formula. This represents the total impedance of the current load transfer path. Indicates the first in the path The resistance value of the branch circuit, Indicates the first in the path The reactance value of the branch circuit, This indicates that all branches in the current path will be traversed.

[0209] In step S6, high-reliability prediction of distribution network conditions is achieved based on long short-term memory networks and the Weibull proportional hazards model. The specific operation is as follows:

[0210] Step S6-1. Construct an LSTM load forecasting model: Input features include load data of the past 168 hours, average daily temperature, and holiday labels, where 1 represents a holiday and 0 represents a workday. Output the annual maximum load and peak-valley difference for the next 1-5 years.

[0211] The LSTM load forecasting model uses a three-layer LSTM hidden layer, where the first layer has 64 neurons, dropout is set to 0.2, and L2 regularization is 0.001 to capture intraday load fluctuations.

[0212] The second layer consists of 32 neurons, with parameters basically the same as the first layer, extracting cross-sky patterns;

[0213] The third layer consists of 16 neurons, focusing on the core temporal pattern. Each LSTM layer is followed by a batch normalization layer connected to an intermediate fully connected layer with ReLU activation. The L2 regularization parameter is set to 0.001 and the output layer of the Linear activation function is used. The first 5 dimensions represent the maximum load and the last 5 dimensions represent the peak-to-valley difference.

[0214] The loss function was set to mean squared error, the optimizer was Adam, and the learning rate was set to 0.001.

[0215] Historical load and related data of the power distribution network over the past 5 years were collected and divided into training set, validation set and test set in a ratio of 7:2:1.

[0216] Verify model accuracy using a test set: Calculate the true power for N test samples. With predicted power Substitute the deviation into the formula: Ensure that the model's prediction accuracy meets the requirement of MAPE being less than or equal to 8%. If it does not meet this requirement, readjust the model's input features or hyperparameters until the target is met.

[0217] Step S6-2. Equipment failure prediction and reliability index calculation based on Weibull proportional hazards model;

[0218] Collect operational data from the past 5 years for transformers, lines, and switches in the distribution network, including the equipment's service life. Real-time load rate Ambient temperature and historical fault records;

[0219] The parameters of the Weibull proportional hazards model are fitted using maximum likelihood estimation. The model formula is as follows:

[0220] ;

[0221] in, As the baseline failure rate, Characteristic lifetime, Shape parameters This indicates that the failure rate increases over time. , , These are the weighting coefficients for each influencing factor;

[0222] Reliability is assessed using the SAIDI / SAIFI metrics.

[0223] ;

[0224] ;

[0225] Where t represents the duration of the power outage, n represents the number of affected users, m represents the number of power outages, and N represents the total number of users.

[0226] In step S7, a genetic algorithm is used to optimize the distribution network condition for high reliability. The specific operation is as follows:

[0227] Step S7-1. Construction of the multi-objective optimization objective function: Set three main optimization objectives:

[0228] Reliability objective: Minimize the average system outage duration ;

[0229] Economic objective: Minimize the investment cost of power distribution network planning. ;

[0230] Energy efficiency target: Minimize distribution network line loss rate ;

[0231] Where x represents the distribution network planning scheme to be optimized, including line routes, transformer capacity, distributed power source access parameters, etc.

[0232] The multi-objective optimization is transformed into a single-objective optimization using the weighted summation method. The fitness function formula is as follows:

[0233] ;

[0234] in, , , These are the current SAIDI benchmark value, investment cost benchmark value, and line loss rate benchmark value of the distribution network, respectively, with weighting coefficients satisfying the following conditions: , , ;

[0235] Step S7-2. Preliminary optimization based on genetic algorithm: Generate the Pareto optimal solution set. First, convert the planning scheme into chromosome codes that the algorithm can process, denoted as... ,in This is a binary encoding layer, where m represents the total number of devices or lines to be modified. 1 indicates that new construction or expansion is needed, and 0 indicates that the current status should be maintained;

[0236] This is a real-number encoding layer, where p is the total number of modification parameters. , These are the upper and lower limits of the parameter;

[0237] The initial population was generated using random sampling. N is the population size; then, iterative optimization is performed using three genetic operators, with the selection operator randomly selecting 3 individuals from the population each time. Calculate its fitness value Select the individual with the highest fitness to enter the offspring population, and repeat this process N times to generate the offspring selection population. ;

[0238] Crossover operator pairs Individuals are randomly paired with a probability of 0.8.

[0239] The binary layer uses single-point intersection, with the intersection point randomly selected. ,implement Swap bit values;

[0240] The real number layer uses arithmetic crossover, according to... , Generate new values, where, , The values ​​are the real values ​​of the parent individuals, obtained after crossover. ;

[0241] Mutation operator pairs Individuals mutate with a probability of 0.05, executed at the binary layer. Bit flip;

[0242] Gaussian mutation is used for real number layers, according to Generate new values, where, To vary the asynchronous length, take 5%-10% of the parameter range. Following a standard normal distribution, the offspring population obtained after mutation... ;

[0243] After 50 to 100 iterations, a local optimum is obtained, i.e., the Pareto optimum set;

[0244] Step S7-3. Secondary optimization based on JADE differential evolution algorithm to refine the optimal solution: The Pareto optimal solution set obtained by the genetic algorithm is used as the initial solution; the individual encoding completely follows the two-layer structure of "binary and real number" in step S7-2;

[0245] An elite mutation strategy is employed to modify individuals in the g-th generation population. Two different ordinary individuals were randomly selected. , and one elite individual Mutant individuals are randomly selected from the top 10% of the population and generated using the following formula: ;

[0246] In the formula, This is a scaling factor, initially set to 0.5-1.0, used to control the magnitude of variation;

[0247] Subsequently, crossover operations were used to generate experimental individuals. The mutated individuals were randomly paired with an 80% probability. Crossover was performed separately for the two-layer coding. One crossover point was randomly selected by identifying the modified object through the binary coding layer, and the binary bits after the crossover point of the paired individuals were swapped.

[0248] The modification parameters are recorded through a real-number encoding layer and then arithmetically cross-referenced. The parameters of paired individuals are then weighted and fused using a random coefficient α of 0.2-0.8. The formula is as follows: ;

[0249] Experimental individuals with fused parental traits were generated through cross-generation. ;

[0250] Selection of high-quality individuals is achieved through a selection process: The greedy selection criterion is used to calculate the number of experimental individuals. With the original individual fitness value;

[0251] like fitness ≥ The fitness level is retained. As the next generation of individuals Otherwise, retain the original individual. To ensure that the overall performance of the population does not degrade;

[0252] To optimize the dynamic matching process, after each iteration, only the effective scaling factor F and crossover probability CR that "can generate better individuals" are collected, forming sets S_F and S_CR respectively.

[0253] Update the global baseline parameters using the following formula: , ;

[0254] The next generation of individuals' F from Cauchy ( ,0.1) distribution sampling, CR from Normal ( The parameters are sampled in a 0,1) distribution and truncated to the intervals [0,1.2] and [0,1] respectively to achieve adaptive parameter adjustment;

[0255] Calculate the maximum fitness of the population in each generation. with minimum fitness When 5 consecutive generations meet the requirements This indicates that the solution has stabilized, so stop iterating and output the final optimal planning solution;

[0256] Step S7-4. Reliability Verification of the Optimized Scheme: Substitute the optimal planning scheme output in Step S7-3 into the LSTM load forecasting model in Step S6-1 and the Weibull fault prediction model and reliability index calculation formula in Step S6-2, and recalculate the final reliability index corresponding to the scheme. and ;

[0257] The verification process ends when the results meet the requirements.

[0258] minutes / household / year and per household per year

[0259] If the above criteria are not met, adjust the optimization algorithm parameters as follows: multiply the number of iterations of the genetic algorithm by 1.5, and adjust the reliability weight. Increase by 0.1, and re-execute steps S7-1 to S7-3 to optimize the process until the verification criteria are met.

[0260] This method achieves accurate training and efficient optimization of distribution network models through a closed-loop process of graph-model fusion data support, predictive model training, and multi-algorithm optimization, providing core technical support for high-reliability planning.

[0261] First, the graph-model fusion in step S5 lays the foundation for data consistency in model training: the hash value of the device ID and the graphic SVG element ID is calculated by hashing algorithm, and a bidirectional index is established in the Redis database to realize fast mapping and query of graphic and model data;

[0262] When the device status changes, the backend pushes a message to the frontend via the MQTT protocol. After the frontend updates, it checks the time difference. If it exceeds 1 second, it pulls the latest model data and redraws the graph to ensure that the frontend and backend status are synchronized.

[0263] The topological signature difference between the calculation model and the graph is calculated. If it exceeds 0.5%, redundant nodes and missing edges in the graph are corrected according to priority to ensure the consistency of the topological structure.

[0264] When a fault occurs, the Dijkstra algorithm is used to select the load transfer path with the minimum total impedance based on the fused data, providing the model with a decision-making basis under real fault scenarios.

[0265] Secondly, step S6 completes the prediction model training: construct a three-layer LSTM load prediction model, input the load of the past 168 hours, average daily temperature, and holiday labels, and output the annual maximum load and peak-valley difference for the next 1-5 years. Use the historical data of the past 5 years to divide the training / validation / test sets in a 7:2:1 ratio, and verify the effectiveness of the model by meeting the accuracy requirement of MAPE≤8%.

[0266] We use maximum likelihood estimation to fit the Weibull proportional hazards model, and combine factors such as equipment operating years, real-time load rate, and ambient temperature to predict failures. We calculate the SAIDI / SAIFI reliability index to provide target benchmark values ​​for optimization.

[0267] Finally, step S7 achieves multi-objective optimization through genetic algorithm and JADE differential evolution: construct a multi-objective function with the objectives of minimizing SAIDI, investment cost, and line loss rate, and transform it into a single-objective fitness function using a weighted summation method;

[0268] The genetic algorithm encodes the planning scheme into a two-layer structure: binary for identifying the modification object and real numbers for recording the modification parameters. It generates the Pareto optimal solution set by randomly selecting 3 individuals to select the best, cross-linking binary single-point swapping, real number arithmetic fusion, mutation binary bit flipping, and real number Gaussian mutation operator iterating 50-100 times.

[0269] The JADE algorithm uses this solution set as the initial scheme, adopts an elite mutation strategy to generate new individuals, uses cross-fertilization of parent features, retains high-quality individuals through greedy selection, adaptively adjusts the scaling factor F and crossover probability CR, stops iterating after 5 generations of stabilization, and outputs the optimal scheme. Finally, it verifies that the scheme's SAIDI ≤ 5 minutes / household·year and SAIFI ≤ 5 times / household·year. If these conditions are not met, the algorithm parameters are adjusted and re-optimized.

[0270] The entire process, from ensuring data consistency to training the predictive model, and then to multi-algorithm collaborative optimization and reliability verification, forms a complete model training and optimization system, ensuring the scientific nature and high reliability of the planning scheme.

[0271] The visualization planning platform is specifically designed as follows: it includes a data visualization module, a graphic-model fusion visualization module, a planning scheme visualization module, a simulation module, and a platform interaction and compatibility module. The construction method is as follows:

[0272] The data visualization module uses a dashboard as its core, integrating real-time load heatmaps, equipment failure rate statistics pie charts, and key indicator components such as SAIDI / SAIFI trend line charts; it also supports multi-condition filtering by region and time dimension.

[0273] The graphic model fusion visualization module integrates GIS geographic data with power distribution network topology models to achieve linked display of geographic wiring diagrams and topology diagrams; clicking on graphic elements allows you to view detailed model data of the corresponding equipment, and modifying model parameters will synchronously update the graphic status; the built-in fault simulation function allows the system to automatically calculate and visualize the optimal load transfer path and power outage range after setting equipment faults.

[0274] The planning scheme visualization module provides route adjustment maps, transformer capacity configuration comparison tables, and reliability index improvement bar charts. In addition, it realizes the 3D preview function of the scheme based on GIS terrain data, which shows the spatial layout relationship of lines and transformers from a three-dimensional perspective, and helps to evaluate the engineering feasibility of the planning scheme.

[0275] The simulation module supports the simulation of distribution network operation status in multiple scenarios, including extreme weather and typical load growth scenarios; it uses visualization to present the voltage drop range and line loss hotspots in the scenarios, helping users to verify the adaptability and stability of the planning scheme in extreme or future scenarios;

[0276] The platform's interaction and compatibility modules support drag-and-drop scheme adjustments; provide multi-format export functionality, generating CAD drawings, Excel parameter tables, and PDF reports; are compatible with tablets and mobile devices; and feature a three-tiered permission system: administrator, planner, and approver. Administrators are responsible for system configuration and data management, planners can edit schemes, and approvers are responsible for approval.

[0277] In one embodiment, refer to Figure 2 , Figure 3 and Figure 4 As can be seen, this method demonstrates significant advantages over traditional methods in three core dimensions: consistency of distribution network planning map and model fusion, accuracy of load forecasting, and efficiency of optimization algorithms, providing key technical support for high-reliability distribution network planning.

[0278] First, regarding topology discrepancy, as the number of distribution network devices increases from 100 to 500, the topology discrepancy of traditional methods rapidly climbs to over 2.5%, while this method consistently remains at an extremely low level below 0.5%. This indicates that this method, through hash bidirectional indexing and fusion verification mechanisms, effectively solves the problem of inconsistent graphical and model data in large-scale device scenarios, ensuring real-time synchronization between graphical elements and model data, and providing a reliable topology foundation for subsequent analysis and decision-making. Second, regarding load forecasting accuracy, compared with traditional models such as ARIMA and BP neural networks, the LSTM forecasting model of this method consistently maintains an average absolute percentage error (MAPE) below 7.5% over a forecasting period of 1-5 years, with minimal increase in MAPE with increasing forecasting time. In contrast, the MAPE of the ARIMA model rises from approximately 3% to 15%, and the BP model exceeds 15%, fully demonstrating the stability and high accuracy of this method's load forecasting model in long-term forecasting. This provides accurate load data support for distribution network capacity planning and line upgrades over the next 5 years, avoiding resource waste or power shortages due to forecasting errors.

[0279] Finally, in terms of optimizing algorithm efficiency, the combination of genetic and JADE differential evolution algorithms used in this method converges much faster than the simple genetic algorithm: when iterated to the 20th generation, the fitness of this method is close to 0.9, while that of the simple genetic algorithm is only about 0.6; when iterated to the 50th generation, the fitness of this method is stable above 0.95, while that of the simple genetic algorithm has not yet broken through 0.85.

[0280] This demonstrates that the proposed optimization process can quickly converge to the optimal solution, significantly shortening the calculation time for planning schemes and improving planning efficiency, making it particularly suitable for multi-objective, large-scale distribution network planning scenarios. In summary, through technological innovations in graph-model fusion, prediction models, and optimization algorithms, this method effectively addresses the pain points of traditional distribution network planning, such as data inconsistency, low prediction accuracy, and slow optimization efficiency, providing strong support for building a distribution network that combines high reliability, economy, and energy efficiency.

[0281] Working principle: This method is based on multi-source data acquisition and processing, supported by graph-model fusion as the core technology, combined with prediction and optimization algorithms, and finally realizes high-reliability distribution network planning through a visualization platform.

[0282] First, real-time, static, and unstructured data are collected from systems such as scheduling automation, metering automation, and SCADA. The unstructured data is then converted into a structured format through OCR recognition and NLP. Next, the data is cleaned, missing values ​​are filled, a unified format is converted and normalized, calibration parameters are corrected and removed, invalid data is deleted, and the data is integrated to form a unified data pool associated with the unique ID of the equipment.

[0283] Then, a power distribution network topology diagram containing device nodes, electrical connection edges, timestamps, and operating scenarios is constructed and imported into the Neo4j database. At the same time, a unified data model is established to quantify the static attribute types, capacity, etc. of devices and the dynamic state power, voltage, etc., and stored in PostgreSQL and InfluxDB databases respectively.

[0284] Then, a bidirectional hash index is used to map the graph to the model ID, the MQTT protocol is used to ensure the synchronization of the front-end and back-end states, and topology difference quantification correction is used to ensure graph-model consistency. The optimal load transfer path can be calculated in case of failure. Then, the LSTM model is used to predict the load for the next 5 years, and the Weibull proportional hazards model is used to predict equipment failure and calculate the SAIDI / SAIFI reliability index. Subsequently, the Pareto optimal solution set is generated by the genetic algorithm, and the planning scheme that meets the reliability requirements is obtained by JADE differential evolution secondary optimization.

[0285] Finally, a visualization platform was developed to display data indicators, interactive graphics, comparison of planning schemes, and simulation results, thereby enabling visualization and efficient decision-making in power distribution network planning.

[0286] This method addresses the problems of inconsistent map and model data, low prediction accuracy, slow optimization efficiency, poor data quality, and insufficient visualization in power distribution network planning, and achieves a systematic solution through full-process technological innovation.

[0287] First, to address the issues of scattered multi-source data and the difficulty in utilizing unstructured data, real-time / static data is collected from systems such as scheduling automation, metering automation, and SCADA. Maintenance reports and meteorological texts are converted into structured data using OCR (EAST text detection) and CRNN recognition. Then, the data undergoes cleaning (filling missing values ​​with mean or median), conversion (unifying time format, geographic coordinates, units, and normalization), correction and removal (calibrating nameplate parameters), deletion of invalid data, and integration (associating data with unique equipment IDs to form a unified data pool of "equipment-geography-user-load"), thus resolving the problems of poor data quality and inconsistent formats.

[0288] Secondly, to address the issues of asynchronous graph and model data and inconsistent topology, a topology graph containing device nodes, electrical connection edges, timestamps, and operating scenarios is constructed and imported into the Neo4j database. A bidirectional hash index is used to achieve fast mapping between the graph and the model ID. Status change messages are pushed via the MQTT protocol. If the time difference exceeds 1 second, the data is pulled and redrawn. Topology differences are quantitatively corrected. When the difference is >0.5%, redundant nodes are deleted and missing edges are added to ensure real-time synchronization of graph and model data and topology consistency.

[0289] Furthermore, to address the long-term issues of low load forecasting accuracy and inaccurate fault prediction, a three-layer LSTM load forecasting model is constructed. Inputting 168-hour load, daily average temperature, and holiday labels ensures high accuracy of MAPE ≤ 8%. The Weibull proportional hazards model is used in conjunction with equipment operating years, load rate, and ambient temperature to predict faults, and the SAIDI / SAIFI reliability index is calculated to provide a reliable basis for optimization.

[0290] Then, to address the issues of low efficiency in multi-objective optimization and difficulty in balancing reliability and economy, a genetic algorithm is used to generate the Pareto optimal solution set, followed by secondary optimization using the JADE differential evolution algorithm, an elite mutation strategy, and adaptive parameter adjustment to ensure that the optimized solution meets the reliability requirements of SAIDI≤5 minutes / household·year and SAIFI≤5 times / household·year.

[0291] Finally, to address the issues of a lack of intuitiveness and poor interactivity in the planning process, a visualization platform was developed. This platform supports real-time load heat maps and fault simulations, allowing users to intuitively view load transfer paths, compare planning schemes, adjust lines, configure capacity, improve reliability indicators, and preview in 3D. It enables two-way interaction between graphics and data, thereby improving planning efficiency and decision-making transparency.

[0292] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A high-reliability power distribution network visualization planning method based on graph-module fusion technology, characterized in that: Includes the following steps: Step S1: Multi-source data acquisition, obtaining real-time / static / unstructured data from multiple system sources; Step S2: Perform data preprocessing using data cleansing methods; Step S3: Construct the distribution network topology; Step S4: Construct a unified data model for the power distribution network; Step S5: Achieve topology mapping and fusion of the unified data model of the distribution network based on fusion verification and topology analysis; Step S6: Achieve high-reliability prediction of distribution network conditions based on long short-term memory networks and the Weibull proportional hazards model; Step S7: Utilize a genetic algorithm to achieve high reliability optimization of the power distribution network conditions; Step S8: Generate a visual planning platform.

2. The highly reliable distribution network visualization planning method based on graph-model fusion technology according to claim 1, characterized in that: In step S1, multi-source data acquisition is performed to obtain real-time / static / unstructured data from multiple system sources. The specific operation is as follows: Step S1-1: Obtain real-time data from external dispatch automation system, metering automation system, monitoring and data acquisition system, and achieve second-level real-time transmission using IEC61850 standard interface; Step S1-2: Obtain static data from the production management information system and geographic information system, and use ETL tools to achieve periodic synchronization on a daily or weekly basis; Steps S1-3: Obtain unstructured data from maintenance reports and meteorological text information, and convert it into structured data through OCR recognition and natural language processing to ensure that the data format is consistent for access; OCR recognition stage: For scanned copies of paper maintenance reports or electronic image format text, an adaptive binarization algorithm is first used to process the image, and then the text region is located by the EAST text detection model; The EAST text detection model uses PVANet as its backbone network and outputs the coordinates and confidence of the text region after fusing feature maps of different scales. Define the text region set as ,in This represents the top-left corner of the k-th text region. and bottom right corner coordinate; Subsequently, a CRNN recognition model is used to extract the text content of each region. The CRNN recognition model consists of a convolutional layer for extracting image features, a bidirectional LSTM layer for capturing sequence dependencies, and a transcription layer for mapping features to characters, resulting in an unstructured text sequence. ,in For the region The corresponding recognized text; NLP Structured Transformation Stage: A BIO annotation system is constructed based on a power distribution network domain dictionary. Entity annotation is performed on the text sequence T, and a BiLSTM-CRF model is used to identify key entities. The BiLSTM-CRF model includes an embedding layer that converts text into vectors, a bidirectional LSTM layer that captures contextual semantics, and a CRF layer that optimizes the rationality of the annotation sequence. The entity recognition probability calculation formula is as follows: ； in For input text features, For labeled sequences, Here is the state transition matrix. Here is the emission probability matrix. Normalization factor; Key entities identified include device ID, fault type, maintenance time, and meteorological factors; Finally, structured data is generated through entity-attribute mapping rules, in the format {Device ID: [Fault Type, Maintenance Time, Meteorological Parameter 1, Meteorological Parameter 2, ...]}.

3. The highly reliable distribution network visualization planning method based on graph-model fusion technology according to claim 1, characterized in that: In step S2, data preprocessing is performed using a data cleansing method. The specific operation is as follows: Step S2-1. Data cleaning: For missing values ​​in the data, continuous data is filled with the mean or median and corrected by combining it with historical data from the same period of the device; discrete data is filled with a pattern and matched with the default parameters of the same type of device. Step S2-2. Data Conversion; The collected data is standardized in terms of format and dimensions: the time format is unified to "YYYY-MM-DDHH:MM:SS", the geographic coordinates are unified to the WGS84 coordinate system, and the units of equipment parameters are unified. Dimensional normalization: The min-max normalization method is used to normalize data of different magnitudes to the [0,1] interval. The normalization formula is as follows: ,in The original data, For normalized data, The minimum value of this type of data. The maximum value of this type of data; Step S2-3. Data correction and rejection: Compare the equipment nameplate parameters with the collected data. If the deviation exceeds 5%, use the nameplate data as the benchmark for calibration. The route length is corrected by combining GIS terrain data. First, the starting point of the route is extracted through the GIS system. ,end The WGS84 coordinates, after being converted to planar coordinates using the Gauss-Kruger projection, are substituted into the formula. Calculate the horizontal distance; Then, the terrain type of the area traversed by the line is obtained by GIS, and the correction coefficient k for each terrain type is determined according to the power distribution network engineering regulations, ultimately determining the actual length of the line. ; Delete duplicate data, data outside the reasonable range, and irrelevant data; Step S2-4. Data Integration: Using "Unique Device ID" as the key field, link data from multiple systems to form a unified data pool that links "device, geography, user, and load".

4. The highly reliable distribution network visualization planning method based on graph-model fusion technology according to claim 1, characterized in that: The specific steps for constructing the distribution network topology in step S3 are as follows: Step S3-1: Identify the core elements of the distribution network and construct a topology diagram from the data using the following formula: ； in, This represents a set of equipment nodes, including physical equipment such as transformers, lines, and switches. Represents a set of topological edges, characterizing the electrical connections between devices. Represents a set of timestamps. This represents a set of operational scenarios, including normal operation, emergency repair, and maintenance plans. Step S3-2. Import each element in the topology graph into the Neo4j graph database, construct the topology network in the form of "node-relationship-node", and support topology traversal query; Step S3-3. Topology Consistency Verification: Calculate the model topology signature, count the number of connected edges for each node, and determine the node degree. ,in Represent and sort the i-th device node in the topology graph G; Obtain the connected region number to which each node belongs, where the electrical island... ,in Mapping functions for numbering connected components; Substitute into the formula ; Generate model signature, where The model topological signature representing the topological graph G. This is an ascending sorting function. The set consisting of the degrees of all nodes. A set consisting of the numbered connected regions of all nodes; "+" indicates that the sets are concatenated. Generate a comparison graph topology signature. If the Hamming distance between the two graphs is greater than 2, indicating two or more topological differences, then automatically perform the correction: delete isolated nodes in the graph that do not have a corresponding model ID, add nodes or edges that exist in the model but are missing from the graph, and generate a correction audit log.

5. The highly reliable distribution network visualization planning method based on graph-model fusion technology according to claim 1, characterized in that: In step S4, the unified data model of the distribution network is constructed, and the specific operation is as follows: Step S4-1. Use static attributes Dynamic state For device nodes in the topology graph Data quantification; based on the device's unique ID. Collect device types one by one Rated capacity GIS coordinates Annual failure rate ,pass ; Implement structured storage; Real-time acquisition of equipment active power through SCADA / metering automation system reactive power Bus voltage Switch status Equipment temperature Using the formula with a 1-minute step size: Record dynamic changes; Topological edges in a topological graph Includes sampling resistor Reactance susceptance ,distance Rated current carrying capacity Through formula Integration; Step S4-2. Abnormal Data Processing: Calculate Real-time Active Power Second-order difference: ;in, This is the current timestamp. Timestamps are adjacent 1-minute intervals; Calculate the standard deviation of the device's 72-hour power data. Substitute into the formula If the condition is met, the data is marked as abnormal, triggering the SCADA system to issue a real-time re-sampling command for that node. Step S4-3. Set the node's static attributes edge parameters Store the device ID as the primary key in a PostgreSQL relational database and create an index to support batch updates; Dynamic state of nodes Stored in the InfluxDB time-series database by partitioning by "Device ID and Timestamp".

6. The highly reliable distribution network visualization planning method based on graph-model fusion technology according to claim 1, characterized in that: In step S5, the unified data model of the distribution network is realized based on fusion verification and topology analysis to achieve topology mapping and fusion. The specific operation is as follows: Step S5-1. Implement bidirectional ID indexing based on hashing: Substitute the model-side ID and the front-end SVG element ID of each device ID into the formula. In the formula, The output hash value is represented by id, which represents the model-side device ID or the front-end SVG element ID of the input graphics ID. mod represents the modulo operation, where 65521 is less than 1. Find the largest prime number and calculate its hash value. Create a two-way index in the Redis database using hash value, model ID, graph ID, and last synchronization timestamp to enable mapping queries. Clicking on a graph element will allow you to quickly retrieve the corresponding model data using the hash value. Step S5-2. Fusion verification to achieve front-end and back-end state synchronization: When the device state changes, the back-end pushes a message to the front-end via the MQTT protocol, and the message carries a state change timestamp. ; After receiving the message, the frontend updates the state of the graphic element and records the graphic refresh timestamp. Substitute into the formula If the threshold is exceeded, it is judged as a synchronization abnormality, and the front end calls the MAPI interface to fetch the latest model data again and perform graphics redraw. Step S5-3. Topological Difference Quantization: Calculate the topological signature of each model. and graph topology signature ; Substitute into the formula ; Calculate the topological dissimilarity, where This indicates the degree of difference between the model and the graph topology. Represents the L2 norm; like The correction is performed according to the following priority: first delete redundant nodes in the graph, then add missing edges to the model, and after correction, update the graph signature and synchronize it to the model side. Step S5-4. When a line is disconnected or a transformer is out of service in the distribution network, for areas where the N-1 safety check is not met, Dijkstra's algorithm is used to traverse all possible load transfer paths based on the fused graph data. For each path, substitute into the formula Calculate the total impedance and select... The path with the shortest path is taken as the optimal load transfer path, in the formula. This represents the total impedance of the current load transfer path. Indicates the first in the path The resistance value of the branch circuit, Indicates the first in the path The reactance value of the branch circuit, This indicates that all branches in the current path will be traversed.

7. The highly reliable distribution network visualization planning method based on graph-model fusion technology according to claim 1, characterized in that: In step S6, high-reliability prediction of distribution network conditions is achieved based on long short-term memory networks and the Weibull proportional hazards model. The specific operation is as follows: Step S6-1. Construct an LSTM load forecasting model: Input features include load data of the past 168 hours, average daily temperature, and holiday labels, where 1 represents a holiday and 0 represents a workday. Output the annual maximum load and peak-to-valley difference for the next 1-5 years. The LSTM load forecasting model uses a three-layer LSTM hidden layer, where the first layer has 64 neurons, dropout is set to 0.2, and L2 regularization is 0.001 to capture intraday load fluctuations. The second layer consists of 32 neurons, with parameters basically the same as the first layer, extracting cross-sky patterns; The third layer consists of 16 neurons, focusing on the core temporal pattern. Each LSTM layer is followed by a batch normalization layer connected to an intermediate fully connected layer with ReLU activation. The L2 regularization parameter is set to 0.001 and the output layer of the Linear activation function. The first 5 dimensions represent the maximum load, and the last 5 dimensions represent the peak-to-valley difference. The loss function was set to mean squared error, the optimizer was Adam, and the learning rate was set to 0.

001. Historical load and related data of the power distribution network over the past 5 years were collected and divided into training set, validation set and test set in a ratio of 7:2:

1. Verify model accuracy using a test set: Calculate the true power for N test samples. With predicted power Substitute the deviation into the formula: Ensure that the model's prediction accuracy meets the requirement of MAPE being less than or equal to 8%. If it does not meet this requirement, readjust the model's input features or hyperparameters until the target is met. Step S6-2. Equipment failure prediction and reliability index calculation based on Weibull proportional hazards model; Collect operational data from the past 5 years for transformers, lines, and switches in the distribution network, including the equipment's service life. Real-time load rate Ambient temperature and historical fault records; The parameters of the Weibull proportional hazards model are fitted using maximum likelihood estimation. The model formula is as follows: ； in, As the baseline failure rate, Characteristic lifetime, Shape parameters This indicates that the failure rate increases over time. , , These are the weighting coefficients for each influencing factor; Reliability is assessed using the SAIDI / SAIFI metrics. ； ； Where t represents the duration of the power outage, n represents the number of affected users, m represents the number of power outages, and N represents the total number of users.

8. The highly reliable distribution network visualization planning method based on graph-model fusion technology according to claim 7, characterized in that: In step S7, a genetic algorithm is used to optimize the distribution network condition for high reliability. The specific operation is as follows: Step S7-1. Construction of the multi-objective optimization objective function: Set three main optimization objectives: Reliability objective: Minimize the average system outage duration ; Economic objective: Minimize the investment cost of power distribution network planning. ; Energy efficiency target: Minimize distribution network line loss rate ; Where x represents the distribution network planning scheme to be optimized, including line routes, transformer capacity, distributed power source access parameters, etc. The multi-objective optimization is transformed into a single-objective optimization using the weighted summation method. The fitness function formula is as follows: ； in, , , These are the current SAIDI benchmark value, investment cost benchmark value, and line loss rate benchmark value of the distribution network, respectively, with weighting coefficients satisfying the following conditions: , , ; Step S7-2. Preliminary optimization based on genetic algorithm: Generate the Pareto optimal solution set. First, convert the planning scheme into chromosome codes that the algorithm can process, denoted as... ,in This is a binary encoding layer, where m represents the total number of devices or lines to be modified. 1 indicates that new construction or expansion is needed, and 0 indicates that the current status should be maintained; This is a real-number encoding layer, where p is the total number of modification parameters. , These are the upper and lower limits of the parameter; The initial population was generated using random sampling. N is the population size; then, iterative optimization is performed using three genetic operators, with the selection operator randomly selecting 3 individuals from the population each time. Calculate its fitness value Select the individual with the highest fitness to enter the offspring population, and repeat this process N times to generate the offspring selection population. ; Crossover operator pairs Individuals are randomly paired with a probability of 0.

8. The binary layer uses single-point intersection, with the intersection point randomly selected. ,implement Swap bit values; For the real number layer, use arithmetic crossover, as follows: , Generate new values, where, , The values ​​are the real values ​​of the parent individuals, obtained after crossover. ; Mutation operator pairs Individuals mutate with a probability of 0.05, executed at the binary layer. Bit flip; Gaussian mutation is used for real number layers, according to Generate new values, where, To vary the asynchronous length, take 5%-10% of the parameter range. Following a standard normal distribution, the offspring population obtained after mutation... ; After 50 to 100 iterations, a local optimum is obtained, i.e., the Pareto optimum set; Step S7-3. Secondary optimization based on JADE differential evolution algorithm to refine the optimal solution: The Pareto optimal solution set obtained by the genetic algorithm is used as the initial solution; the individual encoding completely follows the "binary and real number" two-layer structure of step S7-2; An elite mutation strategy is employed to modify individuals in the g-th generation population. Two different ordinary individuals were randomly selected. , and one elite individual Mutant individuals are randomly selected from the top 10% of the population and generated using the following formula: ; In the formula, This is a scaling factor, initially set to 0.5-1.0, used to control the magnitude of variation; Subsequently, crossover operations were used to generate experimental individuals. The mutated individuals were randomly paired with an 80% probability. Crossover was performed separately for the two-layer coding. One crossover point was randomly selected by identifying the modified object through the binary coding layer, and the binary bits after the crossover point of the paired individuals were swapped. The modification parameters are recorded through a real-number encoding layer and then arithmetically cross-referenced. The parameters of paired individuals are then weighted and fused using a random coefficient α of 0.2-0.

8. The formula is as follows: ; Experimental individuals with fused parental traits were generated through cross-generation. ; Selection of high-quality individuals is achieved through a selection process: The greedy selection criterion is used to calculate the number of experimental individuals. With the original individual fitness value; like fitness ≥ The fitness level is retained. As the next generation of individuals Otherwise, retain the original individual. To ensure that the overall performance of the population does not degrade; To optimize the dynamic matching process, after each iteration, only the effective scaling factor F and crossover probability CR that "can generate better individuals" are collected, forming sets S_F and S_CR respectively. Update the global baseline parameters using the following formula: ; ; The next generation of individuals' F from Cauchy ( ,0.1) distribution sampling, CR from Normal ( The parameters are sampled in a 0,1) distribution and truncated to the intervals [0,1.2] and [0,1] respectively to achieve adaptive parameter adjustment; Calculate the maximum fitness of the population in each generation. with minimum fitness When 5 consecutive generations meet the requirements This indicates that the solution has stabilized, so stop iterating and output the final optimal planning solution; Step S7-4. Reliability Verification of the Optimized Scheme: Substitute the optimal planning scheme output in Step S7-3 into the LSTM load forecasting model in Step S6-1 and the Weibull fault prediction model and reliability index calculation formula in Step S6-2, and recalculate the final reliability index corresponding to the scheme. and ; The verification process ends when the results meet the requirements. minutes / household / year and per household per year If the above criteria are not met, adjust the optimization algorithm parameters as follows: multiply the number of iterations of the genetic algorithm by 1.5, and adjust the reliability weight. Increase by 0.1, and re-execute steps S7-1 to S7-3 to optimize the process until the verification criteria are met.

9. The highly reliable distribution network visualization planning method based on graph-model fusion technology according to claim 1, characterized in that: The visualization planning platform in step S8 includes a data visualization module, a graphic-model fusion visualization module, a planning scheme visualization module, a simulation module, and a platform interaction and compatibility module, constructed as follows: The data visualization module uses a dashboard as its core platform, integrating key indicator components such as real-time load heatmaps, equipment failure rate statistics pie charts, and SAIDI / SAIFI trend line charts; it also supports multi-condition filtering by region and time dimension. The graphic model fusion visualization module integrates GIS geographic data with power distribution network topology models to achieve linked display of geographic wiring diagrams and topology diagrams; clicking on graphic elements allows you to view detailed model data of the corresponding devices, and modifying model parameters will synchronously update the graphic status; The built-in fault simulation function allows the system to automatically calculate and visualize the optimal load transfer path and power outage range after a device fault is set. The planning scheme visualization module provides route adjustment diagrams, transformer capacity configuration comparison tables, and reliability index improvement bar charts. In addition, a 3D preview function for the plan is implemented based on GIS terrain data, which displays the spatial layout relationship of lines and transformers from a three-dimensional perspective, and helps to evaluate the engineering feasibility of the planning scheme. The simulation module supports the simulation of distribution network operation status in multiple scenarios, including extreme weather and typical load growth scenarios; it uses visualization to present the voltage drop range and line loss hotspots in the scenarios, helping users to verify the adaptability and stability of the planning scheme in extreme or future scenarios; The platform's interaction and compatibility modules support drag-and-drop scheme adjustments; provide multi-format export functionality, generating CAD drawings, Excel parameter tables, and PDF reports; are compatible with tablets and mobile devices; and feature a three-tiered permission system: administrator, planner, and approver. Administrators are responsible for system configuration and data management, planners can edit schemes, and approvers are responsible for approval.

Images