A method, system, electronic equipment, and medium for generating power outage maintenance plans for distribution networks.

Abstract

This invention relates to the field of distribution network maintenance technology, and particularly to a method, system, electronic equipment, and medium for generating distribution network outage maintenance plans. The method includes: acquiring multi-source distribution network operation data to obtain a distribution network data pool; extracting transformer area features, line layer features, and user layer features from the distribution network data pool, and constructing a three-level risk linkage mapping relationship; constructing a frequent outage risk prediction model based on the three-level risk linkage mapping relationship; using the risk prediction results, real-time distribution network status, and maintenance resource data as input, and generating an adaptive maintenance plan by introducing a dynamic reward mechanism; and generating an adaptive distribution network outage maintenance plan based on the adaptive maintenance plan. This approach solves the technical problem of existing distribution network outage maintenance plan formulation technologies being unable to simultaneously balance multiple constraints in real time, improving the dynamic adaptability, overall coordination, and robustness of the maintenance plan's execution.

Description

Technical Field

[0001] This invention relates to the field of power distribution network maintenance technology, and in particular to a method, system, electronic device and medium for generating power outage maintenance plans for power distribution networks. Background Technology

[0002] With the acceleration of urbanization and the continuous increase in the proportion of renewable energy integration, the scale of the power distribution network continues to expand, the line structure is becoming increasingly complex, and users' electricity demand is showing diversified and differentiated characteristics. As the end link of the power system, the distribution network directly connects various types of electricity users, and its power supply reliability and the rationality of maintenance and scheduling profoundly affect users' electricity experience and socio-economic activities. At the same time, the operating status of the distribution network is affected by multiple factors such as load fluctuations, weather changes, and equipment aging, significantly increasing the risk of frequent power outages. Maintenance plans need to take into account multiple constraints such as the distribution network's power supply capacity, equipment maintenance cycles, and user power supply guarantees. Traditional maintenance and scheduling models are no longer suitable for the complex and ever-changing operational needs of the distribution network.

[0003] In the field of distribution network outage maintenance planning, existing technologies have established a certain foundation. Currently, maintenance plans generally rely on manual experience or fixed templates, lacking the ability to adapt to dynamic factors. Existing technology (application publication number CN114493476A) discloses a method for formulating a coordinated maintenance plan for the main and distribution networks. This method establishes a coordinated maintenance auxiliary decision-making model and generates a multi-level maintenance plan table through steps such as provincial and local power supply and dispatching status analysis, main and distribution network coordination analysis, user impact analysis, and maintenance risk analysis. This method reduces the workload of entirely manual analysis by comprehensively considering the main and distribution network operating conditions and user load data, and reduces the impact of maintenance on production and daily life. However, existing technologies focus more on the coordination and static planning between the main and distribution networks, failing to fully consider the real-time fluctuation characteristics of the distribution network's internal operating status and the dynamic allocation needs of maintenance resources. Specifically, existing technologies suffer from two prominent problems: First, the planning process lacks a dynamic linkage mechanism with risk prediction results, real-time distribution network status, and maintenance resource data, failing to respond in real time to changes in risk, load fluctuations, and resource supply and demand adjustments during distribution network operation, resulting in insufficient adaptability of maintenance plans to actual operating conditions. Second, existing methods lack a three-dimensional overall design encompassing timing, regional priority, and real-time feedback, making it difficult to dynamically optimize maintenance strategies based on user electricity consumption periods, regional risk levels, and resource consumption during plan execution. This can easily lead to problems such as increased impact on highly sensitive users' electricity consumption, delayed maintenance in high-risk areas, or resource misallocation. These limitations are intertwined, making traditional maintenance plans inflexible and rigidly enforced in complex distribution network environments, resulting in a need to improve overall maintenance efficiency and power supply reliability. Therefore, existing distribution network outage maintenance planning technologies suffer from the technical problem of failing to balance multiple constraints in real time. Summary of the Invention

[0004] In view of the above-mentioned shortcomings or disadvantages, the present invention provides a method, system, electronic device and medium for generating power outage maintenance plans for distribution networks, which can solve the technical problem that existing power outage maintenance plan formulation technologies cannot take into account multiple constraints in real time.

[0005] This invention provides a method for generating a power outage maintenance plan for a distribution network, comprising: Obtain multi-source distribution network operation data and perform fusion preprocessing on the multi-source distribution network operation data, including time sequence alignment, outlier handling, and missing value completion, to obtain the distribution network data pool.

[0006] Features at the distribution network level, line level, and user level are extracted from the distribution network data pool, and a three-level risk linkage mapping relationship is constructed based on these features.

[0007] Based on the three-level risk linkage mapping relationship, and combined with the power supply capacity constraints of the distribution network, the equipment maintenance cycle constraints, and the user power supply guarantee constraints, a frequent power outage risk prediction model is constructed.

[0008] By taking risk prediction results, real-time distribution network status, and maintenance resource data as inputs, and using reinforcement learning with a dynamic reward mechanism for iterative optimization, an adaptive maintenance plan is generated.

[0009] Based on the adaptive maintenance plan, an adaptive maintenance plan for power outages in the distribution network is generated, including timing arrangements, regional priorities, and real-time feedback.

[0010] According to a second aspect, the present invention provides a power distribution network outage maintenance plan generation system, comprising: The distribution network data pool creation module is used to acquire multi-source distribution network operation data and perform fusion preprocessing on the multi-source distribution network operation data, including time sequence alignment, outlier handling, and missing value completion, to obtain the distribution network data pool.

[0011] The risk linkage mapping relationship construction module is used to extract distribution area layer features, line layer features and user layer features from the distribution network data pool, and construct a three-level risk linkage mapping relationship based on the distribution area layer features, line layer features and user layer features.

[0012] The frequent power outage risk prediction model construction module is used to construct a frequent power outage risk prediction model based on the three-level risk linkage mapping relationship and combined with the power supply capacity constraints of the distribution network, equipment maintenance cycle constraints, and user power supply guarantee constraints.

[0013] The adaptive maintenance plan generation module takes risk prediction results, real-time distribution network status, and maintenance resource data as input, and iteratively optimizes them through reinforcement learning with a dynamic reward mechanism to generate an adaptive maintenance plan.

[0014] The power outage maintenance plan generation module is used to generate an adaptive power outage maintenance plan for the distribution network based on the adaptive maintenance scheme, including timing, regional priority, and real-time feedback.

[0015] According to a third aspect, the present invention provides an electronic device comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to execute any of the distribution network outage maintenance plan generation methods in the embodiments of the present invention.

[0016] According to another aspect of the present invention, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute any of the distribution network power outage maintenance plan generation methods in the embodiments of the present invention.

[0017] The present invention provides a method for generating power outage maintenance plans for distribution networks. This method is achieved through five core steps: data fusion preprocessing, three-level risk linkage mapping, construction of a multi-constraint risk prediction model, iterative optimization of the scheme based on reinforcement learning, and adaptive plan generation. This process involves several key steps: First, multi-source distribution network operation data undergoes fusion preprocessing, including time-series alignment, outlier handling, and missing value completion, to integrate and standardize the data, laying the foundation for high-quality data analysis. Second, distribution area, line, and user-level features are extracted from the distribution network data pool, and a three-tiered risk linkage mapping relationship is constructed to achieve a global and comprehensive understanding of distribution network risks. Third, a frequent power outage risk prediction model is built based on this three-tiered risk linkage mapping relationship, combined with distribution network power supply capacity constraints, equipment maintenance cycle constraints, and user power supply guarantee constraints, to integrate multiple practical constraints into risk quantification assessment. Fourth, risk prediction results, real-time distribution network status, and maintenance resource data are used as inputs, and iterative optimization is achieved through reinforcement learning with a dynamic reward mechanism to generate adaptive maintenance plans, enabling dynamic adaptation to complex distribution network conditions and resource requirements. Finally, an adaptive power outage maintenance plan, including time-series scheduling, regional priority, and real-time feedback, is generated based on the adaptive maintenance plan to guide maintenance execution with three-dimensional adaptive scheduling capabilities.

[0018] In this technical solution, the present invention addresses the problems mentioned in the background art, such as the reliance on manual experience in traditional maintenance plans, the lack of dynamic linkage between risk prediction results, real-time distribution network status, and maintenance resource data. It constructs a three-level risk linkage mapping relationship and combines reinforcement learning to iteratively optimize and generate adaptive maintenance plans, achieving real-time dynamic matching between maintenance plans and risk changes, load fluctuations, and resource supply and demand, thereby improving the adaptability of the plans to the actual operating conditions. Addressing the problem mentioned in the background art, such as the lack of a three-dimensional overall design of time-series scheduling, regional priority, and real-time feedback in existing maintenance plans, the present invention generates an adaptive plan that includes time-series scheduling, regional priority, and real-time feedback, constructing a multi-dimensional collaborative scheduling framework with time-series, spatial, and process feedback. This solves the defects of conflict between maintenance periods and user electricity demand, maintenance delays in high-risk areas, and rigid plan execution. Finally, addressing the problem mentioned in the background art, such as the inability of traditional methods to respond to dynamic factors leading to low maintenance efficiency and power supply reliability, the present invention continuously optimizes the plan through reinforcement learning based on a dynamic reward mechanism and introduces a real-time feedback mechanism to dynamically adjust the plan, achieving adaptive adjustment of maintenance strategies and closed-loop optimization of the execution process. Therefore, the technical solution of the present invention solves the technical problem that the existing power outage maintenance plan formulation technology cannot take into account multiple constraints in real time, and improves the dynamic adaptability, overall coordination and robustness of the maintenance plan. Attached Figure Description

[0019] Figure 1 This is a flowchart of a distribution network power outage maintenance plan generation method according to an embodiment of the present invention; Figure 2 This diagram illustrates a specific operational flow example of a distribution network power outage maintenance plan generation method according to another embodiment of the present invention. Figure 3 This diagram illustrates a process for three-level feature extraction and linkage mapping of distribution network data according to another embodiment of the present invention. Figure 4 This diagram illustrates a specific architecture and operational example of a frequent power outage risk prediction model according to another embodiment of the present invention. Figure 5 This diagram illustrates a specific operational example of the adaptation maintenance scheme generation and optimization process according to another embodiment of the present invention. Figure 6 This is a schematic diagram of the structure of a power outage maintenance plan generation system according to an embodiment of the present invention; Figure 7 This is a block diagram of an electronic device used to implement embodiments of the present invention. Detailed Implementation

[0020] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0021] During the development of this invention, researchers conducted numerous experiments and data analysis, discovering inherent correlations among the characteristics of the distribution network operation data at the transformer substation, line, and user levels: transformer substation load fluctuations are significantly positively correlated with the number of power outages (correlation coefficient greater than 0.7); the line fault propagation path and power supply radius jointly determine the fault impact range; and user electricity consumption type and the frequency of power outage complaints directly determine the power outage sensitivity level. These correlations collectively constitute a three-level linkage mechanism for frequent power outage risks. Based on this relationship, this invention innovatively proposes this technical solution, utilizing the fusion preprocessing of multi-source distribution network operation data, extracting three-level features and constructing risk linkage mapping relationships, combining three types of constraints—distribution network power supply capacity, equipment maintenance cycle, and user power supply guarantee—as well as iterative optimization through reinforcement learning and dynamic reward mechanisms, thereby achieving adaptive generation and dynamic balancing of distribution network power outage maintenance plans, embodying the core concept of "data-driven, multi-level linkage, and real-time feedback."

[0022] Specifically, through comparative experiments, the invention team discovered that traditional maintenance planning methods relying on manual experience or fixed templates suffer from technical defects such as slow response and rigidity: they cannot dynamically link with risk prediction results, real-time distribution network status, and maintenance resource data, resulting in insufficient adaptability of maintenance plans to actual operating conditions; and they lack a three-dimensional overall design that considers timing, regional priority, and real-time feedback, easily leading to user power conflicts or resource misallocation. In contrast, the reinforcement learning-based adaptive balancing method for distribution network outage maintenance plans proposed in this invention can improve the accuracy and efficiency of maintenance plans; through a three-level risk linkage mapping, it can achieve global perception and quantitative prediction of outage risks; through a dynamic reward mechanism and reinforcement learning iteration, it can ensure that the optimization process quickly converges to the suitable solution; and through a real-time feedback mechanism, it can achieve dynamic adjustments and optimized resource allocation during plan execution.

[0023] Therefore, this invention provides a method for generating power outage maintenance plans for distribution networks, which can be applied to a distribution network dispatching system (hereinafter referred to as the "system"). This system can be deployed centrally in the cloud or distributed at the edge within a power dispatching and control environment to automate the entire process from data acquisition to plan generation. Specifically, this system can be deployed in various hardware environments, including but not limited to: cloud server clusters, edge computing gateways, distribution network dispatching terminal equipment, and mobile inspection devices. This flexible deployment architecture allows the system to meet both the data integration needs of centralized processing in large-scale distribution network centers and the low-latency requirements of localized real-time decision-making by field operation units.

[0024] like Figure 1 As shown, the method may include: Step S110: Obtain multi-source distribution network operation data, and perform fusion preprocessing on the multi-source distribution network operation data, including time sequence alignment, outlier handling, and missing value completion, to obtain the distribution network data pool.

[0025] Among them, multi-source distribution network operation data refers to a heterogeneous data set collected from different monitoring nodes of the distribution network, including low-voltage user power outage records, transformer equipment operation parameters, line load data, maintenance history records and meteorological image data; fusion preprocessing refers to the process of integrating multi-source data through standardized procedures, which aims to improve data consistency and quality; distribution network data pool refers to a unified, structured data storage repository formed after preprocessing, used to support subsequent feature extraction and risk analysis.

[0026] Specifically, the system can acquire multi-source data in real time through a distributed data acquisition interface, and utilize a preprocessing pipeline to perform time-series alignment (using linear interpolation to unify timestamps) and outlier handling (based on...). The principles include filtering outliers and imputing missing values ​​(filling in missing data using an LSTM prediction model). The LSTM (Long Short-Term Memory) prediction model is a variant of the Recurrent Neural Network (RNN) used to process time-series data. Its core feature is the introduction of gating mechanisms (input gate, forget gate, output gate) to control the flow of information, thereby effectively capturing long-term dependencies. The principle is a statistical outlier detection and handling guideline based on the assumption of normal distribution. Its core idea is that in a normal distribution, 99.73% of the data will fall within the mean. plus or minus three standard deviations Within the range.

[0027] For example, the system collects transformer load data (sampling interval 5 minutes), line current data (sampling interval 10 seconds), and user power outage records (event-driven collection) for a certain area within 24 hours. Time alignment is used to unify the data to a 1-minute granularity. In outlier filtering, load data exceeding the historical mean ± 3 standard deviations (e.g., a historical mean of 50 kW and a standard deviation of 10 kW would filter data exceeding 20-80 kW) are removed. For missing value imputation, for data segments with consecutive missing values ​​exceeding 1 hour, a pre-trained LSTM model (128 hidden layer units) is used for prediction and imputation, ultimately forming a dimensionless... Distribution network data pool (1440 time points, 5 types of data).

[0028] Step S120: Extract the transformer area layer features, line layer features, and user layer features from the distribution network data pool, and construct a three-level risk linkage mapping relationship based on the transformer area layer features, line layer features, and user layer features.

[0029] Among them, the transformer substation characteristics refer to the core indicators reflecting the operating status of the transformer substation, such as the correlation coefficient between load fluctuation and the number of power outages; the line substation characteristics refer to the parameters describing the propagation characteristics of line faults, such as the correlation between the fault impact range and the power supply radius; the user substation characteristics refer to the attributes characterizing the user's power outage sensitivity, such as the classification of electricity consumption type and complaint frequency; the three-level risk linkage mapping relationship refers to the risk transmission logic between the transformer substation-line-user levels established through quantitative thresholds and rules.

[0030] Specifically, the system can extract features at each level through the feature calculation module: at the transformer substation level, the Pearson correlation coefficient is used to calculate the correlation between load fluctuations and the number of power outages (formula: ),in For the sample size, For load fluctuation range, The system measures the number of power outages; the line layer analyzes fault propagation paths using graph theory algorithms; and the user layer classifies sensitivity levels based on weighted scoring (electricity type weight 60%, complaint frequency weight 40%).

[0031] For example, using 30 days of data from a certain transformer area, the correlation coefficient between load fluctuation and the number of power outages is calculated. ( , This represents the daily load fluctuation value. (Daily power outage count), marking loads exceeding 80% as high-risk zones; line-level analysis of a 10-node topology, fault propagation path weights calculated based on historical fault rates, and the impact range formula is... ,in , The load deviation rate; at the user level, hospital users (electricity type score 100) and those with 3 complaints per month (score 50) are combined to obtain a sensitivity score of 80. It is classified as a high-sensitivity level.

[0032] Step S130: Based on the three-level risk linkage mapping relationship, and combined with the power supply capacity constraints of the distribution network, the equipment maintenance cycle constraints, and the user power supply guarantee constraints, a frequent power outage risk prediction model is constructed.

[0033] Among them, the power supply capacity constraint of the distribution network refers to the safe operating threshold that limits the maximum load of the distribution network; the equipment maintenance cycle constraint refers to the maximum time limit for continuous operation of equipment; the user power supply guarantee constraint refers to the time period limit for guaranteeing power supply to highly sensitive users; and the frequent power outage risk prediction model refers to a machine learning model that outputs the risk level by fusing multiple constraints.

[0034] Specifically, the system can use a constraint ensemble module to input the three-level mapping relationship with the three types of constraints into the risk prediction model (such as the XGBoost and LSTM fusion model), and update the training data using a sliding time window (window length 30 days, sliding step size 7 days), while setting an error threshold. Triggering model retraining. XGBoost is a high-performance, scalable machine learning algorithm and open-source software library based on the gradient boosting decision tree principle.

[0035] For example, the model uses high-risk areas of the transformer substation, the range of influence of the line, and user sensitivity as input features, and the constraint weights are dynamically adjusted (e.g., when the load exceeds 90%, the power supply capacity weight increases to 0.6); when the prediction error exceeds 0.15 for three consecutive times, the features are automatically re-extracted and the model parameters are adjusted (the learning rate is reduced from 0.01 to 0.001).

[0036] Step S140: Using the risk prediction results, real-time status of the distribution network, and maintenance resource data as input, iterative optimization is performed through reinforcement learning with a dynamic reward mechanism to generate an adaptive maintenance plan.

[0037] Among them, the dynamic reward mechanism refers to the set of rules for dynamically calculating reward values ​​based on core indicators (risk prediction error, overload, and resource idleness); the adaptive maintenance plan refers to the optimized maintenance strategy that meets the real-time status and resource conditions.

[0038] Specifically, the system can be iteratively trained using reinforcement learning agents (such as the DQN algorithm): in reward calculation, a risk prediction error below 0.05 results in a +5 point bonus, a load exceeding the limit by more than 10% deducts 3 points, and a resource idle rate exceeding 30% deducts 2 points; after each iteration, the target achievement rate (risk accuracy, state stability, and resource utilization) is evaluated, and the process terminates when all targets are met (≥90%). The DQN (Deep Q-Network) algorithm is a deep learning-based reinforcement learning algorithm. Its core is to approximate the Q-function in reinforcement learning using a deep neural network, thereby solving sequential decision-making problems in high-dimensional state spaces.

[0039] For example, in a certain round, the input risk error is 0.06, the load exceeds the limit by 15%, and the resource idle time is 20%. After adjusting the strategy network parameters, continue iterating until the target achievement rate exceeds 90%.

[0040] Step S150: Based on the adaptive maintenance plan, generate an adaptive plan for power outage maintenance of the distribution network, including timing, regional priority, and real-time feedback.

[0041] Among them, the timing arrangement refers to the sequence of maintenance periods based on the load curve; the regional priority refers to the processing order determined by risk, load, and resource scores; and the real-time feedback refers to the mechanism for dynamically adjusting the plan based on deviations during the execution process.

[0042] Specifically, the system can divide time periods (peak 7:00~10:00, off-peak 10:00~17:00, low-peak 22:00~7:00) through the planning generation module, construct a priority scoring system (risk weight 0.4, load weight 0.3, resource weight 0.3), and continuously collect deviation data (such as actual maintenance time deviating from the plan by more than 1 hour) to trigger adjustments.

[0043] For example, the plan schedules maintenance for a certain hospital area (highly sensitive) during off-peak hours, with a priority score of 85 points (50 points for high risk + 30 points for load pressure + 5 points for resource coverage). If the load suddenly increases by more than 10% during execution, the maintenance will be delayed and resources will be reallocated through a feedback chain.

[0044] In other embodiments, such as Figure 2 This example demonstrates a specific operational flow of the distribution network outage maintenance plan generation method based on an embodiment of the present invention. The example uses the maintenance plan formulation for a distribution network in a new urban area during the summer peak electricity consumption period as an application scenario, specifically illustrating the collaborative operational logic of the entire process from data acquisition to plan generation. This operational flow follows the sequence of steps S1 to S5 in the diagram, and the execution content of each step is as follows: S1: Collect and merge multi-source data to form a distribution network data pool. In this step, the system acquires multi-source operational data of the new area's distribution network within a statistical period (e.g., July 1 to July 30, 2023) through a data acquisition interface. Specific data includes: power outage records for 5,000 low-voltage users obtained from the electricity information collection system; operating parameters (including load and temperature, with a 5-minute sampling interval) for 200 transformer substations and load data for 50 feeders (with a 10-second sampling interval) obtained from the distribution automation system; maintenance history records (including work orders, time consumption, and personnel) obtained from the Production Management System (PMS); and hourly meteorological imagery data (including temperature, humidity, and wind speed) obtained from the meteorological department interface. The system performs fusion preprocessing on this multi-source data: using a time-series alignment module to unify all data to a 1-minute timestamp (using linear interpolation); and applying... Outliers were filtered out based on the principle of filtering data points exceeding 20-80 kW if the historical mean was 50 kW and the standard deviation was 10 kW. Missing line current data were filled using an LSTM model (128 hidden layer units). After preprocessing, a dimension of [missing data] was formed. The distribution network data pool (43,200 time points, 5 types of data) is used for subsequent analysis.

[0045] S2: Extract frequent power outage features across three dimensions and establish a three-level risk linkage mapping relationship. In this step, the system extracts features at the distribution network data pool, including transformer substations, lines, and users, in parallel. For the transformer substation level, the system calculates the Pearson correlation coefficient between load fluctuations and the number of power outages for each substation, using the following formula: For example, the 30-day data for transformer area A is calculated as follows: The load factor exceeded the threshold of 0.7, therefore, the interval with a load rate exceeding 75% was marked as a high-risk load interval, and the first frequent power outage threshold was determined to be 75% load rate. For the line layer, the system analyzes the line topology based on graph theory algorithms, calculates the fault propagation path, and combines this with the power supply radius formula. Determine the scope of impact. For example, line L1 is designed with a power supply radius of 5 kilometers, and the actual load deviation rate is 20%. Calculated The distance is set as the second most frequent power outage threshold. For the user level, the system uses a weighted scoring method to classify sensitivity levels. For example, if a hospital user's electricity usage type score is 100, and two complaints per month correspond to a complaint frequency score of 60, then... If the score exceeds 80 points, it is classified as extremely sensitive, and a third frequent power outage threshold of 80 points is determined. Finally, the system constructs a three-level risk linkage mapping relationship based on the three thresholds and linkage rules (such as "when the load of the transformer area exceeds 75%, the impact range of the line increases by 20%).

[0046] S3: Adapting to Three Types of Constraints to Build a Frequent Power Outage Risk Prediction Model. In this step, the system incorporates three types of constraints based on the three-level risk linkage mapping relationship established in S2: distribution network power supply capacity constraint (load rate limit of 90%), equipment maintenance cycle constraint (maximum transformer operating time of 720 hours), and user power supply guarantee constraint (hospitals must not experience power outages for 24 hours). The system sets a dynamic weight adjustment mechanism for the three types of constraints, with initial weights of 0.4, 0.3, and 0.3, respectively. On peak summer electricity consumption days, due to increased load pressure, the system automatically increases the weight of the power supply capacity constraint to 0.5. Based on this, the system builds a frequent power outage risk prediction model (e.g., a fusion model of XGBoost and LSTM), uses a sliding time window (window length of 30 days, sliding step size of 7 days) to update training data, and sets a risk prediction error threshold (mean squared error MSE = 0.15). When the prediction error of the model exceeds 0.15 in two consecutive sliding windows, the system automatically triggers feature re-extraction and model parameter tuning.

[0047] S4: Input relevant data and use reinforcement learning to generate adaptive maintenance plans. In this step, the system takes the risk prediction results output by the S3 model (e.g., the probability of power outage in area X in the next 24 hours is 0.15), the current real-time status of the distribution network (load rate 82%, voltage qualification rate 99%), and maintenance resource data (10 available personnel and 3 vehicles) as input and feeds them into the reinforcement learning agent for iterative optimization. The system initializes the policy network parameters of the Deep Q-Network (DQN) and sets the maximum number of iterations to 100. Training data is loaded in batches (72 hours of data per batch). The reward for each step is calculated based on a dynamic reward mechanism: if the risk prediction error is less than 0.05, the reward is +5 points; if the load rate exceeds 80%, 3 points are deducted; if the resource idle rate exceeds 30%, 2 points are deducted. The policy network outputs candidate maintenance plans (e.g., "arrange 2 personnel to maintain line L1 during the period from 22:00 to 02:00 in area X"). After each iteration, the achievement rate of the optimization target is calculated. When the achievement rate exceeds 100% for 5 consecutive rounds, convergence is determined, and the final adaptability maintenance plan is output.

[0048] S5: Generate an adaptive plan for distribution network outage maintenance. In this step, based on the adaptive maintenance scheme generated in S4, the system further generates a final plan including timing, regional priority, and real-time feedback. First, based on the current daily load curve, peak (07:00~10:00, 17:00~22:00), off-peak (10:00~17:00), and off-peak (22:00~07:00) periods are divided, and the maintenance execution time is scheduled during the off-peak period. Second, a regional priority scoring system is constructed by combining risk distribution, load pressure, and resource coverage. For example, if region X has a risk score of 80, a load pressure score of 100, and a resource coverage score of 60, with a total score of 80, it is ranked first and prioritized. Finally, during plan execution, the system continuously collects actual data (such as maintenance start time deviation and load abrupt changes), identifies the type of deviation, and dynamically adjusts subsequent strategies (such as automatically postponing subsequent tasks by 30 minutes due to delays in previous tasks), ultimately outputting an executable adaptive plan for distribution network outage maintenance. This running example fully demonstrates Figure 2 The five steps involved in generating a dynamically adapted maintenance plan, based on actual data and rules, demonstrate the feasibility and effectiveness of this method.

[0049] In other embodiments, such as Figure 3 This document demonstrates an example of a three-level feature extraction and linkage mapping process for distribution network data based on an embodiment of the present invention. Using historical operational data from a municipal-level distribution network as a foundation, this example illustrates the data processing logic from the raw data pool to the generation and linkage of three-level risk quantification indicators. The process follows the left-to-right, top-to-bottom order shown in the diagram, demonstrating the parallel processing and final linkage process across the transformer substation, line, and user layers. The execution content of each step is as follows: 1. Data Input and Feature Extraction: The process begins with a distribution network data pool containing a standardized dataset that has undergone fusion and preprocessing. The data is then fed in parallel into three feature extraction modules.

[0050] Transformer Area Feature Extraction: Based on the operating parameters of transformer area equipment (such as load curves) and power outage records, the system calculates the correlation coefficient between load fluctuations and the number of power outages. Specifically, the system uses the Pearson correlation coefficient formula: Quantitative analysis was conducted, among which For the number of days in the sample, Daily load fluctuation range (unit: kilowatts). This represents the number of power outages per day. For example, the correlation coefficient is calculated based on 90 days of data for transformer substation number T-101. If the load is significantly higher than the preset threshold of 0.7, the system will immediately mark the high-risk load range as the running segment where the load rate is consistently higher than 75%.

[0051] 2. Line Layer Feature Extraction: Based on line load data and historical fault records, the system analyzes the relationship between fault propagation paths and power supply radius. Specifically, the system uses graph theory algorithms (such as Dijkstra's shortest path algorithm) to construct a line topology model and combines this with the power supply radius parameter to calculate the fault impact range. The formula is as follows: ,in Design the power supply radius for the line (unit: kilometers). This is the load impact factor (typically ranging from 0.1 to 0.3). This refers to the load deviation rate. For example, for line L-205, the design power supply radius... kilometers, current load deviation rate ,Pick The scope of its fault impact was calculated. The system determines the fault impact range of a given line level as kilometers. Dijkstra's shortest path algorithm is a classic graph theory algorithm used to compute single-source shortest paths in weighted directed or undirected graphs. Its core idea is to iteratively find the shortest path from a given starting point to all other vertices in the graph by progressively expanding the set of vertices with known shortest paths.

[0052] 3. User-level feature extraction: Based on low-voltage user power outage records, the system classifies user sensitivity levels according to electricity usage type and complaint frequency. Specifically, the system uses a weighted scoring method: The weighting of electricity usage type is set according to industry importance (e.g., hospitals 100 points, schools 80 points, commercial 60 points, residential 40 points); the weighting of complaint frequency is mapped according to the number of complaints per month (e.g., 0 complaints = 100 points, 1-2 complaints = 70 points, 3 or more complaints = 40 points). For example, user U-3001 is a hospital user with an electricity usage type score of 100 points. If they have 2 complaints in the past month, their complaint frequency score is 70 points. Based on the threshold (extremely high sensitivity ≥ 85 points, high sensitivity 70~84 points), the system outputs that the user is at the extremely high sensitivity level.

[0053] Three-level linkage and result output: The feature extraction results from the three dimensions are ultimately linked and integrated. The system maps the high-risk load range output from the distribution area layer to the regional risk level (such as high risk, medium risk), superimposes the fault impact range output from the line layer with the risk level, comprehensively determines the affected physical area, and combines the user sensitivity level distribution output from the user layer to identify the distribution of sensitive users within the impact range.

[0054] Therefore, according to Figure 3The illustrated process demonstrates how the system can transform raw distribution network operation data into structured risk quantification indicators (risk level, impact range, and sensitive user distribution) through parallel and hierarchical data processing. This provides a clear and operable data foundation for subsequently constructing a precise three-tiered risk linkage mapping relationship. The process visually illustrates the complete chain of data processing from multi-source fusion to feature extraction and then to comprehensive analysis.

[0055] In other embodiments, such as Figure 4 This document demonstrates a specific architecture and operational example of a frequent power outage risk prediction model based on an embodiment of the present invention. Using the risk prediction task of an industrial park's power distribution network in the first quarter of 2024 as a case study, the example illustrates the end-to-end operational flow of the model from data input, processing, fusion calculation to result output. The architecture follows the top-down data flow shown in the diagram, and the specific operations at each level are as follows: Input Layer: The model receives two types of input data. The first type is the result of the three-level risk linkage mapping, which is generated by... Figure 3 The process generation shown includes: three high-risk load zones marked at the distribution area level (corresponding to distribution areas with a load rate >78%), the expanded fault impact range of two lines calculated at the line level (5.1 km and 4.8 km respectively), and five extremely sensitive user nodes identified at the user level (such as data centers and chemical plants). The second category is distribution network runtime timing data, including minute-level load, voltage, and current timing curves of the distribution network within the statistical period. The system uniformly encapsulates this structured and unstructured data and inputs it to the data processing layer.

[0056] Data Processing Layer: This layer performs temporal organization and quality checks on the input data. The core mechanism is the use of a sliding time window for dynamic management of training data. The system sets the window length to 30 days and the sliding step to 7 days. For example, on March 1, 2024, the system loads training data with a time window from January 31, 2024 to March 1, 2024 (30 days in total); 7 days later (March 8), the window automatically slides to February 7, 2024 to March 8, 2024, removing the earliest 7 days of data and adding the latest 7 days of data, thus achieving rolling updates of training data. Simultaneously, this layer performs consistency checks to ensure the temporal continuity and dimensional uniformity of the data within the window.

[0057] Constraint Layer: This layer injects three types of business rule constraints into the model and has a dynamic constraint weight adjustment mechanism. Specific constraints include: (1) User power supply guarantee constraint, which stipulates that the power supply to highly sensitive users such as data centers cannot be interrupted for 24 hours a day; (2) Equipment maintenance cycle constraint, which limits the maximum continuous operation time of the main transformer to 720 hours; (3) Distribution network power supply capacity constraint, which sets the safe upper limit of the load rate of the main line of the park to 85%. The weight adjustment mechanism dynamically balances the priorities of the three according to the real-time operating status. For example, on the first day of resumption of work after the Spring Festival, the load of the park increased sharply, and the system automatically increased the weight of the distribution network power supply capacity constraint from the basic value of 0.4 to 0.5 to prioritize the stability of the power grid.

[0058] Core Layer: This layer is the computational hub of the entire architecture, employing a fusion algorithm of XGBoost (Extreme Gradient Boosting Tree) and LSTM (Long Short-Term Memory). The specific workflow is as follows: Time-series data updated by the processing layer and rule-based features weighted by the constraint layer are input into the fusion model. The XGBoost algorithm excels at capturing non-linear relationships and importance between features (such as the correlation between load rate and risk level), while the LSTM algorithm excels at learning long-term dependency patterns in time series (such as the impact of periodic load fluctuations on future risks). Both are fused at the feature level through an attention mechanism to jointly complete risk prediction. This layer also integrates an error assessment module and a parameter adjustment module. The system sets the error threshold for risk prediction to a mean squared error (MSE) of 0.15. The error assessment module continuously monitors the deviation between the model output and the actual value. When the MSE is greater than 0.15 for three consecutive prediction periods (i.e., error > threshold), the parameter adjustment module is triggered, automatically adjusting the XGBoost parameters. Hyperparameters such as the learning rate (e.g., adjusted from 0.01 to 0.005) or LSTM dropout (random inactivation rate) can be adjusted, and may trigger partial feature re-extraction to optimize model performance.

[0059] Output Layer: This layer receives and formats the calculation results from the core layer of the model, generating the final prediction result for frequent power outage risks. The output result is the power outage risk level of a specified area (such as a feeder or a transformer substation) within the next 24 hours or the next week, categorized into four levels: extremely high, high, medium, and low. For example, the model might output "10kV feeder F5 in Zone A of the Industrial Park, power outage risk level in the next 24 hours: high," which will directly serve as a key input for generating an adaptive maintenance plan.

[0060] Therefore, according to Figure 4The architecture and operational example shown demonstrate that this system can achieve accurate and adaptive prediction of the risk of frequent power outages in the distribution network through an automated pipeline that integrates data processing, dynamic constraints, and advanced fusion algorithms, providing a reliable core judgment basis for subsequent intelligent maintenance decisions.

[0061] In other embodiments, such as Figure 5 This example demonstrates a specific operational instance of the adaptive maintenance plan generation and optimization process based on embodiments of the present invention. Using the task of developing a maintenance plan for a power distribution network in an urban area during peak summer demand as a case study, this example illustrates the closed-loop optimization logic of the power outage maintenance plan generation method, from initialization and iterative training to the final output plan. The process follows the top-to-bottom order shown in the diagram, and the specific execution content and data examples for each step are as follows: Step 1: Initialization. After the process begins, the system first performs initialization. The policy network parameters are initialized using the Xavier method, randomly initializing the weights of the policy network (e.g., a deep Q-network with two hidden layers and 128 neurons per layer), and setting the bias term to zero. Simultaneously, the dynamic reward mechanism weights are initialized, setting initial weights for the three optimization objectives: risk prediction accuracy, distribution network stability, and resource utilization. For example, the weight vector is... .

[0062] Step 2: Batch Input of Training Data. After initialization, the system loads training data in batches. The input data consists of integrated risk prediction results, real-time distribution network status, and maintenance resource data. For example, a training batch might include the probability of power outage risk in a certain area over the past 72 hours (risk prediction results), minute-level load rate and voltage data (real-time distribution network status), and available maintenance teams and vehicle location information (maintenance resource data), forming a batch with the following dimensions: The three-dimensional tensor.

[0063] Step 3: Preliminary Reinforcement Learning Training and Output Scheme. Based on the input data and the initialized dynamic reward mechanism, the system initiates reinforcement learning training. The agent (e.g., DQN) selects an action (e.g., "perform maintenance in area A from 00:00 to 04:00") according to the current state (input data), and the environment provides feedback according to the reward rules. For example, if the action reduces the risk prediction error... If it results in a +5 bonus, then a +5 bonus will be awarded; otherwise, it will cause the load to exceed the limit. If not, 3 points will be deducted. After multiple rounds of exploration and learning, the policy network outputs a preliminary maintenance plan, such as a plan vector that includes specific maintenance time, area, personnel, and equipment allocation.

[0064] Step 4: Calculate the achievement rate of optimization goals. The system calculates the achievement rate of three core indicators to evaluate the quality of the preliminary plan: risk prediction accuracy achievement rate (e.g., target...). ,actual ,but Distribution network stability achievement rate (e.g., target load over-limit time percentage <5%, actual percentage 3%) Resource utilization rate achievement rate (e.g., target resource utilization rate > 85%, actual rate is 88%) Then, the system determines whether all three achievement rates have met the target, that is, whether they are all greater than or equal to the preset threshold of 0.9.

[0065] Step 5: Decision and Output or Adjustment. Based on the judgment result of Step 4, the process branches out: If all criteria are met: The system determines that the current solution meets the optimization objectives and directly outputs an adaptive maintenance solution. For example, the output solution is: "During the off-peak hours (23:00~03:00) in region X, deploy 2 maintenance teams and 1 live-line work vehicle to prioritize the maintenance of line L-108."

[0066] If not all criteria are met (No): The system will enter the parameter adjustment phase. Specific adjustments include: adjusting the dynamic reward mechanism's score range, for example, changing the penalty range for exceeding load limits from... The 3-point deduction has been adjusted to Deduct 4 points to strengthen the penalty for load spikes; and adjust the learning rate of the policy network within a preset range of 0.001 to 0.01, for example, reducing it from the current 0.005 to 0.003 to make the learning process more stable. After the adjustment is completed, the process returns to step 2 (batch input of training data) and uses the new parameters for the next round of iterative training until the achievement rate meets the threshold requirements.

[0067] Therefore, according to Figure 5 The illustrated process embodiment shows that the system, through a closed-loop feedback mechanism of "initialization-training-evaluation-adjustment", can continuously optimize the reinforcement learning strategy based on dynamic rewards, automatically generate an adaptive maintenance plan that achieves a balance between risk prediction accuracy, power grid operation stability and resource utilization efficiency, and significantly improve the intelligence level and plan quality of maintenance plan formulation.

[0068] Therefore, according to the above implementation method, the system achieves its goals through five core steps: data fusion preprocessing, three-level risk linkage mapping, construction of multi-constraint risk prediction model, iterative optimization of scheme based on reinforcement learning, and adaptive plan generation. This process involves several key steps: First, multi-source distribution network operation data undergoes fusion preprocessing, including time-series alignment, outlier handling, and missing value completion, to integrate and standardize the data, laying the foundation for high-quality data analysis. Second, distribution area, line, and user-level features are extracted from the distribution network data pool, and a three-tiered risk linkage mapping relationship is constructed to achieve a global and comprehensive understanding of distribution network risks. Third, a frequent power outage risk prediction model is built based on this three-tiered risk linkage mapping relationship, combined with distribution network power supply capacity constraints, equipment maintenance cycle constraints, and user power supply guarantee constraints, to integrate multiple practical constraints into risk quantification assessment. Fourth, risk prediction results, real-time distribution network status, and maintenance resource data are used as inputs, and iterative optimization is achieved through reinforcement learning with a dynamic reward mechanism to generate adaptive maintenance plans, enabling dynamic adaptation to complex distribution network conditions and resource requirements. Finally, an adaptive power outage maintenance plan, including time-series scheduling, regional priority, and real-time feedback, is generated based on the adaptive maintenance plan to guide maintenance execution with three-dimensional adaptive scheduling capabilities.

[0069] Specifically, in the technical solution of this embodiment, addressing the problems mentioned in the background art regarding traditional maintenance plans relying on manual experience and lacking dynamic linkage with risk prediction results, real-time distribution network status, and maintenance resource data, a three-level risk linkage mapping relationship is constructed and combined with reinforcement learning for iterative optimization to generate adaptive maintenance plans. This achieves real-time dynamic matching between maintenance plans and risk changes, load fluctuations, and resource supply and demand, thereby improving the adaptability of the plans to the actual operating status. Addressing the problem mentioned in the background art regarding the lack of a three-dimensional overall design of existing maintenance plans including time-series scheduling, regional priority, and real-time feedback, an adaptive plan incorporating time-series scheduling, regional priority, and real-time feedback is generated. This constructs a multi-dimensional collaborative scheduling framework with time-series, spatial, and process feedback, resolving the defects of conflict between maintenance periods and user electricity demand, maintenance delays in high-risk areas, and rigid plan execution. Addressing the problem mentioned in the background art regarding the inability of traditional methods to respond to dynamic factors, leading to low maintenance efficiency and power supply reliability, a reinforcement learning-based dynamic reward mechanism is used to continuously optimize the plan, and a real-time feedback mechanism is introduced to dynamically adjust the plan, achieving adaptive adjustment of maintenance strategies and closed-loop optimization of the execution process. Therefore, the technical solution of this embodiment solves the technical problem that existing power outage maintenance plan formulation technology cannot take into account multiple constraints in real time, and improves the dynamic adaptability, overall coordination and robustness of the maintenance plan.

[0070] In some embodiments, multi-source distribution network operation data includes low-voltage user outage records, transformer area equipment operating parameters, line load data, maintenance history records, and meteorological image data; the steps of extracting transformer area layer features, line layer features, and user layer features include: Extract features at the transformer substation level, including calculating the correlation between load fluctuations and the number of power outages based on the operating parameters of the equipment in the substation and power outage records, and marking high-risk load intervals.

[0071] Among them, load fluctuation refers to the variation of the load of the distribution transformer area in a time series, usually expressed as the difference between the daily maximum and minimum load values, in kilowatts; power outage frequency refers to the cumulative number of power outage events that occur in the transformer area within the statistical period; correlation refers to the statistical correlation between load fluctuation and power outage frequency, quantified by the Pearson correlation coefficient; high-risk load range refers to the load range where the correlation exceeds a preset threshold (e.g., 0.7), which is marked as an operating state with high power outage risk.

[0072] Specifically, the system can calculate the correlation coefficient between load fluctuations and the number of power outages using the Pearson correlation coefficient, as shown in the formula: ; in, For the sample size, For load fluctuation range, The system calculates the number of power outages and dynamically labels high-risk load zones based on correlation coefficients. For example, it calculates the Pearson correlation coefficient between load fluctuations and the number of power outages for a specific transformer area over 30 days. (sample size) Load fluctuation range The difference between the daily maximum and minimum load values, and the number of power outages. (This is a daily cumulative value), and when the load exceeds 80% of the rated load, it is marked as a high-risk zone.

[0073] Extract line-level features, including analyzing the correlation between fault propagation paths and power supply radius based on line load data and fault history records, and determining the scope of fault impact.

[0074] Among them, fault propagation path refers to the path by which a fault spreads from the point of occurrence to other nodes in the distribution network line; power supply radius refers to the electrical distance from the power source point to the farthest load point, in kilometers; correlation refers to the dependency relationship between fault propagation path and power supply radius, which affects the range of fault spread; fault impact range refers to the area that a fault may affect, expressed by distance or node coverage.

[0075] Specifically, the system can construct a line topology model using graph theory algorithms (such as Dijkstra's shortest path algorithm), analyze fault propagation paths, and calculate the fault impact range by combining the power supply radius parameter. The formula is as follows: ,in To design the power supply radius, The load overload impact factor (values ​​range from 0.1 to 0.3). This refers to the load deviation rate. For example, a power supply radius of 5 kilometers for a certain line is [missing information - likely a specific value]. (The deviation between the actual load rate of 80% and the rated load rate of 60%), take... The calculated range of the fault's impact km.

[0076] Extract user-level features, including classifying user power outage sensitivity levels based on the type of electricity used and the frequency of power outage complaints in low-voltage user power outage records.

[0077] Among them, electricity type refers to the nature of a user's electricity use, such as residential electricity use, commercial electricity use, industrial electricity use, and electricity use for important public services (such as hospitals and schools); power outage complaint frequency refers to the number of times a user files a complaint due to a power outage within a statistical period (such as a calendar month); and user power outage sensitivity level refers to the user's tolerance level to power outages based on electricity type and complaint frequency, which is divided into four levels: extremely high sensitivity, high sensitivity, medium sensitivity, and low sensitivity.

[0078] Specifically, the system can use a weighted scoring method to classify sensitivity levels, with electricity usage type accounting for 60% and complaint frequency accounting for 40%. The sensitivity score is calculated using the following formula: For example, a hospital user's electricity usage type score is 100, and three complaints in a month correspond to a complaint frequency score of 50. It is classified as a high-sensitivity level.

[0079] Therefore, according to the above implementation method, the system can extract three-level features of distribution area, line and user from multi-source data, provide structured input for building a three-level risk linkage mapping relationship, and realize refined perception of power outage risk in distribution network.

[0080] In some embodiments, a three-level risk linkage mapping relationship is constructed based on transformer area features, line layer features, and user layer features, including: Based on the correlation coefficient between load fluctuation and power outage frequency in the transformer area layer characteristics, the first frequent power outage threshold for the corresponding layer of transformer area layer characteristics is determined.

[0081] Among them, the first frequent power outage threshold is a risk threshold value quantified based on the correlation between load fluctuations and the number of power outages in a transformer area, used to identify high-risk load ranges at the transformer area level; the correlation coefficient is a statistical indicator calculated using the Pearson correlation coefficient, reflecting the linear correlation between the load fluctuation amplitude and the occurrence of power outage events.

[0082] Specifically, the system can calculate the correlation coefficient using the Pearson correlation coefficient formula, which is as follows: ; in For the sample size, Daily load fluctuation (unit: kilowatts). The system calculates the daily power outage frequency. When the correlation coefficient exceeds a preset threshold, the corresponding load range is marked as high-risk, and its upper limit is used as the first frequent power outage threshold. For example, for 30 days of data for a certain transformer area, the Pearson correlation coefficient between load fluctuation and the number of power outages is calculated. (sample size) Load fluctuation range The difference between the daily maximum and minimum load, and the number of power outages. (This is the daily cumulative value), with a preset threshold of 0.7. The range where the load exceeds 80% of the rated load is marked as high risk, and 80% load rate is used as the first frequent power outage threshold.

[0083] Based on the correlation between fault propagation paths and power supply radius in the line layer characteristics, the second frequent power outage threshold for the corresponding level of the line layer characteristics is determined.

[0084] Among them, the second frequent power outage threshold refers to the critical distance value determined based on the relationship between the fault propagation range and the power supply radius, which is used to divide the fault impact area at the line level; the fault propagation path refers to the path of the fault point spreading along the line topology; and the power supply radius refers to the electrical distance (unit: kilometers) from the power source point to the farthest load point.

[0085] Specifically, the system can model the line topology using graph theory algorithms (such as Dijkstra's shortest path algorithm), calculate the weights of fault propagation paths, and combine this with the power supply radius parameter, using the formula... Calculate the scope of the fault's impact, where To design the power supply radius, The load overload impact factor (values ​​range from 0.1 to 0.3). The load deviation rate is set as the critical distance of the affected area, which is then set as the second frequent power outage threshold.

[0086] For example, the designed power supply radius of a certain line 1,000 km, actual load rate 80% and rated load rate 60%, then the load deviation rate is... ,Pick Calculated Kilometers, with 5.33 kilometers as the second most frequent power outage threshold.

[0087] Based on the electricity consumption type and the frequency of power outage complaints in the user layer characteristics, the third frequent power outage threshold for the corresponding level of user layer characteristics is determined.

[0088] Among them, the third frequent power outage threshold refers to the scoring threshold based on the user's power outage sensitivity level, which is used to distinguish the user's tolerance for power outages; the types of electricity use include residential electricity, commercial electricity, industrial electricity, and electricity for important public services (such as hospitals and schools), and the frequency of power outage complaints is calculated on a calendar month basis.

[0089] Specifically, the system can calculate the user's sensitivity score using a weighted scoring method, with the following formula: Electricity usage type scores are assigned based on importance (e.g., 100 points for hospitals, 60 points for commercial establishments), and complaint frequency scores are assigned based on the number of complaints (e.g., 100 points for 0 complaints, 50 points for 3 complaints). The threshold for sensitivity scores (e.g., the lowest score for extremely high sensitivity) is set as the third frequent power outage threshold. For example, if a hospital user has an electricity usage type score of 100 and 3 complaints per month corresponding to a complaint frequency score of 50, then... The threshold for the extremely high sensitivity level is set at 80 points, therefore the threshold for the third frequent power outage is 80 points.

[0090] Based on the first frequent power outage threshold, the second frequent power outage threshold, the third frequent power outage threshold, and the linkage rules based on the risk propagation logic between levels, a three-level risk linkage mapping relationship is constructed.

[0091] Among them, the risk propagation logic between levels refers to the risk transmission mechanism between the three levels of distribution area, line, and user. For example, a line fault triggered in a high-risk load area of ​​the distribution area may expand the scope of the impact, thereby increasing the power outage risk for highly sensitive users. The three-level risk linkage mapping relationship is a risk association model formed by integrating the three-level thresholds through quantitative rules.

[0092] Specifically, the system can set linkage rules through the rule engine. For example, when the load of a transformer area exceeds the first frequent power outage threshold, the impact range of the line fault is expanded by 30% based on the second frequent power outage threshold. When the impact range expands, if the sensitivity of users within the range exceeds the third frequent power outage threshold, the user area is marked as extremely high risk. For example, if the load rate of a transformer area is 85% (exceeding the first threshold of 80%), the impact range of the line is triggered to expand from 5.33 kilometers to 6.93 kilometers. If a user's sensitivity score is 90 points (exceeding the third threshold of 80 points) within the expanded area, then that area is mapped as a high-risk area with three-level linkage.

[0093] Therefore, according to the above implementation method, the system can achieve dynamic correlation and accurate mapping of risks at three levels: transformer area, line, and user through quantification thresholds and linkage rules, providing structured input for subsequent prediction of frequent power outage risks.

[0094] In some embodiments, based on a three-level risk linkage mapping relationship and combined with distribution network power supply capacity constraints, equipment maintenance cycle constraints, and user power supply guarantee constraints, a frequent power outage risk prediction model is constructed, including: Based on the three-level risk linkage mapping relationship, and incorporating the constraints of power distribution network supply capacity, equipment maintenance cycle, and user power supply guarantee, a constrained risk prediction basis is obtained.

[0095] Among them, the constrained risk prediction basis refers to the structured data input layer formed by integrating the three-level risk linkage mapping relationship with the three types of constraints, which is used to support the training and inference of the risk prediction model; the power supply capacity constraint of the distribution network limits the load limit of lines and equipment at different times, the equipment maintenance cycle constraint limits the maximum continuous operating time of equipment, and the user power supply guarantee constraint limits the uninterrupted power supply periods of highly sensitive users.

[0096] Specifically, the system can logically associate the high-risk load range of a transformer substation, the impact range of a line fault, and the user sensitivity level in the three-level risk linkage mapping relationship with the threshold conditions of the three types of constraints through the constraint integration module, forming a unified input feature matrix. For example, the feature data of a transformer substation with a load rate of 85% (exceeding the first frequent power outage threshold of 80%), a line impact range of 5.33 kilometers (exceeding the second frequent power outage threshold of 5 kilometers), and a user sensitivity score of 90 (exceeding the third frequent power outage threshold of 80%) are combined with the load limit of 90% of the distribution network power supply capacity constraint, the maximum operating time of 720 hours of equipment maintenance cycle constraint, and the 24-hour guarantee period for hospital power supply constraint to generate a dimension of The eigenvectors serve as the basis for constrained risk prediction.

[0097] A constraint weight adjustment mechanism is set up for the power supply capacity constraints of the distribution network, the equipment maintenance cycle constraints, and the user power supply guarantee constraints, so as to dynamically balance the power supply capacity constraints of the distribution network, the equipment maintenance cycle constraints, and the user power supply guarantee constraints.

[0098] Among them, the constraint weight adjustment mechanism refers to the set of rules that dynamically calculate and adjust the constraint weights based on three dimensions: risk urgency, load criticality, and resource sufficiency; dynamic balancing refers to automatically optimizing the influence ratio of the three types of constraints in decision-making based on the real-time operating status.

[0099] Specifically, the system can perform weight adjustments through a rules engine: first, set basic weights (power supply capacity 0.4, maintenance cycle 0.3, user protection 0.3), and then calculate adjustment coefficients based on real-time indicators. , Furthermore, the sum of the weights of the three categories is normalized to 1.

[0100] For example, if the load rate of a distribution network in a certain area is 92% (criticality score 100) during the peak electricity consumption period in summer, then... The weight is adjusted from 0.4 to 0.48; if the region also has a resource idle rate of 8% (sufficiency score of 30), then The weight was adjusted from 0.3 to 0.318; the user protection constraint weight was automatically adjusted to... .

[0101] A frequent power outage risk prediction model is constructed based on a constrained risk prediction foundation and a constraint weight adjustment mechanism. A sliding time window is used to update the model training data, and a risk prediction error threshold is set.

[0102] Among them, the sliding time window refers to a data management mechanism that updates training data in a rolling manner with a fixed duration; the risk prediction error threshold refers to the accuracy threshold that triggers model retraining.

[0103] Specifically, the system can use a stream processing framework to set a window length of 30 days and a sliding step of 7 days. Every 7 days, it automatically removes the earliest 7 days of data and adds the latest 7 days of data. The risk prediction error threshold is set to a mean squared error (MSE) of 0.15. When the MSE exceeds 0.15 for three consecutive window periods, feature re-extraction and parameter adjustment are automatically triggered. For example, if a model trained during the window period from July 1st to July 30th has an MSE of 0.12, and the window slides to July 8th to August 4th on August 1st, the new window data has an MSE of 0.16. After exceeding the threshold three times consecutively, the system automatically re-extracts the tertiary features and adjusts the model's hyperparameters (e.g., reducing the XGBoost learning rate from 0.01 to 0.005).

[0104] Therefore, according to the above implementation method, the system can construct a risk prediction model that adapts to multiple constraints through constraint integration and dynamic weight adjustment, and ensure continuous optimization of the model by means of sliding window and error monitoring mechanism, thereby improving the accuracy and timeliness of prediction.

[0105] In some embodiments, the dynamic reward mechanism is configured to: Reward rules are set based on risk prediction results, real-time status of distribution network and maintenance resource data, and the reward calculation weight is adjusted based on the dynamic changes of reward rules.

[0106] Among them, the dynamic reward mechanism refers to the adaptive optimization logic that uses the accuracy of risk prediction results, the stability of the real-time status of the distribution network, and the utilization rate of maintenance resources as core indicators, and quantifies the reward value through a piecewise function and dynamically adjusts the weights; the reward rules refer to the set of rules that define the conditions for increasing and decreasing rewards, including preset parameters such as error threshold, load threshold, and idle threshold; the reward calculation weight refers to the proportion of each indicator in the total reward value, which is dynamically balanced through deviation weighting and smoothing factors.

[0107] Specifically, the system can configure the reward calculation logic through the rules engine: the reward value range is set to... The score is calculated based on the mean square error (MSE) of risk prediction results, the load over-limit magnitude and time of real-time distribution network status, and the resource idle rate and time of maintenance resource data, with reward and deduction scores calculated in segments. The weighting of each indicator is dynamically adjusted every 12 hours according to the degree of indicator deviation. The weighting adjustment formula is as follows: ,in As the current weight, For the previous round of weighting, This is a smoothing factor.

[0108] For example, during system initialization, the risk prediction accuracy weight is 0.35, the distribution network status stability weight is 0.4, and the resource utilization rate weight is 0.25. If the distribution network load rate exceeds the limit continuously during a certain period and the deviation increases, the weights are recalculated, increasing the distribution network status stability weight to 0.48, and other weights are adjusted accordingly.

[0109] The reward rules include: If the error in the risk prediction result is lower than the preset error threshold, the reward will be increased.

[0110] The preset error threshold refers to the acceptable upper limit of the mean square error (MSE) of the risk prediction result, which is set to 0.08; the additional reward refers to giving positive feedback incentives when the actual error is lower than the threshold, so as to encourage the model to improve the prediction accuracy.

[0111] Specifically, the system can calculate the reward value using a piecewise function by comparing the risk prediction result MSE with a preset threshold of 0.08: if 5 points will be awarded; if , ;like A score of 0 is awarded. For example, the risk prediction result for a certain period. Substitute into the formula to calculate The system will add 4 points as a reward.

[0112] Rewards will be deducted if the load in the real-time status of the distribution network exceeds a preset load threshold or the resource idle rate in the maintenance resource data exceeds a preset idle threshold.

[0113] Among them, the preset load threshold refers to the safe upper limit of the rated load rate of the distribution network equipment, which is set to 80%; the preset idle threshold refers to the warning line of the idle rate of maintenance resources (such as personnel and equipment), which is set to 30%; the deduction reward refers to the negative feedback penalty applied when the load exceeds the limit or resources are idle, in order to optimize resource allocation and operational stability.

[0114] Specifically, the system can monitor the real-time load rate of the distribution network: if the load rate exceeds 80%, it will adjust the load rate based on the extent of the over-limit. The bonus and score will be deducted for exceeding the time limit T (in hours). For minor over-limit , Used for moderate over-limit ( , Used for severe over-limit Simultaneously monitor resource idle rate: if the idle rate exceeds 30%, adjust the idle rate accordingly. and idle time (Unit: hour) Deduct rewards, deduct points For mild idleness , For moderate idle time , Distributed to heavy idleness .

[0115] For example, the distribution network load rate is 85% during a certain period (exceeding the limit). (Mildly exceeding limits) and lasting for 1.5 hours, points will be deducted. Points will be deducted if the resource idle rate reaches 40% (mild idleness) for 3 consecutive hours. Points; Total deduction of reward is point.

[0116] Therefore, according to the above implementation method, the system can optimize the reinforcement learning strategy in real time through a dynamic reward mechanism, thereby improving the adaptability of the maintenance plan to the distribution network operation status.

[0117] In some embodiments, risk prediction results, real-time distribution network status, and maintenance resource data are used as inputs, and reinforcement learning with a dynamic reward mechanism is used for iterative optimization to generate an adaptive maintenance plan, including: Initialize the policy network parameters of the reinforcement learning model and set the maximum number of iteration rounds.

[0118] Among them, the policy network parameters refer to the set of neural network weights and biases used to output action policies in the reinforcement learning model; the maximum number of iterations refers to the upper limit of the maximum number of loops allowed during training, which is used to prevent overfitting and waste of computational resources.

[0119] Specifically, the system can set the initial weights of the policy network using the Xavier initialization method, with the bias term initialized to zero. The maximum number of iterations is set to 100, and the learning rate adopts a linear decay strategy, with an initial value of 0.01, decaying to 90% of its original value every 10 iterations. For example, the policy network is a fully connected neural network with an input layer dimension of 72 (corresponding to 72 hours of data features), 2 hidden layers, each with 128 units, and an output layer dimension of 10 (corresponding to 10 maintenance actions). The initial learning rate is 0.01, the maximum number of iterations is 100, and the learning rate is adjusted to [value missing] when it reaches the 50th iteration. .

[0120] Training data is loaded in batches, with each batch containing risk prediction results, real-time status of the distribution network, and maintenance resource data for a preset time span.

[0121] Here, "batch" refers to the grouping unit of training data, used for mini-batch gradient descent optimization; "preset time span" refers to the continuous time length covered by a single batch of data, in hours.

[0122] Specifically, the system can use a data loader to divide data into batches in chronological order. Each batch contains 72 hours of risk prediction results (including power outage probability and risk level), real-time distribution network status (including load rate, voltage deviation, and equipment status), and maintenance resource data (including personnel quantity, equipment quantity, and material inventory). The batch size is set to 30 groups. For example, starting from 00:00 on January 1, 2023, the first batch of data is loaded, covering 72 hours of data from 00:00 on January 1 to 23:59 on January 3, with a total of 30 samples; the second batch of data covers 00:00 on January 4 to 23:59 on January 6, and so on.

[0123] The reward value for each step is calculated based on a dynamic reward mechanism, and candidate maintenance solutions are output through a policy network.

[0124] Among them, the reward value refers to the numerical value of environmental feedback quantified according to preset rules, which is used to guide strategy optimization; the candidate maintenance plan refers to the action vector output by the strategy network, including maintenance time, region, and resource allocation plan.

[0125] Specifically, the system calculates the reward for each step using dynamic reward rules: +5 points for a risk prediction error below 0.05, -3 points for a load rate exceeding 80%, and -2 points for a resource idle rate exceeding 30%. The policy network takes the current state as input, outputs the action probability distribution through forward propagation, and samples candidate solutions. For example, the input state for a certain step might be a risk error of 0.04, a load rate of 85%, and a resource idle rate of 25%. After the strategy network outputs the action probability, it selects the action with the highest probability: "Schedule maintenance during the off-peak hours in region A and assign 2 personnel".

[0126] After each iteration, the achievement rate of optimization targets for risk prediction accuracy, distribution network stability, and resource utilization is calculated.

[0127] Among them, the optimization target achievement rate refers to the ratio of the actual performance index to the preset target value, which is used to evaluate the training progress; the risk prediction accuracy is measured by the reciprocal of the mean square error (MSE); the distribution network status stability is measured by the reciprocal of the load over-limit time ratio; and the resource utilization rate is measured by the resource utilization rate.

[0128] Specifically, the system can calculate the achievement rate of each objective using a formula: , , The weights of the three objectives are as follows: The overall achievement rate is a weighted sum. For example, the actual achievement rate after a certain iteration is... The percentage of time exceeding the load limit is 0.1%, and the resource utilization rate is 80%. , , ; .

[0129] In response to the optimization target achievement rate meeting the preset convergence condition, the iteration process is terminated and an adaptive repair plan is output.

[0130] Among them, the preset convergence condition refers to the threshold standard for training termination, which is usually the continuous achievement of the overall achievement rate; the adaptive maintenance plan refers to the final optimized maintenance strategy that conforms to the current distribution network status.

[0131] Specifically, the system can monitor the achievement rate within a window period: if the total achievement rate for 5 consecutive iterations is... If the success rate is 100%, then convergence is determined, training is terminated, and the final solution is output; otherwise, the reward weights or learning rate are adjusted, and iteration continues. For example, when iterating to the 45th round, the achievement rates for the five consecutive rounds are as follows: If all values ​​are ≥1.0, the system terminates the iteration and outputs the maintenance plan generated by the current strategy network.

[0132] Therefore, according to the above implementation method, the system can dynamically generate maintenance plans that match the real-time status of the distribution network through reinforcement learning and iterative optimization, thereby improving the adaptability and efficiency of the maintenance plan.

[0133] In some embodiments, based on the adaptive maintenance scheme, an adaptive plan for power outage maintenance of the distribution network is generated, including timing arrangements, regional priorities, and real-time feedback, comprising: Based on the load curve in the real-time status of the distribution network, peak hours, off-peak hours and low-peak hours are divided, and the execution sequence of the adaptive maintenance plan is arranged based on the division results.

[0134] Among them, the load curve is a graphical representation of the distribution network load rate changing over time, used for visual analysis of electricity consumption patterns; peak electricity consumption period refers to the time period when the load rate exceeds 80%, off-peak period refers to the time period when the load rate is between 50% and 80%, and low-peak period refers to the time period when the load rate is below 50%.

[0135] Specifically, the system can extract historical load data from the real-time status of the distribution network through the load analysis module, automatically divide time period boundaries using clustering algorithms (such as K-means clustering, a classic unsupervised machine learning algorithm used to automatically divide a set of data points into K mutually exclusive clusters, whose core objective is to maximize the similarity of data points within a cluster while minimizing the similarity between different clusters), and optimize the maintenance sequence arrangement based on user power supply guarantee constraints (such as periods when power outages are not allowed for highly sensitive users).

[0136] For example, based on summer load data for a certain region, the system analysis shows that peak hours are 7:00-10:00 and 17:00-22:00 (load rate 85%-95%), off-peak hours are 10:00-17:00 (load rate 60%-75%), and low-peak hours are 22:00-7:00 (load rate 30%-45%). The maintenance schedule for hospital user areas is arranged during the low-peak hours of 22:00-2:00, with each maintenance session lasting no more than 4 hours.

[0137] By combining the distribution of high-risk areas in the risk prediction results, the distribution of load pressure areas in the real-time status of the distribution network, and the resource coverage in the maintenance resource data, a regional priority scoring system is constructed, and the processing order of maintenance areas is determined based on the scoring system.

[0138] Among them, the regional priority scoring system refers to a quantitative model that calculates the urgency of regional maintenance by weighted summation; the high-risk area distribution refers to the geographical range with a risk prediction level of "extremely high risk" or "high risk"; the load pressure area distribution refers to the range of lines or transformer areas where the load rate in the real-time status of the distribution network continuously exceeds 80%; and the resource coverage range refers to the area that maintenance resources (such as personnel and vehicles) can reach within 1 hour.

[0139] Specifically, the system assigns three scores to each region through a scoring calculation module: risk prediction score (weight 0.4, assigned based on risk level, 100 points for extremely high risk, 80 points for high risk), load pressure score (weight 0.3, assigned based on load rate, 100 points for load rate ≥ 80%, 60 points for load rate 60%~80%), and resource coverage score (weight 0.3, assigned based on resource arrival time, 100 points for ≤ 1 hour, 60 points for 1~3 hours). The total score is a weighted sum, and regions are sorted from highest to lowest score. For example, a region might have a risk prediction score of 80 (high risk), a load pressure score of 100 (load rate 85%), and a resource coverage score of 60 (arrival time 1.5 hours). After sorting, this area will be ranked first in priority and will be given priority in the allocation of maintenance resources.

[0140] During the execution of the maintenance plan, deviation data of risk prediction results, real-time status change data of distribution network, and consumption data of maintenance resources are continuously collected. Based on the comparison results of the collected data and the adaptive maintenance plan, the deviation type is identified, and the maintenance strategy is dynamically adjusted to obtain an adaptive plan for power outage maintenance of distribution network.

[0141] Deviation data refers to the set of differences between actual measured values ​​and planned values, including time deviation, regional deviation, and resource deviation; change data refers to real-time fluctuation information such as distribution network load rate and equipment status; consumption data refers to resource consumption records such as man-hours and material usage during maintenance.

[0142] Specifically, the system can acquire three types of data in real time through a data acquisition interface, use a sliding window (window length of 1 hour, step size of 30 minutes) to calculate deviations, and identify the types of deviations through a rule engine: timing deviation (difference between maintenance start / end time and plan exceeds 30 minutes), area deviation (actual maintenance area does not match the planned area), and resource deviation (resource consumption exceeds the planned value by more than 20%). For example, during maintenance, if the system collects data showing that the planned start time for maintenance on a certain line is 8:00, but the actual start time is 8:40, a timing deviation of 40 minutes, after identifying the timing deviation, the system will automatically postpone subsequent maintenance tasks by 40 minutes and reallocate resources to avoid conflicts.

[0143] Therefore, according to the above implementation method, the system can generate a maintenance plan that dynamically adapts to the operation status of the distribution network through timing optimization, priority sorting and real-time feedback adjustment, thereby improving maintenance efficiency and power supply reliability.

[0144] Figure 6 This is a structural block diagram of a power outage maintenance plan generation system according to an embodiment of the present invention.

[0145] like Figure 6 As shown, the power outage maintenance plan generation system for this distribution network includes: The distribution network data pool creation module 210 is used to acquire multi-source distribution network operation data and perform fusion preprocessing on the multi-source distribution network operation data, including time sequence alignment, outlier handling and missing value completion, to obtain the distribution network data pool.

[0146] The risk linkage mapping relationship construction module 220 is used to extract the distribution area layer features, line layer features and user layer features from the distribution network data pool, and construct a three-level risk linkage mapping relationship based on the distribution area layer features, line layer features and user layer features.

[0147] The frequent power outage risk prediction model construction module 230 is used to construct a frequent power outage risk prediction model based on the three-level risk linkage mapping relationship and combined with the power supply capacity constraints of the distribution network, the equipment maintenance cycle constraints, and the user power supply guarantee constraints.

[0148] The adaptive maintenance plan generation module 240 is used to take risk prediction results, real-time status of distribution network and maintenance resource data as input, and perform iterative optimization through reinforcement learning with a dynamic reward mechanism to generate an adaptive maintenance plan.

[0149] The power outage maintenance plan generation module 250 is used to generate an adaptive power outage maintenance plan for the distribution network, including timing, regional priority, and real-time feedback, based on the adaptive maintenance scheme.

[0150] The specific functions and examples of each module and submodule of the device in this embodiment can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0151] According to embodiments of the present invention, the above-described method of the present invention can be applied to an electronic device and a readable storage medium.

[0152] Figure 7 A schematic block diagram of an electronic device 600 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0153] like Figure 7 As shown, the electronic device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 602 or a computer program loaded from a storage unit 608 into a random access memory (RAM) 603. The RAM 603 may also store various programs and data required for the operation of the electronic device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0154] Multiple components in electronic device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of displays, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows electronic device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0155] The computing unit 601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as a distribution network outage maintenance plan generation method. For example, in some embodiments, a distribution network outage maintenance plan generation method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the distribution network outage maintenance plan generation method described above can be performed. Alternatively, in other embodiments, the computing unit 601 may be configured by any other suitable means (e.g., by means of firmware) to perform a distribution network outage maintenance plan generation method.

[0156] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0157] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0158] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0159] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0160] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0161] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0162] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0163] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for generating a power outage maintenance plan for a distribution network, characterized in that, include: Acquire multi-source distribution network operation data, and perform fusion preprocessing on the multi-source distribution network operation data, including time sequence alignment, outlier handling, and missing value completion, to obtain the distribution network data pool; The distribution network data pool extracts transformer area features, line features, and user features, and a three-level risk linkage mapping relationship is constructed based on the transformer area features, line features, and user features. Based on the aforementioned three-level risk linkage mapping relationship, and combined with the power supply capacity constraints of the distribution network, the equipment maintenance cycle constraints, and the user power supply guarantee constraints, a frequent power outage risk prediction model is constructed. The risk prediction results, real-time status of the distribution network, and maintenance resource data are used as inputs. The system is iteratively optimized through reinforcement learning with a dynamic reward mechanism to generate an adaptive maintenance plan. Based on the adaptive maintenance scheme, an adaptive maintenance plan for power outages in the distribution network, including timing arrangements, regional priorities, and real-time feedback, is generated.

2. The method according to claim 1, characterized in that, The multi-source distribution network operation data includes low-voltage user power outage records, transformer area equipment operating parameters, line load data, maintenance history records, and meteorological image data; the steps of extracting transformer area layer features, line layer features, and user layer features include: Extracting features at the transformer substation level, including calculating the correlation between load fluctuations and the number of power outages based on the operating parameters of the equipment in the substation and power outage records, and marking high-risk load intervals; Extracting line layer features includes analyzing the correlation between fault propagation paths and power supply radius based on the line load data and fault history records, and determining the fault impact range; Extracting user-level features includes classifying user power outage sensitivity levels based on the type of electricity used and the frequency of power outage complaints in the low-voltage user power outage records.

3. The method according to claim 2, characterized in that, The construction of a three-level risk linkage mapping relationship based on the characteristics of the distribution area layer, line layer, and user layer includes: Based on the correlation coefficient between load fluctuation and power outage frequency in the transformer area layer characteristics, the first frequent power outage threshold of the corresponding level of the transformer area layer characteristics is determined; Based on the correlation between the fault propagation path and the power supply radius in the line layer characteristics, the second frequent power outage threshold of the corresponding level of the line layer characteristics is determined; Based on the electricity consumption type and the frequency of power outage complaints in the user layer characteristics, the third frequent power outage threshold of the corresponding layer of the user layer characteristics is determined; Based on the first frequent power outage threshold, the second frequent power outage threshold, the third frequent power outage threshold, and the linkage rules based on the risk propagation logic between levels, the three-level risk linkage mapping relationship is constructed.

4. The method according to claim 1, characterized in that, Based on the aforementioned three-level risk linkage mapping relationship, and combined with distribution network power supply capacity constraints, equipment maintenance cycle constraints, and user power supply guarantee constraints, a frequent power outage risk prediction model is constructed, including: Based on the aforementioned three-level risk linkage mapping relationship, and incorporating the constraints of the power distribution network supply capacity, the equipment maintenance cycle, and the user power supply guarantee, a constrained risk prediction basis is obtained. A constraint weight adjustment mechanism is set for the distribution network power supply capacity constraint, the equipment maintenance cycle constraint, and the user power supply guarantee constraint to dynamically balance the distribution network power supply capacity constraint, the equipment maintenance cycle constraint, and the user power supply guarantee constraint. The frequent power outage risk prediction model is constructed based on the constrained risk prediction basis and the constraint weight adjustment mechanism. The model training data is updated using a sliding time window, and a risk prediction error threshold is set.

5. The method according to claim 1, characterized in that, The dynamic reward mechanism is configured to: Reward rules are set based on the risk prediction results, real-time status of the distribution network, and maintenance resource data, and the reward calculation weight is adjusted based on the dynamic changes of the reward rules. The reward rules include: If the error of the risk prediction result is lower than a preset error threshold, the reward is increased; If the load in the real-time status of the distribution network exceeds a preset load threshold or the resource idle rate in the maintenance resource data exceeds a preset idle threshold, the reward will be deducted.

6. The method according to claim 5, characterized in that, The process involves taking the risk prediction results, real-time distribution network status, and maintenance resource data as input, and iteratively optimizing the solution through reinforcement learning with a dynamic reward mechanism to generate an adaptive maintenance plan, including: Initialize the policy network parameters of the reinforcement learning model and set the maximum number of iteration rounds; Training data is loaded in batches, with each batch containing risk prediction results, real-time status of the distribution network, and maintenance resource data for a preset time span; The reward value for each step is calculated based on the dynamic reward mechanism, and candidate maintenance solutions are output through the policy network. After each iteration, the achievement rate of optimization targets for risk prediction accuracy, distribution network stability, and resource utilization is calculated. In response to the optimization target achievement rate satisfying the preset convergence condition, the iteration process is terminated and the adaptation repair scheme is output.

7. The method according to claim 1, characterized in that, The process of generating an adaptive maintenance plan for distribution network outages, based on the aforementioned adaptive maintenance scheme, including timing arrangements, regional priorities, and real-time feedback, includes: Based on the load curve in the real-time status of the distribution network, peak electricity consumption periods, off-peak periods, and low-peak periods are divided, and the execution sequence of the adaptability maintenance plan is arranged based on the division results. Combining the distribution of high-risk areas in the risk prediction results, the distribution of load pressure areas in the real-time status of the distribution network, and the resource coverage in the maintenance resource data, a regional priority scoring system is constructed, and the processing order of maintenance areas is determined based on the scoring system. During the execution of the maintenance plan, deviation data of the risk prediction results, changes in the real-time status of the distribution network, and consumption data of maintenance resources are continuously collected. Based on the comparison results between the collected data and the adaptive maintenance plan, the deviation type is identified, and the maintenance strategy is dynamically adjusted to obtain the adaptive plan for power outage maintenance of the distribution network.

8. A power outage maintenance plan generation system for a distribution network, characterized in that, include: The distribution network data pool creation module is used to acquire multi-source distribution network operation data and perform fusion preprocessing on the multi-source distribution network operation data, including time sequence alignment, outlier handling and missing value completion, to obtain the distribution network data pool. The risk linkage mapping relationship construction module is used to extract transformer area features, line layer features and user layer features from the distribution network data pool, and construct a three-level risk linkage mapping relationship based on the transformer area features, line layer features and user layer features. The frequent power outage risk prediction model construction module is used to construct a frequent power outage risk prediction model based on the three-level risk linkage mapping relationship and combined with the power supply capacity constraints of the distribution network, the equipment maintenance cycle constraints, and the user power supply guarantee constraints. The adaptive maintenance plan generation module is used to take the risk prediction results, real-time status of the distribution network and maintenance resource data as input, and perform iterative optimization through reinforcement learning with a dynamic reward mechanism to generate an adaptive maintenance plan. The power outage maintenance plan generation module is used to generate an adaptive power outage maintenance plan for the distribution network, including timing arrangements, regional priorities, and real-time feedback, based on the adaptive maintenance scheme.

9. An electronic device, characterized in that, include: At least one processor; and a memory that is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, in, Computer instructions are used to cause a computer to perform the method according to any one of claims 1-7.

Images