Risk early warning method and system based on power service sensitive event

Abstract

The invention discloses a risk early warning method and system based on a power service sensitive event, and belongs to the technical field of deep learning, and the method comprises the steps: obtaining a customer service voice text, work order data and weather early warning information, and constructing a heterogeneous space-time diagram containing power distribution equipment and power consumers; inputting the heterogeneous space-time diagram into a graph neural network model to obtain event propagation path probability distribution and potential source nodes; generating an early warning work order based on the event propagation path probability distribution and a preset sensitive word library; triggering a supervision process according to the early warning work order, generating a supervision work order and configuring an elink message sending rule; and sending the elink message to a specified person in charge, feeding back a supervision processing result and updating model parameters. According to the method, leap-type upgrading from passive disposal to active prevention and control, from experience judgment to data intelligence and from single-point analysis to system governance is realized, and the method has remarkable technical advancement, management practicability and engineering popularization value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a risk early warning method and system based on power service sensitive events. BACKGROUND

[0002] With the increasing complexity and diversification of power customer service needs, power service risk management based on big data and artificial intelligence technology has become a key means to improve customer satisfaction, prevent major complaints and negative public opinion. Currently, Guizhou Power Grid Co., Ltd. is committed to building a service risk early warning system with clear management levels, convenient business processes, timely information interaction, and efficient complaint handling. By integrating marketing management systems, voice recognition systems, weather warning systems, and other multi-source data, AI models are used to automatically identify and grade sensitive events. Combined with the supervision process and elink message push mechanism, a control chain from risk perception to disposal is initially formed. The system has planned core functional modules such as early warning work order generation, sensitive word library maintenance, service risk mining model management, supervision process initiation and archiving, and has the ability to connect and visualize model operation display within minutes, laying a foundation for intelligent and refined management of power service risks.

[0003] However, the existing system still has the following key problems in terms of technical implementation and business closure: the current AI model output is a one-time judgment, lacking a feedback channel with the subsequent supervision and disposal results, which prevents the model from optimizing itself according to the actual disposal effect, and may result in an increase in false positives and a decrease in recognition sensitivity over time; the risk mining model relies on static rules and sensitive word libraries, lacks online learning capabilities based on historical disposal effectiveness, and is difficult to adapt to the evolution of new complaint patterns and complex complaint scenarios; the system can generate early warning and supervision work orders, but lacks a quantitative evaluation mechanism for the effectiveness of early warning strategies, making it impossible to determine whether intervention measures such as "early call" and "upgrade response" have truly reduced the risk of complaints, resulting in a lack of data support for decision-making. To address the above technical problems, the present application proposes a deep learning identification and risk early warning method that integrates graph neural networks, causal reasoning, and feedback optimization. By using supervision and disposal results to update the training parameters of the graph neural network model, a closed-loop mechanism is established to achieve continuous evolution of the model and dynamic improvement of risk management efficiency. SUMMARY

[0004] In view of the above problems, the present application is proposed.

[0005] To solve the above technical problems, the present application provides the following technical solutions: a risk early warning method based on power service sensitive events, comprising: acquiring customer service voice text, work order data, and weather warning information, and constructing a heterogeneous spatio-temporal graph containing power distribution equipment and power users;

[0006] inputting the heterogeneous spatio-temporal graph into a graph neural network model to obtain an event propagation path probability distribution and a potential source node;

[0007] generating an early warning work order based on the event propagation path probability distribution and a preset sensitive word library;

[0008] triggering a supervision process according to the early warning work order, generating a supervision work order and configuring an elink message sending rule;

[0009] sending the elink message to designated responsible personnel, feeding back a supervision processing result and updating model parameters.

[0010] As a preferred scheme of the risk early warning method based on the power service sensitive event, the heterogeneous spatio-temporal graph containing the power distribution equipment and the power user is constructed, and the construction includes:

[0011] obtaining customer service work order data from a marketing management system, the customer service work order data including call content, a person who makes a complaint, a business type, a transformer area name, a line name and a complaint time;

[0012] obtaining customer service voice text data from a voice recognition system, the customer service voice text data being generated by processing customer service call audios by voice-to-text;

[0013] obtaining weather warning information from a weather warning system, the weather warning information including a weather date, weather content and a weather risk level;

[0014] aligning the customer service work order data, the customer service voice text data and the weather warning information according to timestamps and geographical positions to form a multi-source fusion data set;

[0015] determining power distribution equipment nodes based on power grid topology data, determining power user nodes based on user numbers and transformer area attribution, and constructing a graph structure containing the equipment nodes and the user nodes;

[0016] in the graph structure, edges between the equipment nodes are established according to power grid physical connection relationships, and edges between the user nodes are established according to user geographical proximity and complaint correlation;

[0017] mapping the multi-source fusion data set to corresponding nodes of the graph structure to generate a heterogeneous spatio-temporal graph containing attribute features.

[0018] As a preferred scheme of the risk early warning method based on the power service sensitive event, the event propagation path probability distribution and the potential source node are obtained, and the obtaining includes:

[0019] For each node in the heterogeneous spatio-temporal graph, a node feature vector is initialized, the node feature vector comprising a node type, a historical work order quantity, an emotional intensity score, and a weather risk level;

[0020] The node feature vector is input into a graph neural network layer based on a gated temporal attention mechanism, and embedding representation of the current node is updated by neighbor node feature weighted aggregation;

[0021] A plurality of graph attention layers are stacked in the time dimension to capture the propagation dynamics of events at different time points, and output state probability distribution of each node at subsequent time points;

[0022] According to the diffusion intensity and propagation breadth of the node state evolution trajectory, an event propagation path probability distribution is generated, and a node with the highest propagation starting intensity is determined as a potential source node.

[0023] As a preferred scheme of the power service sensitive event risk early warning method, the propagation dynamics of events at different time points comprises:

[0024] The graph attention layers are stacked in time sequence, and each layer corresponds to a graph structure state of a time step;

[0025] The node state is updated in time sequence through a gated recurrent mechanism, historical propagation information is retained, and irrelevant fluctuations are filtered;

[0026] The state evolution trajectory of each node at future time steps is output.

[0027] As a preferred scheme of the power service sensitive event risk early warning method, the generation of the early warning work order comprises:

[0028] According to the event propagation path probability distribution, the event influence range and diffusion rate are determined, and an event risk intensity index is generated;

[0029] The event propagation path probability distribution is associated with the work order content and a preset sensitive word library, and the matching sensitive words and their occurrence frequencies are identified;

[0030] According to the event risk intensity index and the sensitive word matching result, the early warning level is determined, the early warning level comprising first, second, third, and fourth levels;

[0031] According to the early warning level and the configuration rules in the early warning business library, an early warning work order containing the early warning level, early warning content, responsible unit, and monitoring period is generated.

[0032] As a preferred scheme of the power service sensitive event risk early warning method, the generation of the supervision work order comprises:

[0033] parsing the early warning level, the responsible unit and the early warning content in the early warning work order, and determining whether the supervision trigger condition is met;

[0034] According to the early warning work order that meets the supervision trigger condition, a supervision work order containing a supervision number, an associated early warning work order number, a responsible unit and a processing time limit is generated;

[0035] According to the responsible unit and the processing role of the supervision work order, an elink message receiver list is obtained from the supervision short message receiver maintenance library;

[0036] According to the preset elink message template and the receiver list, the elink message sending rule is configured, and the message sending task is added to the execution queue.

[0037] As a preferred scheme of the risk early warning method based on the power service sensitive event, the feedback supervision processing result and the model parameter updating include:

[0038] The elink message sending task is obtained from the execution queue, and the integrity of the receiver list and the validity of the message template are verified;

[0039] The elink message service interface is called, and the elink message containing the supervision number, the early warning level and the processing time limit is sent to the designated responsible personnel;

[0040] The processing result of the supervision work order is received and parsed, and the processing result includes the work order state, the processing opinion and the attachment information;

[0041] According to the supervision processing result, the training parameters of the graph neural network model are updated, and the online feedback optimization of the model is realized.

[0042] The application provides a risk early warning system based on a power service sensitive event.

[0043] To solve the above technical problems, the application further provides the following technical scheme: a risk early warning system based on a power service sensitive event, comprising: a heterogeneous spatiotemporal graph construction module, configured to obtain customer service voice text, work order data and weather warning information, and construct a heterogeneous spatiotemporal graph containing power distribution equipment and power users;

[0044] A graph neural network inference module is configured to input the heterogeneous spatiotemporal graph into a graph neural network model to obtain an event propagation path probability distribution and a potential source node;

[0045] An intelligent early warning generation module is configured to generate an early warning work order based on the event propagation path probability distribution and a preset sensitive word library;

[0046] The supervision process configuration module is configured to trigger a supervision process according to the early warning work order, generate a supervision work order, and configure an elink message sending rule.

[0047] The push closed-loop feedback module is configured to send the elink message to designated responsible personnel, feed back the supervision processing result, and update the model parameters.

[0048] The application provides a computer device, including a memory and a processor, the memory stores a computer program, characterized in that the processor implements the steps of the risk early warning method based on the power service sensitive event when executing the computer program.

[0049] The application provides a computer readable storage medium, which stores a computer program, characterized in that the computer program is executed by a processor to implement the steps of the risk early warning method based on the power service sensitive event.

[0050] The application has the beneficial effects that: through multi-source data fusion and heterogeneous spatiotemporal graph construction, the customer service voice text, work order data and weather warning information are aligned in time and space dimensions, and combined with power grid topology and user association relationship, global semantic modeling of customer service events and power equipment states is realized, and the context integrity and context expression ability of event perception are significantly improved; further, through the introduction of a graph neural network with a gating time attention mechanism, neighbor node information is dynamically aggregated and the propagation evolution path of events at different times is captured, the probabilistic deduction of event diffusion trend and the accurate positioning of source nodes are realized, and the proactivity and scientificity of early warning are enhanced; on this basis, the event propagation risk intensity and sensitive word semantic recognition are fused to generate a graded early warning work order, so that technical judgment and social influence evaluation are organically combined, and the accuracy and business adaptability of early warning decision are improved; through automatic triggering of a supervision process, generation of a structured supervision work order and configuration of an elink message sending rule, the responsibility to person and forced closed-loop management of high-risk events are realized, and the emergency response efficiency and execution force are greatly improved; finally, the model parameters are updated in reverse by using actual processing results, and a self-evolution closed-loop mechanism of "perception - decision - execution - optimization" is constructed, so that the system has the ability of continuous learning and dynamic adaptation. Overall, the scheme breaks through the bottlenecks of information island, response lag and execution disconnection in traditional power service early warning, realizes the leapfrog upgrade from passive disposal to active prevention and control, from experience judgment to data intelligence, and from single-point analysis to system governance, and has significant technical advancement, management practicability and engineering popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0052] Figure 1 The overall flowchart of a risk early warning method based on power service sensitive events provided by an embodiment of the present application is shown in FIG. 1.

[0053] Figure 2 The computer device diagram of a risk early warning method based on power service sensitive events provided by an embodiment of the present application is shown in FIG. 2. DETAILED DESCRIPTION

[0054] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings in the specification. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should be within the scope of protection of the present application.

[0055] Embodiment 1, refer to Figure 1 For the first embodiment of the present application, the embodiment provides a risk early warning method based on power service sensitive events, which comprises the following steps:

[0056] S1: Obtain customer service voice text, work order data and weather warning information, and construct a heterogeneous space-time graph containing power distribution equipment and power users;

[0057] S2: Input the heterogeneous space-time graph into a graph neural network model to obtain event propagation path probability distribution and potential source nodes;

[0058] S3: Based on the event propagation path probability distribution and the preset sensitive word library, generate an early warning work order;

[0059] S4: Trigger the supervision process according to the early warning work order, generate a supervision work order and configure an elink message sending rule;

[0060] S5: Send the elink message to the designated responsible personnel, feed back the supervision processing result and update the model parameters.

[0061] It should be noted that the power service sensitive event has the characteristics of multi-source heterogeneity, strong transmission and complex causes. The sources of customer voice text, customer service order, power failure information, weather warning and public opinion data are scattered, the semantics are various, and the traditional rule-based early warning method is difficult to fully capture the early signs and evolution trend of the event. At the same time, the sensitive event has the characteristics of cascading propagation in the power grid topology and user group, and the abnormality of a single node may cause regional service risk, but the existing system lacks the modeling capability of event propagation path. In addition, after the order early warning, there are often problems such as treatment lag, unclear responsibility and incomplete closed loop, and the elink message sending and supervision process is mostly triggered passively, lacking active intervention strategy generation and feedback optimization mechanism based on event evolution deduction.

[0062] Therefore, in order to solve the problems of event recognition lag, lack of attribution ability, static early warning strategy and insufficient closed loop feedback, through steps S1-S5, a deep learning recognition model based on heterogeneous spatio-temporal graph is constructed to realize accurate modeling of sensitive event propagation path and source node; combined with preset sensitive word library and event propagation characteristics, the automatic generation of hierarchical early warning order is realized; through the linkage configuration of supervision process and elink message rule, the accurate pushing of risk disposal instruction is realized; and based on the feedback update of model parameters according to the supervision processing result, a closed loop management and control mechanism of "recognition-early warning-disposal-optimization" is formed, which improves the timeliness, accuracy and initiative of power service risk early warning.

[0063] Embodiment 2, refer to Figure 1 For the second embodiment of the application, a risk early warning method based on power service sensitive event is provided.

[0064] S1: Obtain customer service voice text, order data and weather warning information, and construct a heterogeneous spatio-temporal graph containing distribution equipment and power users;

[0065] Specifically, by calling the data interface of marketing management system, voice recognition system and weather warning system, customer service order data, voice transcription text and meteorological risk information are obtained, and the area name, line name and user number in the customer service order data are taken as the geographical correlation field, combined with the power grid topology data to generate a graph structure containing distribution transformers, line nodes and power user nodes, wherein the edges between device nodes are established by physical connection, and the social correlation edges between user nodes are established based on the same area belonging and the proximity of the appeal time, forming an initial heterogeneous graph. Step S1 includes steps S11 to S17.

[0066] S11: Obtain customer service order data from the marketing management system, and the customer service order data includes call content, appeal person, business type, area name, line name and appeal time.

[0067] Specifically, customer service work order data is obtained from the marketing management system via an interface, triggered by proactive requests in real time. The call content consists of the text of the user's request submitted through the hotline. The caller corresponds to the user ID in the customer profile. Service types include four categories: fault reporting, electricity inquiries, complaints / reports, and suggestions / commendations. The transformer area name and line name are used for subsequent geographic matching. The request time is the work order creation timestamp, accurate to the second, serving as a time benchmark for spatiotemporal alignment.

[0068] Furthermore, customer service work order data also includes work order number, power supply unit, processing status, and business subclass. The business subclass is coded according to the "95598 Business Classification Standard" and is used for label alignment in the subsequent multi-level classification model of work orders.

[0069] S12: Obtain customer service voice text data from the speech recognition system. The customer service voice text data is generated by processing customer service call recordings into text.

[0070] Specifically, the speech recognition system performs automatic speech recognition (ASR) on customer service call recordings, generating text data corresponding to the call content. This voice-text data includes the dialogue text between the two parties, voice emotion recognition results, and keyword tags, and is transmitted to the risk control system via an interface. The voice-text data is then matched with work order data in S11 using user ID and call time to supplement incomplete details of the user's requests in the work order, forming a more complete expression of the user's needs.

[0071] Furthermore, the voice emotion recognition results are output through the emotion analysis model in the IVR system, including three categories of labels: "calm," "anxious," and "angry," and assigned an emotion intensity score ranging from 0.0 to 1.0, which is used for risk weighting in subsequent sensitive events.

[0072] S13: Obtain weather warning information from the weather warning system. The weather warning information includes the weather date, weather content, and weather risk level.

[0073] Specifically, weather warning information is obtained from the weather warning system via an interface, triggered by an active request, and triggered in real time. The weather date is the effective date of the warning, and the weather content includes disaster types such as heavy rain, thunderstorms, strong winds, and hail. The weather risk level is divided into four levels: blue, yellow, orange, and red, corresponding to different levels of power grid operation risk. Weather warning information is matched on a per-power supply unit basis for subsequent assignment of environmental risk attributes to equipment nodes in the relevant area.

[0074] Furthermore, the weather risk level is quantitatively mapped according to the disaster type and the scope of impact. Red corresponds to a risk value of 4, orange to 3, yellow to 2, and blue to 1. This quantitative value is used as the node attribute input to the graph neural network model.

[0075] S14: Align the customer service order data, customer service voice text data, and weather warning information by timestamp and geographic location to form a multi-source fusion data set.

[0076] Specifically, with a minute-level time granularity as the alignment window, the complaint time of the order in S11, the call end time of the voice text in S12, and the effective time of the weather warning in S13 are uniformly reduced to the nearest whole minute time. The geographic location is matched by the four-level coding of the power supply unit, city, county, and transformer area. All user orders and voice texts in the same transformer area are collected and associated with the weather warning information of the area. After alignment, a multi-source fusion data set indexed by "time-transformer area" is formed, and each record contains three types of attributes: order content, voice text, and weather risk level.

[0077] Further, for orders and voice records across the minute boundary, a time-weighted fusion strategy is adopted, and the weight is inversely proportional to the time difference from the window center to ensure that high-timeliness data dominates in the fusion.

[0078] S15: Determine the distribution equipment nodes based on the power grid topology data, determine the power user nodes based on the user number and transformer area attribution, and construct a graph structure containing the equipment nodes and user nodes.

[0079] Specifically, the power grid topology data comes from the marketing system and distribution automation system, including the connection relationship of substations, distribution lines, distribution transformers, switches, and other equipment. Each distribution transformer and each line segment is regarded as a distribution equipment node, and each power user is regarded as a power user node. The user number is used to query the transformer area it belongs to, and the attribution relationship between the user node and the distribution equipment node is established. All nodes are included in a unified graph structure, and the node type is distinguished by a label. The equipment node label is "Device", and the user node label is "User".

[0080] Further, the distribution equipment nodes also include the low-voltage line segments and concentrator equipment under the transformer area, which are used to refine the equipment hierarchy and improve the perception ability of the graph model to low-voltage side events.

[0081] S16: In the graph structure, the edges between the equipment nodes are established according to the physical connection relationship of the power grid, and the edges between the user nodes are established according to the geographic proximity and complaint association.

[0082] Specifically, the edges between device nodes are established according to the line connection relationship in the power grid topology. If device A and device B are connected through the same feeder, an undirected edge (A, B) is established in the graph, and the edge type is marked as "PhysicalLink". The edges between user nodes are established according to geographical proximity and complaint correlation: if two users belong to the same area, an edge (U1, U2) is established, and the edge type is marked as "GeoProximity"; if two users submit work orders for the same problem (such as low voltage, frequent power outages) within ± 30 minutes, an edge (U1, U2) is established, and the edge type is marked as "ComplaintCorrelation". The edges can have weights, and the initial value is set to 1.

[0083] Further, the weight of the "ComplaintCorrelation" edge is calculated according to the product of the number of sensitive words matched in the work order and the emotional intensity score, and the formula is:

[0084]

[0085] where w ij is the weight of the "ComplaintCorrelation" edge, s k is the weight coefficient of the kth sensitive word, e i , e j are the emotional intensity scores of users i and j, respectively.

[0086] S17: Map the multi-source fusion data set to the corresponding nodes of the graph structure to generate a heterogeneous spatio-temporal graph containing attribute features.

[0087] Specifically, each record in the aligned multi-source fusion data set in S14 is mapped to the corresponding node in the graph structure. The work order content and voice text are assigned as text attributes to the corresponding power user node, and the attribute field is named "complaint_text"; the weather risk level is assigned as an environmental attribute to all power distribution device nodes under the area, and the attribute field is named "weather_risk_level"; the complaint time is stored as the timestamp attribute "timestamp" of the node. The final generated heterogeneous spatio-temporal graph is stored in the format of a graph database, supporting time slicing queries, and used as input for subsequent graph neural network models.

[0088] Further, the text attributes are pre-processed by a pre-processing module before mapping, and the annotation results are embedded as auxiliary feature fields "sensitive_tags" in the user nodes, which are called by the subsequent risk mining model.

[0089] S2: Input the heterogeneous spatio-temporal graph into the graph neural network model to obtain the event propagation path probability distribution and the potential source node;

[0090] Specifically, the heterogeneous spatio-temporal graph is subjected to multi-layer node state updating and aggregation by a graph neural network model to extract dynamic dependency between nodes, generate path probability distribution containing event propagation intensity and direction information, and identify key nodes with high propagation influence as potential sources. Step S2 includes steps S21 to S24.

[0091] S21: For each node in the heterogeneous spatio-temporal graph, initialize a node feature vector including node type, historical work order quantity, emotion intensity score, and weather risk level.

[0092] Specifically, the node type of a device node is marked as "Device", and the node type of a power user node is marked as "User"; the historical work order quantity is the total number of work orders associated with the node in the past 24 hours; the emotion intensity score is obtained from the voice text emotion analysis result associated with the node, and is set to 0.0 if there is no voice data; the weather risk level is mapped to a numerical feature according to the weather warning information of the region, with red warning corresponding to the highest risk level and blue warning corresponding to the lowest risk level. All features are normalized and concatenated into an initial feature vector.

[0093] S22: Input the node feature vector into a graph neural network layer based on the gating time attention mechanism, and update the embedding representation of the current node by weighted aggregation of neighbor node features.

[0094] Specifically, the graph neural network layer calculates the attention weight between the current node and the neighbor node at each propagation step, and the weight is determined according to the node feature similarity and the edge type. For a device node, it is preferred to focus on the downstream user nodes directly connected to it; for a user node, it is preferred to focus on the neighboring users in the same area and the associated distribution equipment. The weighted aggregation generates a new node embedding representation reflecting the state evolution at the current time.

[0095] Further, the calculation of the attention weight considers a time decay factor, which gives higher weight to the recent neighbor node state changes to capture the timeliness feature of event propagation.

[0096] Further, an external event trigger signal is introduced in the weight calculation. When a new first-level warning work order or a red weather warning is detected, the attention weight of the related node is dynamically increased to enhance the response ability of the model to sudden risks.

[0097] S23: Stack multiple graph attention layers in the time dimension to capture the propagation dynamics of events at different times, and output the state probability distribution of each node at the subsequent time.

[0098] Specifically, the graph attention layer in S22 is stacked in time sequence, and each layer corresponds to the graph structure state of a time step. The node state is updated in time sequence through a gating cycle mechanism, and the historical propagation information is retained and irrelevant fluctuations are filtered. Finally, the state evolution trajectory of each node at multiple future time steps is output, which is used to analyze the event diffusion trend.

[0099] Further, an external event trigger signal is introduced in each time step. When a new primary warning work order or a red weather warning is detected, the activation threshold of the related node is dynamically adjusted to improve the response speed of the model to sudden events.

[0100] Further, the state evolution trajectory updating process supports parallel computing by power supply unit, improving the processing efficiency of the model in the provincial system.

[0101] S24: According to the diffusion intensity and propagation breadth of the node state evolution trajectory, generate an event propagation path probability distribution, and determine the node with the highest propagation starting intensity as the potential source node.

[0102] Specifically, the event propagation path probability distribution is generated by analyzing the time of the first significant change in the state of each node and the impact range. The potential source node is identified by the following method: among all nodes, the node whose state first changes significantly and can trigger the most downstream node response is selected as the potential source node

[0103] Further, the identification result of the potential source node is compared with the historical major event records in the warning business library. If a match is successful, the corresponding emergency plan number is automatically associated for subsequent rapid disposal of warning work orders.

[0104] Further, the event propagation path probability distribution and the potential source node information are encapsulated as structured data, which are used as input basis for generating hierarchical warning work orders.

[0105] S3: Based on the event propagation path probability distribution and the preset sensitive word library, generate a warning work order;

[0106] Specifically, by analyzing the intensity and coverage range of the event propagation path, combining the keyword matching results in the preset sensitive word library, determining the risk level of the event, and triggering the corresponding warning process according to the risk level, a structured warning work order is generated. Step S3 includes steps S31 to S34.

[0107] S31: According to the event propagation path probability distribution, determine the event impact range and diffusion rate, and generate an event risk intensity indicator.

[0108] Specifically, the event influence range is determined by counting the number of user nodes with significant changes in state within a preset time window, the diffusion rate is obtained by calculating the time interval required for the event to propagate from the source node to the Nth downstream node, and the event risk intensity indicator is a weighted combination of the influence range and the diffusion rate, which is used to quantify the potential harm degree of the event.

[0109] Further, the time window is dynamically adjusted according to the event type, and is set to 30 minutes for power outage events and 60 minutes for voltage anomaly events to adapt to the evolution characteristics of different types of events.

[0110] S32: Match the work order content associated with the event propagation path probability distribution with the preset sensitive word library to identify the matched sensitive words and their occurrence frequencies.

[0111] Specifically, the preset sensitive word library includes high-risk keywords such as "complaint", "exposure", "government", "media", "collective", and "threat", which are classified into "regulatory", "public opinion", and "livelihood" categories. The user work order content and voice text associated with the event are processed by word segmentation and compared with the sensitive word library one by one, and the matching times and categories of each sensitive word are recorded.

[0112] Further, for sensitive words that repeatedly appear across work orders, a sensitive word aggregation view at the substation level or line level is generated for evaluating the overall risk level of the region.

[0113] S33: Determine the warning level according to the event risk intensity indicator and the sensitive word matching result, and the warning level includes levels one, two, three, and four.

[0114] Specifically, the event risk intensity indicator is compared with a preset threshold value, and if it exceeds the first threshold value, it is preliminarily determined as high risk; at the same time, if "regulatory" or "public opinion" sensitive words are matched, the warning level is increased by one based on the preliminary determination. The final warning level is determined according to a comprehensive scoring table, where level one is the highest risk and level four is the lowest risk.

[0115] Further, when the same substation triggers a level three or above warning more than three times within 24 hours, it is automatically upgraded to a level one warning and marked as a "repeat high-risk area".

[0116] S34: According to the warning level and the configuration rules in the warning business library, generate a warning work order containing the warning level, warning content, responsible unit, and monitoring period.

[0117] Specifically, the early warning service library is pre-configured with a corresponding disposal process, a responsible role, a response time limit, and a message template for each early warning level. According to the determined early warning level, the corresponding configuration is called to fill in the early warning work order field and generate a structured early warning work order. The early warning work order is published through a system internal message queue and used to trigger subsequent elink message sending, supervision process, or outbound call tasks.

[0118] Further, when the early warning work order is generated, the original work order number, event propagation path map, and sensitive word matching details are automatically associated as attachments embedded in the work order for subsequent audit and traceability.

[0119] S4: Trigger the supervision process according to the early warning work order, generate a supervision work order, and configure an elink message sending rule;

[0120] Specifically, by analyzing the early warning level, responsible unit, and monitoring period of the early warning work order, the corresponding supervision process is triggered, a structured supervision work order is generated, and an elink message sending strategy is configured according to the preset message template and recipient rule. Step S4 includes steps S41 to S44.

[0121] S41: Analyze the early warning level, responsible unit, and early warning content in the early warning work order to determine whether the supervision trigger condition is met.

[0122] Specifically, the supervision trigger condition is determined according to the configuration rule in the early warning service library. When the early warning level is level one or level two and the responsible unit is clear, it is determined that the supervision trigger condition is met; for level three or level four early warning work orders, the supervision process is started only when it is repeatedly triggered or associated with major public opinion.

[0123] Further, the configuration rule in the early warning service library includes a "whether to supervise" field, and the system determines whether to generate a supervision work order according to the Boolean value of this field.

[0124] S42: Generate a supervision work order containing a supervision number, an associated early warning work order number, a responsible unit, and a processing time limit according to the early warning work order that meets the supervision trigger condition.

[0125] Specifically, the supervision number is automatically generated in the format "DW-YYYYMMDD-XXXX", the associated early warning work order number records the unique identifier of the original early warning work order, the responsible unit is determined according to the power supply unit in the early warning work order, and the processing time limit is dynamically set according to the early warning level, with 2 hours for level one and 4 hours for level two.

[0126] Further, the event propagation path map and sensitive word matching details are automatically associated as attachments for subsequent processing personnel to review when the supervision work order is generated.

[0127] S43: Obtain the elink message receiver list from the supervision short message receiver maintenance library according to the responsibility unit and the processing role of the supervision order.

[0128] Specifically, the elink message receiver list is matched according to the organization level and the business role of the responsibility unit. The provincial responsibility unit corresponds to the provincial service dispatcher, the city responsibility unit corresponds to the city service dispatcher and the customer service engineer; the list information is derived from the contact data pre-configured in the “supervision short message receiver maintenance” function module.

[0129] Further, when the responsibility unit is multiple, the system automatically combines the receivers of all related units into one sending list.

[0130] S44: Configure the elink message sending rule according to the preset elink message template and the receiver list, and add the message sending task to the execution queue.

[0131] Specifically, the elink message template is called from the message template configuration library according to the warning level and the event type. The first-level warning uses the “urgent supervision” template, and the second-level warning uses the “important supervision” template. The message content includes the supervision number, the warning level, the event summary and the processing time limit. After the sending rule is configured, the system generates a message sending task and adds it to the message queue, waiting for the message service to execute asynchronously.

[0132] Further, if the message sending fails, the system will record the failure log and try to resend after a specified time.

[0133] S5: Send the elink message to the designated responsibility personnel, feed back the supervision processing result and update the model parameters.

[0134] Specifically, by executing the elink message sending task, the instant reach of risk warning information is realized, and the risk identification model is fed back and optimized according to the processing result of the supervision order. Step S5 includes steps S51 to S54.

[0135] S51: Obtain the elink message sending task from the execution queue, verify the integrity of the receiver list and the validity of the message template.

[0136] Specifically, the system scans the message execution queue regularly, extracts the elink message task to be sent; checks whether the elink account status of each responsibility personnel in the receiver list is valid, and records the alarm log if there is an invalid account; at the same time, it verifies whether the used message template is in the “enabled” state, if the template has been disabled, replace it with the default template or stop sending.

[0137] Further, for the elink message corresponding to the first level warning, the system automatically adds a "urgent" label and improves the message push priority.

[0138] S52: Call the elink message service interface, and send the elink message containing the supervision number, warning level and processing time limit to the designated responsible person.

[0139] Specifically, the elink message service interface is provided by the wisdom coordination middle platform, and the system sends the message content, recipient list, message type and other parameters after encapsulation through the HTTP request mode; the message is displayed in the form of a card in the elink group, and one-key jump to the supervision work order detail page is supported; after sending, the system records the sending success state and timestamp.

[0140] Further, if the elink message service interface returns a sending failure, the system will reinsert the task into the execution queue and set an exponential backoff retry strategy.

[0141] S53: Receive and analyze the processing result of the supervision work order, which includes the work order state, processing opinion and attachment information.

[0142] Specifically, when the supervision work order is completed, audited or archived, the system obtains the final processing result from the supervision work order management module; the work order state includes "closed loop", "backed off", "transferred"; the processing opinion is filled by the responsible person, which is used to analyze the effectiveness of the disposal measures; the attachment information includes on-site photos, communication records and other supplementary materials.

[0143] Further, the system extracts keywords from the processing opinion and identifies positive feedback labels such as "notified", "repaired" and "user understood".

[0144] S54: According to the supervision processing result, update the training parameters of the graph neural network model to realize online feedback optimization of the model.

[0145] Specifically, the event samples of supervision closed loop are used as positive samples, and the events that are not closed loop or upgraded to complaints are used as negative samples for model retraining; the update content includes the attention weight matrix of the graph neural network and the initialization strategy of the node embedding vector; the parameter update is completed through incremental learning to avoid full retraining.

[0146] Further, the system regularly generates a model performance comparison report to evaluate the accuracy and false positive rate changes after feedback optimization, which is used for model version management.

[0147] Embodiment 3 is a third embodiment of the present application, which provides a risk warning system based on power service sensitive events, comprising:

[0148] A heterogeneous spatio-temporal graph construction module is configured to acquire customer service voice text, work order data and weather warning information, and construct a heterogeneous spatio-temporal graph containing power distribution equipment and power users;

[0149] A graph neural network inference module is configured to input the heterogeneous spatio-temporal graph into a graph neural network model to obtain an event propagation path probability distribution and a potential source node;

[0150] An intelligent early warning generation module is configured to generate an early warning work order based on the event propagation path probability distribution and a preset sensitive word library;

[0151] A supervision process configuration module is configured to trigger a supervision process according to the early warning work order, generate a supervision work order and configure an elink message sending rule;

[0152] A push closed-loop feedback module is configured to send an elink message to designated responsible personnel, feed back a supervision processing result and update model parameters.

[0153] Embodiment 4, refer to Figure 2 The fourth embodiment of the present application is different from the first three embodiments in that the function, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application or the part that contributes to the prior art or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device) to execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk and various program code storage media.

[0154] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logical functions, and can be specifically embodied in any computer-readable medium for use by an instruction execution system, device or apparatus, such as a computer-based system, a system including a processor or other system that can fetch and execute instructions from an instruction execution system, device or apparatus, or in conjunction with these instructions execution system, device or apparatus. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by an instruction execution system, device or apparatus or in conjunction with these instruction execution system, device or apparatus.

[0155] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer readable medium can be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via an optical scanner, then compiled, interpreted, or otherwise processed, as necessary, and stored in a computer memory.

[0156] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the embodiments described above, various steps or methods can be implemented, for example, by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following techniques, which are well known in the art of hardware implementation, can be used: a combination of logic gates in a logic circuit, a combination of processor(s) and memory(ies) that executes software, programmable gate arrays (PGA), field programmable gate arrays (FPGA), etc.

[0157] It should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, but not limit the present application, and although the present application is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or equivalent replaced, without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A risk early warning method based on power service sensitive events, characterized in that: The method comprises the steps of: obtaining customer service voice text, work order data and weather warning information, and constructing a heterogeneous space-time graph containing power distribution equipment and power users; inputting the heterogeneous space-time graph into a graph neural network model to obtain event propagation path probability distribution and potential source nodes; generating a warning work order based on the event propagation path probability distribution and a preset sensitive word library; triggering a supervision process according to the warning work order, generating a supervision work order and configuring an elink message sending rule; sending the elink message to designated responsible personnel, feeding back the supervision processing result and updating the model parameters.

2. The method of claim 1, wherein the method comprises: The method comprises the steps of: obtaining customer service work order data from a marketing management system, the customer service work order data including call content, complainant, business type, transformer area name, line name and complaint time; obtaining customer service voice text data from a voice recognition system, the customer service voice text data being generated by processing customer service call recording through voice-to-text; obtaining weather warning information from a weather warning system, the weather warning information including weather date, weather content and weather risk level; aligning the customer service work order data, the customer service voice text data and the weather warning information according to time stamp and geographical location to form a multi-source fusion data set; determining power distribution equipment nodes based on power grid topology data, and determining power user nodes based on user number and transformer area attribution relationship, and constructing a graph structure containing equipment nodes and user nodes; in the graph structure, establishing edges between equipment nodes according to power grid physical connection relationship, and establishing edges between user nodes according to user geographical proximity and complaint correlation; mapping the multi-source fusion data set to the corresponding nodes of the graph structure to generate a heterogeneous space-time graph containing attribute characteristics.

3. The method of claim 2, wherein the method further comprises: The method comprises the steps of: initializing node feature vectors for each node in the heterogeneous space-time graph, the node feature vectors including node type, historical work order quantity, emotional intensity score and weather risk level; inputting the node feature vectors into a graph neural network layer based on a gating time attention mechanism, and updating the embedding representation of the current node through neighbor node feature weighted aggregation; stacking multiple graph attention layers in the time dimension to capture the propagation dynamics of events at different times, and outputting the state probability distribution of each node at subsequent time; generating event propagation path probability distribution according to the diffusion intensity and propagation breadth of node state evolution trajectory, and determining the node with the highest propagation starting intensity as the potential source node.

4. The risk early warning method based on power service sensitive events as described in claim 3, characterized in that: The method comprises the steps of: stacking the graph attention layers in time sequence, each layer corresponding to the graph structure state of one time step; updating the node state in time sequence through a gating cycle mechanism to retain historical propagation information and filter irrelevant fluctuations; outputting the state evolution trajectory of each node at multiple future time steps.

5. The risk warning method based on power service sensitive events according to claim 4, wherein: the method comprises the steps of: According to the event propagation path probability distribution, the event influence range and diffusion rate are determined, and an event risk intensity index is generated; The event propagation path probability distribution associated ticket content is matched with a preset sensitive word library to identify the matching sensitive words and their frequency of occurrence; According to the event risk intensity index and the sensitive word matching result, the warning level is determined, which includes levels one, two, three and four; According to the warning level and the configuration rules in the warning business library, a warning ticket containing the warning level, warning content, responsible unit and monitoring period is generated.

6. The method of claim 5, wherein the method further comprises: The generation of the supervision ticket includes: Analyzing the warning level, responsible unit and warning content in the warning ticket to determine whether the supervision trigger condition is met; According to the warning ticket that meets the supervision trigger condition, a supervision ticket containing the supervision number, associated warning ticket number, responsible unit and processing time limit is generated; According to the responsible unit and processing role of the supervision ticket, the elink message receiver personnel list is obtained from the supervision short message receiver maintenance library; According to the preset elink message template and the receiver personnel list, the elink message sending rule is configured, and the message sending task is added to the execution queue.

7. The method of claim 6, wherein the method further comprises: determining a risk level of the power service based on the power service sensitive event; and providing a risk alert based on the risk level of the power service. 7 The feedback of the supervision processing result and the update of the model parameters include: Get the elink message sending task from the execution queue, verify the integrity of the receiver personnel list and the validity of the message template; Call the elink message service interface to send the elink message containing the supervision number, warning level and processing time limit to the designated responsible personnel; Receive and analyze the processing result of the supervision ticket, which includes the ticket state, processing opinion and attachment information; According to the supervision processing result, the training parameters of the graph neural network model are updated to realize the online feedback optimization of the model.

8. A risk warning system based on power service sensitive events, applying a risk warning method based on power service sensitive events according to any one of claims 1-7, characterized in that, It includes: A heterogeneous spatio-temporal graph construction module is used to obtain customer service voice text, ticket data and weather warning information, and to construct a heterogeneous spatio-temporal graph containing power distribution equipment and power users; A graph neural network inference module is used to input the heterogeneous spatio-temporal graph into a graph neural network model to obtain an event propagation path probability distribution and a potential source node; An intelligent warning generation module is used to generate a warning ticket based on the event propagation path probability distribution and a preset sensitive word library; A supervision process configuration module is used to trigger the supervision process according to the warning ticket, generate a supervision ticket and configure an elink message sending rule; A push closed-loop feedback module is used to send the elink message to the designated responsible personnel, feedback the supervision processing result and update the model parameters. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the risk warning method based on the power service sensitive event in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the risk warning method based on the power service sensitive event in any one of claims 1 to 7.

Images