Visual monitoring and shooting and intelligent identification external damage early warning method and system

Abstract

The invention discloses a visual monitoring and shooting and intelligent identification external damage early warning method and system, and relates to the field of power transmission line safety monitoring. The method comprises the following steps: acquiring external damage accident information; reversely tracing from the power grid topology database based on the accident occurrence position to determine an affected user list; constructing a causal inference model, and carrying out anti-factual estimation on the power consumption of the affected user when no external damage is assumed to occur in the accident influence time period based on the power consumption characteristic data of the affected user in the historical accident-free time period; acquiring actual electricity consumption of a user; determining an electric quantity loss causal effect value according to a difference value between the actual electricity consumption and the anti-factual electricity consumption; and according to the effect value and user electricity price information obtained from the marketing system, generating an external damage follow-up loss list containing the electricity loss cost. According to the method, accurate accounting of the electric quantity loss is realized through a causal inference technology, and a full-link closed-loop system from pre-event early warning to post-event follow-up is constructed in combination with power grid topology analysis and multi-system data fusion.

Description

A Visual Monitoring and Intelligent Identification Method and System for External Damage Early Warning Technical Field

[0001] This application belongs to the field of power transmission line safety monitoring, specifically involving a method and system for visual monitoring and intelligent identification of external damage early warning. Background Technology

[0002] With the continuous expansion of power grid construction, the geographical environment of transmission and distribution lines is becoming increasingly complex. Early warning and prevention of external damage have become core tasks for ensuring the safe and stable operation of the power system. Traditional power line inspections mainly rely on manual patrols or basic video surveillance, aiming to achieve a preliminary understanding of the surrounding environment through real-time image transmission. In the context of digital transformation, visual monitoring technology, by deploying high-precision sensing equipment, can acquire massive amounts of on-site image data, providing a necessary data foundation for the timely detection of potential external damage hazards such as crane violations or accidental mechanical collisions.

[0003] Among them, the visual monitoring and intelligent identification technology for external damage early warning is a key direction for refined operation and maintenance of the power grid. Its core lies in using digital spatial technology and artificial intelligence algorithms to deeply analyze monitoring images. This technology aims to build an intelligent prevention and control system that integrates hazard identification, risk assessment, and early warning triggering through real-time monitoring of the dynamic environment around the power lines. This allows for early intervention in external damage risks before physical contact occurs, reducing the negative impact of external damage accidents on power grid assets and power supply reliability.

[0004] However, existing technologies still have significant limitations in addressing external damage hazards in complex field environments. First, traditional monitoring methods primarily focus on physical vibration sensing or basic feature extraction, lacking deep AI-powered identification capabilities for specific external damage sources such as agricultural drones and harvesters, making them prone to false alarms under complex lighting conditions. Second, the early warning system lacks deep coupling with the power grid topology, making it difficult to quickly pinpoint affected user groups and accurately assess the impact range after an incident. Furthermore, existing methods for calculating power loss and economic losses caused by external damage often employ linear estimations, failing to accurately simulate nonlinear energy fluctuations and lacking closed-loop linkage with insurance claims systems. This results in cumbersome claims processes and insufficient loss calculation accuracy, failing to meet the needs of modern power grids for comprehensive claims management.

[0005] Therefore, a method and system for early warning of external damage using visual monitoring and intelligent identification is desired. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for visual monitoring and intelligent identification of external damage early warning, which can effectively solve the problems in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for early warning of external damage through visual monitoring and intelligent identification, comprising the following steps: acquiring accident information of an external damage accident, the accident information including at least the location of the accident; based on the location of the accident, tracing back from the power grid topology database to determine the power outage area affected by the external damage accident and the list of affected users; constructing a causal inference model, based on the electricity consumption characteristic data of the affected users during historical accident-free periods, and counterfactually estimating the counterfactual electricity consumption of the affected users during the period affected by the external damage accident, assuming no external damage occurred; acquiring the actual electricity consumption of the affected users during the period affected by the external damage accident; determining the causal effect value of the power loss caused by the external damage accident based on the difference between the actual electricity consumption and the counterfactual electricity consumption; and generating a list of external damage recovery losses including electricity loss fees based on the causal effect value of the power loss and the electricity price information of the affected users obtained from the marketing system.

[0008] Preferably, the step of tracing back from the power grid topology database to determine the outage area and the list of affected users affected by the external failure includes: mapping the location of the failure to the corresponding physical power grid node in the digital power grid topology data; starting from the physical power grid node, using a graph search algorithm to traverse upstream along the power grid topology until a power source is found to determine the starting boundary of the fault segment, and traversing downstream along the power grid topology until a tie switch or the end of the line is reached to delineate the fault impact range; based on the fault impact range, associating and generating a list of affected users from the electricity consumption information collection system or marketing system.

[0009] Preferably, the step of constructing a causal inference model and counterfactually estimating the counterfactual electricity consumption of the affected users during the period of the external damage incident includes: constructing a causal structure diagram describing the causal relationship between variables affecting electricity consumption, wherein the variables include at least the external damage incident intervention variable, historical load characteristics, meteorological characteristics, time characteristics, and electricity consumption as the outcome variable; training a structural equation model using historical data based on the causal structure diagram to learn the nonlinear mapping relationship between variables; for each time point during the period of the external damage incident, forcibly setting the intervention variable to a state where no external damage has occurred, and inputting the historical load characteristics, meteorological characteristics, and time characteristics corresponding to that time point into the trained structural equation model to calculate the counterfactual electricity consumption.

[0010] Preferably, after determining the causal effect value of power loss caused by the external damage accident based on the difference between the actual power consumption and the counterfactual power consumption, the method further includes: constructing a deep reinforcement learning power loss analysis model, wherein the state space of the deep reinforcement learning power loss analysis model includes historical power consumption sequences, meteorological characteristics, and temporal characteristics, and the action space is the adjustment coefficient for the predicted power consumption value for the next time period; when training the deep reinforcement learning power loss analysis model, the causal effect value of power loss is introduced as a dynamic penalty term into the reward function to guide the model to learn predictions that conform to causal laws; and using the trained deep reinforcement learning power loss analysis model to perform refined prediction of time-series power consumption in scenarios without external damage.

[0011] Preferably, the step of introducing the causal effect value of power loss as a dynamic penalty term into the reward function specifically includes: constructing a dynamic penalty term based on the causal effect value of power loss; combining the original reward function with the dynamic penalty term to form an effective reward function, wherein the original reward function is constructed based on the prediction error, and the effective reward function is used to penalize the deviation of the causal effect implied in the prediction result from the causal effect value of power loss while optimizing the prediction accuracy during model training.

[0012] Preferably, the step of generating an external damage recovery loss list containing electricity loss costs based on the causal effect value of electricity loss and the electricity price information of the affected users obtained from the marketing system includes: obtaining the typical load curve of each affected user in the same period before the fault occurred from the electricity information collection system through a standardized interface based on the list of affected users; calculating the proportion of the typical electricity consumption of each affected user during the period affected by the external damage accident to the total typical electricity consumption of all affected users based on the typical load curve of each affected user; multiplying the proportion by the causal effect value of electricity loss to allocate the total electricity loss to each affected user to obtain the electricity loss value of each user; obtaining the tiered electricity price information of each affected user from the marketing system; calculating the electricity loss cost of each user based on the electricity loss value of each user and its corresponding tiered electricity price, and summarizing to generate the external damage recovery loss list.

[0013] Preferably, the step of generating an external damage recovery loss list that includes electricity loss costs further includes: obtaining asset information of damaged power equipment from the Enterprise Resource Planning (ERP) system through a standardized data interface, wherein the asset information includes the original purchase cost and accumulated depreciation; calculating power facility loss costs based on the asset information of the damaged power equipment; and integrating the power facility loss costs into the external damage recovery loss list.

[0014] Preferably, after calculating the power facility loss cost based on the asset information of the damaged power equipment, the method further includes: correcting the power facility loss cost using preset factor coefficients according to the compliance requirements of insurance claims to generate power facility loss costs that conform to the insurance claim criteria; and automatically pushing the external damage recovery loss list, which includes the power loss cost and the corrected power facility loss cost, to the external insurance claims system through a RESTful API to initiate the claims process.

[0015] Preferably, the steps for obtaining accident information of external damage accidents specifically include: receiving early warning information from a monitoring terminal, wherein the early warning information is triggered by the monitoring terminal through real-time analysis of the scene image by an integrated image recognition processing engine, wherein the image recognition processing engine uses a deep learning object detection network to identify the hazard source and uses a recurrent neural network to encode the motion trajectory of multiple consecutive frames to match a preset dangerous operation action mode.

[0016] Secondly, a visual monitoring and intelligent identification system for external damage early warning includes: an information acquisition module for acquiring accident information of external damage accidents, the accident information including at least the location of the accident; a topology analysis module for tracing back from the power grid topology database based on the accident location and determining the power outage area affected by the external damage accident and a list of affected users; a causal inference module for constructing a causal inference model, based on the electricity consumption characteristic data of the affected users during historical accident-free periods, and counterfactually estimating the counterfactual electricity consumption of the affected users during the period affected by the external damage accident, assuming no external damage occurred; a data acquisition module for acquiring the actual electricity consumption of the affected users during the period affected by the external damage accident; a causal effect calculation module for determining the causal effect value of the electricity loss caused by the external damage accident based on the difference between the actual electricity consumption and the counterfactual electricity consumption; and a loss list generation module for generating an external damage recovery loss list including electricity loss costs based on the causal effect value of the electricity loss and the electricity price information of the affected users obtained from the marketing system.

[0017] In summary, this application includes at least one of the following beneficial technical effects: 1. By constructing a causal inference model, this invention estimates the electricity consumption when no external damage occurs through counterfactual estimation, accurately calculates the causal effect value of electricity loss caused by the external damage accident, overcomes the defect of traditional linear estimation that cannot simulate nonlinear energy fluctuations, provides scientific and accurate data support for loss tracking, and significantly improves the accuracy of accounting.

[0018] 2. Based on the location of the accident, this invention uses a graph search algorithm to trace the fault segment backward from the power grid topology database and search the affected area forward, locking the list of affected users in seconds. This enables accurate impact assessment from the external fault point to specific users, significantly improving the efficiency and accuracy of accident impact analysis and laying the foundation for subsequent loss accounting.

[0019] 3. This invention automatically obtains equipment asset information and user electricity price data by connecting with ERP, marketing and external insurance systems, generates a loss list that conforms to the insurance claim criteria and automatically pushes it to the insurance company via RESTful API, and builds a closed-loop management of the entire life cycle from pre-warning to post-claims, which greatly shortens the claims cycle and reduces human error. Attached Figure Description

[0020] Figure 1 is a flowchart illustrating the external damage early warning method based on visual monitoring and intelligent identification in this invention; Figure 2 is a flowchart illustrating the power loss analysis based on deep reinforcement learning and causal inference algorithms in this invention; Figure 3 is a flowchart illustrating the stages of visual monitoring, artificial intelligence identification, and automatic early warning of external damage hazards in this invention; Figure 4 is a diagram illustrating the multi-level interaction between the power grid production / marketing system and the external insurance claims system in this invention; Figure 5 is a flowchart illustrating the user impact analysis and tiered electricity price loss calculation based on grid topology relationships in this invention. Detailed Implementation

[0021] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of specific embodiments based on the present invention is provided in conjunction with the accompanying drawings and preferred embodiments.

[0022] Referring to Figures 1 to 5, a method for early warning of external damage using visual monitoring and intelligent identification includes the following steps: Step S1, performing early warning of external damage accidents to power transmission and distribution lines.

[0023] In this embodiment, each transmission and distribution line is assigned one or more maintenance personnel in advance, referred to as the area manager, who is responsible for the daily inspection and emergency response of the line, and their contact information is pre-entered into the system.

[0024] Step S101: First, construct a visualization monitoring architecture for transmission and distribution lines based on the digital space of the power grid to achieve comprehensive and real-time monitoring of lines with voltage levels of 10 kVA and above in a specific area.

[0025] In the specific implementation process, the Geographic Information System (GIS) is used as the underlying support platform. Each monitoring terminal is digitally modeled, and its spatial coordinates are accurately mapped to the corresponding position on the GIS base map based on the equipment ledger or on-site measurement data. At the same time, a monitoring terminal list management module is established. This module is linked with the equipment ledger database in real time, supports fuzzy search by line name, and displays a complete list of physical towers to which the line belongs in the search results.

[0026] When a monitoring terminal is identified as being installed on a physical pole through the equipment ledger, a monitoring terminal identifier will be displayed next to the pole name in the form of a specific icon, making it easy for the area manager to quickly locate the pole with monitoring resources.

[0027] Step S102 provides a multi-dimensional visual monitoring canvas management function. The area manager can freely set the video window splitting mode on the operation interface according to the specific monitoring task requirements. It supports switching between three window layouts: 1 grid, 2 grids, and 4 grids, and realizes the synchronous real-time display of multiple monitoring images.

[0028] When a user switches the window to single-frame full-screen display mode, the current screen can be zoomed in and out without stepless zooming. The image scaling algorithm maintains the image clarity so that the substation manager can capture the microscopic details of the surrounding environment, such as the specific model of the construction machinery or the operation behavior of the workers.

[0029] Step S103: Integrate artificial intelligence recognition technology into the image recognition processing engine in the digital space to perform real-time streaming analysis on the monitoring images transmitted from the front end.

[0030] The artificial intelligence recognition module adopts a deep learning-based target detection network. In this embodiment, YOLOv5 is selected as the basic framework. This network automatically extracts multi-scale features from images through multi-layer convolution and pooling operations. For preset hazardous source object categories such as agricultural drones, harvesters, cranes, excavators, and pile drivers, on-site images containing various construction machinery under different lighting, angles, and backgrounds are collected in advance. The target objects in each image are manually labeled to form a training sample set. Then, the YOLOv5 network is trained using this sample set through transfer learning. After training, the network model is deployed in the image recognition processing engine for real-time image analysis.

[0031] In step S104, during the real-time recognition process, the real-time video stream is decoded frame by frame and scaled to the network input size, then input into the trained object detection network. The network outputs the bounding box coordinates, confidence scores, and category labels of the detected objects.

[0032] Further post-processing of the detection results: Non-maximum suppression algorithm is used to remove redundant overlapping boxes and retain the detection results with the highest confidence; Kalman filter algorithm is used to track the center point of the bounding box of the same target in multiple consecutive frames to form the target's motion trajectory; at the same time, the geometric proportion features of the target are extracted based on the aspect ratio of the bounding box.

[0033] The real-time features mentioned above are compared item by item with the standard feature data of various types of machinery pre-stored in the hazard source feature library. The hazard source feature library is constructed as follows: for each type of hazardous machinery, sample images under multiple angles and postures are collected in advance, and its depth feature vector is extracted through the same convolutional neural network and stored as a reference template; at the same time, for typical hazardous operation actions such as crane boom rotation and excavator continuous digging, the motion trajectory of multiple consecutive frames is encoded using a recurrent neural network to form an action pattern template.

[0034] During the comparison, the similarity between the current target depth features and each baseline template is calculated. If the highest similarity exceeds a preset threshold, the category is determined to match. At the same time, the motion trajectory is matched with the action pattern template. If the matching distance is less than the threshold, the behavior is determined to conform to the dangerous mode.

[0035] Step S105: Once the system comprehensively determines that the object in the current image belongs to a suspected externally damaged machinery and its behavior characteristics conform to the preset dangerous operation mode, the early warning mechanism is immediately triggered.

[0036] The early warning information generation module automatically assembles early warning content, which precisely includes the geographical coordinates (latitude and longitude), physical tower number, precise time of the incident, and on-site images captured by the monitoring terminal that identify the type of hazard. Subsequently, through a standardized SMS service interface, this early warning information is automatically pushed to the mobile terminal of the pre-set responsible area manager for the line, ensuring that relevant personnel can be informed of the on-site risks as soon as possible and take swift intervention measures.

[0037] Step S2: Establish a standardized online prosecution process.

[0038] In this embodiment, the online prosecution process involves two types of system users: area managers and auditors. The area manager is the aforementioned area manager, who is responsible for entering accident information and initiating prosecutions; the auditor is responsible for reviewing and confirming the preliminary loss report generated by the system.

[0039] Step S201: Provide a standardized accident entry template in the external damage investigation subspace. The template includes basic information fields such as accident name, accurate accident time, type of affected power grid facilities, and preliminary cause of accident. Types of power grid facilities include cables, overhead lines, and towers.

[0040] The district manager fills in and submits basic information about the external damage accident using this template, thus completing the digital registration of the accident.

[0041] Step S202: Display the details of the current accident in a floating window on the GIS map. At the same time, allow the area manager to accurately mark the specific location coordinates of the external break point on the map by clicking with the mouse, and automatically capture the latitude and longitude data of the point and store it in the accident record.

[0042] Step S203: Based on the calibrated location of the external fault point, the spatial calculation engine of the digital space is invoked, and combined with the pre-stored power grid topology data, the reverse tracing and forward search of the faulty line are automatically performed.

[0043] The power grid topology data is organized using a graph structure, where each physical tower or device is abstracted as a node, and lines are abstracted as edges.

[0044] Starting from the physical tower node where the calibration point is located, a depth-first search algorithm is used to traverse upstream along the power grid topology until the power source point is found, and the starting boundary of the fault section is determined. At the same time, all connected lines and equipment are traversed downstream until the tie switch or the end of the line, and the fault range is delineated. The affected fault section line and the number of all users associated with the line downstream can be identified within seconds, and the results are presented in the form of a list.

[0045] Step S204: Call the pre-trained neural network loss prediction model. This model adopts a multi-layer feedforward neural network structure to quickly estimate the amount of various losses based on the fault characteristics.

[0046] The input features of the neural network include the number of affected users, the length of the faulty line segment, the type and number of damaged equipment, and the time period of the fault. The network consists of an input layer, two hidden layers, and an output layer. The output layer has three neurons, which correspond to the preliminary estimates of the amount of electricity loss, the amount of power facility loss, and other related costs.

[0047] The training dataset for the neural network loss prediction model comes from a historical external damage case database. Each case sample contains the above-mentioned input features and the corresponding actual loss amount label. During training, mean squared error is used as the loss function, and the Adam optimizer is used to optimize the parameters. After training, the model is used for online prediction.

[0048] Furthermore, by obtaining the tiered electricity price data from the affected user information ledger through a standardized data interface, and combining it with the loss amount output by the model, a preliminary loss report is generated, which includes a list of electricity loss, a list of power facility loss, and other expenses.

[0049] Step S205: The system has a built-in insurance claims integration module. This module establishes a secure connection with external insurance claims systems through a standardized, expressive state passing application programming interface (RESTful API). External insurance claims systems refer to the claims business systems of insurance companies, such as the Yingda Property Insurance system. This module uses JavaScript object notation in JSON format to encapsulate claims data and supports two-way authentication and encrypted data transmission to ensure the security and compliance of data exchange.

[0050] In step S206, the preliminary loss report generated in step S204 is converted into a claim list format that meets the requirements of the insurance company and submitted to the internal review process. The relevant reviewers can view the claim list and related supporting documents online through the system and perform the review operation. If the review is approved, the final claim list will be automatically pushed to the insurance company through the interface established in step S205.

[0051] The entire claims process is compressed into a standard cycle of 7 to 15 working days, achieving closed-loop management of the entire lifecycle of external damage accidents from occurrence to claims settlement.

[0052] Step S3: Perform equipment loss calculation and analysis.

[0053] Step S301: First, a refined asset association model is established. This model establishes a real-time linkage relationship between the equipment's unique identifier and the asset inventory list in the Enterprise Resource Planning (ERP) system. The asset inventory list records the equipment's original purchase cost and accumulated depreciation, among other financial information.

[0054] In practice, it connects with the ERP system through a standardized interface, supporting real-time querying or timed synchronization of asset data. When an external damage accident occurs, the system automatically retrieves the original purchase cost, accumulated depreciation rate, and current net asset value calculated by subtracting accumulated depreciation from the original purchase cost from the ERP system based on the identification code of the damaged equipment. This forms the basis for subsequent assessment of the equipment's value loss.

[0055] In step S302, on the operation interface, the substation manager selects the name of the line that has been damaged and the specific damaged tower number by clicking. After receiving the selection instruction triggered by the mouse click operation, the system automatically retrieves and lists all the auxiliary facilities associated with the tower from the equipment ledger database. The list includes pole-mounted transformers, disconnect switches, hardware, crossarms, insulators, and conductor segments of different types.

[0056] Each facility entry is accompanied by its physical ID, specifications, and current status information, making it easy for the area manager to confirm the damage situation one by one. The equipment ledger database is kept synchronized with the production management system PMS3.0 to ensure the accuracy of the information.

[0057] Step S303: For calculating material costs, based on the actual quantity of each type of facility selected and filled in by the area manager in the list, the total material cost is automatically multiplied by the latest purchase price of the material obtained synchronously from the ERP system. The calculation formula is: the total material cost equals the sum of the products of the quantity of each damaged piece of equipment and its corresponding unit price. This process requires no manual intervention, ensuring the accuracy and timeliness of the calculation.

[0058] Step S304 provides a dedicated manual intervention interface for non-standardized cost items, including on-site repair personnel hiring costs, repair vehicle usage costs, and compensation costs for crop damage caused to third parties.

[0059] The district manager can manually enter the specific amount in the corresponding input box based on the actual work order records, and upload relevant supporting documents as attachments. These manually entered data and automatically calculated material costs will be stored together in the accident database.

[0060] In step S305, the cost calculation engine starts the integrated calculation. The engine adds up the material costs automatically calculated in step S303 and the various non-standardized costs manually entered in step S304 to generate a preliminary list of external damage to power facilities. This list details the cost categories, quantities, unit prices, amounts, and total costs in tabular form.

[0061] Step S306: To meet the compliance requirements of insurance claims, the preliminary list is automatically revised according to the pre-configured factor coefficients. The factor coefficients are provided by the insurance company on a regular basis and preset in the system, such as depreciation discount coefficients and residual value recovery coefficients for different equipment types.

[0062] The costs of each material in the preliminary list are multiplied by the corresponding factor coefficient to calculate the power facility loss costs that meet the insurance claim criteria. The revised list ensures that the data for investigation is consistent with the data submitted for subsequent claims, while retaining the original costs as the basis for internal investigation.

[0063] Step S4: Conduct user impact analysis.

[0064] Step S401: Relying on the high-precision image map of the digital space as the visualization base, the built-in spatial analysis algorithm is called to determine the power outage affected area. The topology data of the power grid is pre-stored in a graph structure, in which equipment such as poles, switches, and transformers are abstracted as nodes, and lines are abstracted as directed edges, with the direction set according to the actual power flow.

[0065] Starting from the physical tower where the external break point is located, a depth-first search algorithm is used to traverse all connected lines and equipment downstream along the power supply direction until a tie switch or the end of the line is encountered.

[0066] After the traversal is completed, all lines and equipment in the affected area are rendered with dynamic highlight colors on the geographic information system (GIS) map to intuitively show the scope of the power outage.

[0067] Step S402 allows the area manager to interact with the map via mouse hover. When the mouse hovers over a section of line within the highlighted area, the current load current, active power, tie switch status, and upstream and downstream topology connections of that section of line are obtained in real time from the control cloud platform or production management system PMS3.0. The data is then dynamically refreshed and displayed in a table or curve format in the hover window, making it easier for the area manager to quickly grasp the on-site operating conditions.

[0068] Step S403: Based on the power grid topology and user profile data from the electricity consumption information collection system, automatically count the number of transformer substations in the affected area. During the count, first select the distribution transformer nodes from the affected area, obtain the unique identifier of each substation, and then use the identifier to associate with the user information ledger in the marketing system to calculate the total number of residential and enterprise users connected to each substation.

[0069] The statistical results are presented in a list format in the sidebar of the operation interface, including key fields such as station name, station number, and total number of users.

[0070] Step S404: When the area manager clicks on the icon of any affected transformer on the GIS map, the geographical location of the area is immediately highlighted with a flashing animation, and an information pop-up window appears.

[0071] The floating window displays the ledger information for the transformer area, including the area name, equipment model, commissioning date, and operating status, including current voltage, current, and load rate. The historical load curve is drawn by calling the load data of the past 24 hours or week from the electricity information collection system. It also lists brief information about each household or enterprise associated with the transformer area, such as household number, household name, and electricity address.

[0072] The regional manager can click on a specific user to view further details.

[0073] Step S405: By establishing a data association with the user information ledger of the marketing system, the current accurate tier price of electricity for each affected user is automatically obtained.

[0074] The system has a built-in tiered electricity pricing rule library, which automatically matches the specific values ​​of the first, second, or third tier electricity price based on the user's location, user category (residential or non-residential), and cumulative monthly electricity consumption.

[0075] For non-residential users, the system directly adopts their applicable electricity price category and standard electricity price. This tiered electricity price data is stored as the underlying input for the subsequent electricity loss calculation module, ensuring that the electricity loss calculation for each household is accurate down to the individual.

[0076] Step S5: Perform power loss analysis.

[0077] Step S501: Construct a deep reinforcement learning power consumption analysis model based on an improved Actor-Critic architecture. The Actor-Critic architecture is a multi-participant commentator architecture. The policy network Actor is responsible for generating actions that adjust the prediction parameters according to the current environmental state, and multiple parallel value networks Critic are responsible for evaluating the expected cumulative reward of the state-action pair. By integrating the outputs of multiple Critics, the estimation bias is reduced and the training stability is improved.

[0078] During the model training phase, a penalty function based on the causal relationship of electricity consumption is introduced. This function uses the causal effect obtained by counterfactual inference to correct the prediction error, forcing the model to follow the causal mechanism behind the data while fitting historical data, thereby improving the prediction accuracy in accident-free scenarios.

[0079] The state space is defined as the variables that can characterize the regional electricity consumption characteristics and their influencing factors, specifically including historical electricity consumption sequences, meteorological characteristics, temporal characteristics, and flags indicating whether external damage accidents have occurred; the action space is defined as the adjustment coefficients for the electricity prediction value of the next period.

[0080] The reward function is designed based on prediction accuracy and will be modified by incorporating a causal penalty term in subsequent steps.

[0081] Step S502: Construct a counterfactual analysis scenario using historical data to quantify the true causal effect of external damage accidents on electricity consumption. The construction of this scenario is based on a causal inference framework and specifically includes the following sub-steps: S5021: Construct a causal structure graph, a directed acyclic graph (hereinafter referred to as DAG).

[0082] Based on knowledge in the power sector and expert experience, the main variables affecting regional electricity consumption and their causal relationships were identified.

[0083] Whether an external rupture accident occurs is used as an intervention variable, denoted as E, with a value of 1 indicating that an external rupture has occurred and 0 indicating that it has not occurred.

[0084] The total electricity consumption of the region is denoted as Y as the result variable. Three types of confounding variables are introduced: historical load characteristics L, which include electricity consumption values ​​with a 1-hour lag, a 2-hour lag, and a 24-hour lag; meteorological characteristics W, which include temperature, humidity, and wind speed; and time characteristics T, which include hour, day of the week, and whether it is a holiday.

[0085] In a DAG, directed edges represent causal relationships: W pointing to Y indicates that weather affects electricity consumption, L pointing to Y indicates that historical load affects current load, and E pointing to Y indicates that external damage directly affects electricity consumption.

[0086] At the same time, it is ensured that all backdoor paths from E to Y are blocked by L, W, and T, that is, the backdoor criterion is satisfied, so that the causal effect of E on Y can be estimated by adjusting these confounding variables.

[0087] S5022, Establish a structural equation model (hereinafter referred to as SEM).

[0088] Based on the DAG constructed in step S5021, a nonlinear structural equation model is established to describe the relationship between variables. In this embodiment, a structural equation model based on gradient boosting tree is used because gradient boosting tree can automatically capture nonlinear relationships and does not require a preset function form.

[0089] Taking the total regional electricity consumption Y as an example, its structural equation is expressed as: Y=f(L,W,T,E)+∈; where f is a gradient boosting tree regression model, the input is the values ​​of variables L, W, T, and E, and the output is the predicted value of Y.

[0090] Structural equation modeling learns the mapping relationship of f by training on historical data, which includes samples from accident days and non-accident days. Each sample contains the actual observed values ​​of each variable. During training, the mean squared error is used as the loss function, and the prediction error is minimized by iteratively adding decision trees.

[0091] The main hyperparameters of the gradient boosting tree are set as follows: learning rate 0.1, maximum tree depth 5, and number of subtrees 100.

[0092] After training, the form of f is fixed and can be used for subsequent counterfactual prediction; ∈ is the random error term, which is assumed to follow a normal distribution with a mean of zero, but is usually ignored in the prediction stage, and the output of f is directly used as an estimate of Y.

[0093] S5023, Perform counterfactual inference.

[0094] To estimate the expected electricity consumption Y0 under ideal conditions without external damage accidents, i.e. after intervention, the do operator indicates intervention on the variable, i.e., forcibly setting E to 0, and the intervention operation is performed using a trained structural equation model.

[0095] The specific approach is as follows: For each time point t within the period affected by the accident, the actual characteristics L at that moment are known. t W t T t And an external failure actually occurred, i.e., E equals 1, at which point the actual observed value Y1 t Input L from the model t W t T t 1.

[0096] In order to obtain the counterfactual result Y0 t Keep L t W t T t Without changing E, force it to be 0, then input it into model f to get the predicted value, i.e., Y0. t equal to L t W t T t 0.

[0097] It should be noted that the implicit assumption here is that external failure events only change the electricity consumption by directly affecting Y, and will not change the values ​​of other variables, in order to conform to the constructed DAG structure.

[0098] S5024, Calculate causal effects.

[0099] For each time point t within the period of the accident's impact, the instantaneous causal effect caused by the external failure is defined as the difference between the actual observed value and the counterfactual estimate: CausalEffect t =Y1 t -Y0 t ; where Y1 t This is the actual electricity consumption, Y0 t This is an estimate of electricity consumption under the condition of no external damage.

[0100] The total power loss causal effect of the entire accident is the sum of the instantaneous effects at all points in time: The sum represents the total power loss caused by the external damage accident. These causal effect values ​​will be used in the design of the penalty function in subsequent steps to guide the deep reinforcement learning model to learn predictions that conform to causal laws.

[0101] Step S503: Incorporate the causal effect calculation results into the design of the reward function for deep reinforcement learning.

[0102] The instantaneous causal effect value calculated in step S5024 is used as the basis for the error penalty term. The relationship between the model fitting accuracy and causal consistency is dynamically balanced through an adaptive penalty mechanism, thereby guiding the model to learn power loss prediction that conforms to causal laws.

[0103] S5031, Construct dynamic penalty terms.

[0104] In the training process of a deep reinforcement learning model, each complete training iteration is called an epoch, denoted as τ. For each training sample i sampled from the experience pool, a penalty term is constructed based on the absolute value of the instantaneous causal effect corresponding to that sample. The calculation formula is as follows: Where t represents the current time, i.e., the time point corresponding to the sample; CausalEffectt Let τ represent the absolute value of the instantaneous causal effect at that moment, and λ represent the current training epoch. τ is the dynamic equilibrium coefficient at round τ. This coefficient is adaptively adjusted based on the causal consistency performance of the current model on the validation set, and is calculated using the following formula: In the formula: λ base The base penalty coefficient is a preset normal number used to control the base strength of the penalty term; in this embodiment, it is set to 0.1. α is an adjustment factor used to control the magnitude of adaptive adjustment; in this embodiment, it is set to 0.5. C target The target causal consistency threshold is a preset constant representing the desired level of causal consistency the model should achieve; in this embodiment, it is set to 0.9. val,τ Let C be the current causal consistency index obtained by evaluating the model on the validation set at the τth epoch. The specific calculation method for this index is as follows: For each sample in the validation set, the current model is used to make a prediction, and the implied causal effect of the predicted value is calculated, i.e., the difference between the predicted scenario with external failure and the scenario without external failure. Then, the Pearson correlation coefficient is calculated with the true causal effect value obtained in step S5024, and this correlation coefficient is used as the current causal consistency index C. val,τ .

[0105] When C val,τ Below the target threshold C target When the term inside the parentheses is greater than 1, λ τ Increase λ to strengthen punishment; conversely, when causal consistency reaches or exceeds the target, λ... τ Reduce or weaken the punishment.

[0106] In this way, the penalty intensity is dynamically adjusted as the causal consistency of the model changes, which not only ensures the model's fitting accuracy to historical data, but also encourages the model to learn stable causal mechanisms.

[0107] S5032, Reconstruct the reward function.

[0108] In the standard framework of deep reinforcement learning, the agent, at each decision time t, determines the current state s based on the current state. t Perform action a t Instant rewards for environmental feedback t and transition to the next state s t+1 In this embodiment, the instant reward r t Defined as a negative value based on the prediction error, specifically using the negative mean absolute percentage error, i.e.: ;in, Let y be the model's predicted electricity consumption at time t. t Let t be the actual electricity consumption at time t. The larger the value of this reward function, the more accurate the prediction.

[0109] In order to incorporate causal constraints into the learning process, the dynamic penalty term p constructed in step S5031 will be... t With instant rewards t Combined, to form a new effective reward r' t The calculation formula is: r' t =r t -p t ; where p t The penalty term defined above is equal to the absolute value of the instantaneous causal effect extracted from the current time sample multiplied by the current dynamic equilibrium coefficient.

[0110] New effective reward r' t Subtracting the causal deviation penalty from the original prediction accuracy reward means that if the model's prediction results lead to a significant difference between its implicit causal effect and the true causal effect obtained from counterfactual analysis, it will be penalized even if the prediction itself fits historical data.

[0111] In this way, the optimization algorithm, while maximizing the cumulative reward, will simultaneously take into account the prediction accuracy and causal consistency, forcing the agent to learn the stable causal relationships behind the data, thereby making more accurate predictions in normal scenarios where no external disruptions occur.

[0112] Step S504: The model integrates two sampling mechanisms, priority experience replay and follow-up experience replay, to efficiently extract training samples from the historical experience pool. Priority experience replay allocates sampling probabilities based on the magnitude of the temporal difference error (TD-error) of the samples, making the model pay more attention to those samples with larger prediction biases, thereby accelerating convergence.

[0113] In practice, the absolute value of TD-error is calculated for each experience tuple, and the sampling probability is proportional to this absolute value. The subsequent experience replay generates additional training samples by reconstructing the target state, which is particularly suitable for sparse reward scenarios. The system replaces the target in the original experience with the actual achieved state with a certain probability, thereby generating virtual success samples and increasing the number of positive experiences.

[0114] In addition, the model is equipped with a sparse reward method, which only gives positive rewards when the prediction error is lower than a preset threshold. In this embodiment, the threshold is set to 5% of the mean absolute percentage error. The reward value is one minus the error percentage. Otherwise, the reward is zero. This design can enhance the learning effect of the model at key convergence nodes.

[0115] Step S505: Input multidimensional data for deep iterative learning. The input data includes the historical electricity consumption sequence of the affected area, which is collected at 15-minute intervals, selecting data from the 30 consecutive days before the fault occurred; at the same time, input the time-series electricity revenue record, also at 15-minute intervals, calculated by multiplying the electricity consumption by the corresponding time period electricity price.

[0116] In addition, spatial numerical meteorological data, including three covariates of temperature, humidity and wind speed, are input, and these data are aligned with the electricity consumption series in time.

[0117] All input data are normalized before being fed into the model to eliminate the influence of different units. After multiple rounds of iterative training, the model finally outputs the time-series power loss density distribution every fifteen minutes. This distribution is presented in the form of mean and variance, reflecting the probability distribution of power loss at each time point under the condition of no external damage accident, and providing a refined estimate for subsequent electricity cost loss calculation.

[0118] Step S506: During model training, a specific loss function is used to optimize the parameters of the policy network and the value network. This loss function is based on the standard form of deep Q-networks, derived from the Bellman optimal equation, and introduces a causal error penalty term into the target value. The expression for the loss function is: Where L(θ) represents the loss function value under the current policy network parameters θ, s and a represent the current state and the action taken, respectively, r is the immediate reward, p is the causal error penalty term introduced in step S503, γ is a discount factor used to balance the importance of current reward and future reward, and Q(s;a;θ) is the value network's estimate of the state-action pair, θ - Here are the parameters of the target value network, which are periodically copied from θ to maintain training stability, maxQ(s';a';θ). - ) represents the optimal action value evaluated by the target network in the next state s'; expectation E represents the average value calculated for all samples drawn from the experience pool according to the mechanism of step S504.

[0119] By minimizing the loss function and updating the parameters using the optimizer, the model can fit historical data while adhering to the constraints of causal laws, eventually converging into a stable and accurate prediction model.

[0120] Step S507: By minimizing the loss function in step S506, the parameters of the policy network and value network are iteratively updated using the Adam optimizer. To ensure training stability and generalization ability, the following measures are taken: S5071, Network structure and initialization.

[0121] Both the policy network and the value network use a four-layer fully connected neural network with 512, 256, and 128 neurons in the hidden layers, respectively. The ReLU activation function is used in the hidden layers. The Tanh activation function is used in the output layer of the policy network to limit the action values ​​to between -1 and 1. The output layer of the value network consists of linear neurons. The network weights are initialized using the Xavier method.

[0122] S5072, Experience Playback and Sampling.

[0123] The experience pool capacity is set to one million records, storing experience tuples such as state, action, reward, and next state. A priority experience replay mechanism is adopted, and the sampling probability is allocated according to the absolute value of the temporal difference error of the sample. The larger the error, the higher the sampling probability. At the same time, the subsequent experience replay is introduced, which reconstructs the target state with a 30% probability, generates additional samples, and improves the sample utilization efficiency.

[0124] S5073, training hyperparameters.

[0125] The batch size is set to 256, the discount factor γ is 0.99, the learning rate is 10 to the power of negative 4, and the target network parameters are soft-updated from the current network every 500 steps with an update coefficient τ of 0.01.

[0126] S5074, Convergence determination.

[0127] The model is required to have an average absolute percentage error of less than 5% on the validation set, and the Pearson correlation coefficient between the external causal effect estimated by the model and the true causal effect calculated in step S5024 is not less than 0.85.

[0128] When both conditions are met simultaneously for ten consecutive training rounds, the model is considered to have converged, and the parameters are saved for subsequent power loss analysis.

[0129] Step S6: Perform electricity price calculation and analysis.

[0130] Step S601: Based on the list of affected users determined in step S4, establish a connection with the electricity marketing system through a standardized data interface. This interface uses the user's unique identifier as the query condition to automatically obtain the electricity consumption data of each affected user, including the user's historical electricity consumption records within a specific time period before the fault occurred and the cumulative electricity consumption for the current month. This data is used for subsequent determination of electricity price tiers and allocation of electricity loss.

[0131] Step S602: Based on the tiered electricity pricing policy rules of the user's region, automatically determine the electricity price tier that each user is in at the time of the accident.

[0132] The system has a built-in tiered electricity pricing rule library, which is pre-configured based on the electricity price documents issued by various regions. For residential users, the system automatically matches the specific values ​​of the first, second, or third tier electricity price based on the user's cumulative electricity consumption in the current month and the rules of their region. For non-residential users, the system directly adopts the electricity price category and standard electricity price that are applied to them.

[0133] Step S603: Call the power loss analysis results output in step S5. The results are time-series power loss values ​​at 15-minute intervals, representing the power loss of the entire affected area at each time point during the fault period.

[0134] To achieve precise accounting at the user level, the total power loss value is proportionally allocated to each specific user. The allocation method is as follows: obtain the average load curve of each user for the same period in the week before the fault occurs from the power consumption information collection system, calculate the proportion of each user's typical power consumption during the fault period to the total typical power consumption of all affected users, and multiply this proportion by the total power loss value output in step S5 to obtain the power loss sequence that each user should bear during the fault period.

[0135] Step S604: Perform a special electricity price loss calculation for industrial users and obtain the peak-valley-flat electricity price period division rules for industrial users from the marketing system, including the start and end times of peak hours, off-peak hours, and normal hours and their corresponding electricity prices.

[0136] Based on the user's electricity consumption value every fifteen minutes allocated in step S603, determine the electricity price type of each time period, and multiply it by the corresponding peak time electricity price, valley time electricity price, or normal time electricity price respectively, and sum them up to obtain the user's electricity price loss amount.

[0137] For residential users, the allocated power loss value is multiplied by the tiered electricity price determined in step S602 to obtain the electricity price loss for the residential user.

[0138] Step S605: Summarize the electricity price loss amount of all affected users and generate a detailed summary table of electricity cost loss caused by external damage. This summary table includes the total loss cost and classifies the loss amount and percentage according to the electricity consumption nature such as residential, agricultural, and industrial and commercial. The detailed table can be exported as a standardized document, which serves as key evidence for external damage prosecution and insurance claims, ensuring the accuracy and legal validity of loss accounting.

[0139] Step S7: Execute the accessibility application.

[0140] Step S701: The built-in Geographic Information System (GIS) engine supports loading and switching various types of map service layers, such as vector maps, image maps, and high-precision topographic maps. The area manager can freely select the currently displayed base map type through the layer switching control on the interface. The system background calls the corresponding standard map service interface in real time according to the user's selection to complete the base map replacement, providing the area manager with a rich visualization environment.

[0141] Step S702: Configure the simulation module. This module integrates a power system flow calculation engine. On the three-dimensional grid information layer, the module obtains real-time grid operation data from the control cloud platform, including the voltage, phase angle, and impedance parameters of each node and line.

[0142] Based on this data, the system automatically calls classic power flow algorithms to calculate the active and reactive power flow directions and values ​​for each line. After the calculation is completed, the power flow direction and load rate are displayed intuitively on the three-dimensional grid model using dynamic arrows or flowing light spots, helping the area manager to grasp the real-time operating status of the power grid.

[0143] In step S703, the simulation module also provides analysis functions along the time axis. The area manager can set the time axis slider on the interface and select to view the power flow distribution at a specific point in time or during a specific time period in the past.

[0144] Based on the selected time, the corresponding data is read from the historical operation database and the power flow calculation is re-executed. The result is rendered as a power flow picture at the corresponding time, which is convenient for accident backtracking or operation mode comparison.

[0145] Step S704: Build a statistical analysis function for historical external damage accident data, retrieve all external damage accident records that occurred in the past three years from the accident event database, and summarize the statistics by month. The statistical indicators include the total number of accidents per month, the types of lines involved, and the amount of loss.

[0146] The distribution of these accidents is then displayed on a GIS map in a dot pattern. The size or color depth of the dots represents the frequency or severity of accidents in that area. The area manager can click on the dots to view a specific list of accidents and details.

[0147] Step S705: The system integrates meteorological service management functions. By connecting to the smart power meteorological platform, it can obtain real-time meteorological data such as temperature, wind force, and rainfall within the administrative region and display relevant early warning information on the interface. It can automatically alert for severe weather based on preset thresholds to assist the area manager in predicting environmental risks.

[0148] Step S706: The system has a built-in intelligent assistant module that provides real-time text communication services for the area manager, supporting point-to-point and group message interaction. All communication records can be stored and retrieved, facilitating work collaboration and historical tracing.

[0149] Step S707: Construct a database of successful external damage prosecution cases. Store key information such as the geographical location of the cases, the entire accident process, and the claims list in a structured manner. Support retrieval by region, time, and other conditions. It can also generate promotional brochures containing accident details and warning images and text with one click for safety education to the surrounding community and construction units.

[0150] To further refine the engineering implementation of the above method, this embodiment also relates to a visual monitoring and intelligent identification system for early warning of external damage. At the hardware level, the system consists of monitoring terminals deployed on power transmission and distribution line towers. These monitoring terminals integrate a high-definition video acquisition module, an image processing module with edge computing capabilities, and a communication module connected to a data platform via a 4G or 5G network. At the software architecture level, the system adopts a microservice architecture, consisting of a data access layer, a spatial computing layer, an algorithm model layer, and a business application layer, from bottom to top.

[0151] The data access layer is responsible for connecting to the PMS3.0 production management system, the materials ERP system, the electricity information collection system, and the control cloud platform. Through the extraction, transformation, and loading tools provided by the data platform, the system can obtain and integrate power grid equipment data, video stream data, user information, and operational data from multiple core business tables. The spatial computing layer is based on the Geographic Information System (GIS) to complete the construction of the power grid topology and various spatial query tasks. The algorithm model layer carries the aforementioned convolutional neural network recognition algorithm, deep reinforcement learning power prediction algorithm, and causal inference model. The business application layer provides users with functional interfaces such as visual monitoring, online tracking, and loss analysis through the web or mobile APP.

[0152] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0153] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for early warning of external damage based on visual monitoring and intelligent identification, characterized in that, Includes the following steps: Obtain accident information about external damage accidents, wherein the accident information includes at least the location where the accident occurred; Based on the location of the accident, the power outage area and the list of affected users affected by the external damage accident are determined by tracing back from the power grid topology database; a causal inference model is constructed, and based on the electricity consumption characteristics data of the affected users during historical accident-free periods, the counterfactual electricity consumption of the affected users during the period affected by the external damage accident is estimated under the assumption that no external damage occurred; Obtain the actual electricity consumption of the affected users during the period of the external damage accident; determine the causal effect value of the electricity loss caused by the external damage accident based on the difference between the actual electricity consumption and the counterfactual electricity consumption. Based on the causal effect value of the power loss and the electricity price information of the affected users obtained from the marketing system, a list of external damage claims losses, including the cost of power loss, is generated.

2. The method for early warning of external damage based on visual monitoring and intelligent identification according to claim 1, characterized in that, The steps of tracing back from the power grid topology database to determine the outage area and the list of affected users affected by the external failure include: mapping the location of the failure to the corresponding physical power grid node in the digital power grid topology data; starting from the physical power grid node, traversing upstream along the power grid topology using a graph search algorithm until a power source is found to determine the starting boundary of the fault segment, and traversing downstream along the power grid topology until the tie switch or the end of the line to delineate the fault impact range; and based on the fault impact range, associating and generating a list of affected users from the electricity consumption information collection system or marketing system.

3. The method for early warning of external damage based on visual monitoring and intelligent identification according to claim 1, characterized in that, The steps of constructing a causal inference model and counterfactually estimating the counterfactual electricity consumption of the affected users during the period of the external damage incident include: constructing a causal structure diagram describing the causal relationship between variables affecting electricity consumption, wherein the variables include at least the external damage incident intervention variable, historical load characteristics, meteorological characteristics, time characteristics, and electricity consumption as the outcome variable; training a structural equation model using historical data based on the causal structure diagram to learn the nonlinear mapping relationship between variables; for each time point during the period of the external damage incident, forcibly setting the intervention variable to a state where no external damage has occurred, and inputting the historical load characteristics, meteorological characteristics, and time characteristics corresponding to that time point into the trained structural equation model to calculate the counterfactual electricity consumption.

4. The method for early warning of external damage based on visual monitoring and intelligent identification according to claim 3, characterized in that, After determining the causal effect value of power loss caused by the external damage accident based on the difference between the actual power consumption and the counterfactual power consumption, the method further includes: constructing a deep reinforcement learning power loss analysis model, wherein the state space of the deep reinforcement learning power loss analysis model includes historical power consumption sequences, meteorological characteristics, and temporal characteristics, and the action space is the adjustment coefficient for the predicted power consumption value for the next period; when training the deep reinforcement learning power loss analysis model, the causal effect value of power loss is introduced as a dynamic penalty term into the reward function to guide the model to learn predictions that conform to causal laws; and using the trained deep reinforcement learning power loss analysis model to perform refined prediction of time-series power consumption in scenarios without external damage.

5. The method for early warning of external damage based on visual monitoring and intelligent identification according to claim 4, characterized in that, The step of introducing the causal effect value of power loss as a dynamic penalty term into the reward function specifically includes: constructing a dynamic penalty term based on the causal effect value of power loss; combining the original reward function with the dynamic penalty term to form an effective reward function, wherein the original reward function is constructed based on the prediction error, and the effective reward function is used to penalize the behavior of the causal effect implied by the prediction result deviating from the causal effect value of power loss while optimizing the prediction accuracy during model training.

6. The method for early warning of external damage based on visual monitoring and intelligent identification according to claim 1, characterized in that, The step of generating an external damage recovery loss list, which includes electricity loss costs, based on the causal effect value of the power loss and the electricity price information of the affected users obtained from the marketing system, includes: obtaining the typical load curve of each affected user during the same period before the fault occurred from the electricity information collection system through a standardized interface based on the list of affected users; calculating the proportion of the typical electricity consumption of each affected user during the period affected by the external damage accident to the total typical electricity consumption of all affected users based on the typical load curve of each affected user; allocating the total power loss to each affected user by multiplying the proportion by the causal effect value of the power loss, thereby obtaining the power loss value of each user; obtaining the tiered electricity price information of each affected user from the marketing system; calculating the electricity loss cost of each user based on the power loss value of each user and its corresponding tiered electricity price, and summarizing them to generate the external damage recovery loss list.

7. The method for early warning of external damage based on visual monitoring and intelligent identification according to claim 1, characterized in that, The steps of generating an external damage claim loss list that includes electricity loss costs further include: obtaining asset information of damaged power equipment from the Enterprise Resource Planning (ERP) system through a standardized data interface, wherein the asset information includes the original purchase cost and accumulated depreciation; calculating the power facility loss costs based on the asset information of the damaged power equipment; and integrating the power facility loss costs into the external damage claim loss list.

8. The method for early warning of external damage based on visual monitoring and intelligent identification according to claim 7, characterized in that, After calculating the power facility loss cost based on the asset information of the damaged power equipment, the process further includes: adjusting the power facility loss cost using preset factor coefficients according to the compliance requirements of insurance claims to generate power facility loss costs that conform to the insurance claim criteria; and automatically pushing the external damage recovery loss list, which includes the power loss cost and the adjusted power facility loss cost, to the external insurance claims system through the Representational State Transmission Application Programming Interface (RESTful API) to initiate the claims process.

9. The method for early warning of external damage based on visual monitoring and intelligent identification according to claim 1, characterized in that, The steps for obtaining accident information of external damage accidents specifically include: receiving early warning information from the monitoring terminal. The early warning information is triggered by the monitoring terminal through real-time analysis of the scene image by the integrated image recognition processing engine. The image recognition processing engine uses a deep learning object detection network to identify the hazard source and uses a recurrent neural network to encode the motion trajectory of multiple consecutive frames to match the preset dangerous operation action mode.

10. A visual monitoring and intelligent identification external damage early warning system, used to execute the visual monitoring and intelligent identification external damage early warning method as described in any one of claims 1 to 9, characterized in that, include: The information acquisition module is used to acquire accident information of external damage accidents, and the accident information includes at least the location where the accident occurred; The topology analysis module is used to trace back from the power grid topology database based on the location of the accident and determine the power outage area and the list of affected users affected by the external damage accident; the causal inference module is used to construct a causal inference model and, based on the electricity consumption characteristic data of the affected users during historical accident-free periods, to estimate the counterfactual electricity consumption of the affected users during the period affected by the external damage accident, assuming that no external damage occurred. The data acquisition module is used to acquire the actual electricity consumption of the affected users during the period of the external damage accident; the causal effect calculation module is used to determine the causal effect value of the power loss caused by the external damage accident based on the difference between the actual electricity consumption and the counterfactual electricity consumption. The loss list generation module is used to generate an external damage recovery loss list that includes the cost of electricity loss, based on the causal effect value of the power loss and the electricity price information of the affected users obtained from the marketing system.

Images