A power service number map linkage monitoring, management and control early warning method

Abstract

The application relates to a power supply service graph linkage monitoring control early warning method. The method comprises the following steps: performing space-time anchor point processing on power supply service multi-source heterogeneous data corresponding to a target power grid to obtain space-time anchor point event data; performing space-time knowledge fusion on the multi-source correlation of the target power grid according to the space-time anchor point event data to obtain power supply service space-time knowledge graph data; performing multi-task risk prediction on the power supply service risk of the target power grid within a preset prediction time window according to the power supply service space-time knowledge graph data to obtain a dynamic early warning trigger result; performing causal attribution counterfactual reasoning on the risk rising key driving factor of the power supply service risk according to the dynamic early warning trigger result to obtain an interpretable early warning result; and the interpretable early warning result is used for behavior disposal arrangement of the target power grid. The method can realize prospective prediction of the risk.

Description

Technical Field

[0001] This application relates to the field of smart grid technology, and in particular to a method for monitoring, controlling and early warning of power supply services using a digital map linkage mechanism. Background Technology

[0002] In traditional technologies, power supply service monitoring, control, and early warning systems periodically aggregate key indicators such as the number and duration of power outages, low voltage, fault alarms, repair delays, and complaint volume from metering / collection terminals, distribution network automation, and customer service work orders. Anomalies are identified and tiered warnings are triggered based on pre-set thresholds or rules (e.g., low voltage lasting X minutes in a certain area, daily alarm count exceeding Y on a certain line, or a Z% increase in complaints in a certain area compared to the previous period). Subsequently, a notification is displayed on the monitoring interface, and a handling work order is automatically generated or pushed. On-duty personnel then follow the procedures to verify, dispatch, repair, and follow up, achieving a basic closed-loop management of "monitoring—alarming—dispatch—handling." However, traditional power supply service monitoring, control, and early warning technologies rely primarily on fixed thresholds and static rules, making it difficult to adapt to changes in scenarios caused by load fluctuations and the coupling of multiple factors, resulting in a lack of proactive risk prediction. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, device, and computer equipment for power supply service digital-map linkage monitoring, control, and early warning that can achieve forward-looking prediction of risks, addressing the aforementioned technical problems.

[0004] Firstly, this application provides a method for monitoring, controlling, and issuing early warnings based on the linkage between digital maps and power supply services, including: Spatiotemporal anchoring processing is performed on the multi-source heterogeneous power supply service data corresponding to the target power grid to obtain spatiotemporal anchor event data; Based on the spatiotemporal anchor point event data, spatiotemporal knowledge fusion is performed on the multi-source correlation of the target power grid to obtain spatiotemporal knowledge graph data of power supply service; Based on the power supply service spatiotemporal knowledge graph data, multi-task risk prediction is performed on the power supply service risk of the target power grid within a preset prediction time window to obtain dynamic early warning triggering results. Based on the dynamic early warning triggering results, causal attribution counterfactual inference is performed on the risk escalation key driving factors of the power supply service risk to obtain an explainable early warning result; the explainable early warning result is used to orchestrate the behavioral actions of the target power grid.

[0005] Secondly, this application also provides a power supply service digital map linkage monitoring, control and early warning device, including: The spatiotemporal anchor module is used to perform spatiotemporal anchoring processing on the multi-source heterogeneous power supply service data corresponding to the target power grid to obtain spatiotemporal anchor event data. The knowledge fusion module is used to perform spatiotemporal knowledge fusion on the multi-source correlation of the target power grid based on the spatiotemporal anchor point event data to obtain power supply service spatiotemporal knowledge graph data; The risk prediction module is used to perform multi-task risk prediction on the power supply service risk of the target power grid within a preset prediction time window based on the power supply service spatiotemporal knowledge graph data, and obtain dynamic early warning triggering results. The result deduction module is used to perform causal attribution counterfactual deduction on the risk escalation key driving factors of the power supply service risk based on the dynamic early warning triggering result, and obtain an explainable early warning result; the explainable early warning result is used to orchestrate the behavioral handling of the target power grid.

[0006] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of a power supply service data-map linkage monitoring, control and early warning method.

[0007] The aforementioned power supply service data-map linkage monitoring, control, and early warning method, device, and computer equipment, through spatiotemporal anchoring of multi-source heterogeneous power supply service data of the target power grid and forming locationable and traceable spatiotemporal anchor event data under a unified "data-map" coordinate system, and then constructing a power supply service spatiotemporal knowledge graph by spatiotemporal knowledge fusion of multi-source relationships, transforms risk identification from scattered data and static display to reasonable computation based on topology-geography-customer groups-event chains. On this basis, the spatiotemporal knowledge graph is used to carry out multi-task risk prediction within a preset prediction time window and form dynamic early warning trigger results, realizing that the early warning threshold varies with the scenario, its own historical distribution, and risk. The tolerance adaptive adjustment significantly reduces false alarms and missed alarms caused by fixed threshold rules; further, by combining dynamic trigger results, causal attribution and counterfactual inference are performed on key drivers of risk escalation, upgrading the early warning from "result notification" to "explainable causes and quantifiable measures" decision output, and directly transforming explainable early warning results into actionable action arrangements, thereby achieving forward-looking prediction, explainable location and closed-loop linkage control of risks, improving the accuracy of resource allocation and customer outreach, shortening response time, suppressing complaint spillover and public opinion escalation, improving power supply service quality and governance efficiency, and supporting continuous feedback optimization to form a self-evolving monitoring-early warning-response closed loop. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is an application environment diagram of the power supply service data-map linkage monitoring, control and early warning method in one embodiment; Figure 2 This is a flowchart illustrating a power supply service data-map linkage monitoring, control, and early warning method in one embodiment. Figure 3 This is a structural block diagram of a power supply service data-driven linkage monitoring, control, and early warning device in one embodiment; Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0011] This application provides a power supply service digital-map linkage monitoring, control, and early warning method, which can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on other network servers. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0012] In one exemplary embodiment, such as Figure 2 As shown, a method for monitoring, controlling, and issuing early warnings based on a digital map of power supply services is provided, which is then applied to... Figure 1 Taking the server in the example, the explanation includes the following steps 202 to 208. Wherein:

[0013] Step 202: Perform spatiotemporal anchoring processing on the multi-source heterogeneous power supply service data corresponding to the target power grid to obtain spatiotemporal anchor event data.

[0014] Step 204: Based on the spatiotemporal anchor point event data, perform spatiotemporal knowledge fusion on the multi-source correlation of the target power grid to obtain spatiotemporal knowledge graph data of power supply service.

[0015] Step 206: Based on the spatiotemporal knowledge graph data of power supply service, perform multi-task risk prediction on the power supply service risk of the target power grid within the preset prediction time window to obtain dynamic early warning trigger results.

[0016] Step 208: Based on the dynamic early warning triggering results, perform causal attribution counterfactual deduction on the risk escalation key driving factors of power supply service risk to obtain interpretable early warning results.

[0017] The target power grid is a specific power supply area that requires power supply service monitoring, control and early warning. Its boundaries can be determined by administrative region, power supply zone or topological range (set of stations, lines and transformer areas).

[0018] Among them, the multi-source heterogeneous data of power supply service is a collection of power supply service related data from multiple systems with different structural types, including but not limited to power outages / faults, work orders and complaints, equipment ledgers and topology, GIS geography, operating status / power quality, customer profiles and channel reach data.

[0019] Spatiotemporal anchoring is a process that unifies multi-source data into event-based representations with timestamps, spatial location anchors, and topological resource anchors, in order to achieve cross-system alignment, location, and computation.

[0020] Among them, spatiotemporal anchor event data is standardized event data formed after spatiotemporal anchoring processing. Each event includes at least a timestamp, spatial location anchor, topological resource anchor, and necessary event attributes and associated objects.

[0021] Among them, multi-source association is the association link and its weight / direction information generated between different data sources in dimensions such as topology, geography, service affiliation, temporal co-occurrence, homogeneous association or resource accessibility.

[0022] Spatiotemporal knowledge fusion is a process of integrating events, entities and their multi-source relationships under a unified spatiotemporal benchmark to remove duplicates, align semantics, and model relationships, making them searchable and inferable.

[0023] Among them, the spatiotemporal knowledge graph data of power supply services is a computable graph structure data formed by integrating entities such as power grid topology, geospatial, customer groups, service events and disposal resources and their spatiotemporal relationships.

[0024] The preset prediction time window is a future time range ΔT (e.g., the next 30 minutes, 2 hours, or 24 hours) used for forward-looking risk prediction. Risk prediction and trigger determination are output within this time range.

[0025] Among them, power supply service risks are a set of risk types that may lead to a decline in power supply service quality or negative impacts on services, including risks such as power outages, complaints / spillovers, power quality / metering anomalies, and escalation of public opinion.

[0026] Among them, multi-task risk prediction involves simultaneously making joint predictions on two or more power supply service risk tasks under the same model or prediction framework, so as to share information and characterize the coupling relationship between tasks.

[0027] Among them, the dynamic early warning triggering result is an early warning triggering conclusion obtained based on the prediction result and combined with dynamic discrimination rules such as gating parameters / uncertainty / scenario tolerance, and includes at least the triggering object, triggering task and triggering level / time window.

[0028] Among them, the key drivers of risk escalation are the set of core factor variables that have a major causal contribution to the risk escalation, including quantifiable factors such as equipment status, load and weather disturbances, maintenance schedules, resource supply and customer sensitivity.

[0029] Among them, causal attribution counterfactual deduction is an analytical process that first uses causal methods to estimate the causal effect of key driving factors on risk, and then uses intervention (counterfactual scenario) settings to deduce "how the risk will change if a certain action is taken / a certain factor is changed".

[0030] Among them, the explainable early warning result is an early warning output that simultaneously provides "early warning conclusion + explanation of cause + evidence path + comparison of handling effect", and includes at least explanatory information such as key driving factors, causal contribution and risk transmission path.

[0031] Among them, behavior handling orchestration is to transform interpretable early warning results into actionable handling plans and task flows, including task breakdown, responsible parties, execution time limits, resource allocation and customer outreach strategies, as well as process tracking and feedback indicators.

[0032] Specifically, the heterogeneous data of power supply services from multiple sources corresponding to the target power grid (including power outages / faults, work orders and complaints, equipment ledgers and topology, GIS geographic location, operating status / power quality, customer profiles and group tags, etc.) are cleaned, standardized, and aligned with primary keys. Each record is then coded as an event using the three-element anchor point of "timestamp-spatial location-topology resource". This means that the heterogeneous data of power supply services from multiple sources is normalized to a unified time granularity based on the occurrence time of the record, the data is mapped to GIS coordinates / grids / transformer area boundaries using spatial information, and the data is mapped to topological objects such as station-line-transformer area-box-table using resource information. At the same time, fields such as event type, scope of impact, associated customer group tags, and handling status are added. Finally, spatiotemporal anchor point event data that can be located on the "map", calculated on the "data", and associated across systems is formed.

[0033] Based on spatiotemporal anchor point event data, graph elements of "entity-relationship-attribute" are extracted and constructed. These elements are then uniformly identified and deduplicated for power grid topology entities (stations / lines / transformer areas / equipment), geographical entities (grids / administrative regions / communities / key user locations), service entities (work orders, power outage events, emergency repair tasks, proactive notifications), customer entities (users / user groups / profile tags), and resource entities (work teams, vehicles, materials, channels). Furthermore, multi-source associations (such as topology connections, geographical adjacencies, service attribution, causal associations, propagation associations, and resource reachability associations) are generated based on event co-occurrence, temporal adjacency, and topological / geographical reachability. Relationship weights are then updated over time to obtain spatiotemporal knowledge graph data for power supply services that can be used for path reasoning and risk propagation calculations, achieving an integrated knowledge foundation of "data" and "graph".

[0034] Based on the spatiotemporal knowledge graph data of power supply services, a multi-task learning input (including node representation, edge relation weights, event sequence features, and customer group sensitivity features) is constructed for a preset prediction time window ΔT, using geographical or topological units as prediction objects. This input jointly predicts at least one or more risk tasks, such as power outage impact risk, work order / complaint spillover risk, power quality / metering anomaly triggering risk, and public opinion escalation risk, outputting the risk prediction value for each unit within ΔT. Simultaneously, dynamic gating parameters are formed by combining historical risk distribution and risk tolerance, transforming the triggering logic from a "fixed threshold" to "scenario-adaptive gating triggering." When the predicted risk meets the gating conditions (which may include a comprehensive judgment of risk magnitude, growth rate, propagation potential, confidence level, etc.), a corresponding dynamic early warning trigger result is generated, specifying the triggering unit, triggering task, and triggering time window.

[0035] Based on the dynamic early warning triggering results, the risk propagation path is traced back on the power supply service spatiotemporal knowledge graph to identify the candidate driving factor path set, and the key driving factors and their causal effect values ​​are obtained by fitting based on the structural causal model constraints. On this basis, counterfactual intervention simulations are performed on at least two candidate disposal actions (such as resource pre-positioning, transfer / uninterrupted power operation arrangement, proactive notification and group outreach strategy, emergency repair organization optimization, etc.) to calculate the expected reduction in risk, scope of impact, and feasibility constraints of each action within ΔT. Finally, an interpretable early warning result is generated, which includes "risk level, key driving factors and their contribution, risk propagation path, comparison of disposal action effects and recommended strategies". This result is then structured and mapped into behavioral disposal orchestration elements (task list, responsible entity, execution time limit, resource allocation and outreach strategy, process traceability and reinjection indicators) for coordinated control and closed-loop disposal of the target power grid.

[0036] In the aforementioned power supply service data-map linkage monitoring, control, and early warning method, spatiotemporal anchoring of multi-source heterogeneous power supply service data of the target power grid is performed, forming locationable and traceable spatiotemporal anchor event data under a unified "data-map" coordinate system. Then, spatiotemporal knowledge fusion of multi-source relationships is used to construct a power supply service spatiotemporal knowledge graph. This transforms risk identification from scattered data and static display to reasonable computation based on topology-geography-customer groups-event chains. On this basis, the spatiotemporal knowledge graph is used to conduct multi-task risk prediction within a preset prediction time window and generate dynamic early warning trigger results. This allows the early warning threshold to adjust according to the scenario, its own historical distribution, and risk tolerance. Adaptive adjustments significantly reduce false alarms and missed alarms caused by fixed threshold rules; further, by combining dynamic trigger results to perform causal attribution and counterfactual deduction on key drivers of risk escalation, the early warning is upgraded from "result notification" to "explainable causes and quantifiable measures" decision output, and the explainable early warning results are directly transformed into actionable action arrangements, thereby achieving forward-looking prediction, explainable location and closed-loop linkage control of risks, improving the accuracy of resource allocation and customer outreach, shortening response time, suppressing complaint spillover and public opinion escalation, improving power supply service quality and governance efficiency, and supporting continuous feedback optimization to form a self-evolving monitoring-early warning-response closed loop.

[0037] In an exemplary embodiment, based on the spatiotemporal knowledge graph data of power supply services, a multi-task risk prediction is performed on the power supply service risk of the target power grid within a preset prediction time window to obtain a dynamic early warning triggering result, including steps 302 to 306. Wherein:

[0038] Step 302: Based on the spatiotemporal knowledge graph data of power supply service, perform quantile calibration on the historical risk result data and risk tolerance parameters of the target power grid to obtain dynamic early warning triggering gating parameters.

[0039] Step 304: Perform multi-task risk prediction on the spatiotemporal knowledge graph data of power supply services to obtain the risk prediction results within the preset prediction time window.

[0040] Step 306: Based on the risk prediction results and the dynamic early warning triggering gating parameters, perform counterfactual perturbation deduction on the power supply service spatiotemporal knowledge graph data to obtain the dynamic early warning triggering results.

[0041] Among them, historical risk outcome data refers to the output data of various power supply service risks that have occurred or been assessed in the target power grid within a historical period, including risk event labels, risk levels, scope of impact, and corresponding actual outcome indicators.

[0042] Among them, the risk tolerance parameter is a control parameter used to characterize the management's acceptance of false alarms, omissions and risk exposures, which can be expressed as risk budget, service level requirements or false alarm / omission cost weight.

[0043] Among them, quantile calibration is a process of calibrating and correcting thresholds or limits based on the statistical or conditional distribution of historical risk results, according to a given quantile, so that the triggering criteria can adapt to the scene distribution.

[0044] Among them, the dynamic early warning trigger gating parameters are a set of parameters obtained by quantile calibration and used to trigger and judge the predicted risks. They include at least dynamic thresholds, trigger confidence lower limits, and task / scenario weights.

[0045] Among them, multi-task risk prediction involves simultaneously modeling and predicting two or more types of power supply service risk tasks under the same prediction framework, so as to share information and characterize the coupling between tasks.

[0046] Among them, the risk prediction result is the risk prediction value and its necessary additional information (such as risk probability, intensity, increment or confidence level) output for each geographic unit or topological unit within the preset prediction time window.

[0047] Among them, counterfactual perturbation simulation involves applying hypothetical interventions / perturbations to key factors or controllable variables on a knowledge graph and generating multi-scenario simulation results to assess the robustness and sensitivity of risk triggering.

[0048] Specifically, based on the spatiotemporal knowledge graph data of power supply services, the historical risk results data of the target power grid are first statistically sliced ​​at the task level, unit level (geographic unit / topological unit), and scenario level (such as seasonality, weather level, load level, and customer group sensitivity stratification). The empirical distribution or conditional distribution of risk indicators under each slice is calculated. Then, a risk tolerance parameter (which can be represented as the allowable false alarm cost, false alarm cost, risk budget, or service level requirement) is introduced and mapped to the corresponding quantile (e.g., P90 / P95 / P99 or a combination of stratified quantiles). This generates a set of parameters such as dynamic gating thresholds, gating weights, and trigger confidence lower limits for each risk task, each unit, and its scenario slice, thereby obtaining dynamic early warning triggering gating parameters that adapt to changes in the operating environment and service strategy.

[0049] The spatiotemporal knowledge graph data of power supply services is characterized and organized into sequences, using nodes (stations / line areas / geographic grids / customer groups) and edges (topological connections, geographical adjacencies, service associations, historical co-occurrences) as the structural framework, while simultaneously overlaying event time series and attribute features (power outages, work orders, emergency repairs, notifications, weather, public opinion, etc.). Based on this, a multi-task joint prediction framework is employed to simultaneously predict multiple types of power supply service risks within a preset prediction time window ΔT. The framework outputs predicted values ​​for each unit within the preset prediction time window ΔT, including risk probability, risk intensity, and risk increment. It can also simultaneously output inter-task coupling information (e.g., the contribution of one type of risk to the propagation of another). The fusion of these data sets forms the risk prediction result within the preset prediction time window.

[0050] Based on risk prediction results and dynamic early warning triggering gating parameters, a counterfactual perturbation set is constructed on the power supply service spatiotemporal knowledge graph. This involves targeted perturbations of corresponding node attributes, edge weights, or event intensity around key uncertainties and controllable factors (such as weather disturbances, load fluctuations, equipment degradation, resource arrival delays, and triggering success rates), generating multiple counterfactual scenario graphs. These graphs are then used to extrapolate and calculate the risk prediction response distribution under different perturbations (forming a risk envelope or robust risk interval). This robust risk distribution is then used to perform gating discrimination with the gating parameters. When the risk meets the triggering conditions in a robust sense (e.g., exceeding a dynamic threshold and still holding under confidence constraints, or the risk upward trend persists in most counterfactual scenarios), a triggering conclusion is output; otherwise, the alarm is suppressed. This yields a dynamic early warning triggering result containing the triggering object, triggering task, triggering level, and robustness evidence.

[0051] In this embodiment, dynamic early warning triggering gating parameters are obtained by performing quantile calibration on historical risk result data and risk tolerance parameters. This allows the early warning threshold to be adaptively adjusted according to the target power grid operation scenario and management strategy, avoiding false alarms and missed alarms caused by fixed thresholds. Based on this, multi-task risk prediction is carried out on the spatiotemporal knowledge graph data of power supply services. Within a preset prediction time window, multiple types of risks and their coupling relationships, such as the impact of power outages, complaint spillover, and power quality triggering, can be simultaneously characterized, achieving more comprehensive forward-looking identification. Furthermore, counterfactual perturbation inference is performed using the risk prediction results and gating parameters. This allows the robustness of risk triggering under various perturbation scenarios to be evaluated and occasional noise to be filtered, thereby outputting more reliable and controllable dynamic early warning triggering results and improving the accuracy and stability of early warnings.

[0052] In an exemplary embodiment, multi-task risk prediction is performed on the spatiotemporal knowledge graph data of power supply services to obtain risk prediction results within a preset prediction time window, including steps 402 to 408. Wherein:

[0053] Step 402: Perform self-supervised representation learning on the spatiotemporal knowledge graph data of power supply services to obtain general spatiotemporal graph representation data; Step 404: Based on the general representation data of the spatiotemporal graph, perform task conditional subgraph retrieval on each risk task within the preset prediction time window to obtain the targeted propagation subgraph representation data. Step 406: Perform multi-task spatiotemporal joint prediction on the directional propagation subgraph representation data to obtain the multi-task risk prediction value within the preset prediction time window; Step 408: Perform fusion correction based on statistical features on the multi-task risk prediction values ​​to obtain the risk prediction results within the preset prediction time window.

[0054] Self-supervised representation learning is a method that uses the inherent structure and information of the data itself to learn general feature representations by designing unsupervised pre-training tasks.

[0055] Among them, the spatiotemporal graph general representation data is a graph embedding representation containing spatiotemporal information and relational features obtained through self-supervised learning, which can be used for multi-task risk prediction and inter-task information sharing.

[0056] Among them, risk tasks are specific types of power supply service risks that need to be predicted or assessed, such as the impact of power outages, spillover of complaints, and abnormal power quality.

[0057] Among them, task-conditional subgraph retrieval is based on the specific conditions of each risk task (such as task type, prediction target, etc.) to retrieve the most relevant local subgraph in the spatiotemporal knowledge graph, so as to improve task relevance and prediction efficiency.

[0058] Among them, the targeted propagation subgraph representation data is obtained by performing targeted weighting or pruning of subgraphs in the graph according to the task-conditional subgraph retrieval based on the task objective, so as to obtain a structured graph representation suitable for the task.

[0059] Among them, multi-task spatiotemporal joint prediction is based on shared spatiotemporal map representation, and simultaneously performs joint modeling and prediction of multiple risk tasks to improve information sharing and synergy among tasks.

[0060] Among them, the multi-task risk prediction value is the prediction result of each risk task obtained through multi-task spatiotemporal joint prediction, including indicators such as the probability of risk occurrence, intensity, or potential impact.

[0061] Among them, the fusion correction based on statistical features is to align and correct the historical statistical features of the multi-task risk prediction values ​​to ensure the robustness and consistency of the prediction results and improve their credibility.

[0062] Specifically, the power grid topology nodes, geographical unit nodes, customer group nodes, service event nodes, and their edge relationships (topological connections, geographical adjacencies, service affiliations, historical co-occurrences, etc.) in the power supply service spatiotemporal knowledge graph data are uniformly encoded as learnable objects. This is achieved by constructing pre-training tasks that do not require manual annotation (such as time-series-based event prediction, node / edge attribute mask reconstruction, temporal adjacency comparison, and sub-...). Figure 1 Consistency constraints, cross-view Figure 1 (e.g., consistency alignment), learns embedding representations under the common constraints of "structure-time-semantics" in the spatiotemporal dimension, and outputs spatiotemporal graph general representation data containing general node embeddings, edge relation embeddings, and event sequence embeddings.

[0063] Based on the general representation data of the spatiotemporal graph, task conditions are constructed for each type of risk task (which may consist of task identifier, target risk type, preset prediction time window ΔT, and the sensitive customer group / key resource type corresponding to the task, etc.), and the task conditions are used as query vectors for subgraph retrieval in the graph. In the actual retrieval process, nodes and edges that are highly correlated with the target unit in terms of topological reachability, geographical adjacency, historical common causes, or temporal co-occurrence are prioritized for retrieval, while relationships unrelated to the task are suppressed. The retrieved subgraphs are further pruned and weighted according to the propagation direction (e.g., from the disturbance source to the influence recipient, from upstream constraints to downstream spillover), ultimately forming targeted propagation subgraph representation data for each task and each prediction object.

[0064] This method jointly models the structural information (node / edge embedding, relation weights, propagation direction) and temporal information (event sequences, time intervals, periodic features) within the directed propagation subgraph representation data. Task-specific prediction heads are overlaid on a shared spatiotemporal coding backbone to achieve information sharing and coupled characterization among various risk tasks. During the prediction process, competition and collaboration between tasks are considered simultaneously (e.g., the linkage between complaint spillover and power outage impacts). The method outputs multi-task risk prediction values ​​(such as risk probability, risk intensity, risk increment, etc.) for each prediction object within a preset prediction time window ΔT, and can include inter-task coupling contributions or intermediate attention weights.

[0065] The multi-task risk prediction values ​​are aligned with historical statistical characteristics and drift corrections are performed. For example, the outputs of different tasks are calibrated by distribution (temperature scaling / quantile mapping), the baseline differences of different units are corrected by layer (grouping by transformer area type, load level, proportion of sensitive customer groups, etc.), and short-term abnormal fluctuations are robustly processed (based on robust mean / variance constraints based on sliding window). The multi-task outputs are then fused in a consistent manner (ensuring that the risk intensity on the same propagation chain is logically coordinated). Finally, the risk prediction results within the preset prediction time window that meet the requirements of comparability, interpretability and stability are obtained.

[0066] In this embodiment, by performing self-supervised representation learning on the spatiotemporal knowledge graph data of power supply services, general spatiotemporal graph features can be effectively extracted, thereby improving the model's generalization ability in multi-task risk prediction. Based on this, task-conditional subgraph retrieval can extract the most relevant local information for specific risk tasks from the entire graph, further enhancing the accuracy and relevance of predictions. Multi-task spatiotemporal joint prediction can simultaneously handle multiple risk tasks and enhance the synergistic effect of predictions through information sharing, ultimately achieving effective coupling between multiple risk tasks. Statistical feature-based fusion correction ensures the robustness and consistency of various risk prediction results by correcting the biases in the prediction results, thereby providing more accurate and reliable decision support for risk warning and control of power supply services.

[0067] In an exemplary embodiment, multi-task spatiotemporal joint prediction is performed on the directed propagation subgraph representation data to obtain multi-task risk prediction values ​​within a preset prediction time window, including steps 502 to 510. Wherein:

[0068] Step 502: Perform task causal gating structure analysis on the task identification data and propagation direction data in the directed propagation subgraph representation data to obtain task coupling gating graph data; Step 504: Based on the task coupling gating graph data, perform gating graph diffusion encoding on the directed propagation subgraph representation data to obtain the task coupling diffusion representation data; Step 506: Perform multi-scale time window folding on the task coupling diffusion characterization data to obtain multi-scale spatiotemporal sequence characterization data; Step 508: Perform confidence-modulated hybrid decoding on the multi-scale spatiotemporal sequence representation data to obtain candidate multi-task risk prediction values; Step 510: Perform supervised parameterization mapping on the candidate multi-task risk prediction values ​​to obtain the multi-task risk prediction values ​​within the preset prediction time window.

[0069] Among them, task identification data is identification information used to uniquely mark the type of risk prediction task, indicating the risk task corresponding to the data sample (such as the impact of power outages, complaint spillover, abnormal power quality, etc.).

[0070] Among them, propagation direction data is used to characterize the direction information of risk impact propagation from source node to recipient node in the knowledge graph, indicating the upstream-downstream relationship of information flow or causal impact.

[0071] Among them, the task causal gating structure analysis is an analysis process that uses task identification and propagation direction to structurally model the causal dependencies between tasks and generate gating structures to control the strength and channels of cross-task information transmission.

[0072] Among them, task coupling gating graph data is a graph structure data that describes the causal coupling relationship between multiple tasks and their gating weights, where nodes correspond to tasks or task states and edges represent channels of causal influence that can be transmitted.

[0073] Among them, gated graph diffusion coding is a process of diffusing and encoding graph features under the constraints of task-coupled gated graphs, so as to realize the directional aggregation of cross-task information along the gated channel.

[0074] Among them, the task coupling diffusion representation data is the representation data obtained after gating graph diffusion encoding, which includes the fused feature representation of each task after considering the causal coupling between tasks.

[0075] Multi-scale time window folding is a process of aggregating, compressing, and aligning sequence segments at different time scales according to multiple time windows, in order to simultaneously preserve short-term fluctuations and long-term trend information.

[0076] Among them, multi-scale spatiotemporal sequence representation data is a spatiotemporal sequence feature representation formed after multi-scale time window folding, which contains graph structure and temporal evolution information at multiple time scales.

[0077] Among them, confidence-modulated hybrid decoding is a decoding process that introduces predicted confidence to adaptively modulate the decoding weights under a multi-decoding branch or multi-expert decoding framework in order to improve output robustness.

[0078] Among them, the candidate multi-task risk prediction value is the set of preliminary multi-task prediction outputs generated by the confidence-modulated hybrid decoder, which has not yet undergone final supervised calibration.

[0079] Supervised parameterized mapping is a process of parameterizing and mapping candidate predicted values ​​using supervisory signals to reduce bias and obtain the final multi-task risk predicted values.

[0080] Specifically, task identifiers and propagation direction data for each task are extracted from the directed propagation subgraph representation data to identify potential causal relationships between risk tasks. By establishing a causal dependency model between tasks, it is identified which tasks influence the occurrence of other tasks, and the directionality of task propagation is further analyzed (e.g., from equipment failure to power outages, or from weather changes to customer complaints). Based on these causal relationships between tasks, a task coupling gating graph is constructed, with tasks and their impact paths as nodes and edges, ultimately resulting in a graph structure where the gating relationships between each task node reflect the dependencies between tasks, forming task coupling gating graph data.

[0081] Based on task-coupled gated graph data, a graph diffusion algorithm (such as a graph convolutional network or graph propagation algorithm) is used to perform diffusion encoding on the directed propagation subgraph representation data, expanding the influence between task nodes. Specifically, the graph diffusion algorithm propagates the influence of each task to its related task nodes, reflecting the propagation effect of causal relationships between tasks. For each task node, its influence weight in other nodes of the graph is calculated based on the propagation path in the task-coupled gated graph data, and these influence weights are embedded into the representation of the task node through the diffusion process, resulting in task-coupled diffused representation data.

[0082] By employing a multi-scale time window folding method, event information from task coupling diffusion representation data at different time scales is merged to better capture the impact of short-term fluctuations and long-term trends on risk. For example, short-term time windows in task coupling diffusion representation data may focus on the immediate impact of daily load fluctuations on risk, while long-term time windows focus on the gradual impact of seasonal changes on risk. By folding data from different time windows in the task coupling diffusion representation data into the same representation space, a multi-scale spatiotemporal sequence representation data is generated. The feature representation of each time window in the multi-scale spatiotemporal sequence representation data will contain information from multiple scales, giving the model sensitivity to risk changes at multiple time scales.

[0083] The generated multi-scale spatiotemporal sequence representation data is input into a hybrid decoder. This decoder not only considers the dependencies between tasks but also incorporates weighted modulation based on the prediction confidence of each task. Specifically, for each task's prediction, the hybrid decoder calculates its corresponding confidence level (e.g., prediction accuracy based on historical data or the difficulty of the current task), and then adjusts the decoder's weights according to the confidence level. This ensures that high-confidence task predictions receive more attention and that high-confidence task results are given greater importance, ultimately generating candidate multi-task risk prediction values.

[0084] The candidate multi-task risk predictions (or historical candidate multi-task risk predictions of the target power grid) are compared with the actual risk labels of the target power grid in historical data. The correction model parameters are optimized by minimizing the loss function (such as mean squared error or cross-entropy loss) between the predicted values ​​and the true labels. The correction model can reduce the error by adjusting the parameters based on the obtained prediction error. The correction model parameters obtained from training are used to feed back and correct the candidate multi-task risk predictions, ultimately obtaining multi-task risk predictions that meet the requirements within the preset prediction time window, and finally obtaining multi-task risk predictions for various risk tasks within the preset prediction time window.

[0085] In this embodiment, by performing task causal gating structure analysis on the task identifiers and propagation directions in the directional propagation subgraph representation data and forming a task coupling gating graph, the causal dependencies and propagation directions between multiple types of power supply service risk tasks can be explicitly characterized, thereby avoiding negative migration caused by "blindly sharing features" in multi-task prediction. Graph diffusion coding is implemented under the constraints of the gating graph, which can propagate the influence between tasks along a controllable causal channel, improving the effectiveness and interpretability of cross-task information interaction. Furthermore, multi-scale time window folding is adopted, which can simultaneously capture short-term sudden disturbances and medium- to long-term evolution trends, making the prediction more sensitive to risk patterns at different time scales. Combined with confidence modulation hybrid decoding, decoding weights can be adaptively allocated according to the confidence of different tasks and scenarios, improving the robustness and noise resistance of the prediction output. Finally, supervised calibration of candidate predictions is performed through supervised parameterized mapping, which can reduce systematic bias and improve the comparability and consistency of different task outputs, thereby obtaining more accurate and stable multi-task risk prediction values.

[0086] In an exemplary embodiment, based on the risk prediction results and dynamic early warning triggering gating parameters, counterfactual perturbation inference is performed on the power supply service spatiotemporal knowledge graph data to obtain the dynamic early warning triggering results, including steps 602 to 610. Wherein:

[0087] Step 602: Based on the risk prediction results and dynamic early warning triggering gating parameters, risk propagation path is extracted from the power supply service spatiotemporal knowledge graph data to obtain targeted intervention subgraph data; Step 604: Solve the minimum counterfactual perturbation problem on the targeted intervention subplot data to obtain the set of perturbation variables and perturbation amplitude data; Step 606: Based on the set of disturbance variables and the disturbance amplitude data, perform causal intervention sampling and inference on the spatiotemporal knowledge graph data of power supply service to obtain multi-counterfact scenario graph data; Step 608: Perform risk envelope estimation on the multi-counterfact scenario graph data to obtain robust risk prediction distribution data; Step 610: Perform gating discrimination processing on the robust risk prediction distribution data and dynamic early warning triggering gating parameters to obtain the dynamic early warning triggering results.

[0088] Among them, the risk propagation path targeted extraction is a process that, based on risk prediction results and gating parameters, filters and extracts the propagation path that contributes the most to the risk in the spatiotemporal knowledge graph along the upstream-downstream direction of the risk impact.

[0089] Among them, the targeted intervention subgraph data is local subgraph structure data extracted from the risk propagation path, which includes key nodes, key edges and their time-series segments, and is used to carry out subsequent intervention and extrapolation calculations.

[0090] Among them, the minimum counterfactual perturbation solution is an optimization process that finds a set of intervention variables and their values ​​that minimize the perturbation cost required to change the risk state under the condition of satisfying the triggering discrimination boundary.

[0091] The set of perturbation variables is the set of variables selected for intervention in counterfactual perturbations. These variables can come from elements such as node attributes, edge weight parameters, or event intensity.

[0092] Among them, the disturbance magnitude data is the data on the specific amount of change or replacement value applied to each disturbance variable, which is used to describe the intensity and direction of the intervention.

[0093] Among them, causal intervention sampling and extrapolation is a calculation process that involves intervening in the perturbation variable and conditionally sampling the remaining uncertain factors to generate and extrapolate the risk evolution under multiple counterfactual scenarios.

[0094] Among them, the multi-counterfact scenario map data is a data set of multiple post-intervention map snapshots and their event evolution sequences within time windows generated by causal intervention sampling and inference.

[0095] Risk envelope estimation is a process of statistically estimating the upper and lower bounds, quantile intervals, and dispersion of risk output samples obtained in multiple counterfactual scenarios, thereby forming a risk range characterization.

[0096] Among them, robust risk prediction distribution data is risk distribution characterization data obtained by risk envelope estimation, and includes at least robust indicators such as quantile intervals / extreme boundary / confidence level or trigger consistency.

[0097] Among them, the gating discrimination processing is the process of performing distribution-level trigger determination on the robust risk prediction distribution and dynamic gating parameters, and outputting the early warning trigger conclusion and trigger intensity.

[0098] Specifically, based on the risk prediction results, the candidate risk tasks, target prediction units (geographic units or topological units), and preset prediction time windows ΔT are determined. Combined with the gating thresholds and weights of the corresponding tasks / scenarios in the dynamic early warning trigger gating parameters, highly sensitive nodes that are "beyond or close to gating risk" are identified as the set of recipient nodes. On the power supply service spatiotemporal knowledge graph data, with the recipient nodes as the endpoints, a directed backtracking search is performed along edge relationships such as topological connections, geographical adjacencies, service affiliations, and historical co-occurrences. Paths with high edge weights, significant propagation contributions, and task relevance are prioritized and then pruned according to the propagation direction (upstream constraints → downstream influence, disturbance sources → service spillover). The backtracked critical paths and their covered nodes / edges and event sequences are packaged to form targeted intervention subgraph data containing "path structure + temporal segments + task gating weights".

[0099] Variable elements in the targeted intervention subgraph data are variableized to form a disturbance candidate space. Variable elements include at least node attributes (such as equipment health, load level, arrival delay, and reach success rate), edge weight parameters (such as propagation strength and adjacency weight), and event intensity / frequency (such as work order influx rate and failure rate). Based on this, a "minimum disturbance" optimization objective is constructed, which aims to minimize the disturbance cost while satisfying the gating trigger discrimination boundary (e.g., changing the risk from non-triggered to triggered, or from triggered to non-triggered). The disturbance cost can be defined by the L0 / L1 norm (minimum number / magnitude of disturbance variables), business feasibility penalty (penalty for uncontrollable variables), and stability penalty (penalty for disturbance discontinuity). A set of disturbance variables and their corresponding disturbance amplitude data are obtained through constraint optimization or heuristic search, thereby obtaining the "minimum intervention most sensitive near the trigger boundary".

[0100] Based on the set of disturbance variables and the disturbance amplitude data, the generation mechanism of each disturbance variable in the set of disturbance variables is replaced by a do-intervention approach (e.g., fixing a node attribute to the value after disturbance, or adjusting the propagation intensity of a side according to the amplitude). Then, the remaining uncertain factors are sampled on the spatiotemporal knowledge graph of the power supply service after intervention according to their historical condition distribution (e.g., weather level, load fluctuation, repair arrival process, resource response delay, etc.). Repeated sampling generates multiple sets of counterfactual scenarios, each set of scenarios corresponding to a "graph snapshot after intervention + event evolution sequence within the time window ΔT", thus forming multiple counterfactual scenario graph data, which depicts the possible range and distribution pattern of risk output when key variables are disturbed.

[0101] For each scenario in the multi-counterfactual scenario map data, calculate the risk output of the target risk task within ΔT (this can be done by directly calling the same multi-task prediction model or its inference module) to obtain a risk sample set. Perform risk envelope estimation on this risk sample set. The envelope estimation includes at least the quantile interval (e.g., P5–P95), extreme value boundaries (worst / best case), variance or confidence interval, and the "trigger consistency" index (the proportion of scenarios in which the triggering condition is met) to form robust risk prediction distribution data.

[0102] Under the constraints of dynamic thresholds, lower confidence limits, and task / scenario weights given by the dynamic early warning triggering gating parameters, a distribution-level discrimination criterion is used to output triggering conclusions for robust risk prediction distribution data. For example, "high quantile risk exceeds the threshold and trigger consistency is higher than the threshold" can be used as the triggering condition, or "worst-case risk exceeds the threshold" can be used as the conservative triggering condition. Simultaneously, trigger strength (margin from the gating boundary), robustness evidence (quantile intervals and consistency), and recommended attention paths (from the targeted intervention subgraph) are generated to match the triggering conclusions. The final output includes the triggering object, triggering task, triggering level, triggering time window, and its robustness evidence.

[0103] In this embodiment, by extracting the risk propagation path from the spatiotemporal knowledge graph of power supply services based on risk prediction results and dynamic early warning triggering gating parameters, and forming a targeted intervention subgraph, the early warning judgment can be transformed from "coarse-grained triggering of full data" to "precise positioning along key propagation channels," significantly reducing computational redundancy and improving the targeting of triggering. Furthermore, by using minimum counterfactual perturbation to obtain the set of perturbation variables and the perturbation amplitude, the most sensitive key factors and minimum intervention amount near the trigger boundary can be identified, thereby improving the interpretability and controllability of the early warning. On this basis, causal intervention sampling and inference are performed to generate a multi-counterfactual scenario graph, so that the risk assessment covers multiple possible evolutions under perturbation and uncertainty conditions, avoiding the oversensitivity of single-point prediction to noise and occasional fluctuations. By estimating the risk envelope of multi-scenario results, a robust risk prediction distribution is obtained, which can quantify the upper and lower bounds of risk and the consistency of triggering. Then, in the gating judgment stage, the dynamic early warning triggering results are output according to the distribution-level criteria, thereby achieving a more robust and reliable early warning triggering, reducing false alarms and missed alarms, and providing a reliable basis for subsequent handling and scheduling.

[0104] In an exemplary embodiment, based on the dynamic early warning triggering result, causal attribution counterfactual inference is performed on the risk escalation key driving factors of power supply service risk to obtain an interpretable early warning result, including steps 702 to 708. Wherein:

[0105] Step 702: Based on the dynamic early warning triggering results, perform risk propagation path backtracking analysis on the power supply service spatiotemporal knowledge graph data to obtain a set of candidate driving factor paths; Step 704: Perform structural causal model constraint fitting on the candidate driving factor path set to obtain the key driving factor set and causal effect values; Step 706: Based on the set of key driving factors and the causal effect value, perform counterfactual effect deduction on at least two candidate actions for the target power grid to obtain the expected risk reduction effect. Step 708: Convert the power grid early warning rules to reflect the expected effect of risk reduction, and obtain interpretable early warning results.

[0106] Among them, risk propagation path retrospective analysis is an analysis process that uses the early warning trigger object as the risk receptor and traces the source and transmission link of risk upstream along the propagation direction of topology, geography and business in the spatiotemporal knowledge graph.

[0107] Among them, the candidate driver path set is a set of several paths that may lead to an increase in risk obtained from backtracking analysis. Each path contains driver node / event and its relationship link to risk receptor and time series evidence.

[0108] Among them, structural causal model constraint fitting is a fitting process of structural search and parameter estimation of structural causal model under prior constraints such as time sequence, topological direction, and physical accessibility, so as to ensure that the causal relationship is reasonable and interpretable.

[0109] The key driver set is the set of driver variables that have made the main causal contribution to the increase in risk after being screened by the structural causal model.

[0110] Among them, the causal effect value is a quantitative measure of the causal contribution of key driving factors to risk indicators, such as the average treatment effect or conditional causal effect.

[0111] Among them, candidate actions are alternative intervention measures or combinations of measures that can be implemented for the target power grid, used to change key driving factors or their propagation paths to reduce risks.

[0112] Counterfactual effect simulation is a simulation process that applies hypothetical interventions to candidate actions and calculates their impact on risk output within a causal model or graph simulation framework.

[0113] Among them, the expected effect of risk reduction is the quantitative expected result of each candidate action in terms of risk reduction magnitude, probability of achievement, or scope of impact within a preset time window, obtained from counterfactual deduction.

[0114] Among them, the power grid early warning rule conversion is a process of mapping the expected effect of risk reduction and its explanatory information into rule-based expressions such as early warning levels, triggering conditions and handling suggestions that can be executed by power grid operations.

[0115] Specifically, based on the dynamic early warning triggering results, the triggering object (geographic unit / topological unit), triggering task type (such as power outage impact, complaint spillover, etc.), and triggering time window are determined, and the node corresponding to the triggering object is designated as a "risk receptor node" in the power supply service spatiotemporal knowledge graph. A directed backtracking search is performed along the multi-source relationships in the graph, including topological connection edges, geographical adjacency edges, service affiliation edges, and historical co-occurrence edges, according to the propagation direction, to find upstream disturbance sources and intermediate transmission links that may lead to an increase in risk at the receptor node. During the backtracking process, paths are screened for task relevance and ranked by contribution (e.g., based on edge weights, attention weights, historical mutual information, temporal sequence constraints, etc.), and paths that satisfy "consistency with the triggering time window, consistency with task semantics, and consistency with the propagation direction" are retained, ultimately forming a candidate driving factor path set. Each path contains a driving factor node / event, a transmission relationship link, and a corresponding temporal evidence fragment.

[0116] A variable model is performed on the candidate driving factor path set, mapping key node attributes, event intensity, and edge weight parameters on the path to candidate factor variables, and using trigger risk indicators as outcome variables. Simultaneously, structural constraints (e.g., topological prior direction, physical reachability constraints, temporal sequence constraints, and common / shared cause relationship constraints) are extracted from the spatiotemporal knowledge graph to form a constraint set for the structural causal model. The structural causal model is fitted under these constraints, including constraint search and parameter estimation of the causal graph structure, ensuring the model satisfies prior constraints and stability requirements while explaining historical data. After fitting, the causal effect (e.g., average treatment effect, path effect, or conditional causal effect) of each candidate factor variable on the outcome variable is calculated, and the key driving factor set and its corresponding causal effect values ​​are obtained by filtering according to effect size and significance, thus converging "relevant factors" into "key factors with causal contributions."

[0117] Based on the set of key driving factors and causal effect values, at least two candidate actions for the target power grid are formally encoded, mapping these actions to controllable interventions in key driving factors (e.g., reducing fault incidence, shortening response time, improving power transfer success rate, and increasing proactive notification coverage), forming an action-factor intervention vector. Then, counterfactual inference is performed on the fitted structural causal model or its inference engine coupled with a knowledge graph. That is, while keeping other conditions unchanged or sampling according to conditional distribution, different candidate actions are applied, and the counterfactual output of risk indicators within the trigger time window is calculated for each action. Finally, the risk difference between the "no action / baseline action" and "candidate action" scenarios is compared to obtain the expected risk reduction effect (which may include reduction magnitude, coverage, probability of achievement, or confidence interval) for each candidate action.

[0118] The expected risk reduction effect is mapped into actionable early warning elements and rule expressions. For example, an early warning level is generated based on the difference between the risk reduction magnitude and the gating threshold; "cause labels and contribution" are generated based on key driving factors and their causal effects; "evidence links and scope of impact" are generated based on the backtracked propagation path; and "recommended handling strategies and priorities" are generated based on the effect comparison of candidate handling actions. At the same time, structured fields such as actionable trigger conditions, handling suggestions, responsible entities and time limits are generated according to the power grid business rule system. Finally, an interpretable early warning result containing "early warning conclusion - cause explanation - evidence path - handling effect comparison / recommended action" is obtained.

[0119] In this embodiment, by tracing the risk propagation path on the spatiotemporal knowledge graph of power supply services based on the dynamic early warning triggering results, the early warning can be upgraded from "point-based triggering" to "link-level tracing," quickly identifying candidate driving factor paths that may lead to increased risk. On this basis, key driving factors are screened by constraining and fitting with a structural causal model, and the causal effect value is quantified, which can effectively distinguish between correlation and causality, reduce misjudgments, and improve the credibility of the explanation. Furthermore, counterfactual effect deduction is carried out on at least two candidate disposal actions, and the expected effect of risk reduction is output, which can upgrade the early warning output from "risk warning" to "quantified comparison of disposal benefits," improving the operability of disposal decisions and the efficiency of resource allocation. Finally, the expected effect of risk reduction is converted into power grid early warning rules and an interpretable early warning result is formed, which can achieve seamless connection with the existing power grid business rule system, so that the early warning conclusion has causal explanation, evidence path, and actionable disposal suggestions, thereby significantly improving the accuracy, interpretability, and closed-loop linkage control capability of the early warning.

[0120] In an exemplary embodiment, a structural causal model constraint fitting is performed on the candidate driving factor path set to obtain a key driving factor set and causal effect values, including steps 802 to 812. Wherein:

[0121] Step 802: Perform path element variable mapping on the candidate driving factor path set to obtain the candidate factor variable set; Step 804: Based on the spatiotemporal knowledge graph data of power supply service, perform relational semantic constraints extraction on the candidate factor variable set to obtain a set of structural causal constraints; Step 806: Perform multi-environment invariance screening on the candidate factor variable set to obtain the core candidate factor set; Step 808: Based on the set of structural causal constraints, perform constraint structure search on the set of core candidate factors to obtain the structure of the structural causal model. Step 810: Estimate the constraint parameters of the structural causal model to obtain the structural causal model parameters; Step 812: Calculate the causal effect based on the parameters of the structural causal model to obtain the set of key driving factors and the causal effect value.

[0122] Among them, path element variable mapping is the process of transforming node attributes, edge relationship parameters and event elements in candidate driving factor paths into computable causal modeling variables.

[0123] Among them, the candidate factor variable set is the set of causal candidate variables to be tested obtained by variable mapping, which is used to explain risk outcomes and participate in subsequent structure learning and effect calculation.

[0124] Among them, relational semantic constraint extraction is the process of extracting prior constraints such as causal direction and reachability from the topological, geographical and business relational semantics of spatiotemporal knowledge graphs.

[0125] The structural causal constraint set is a prior set of causal structures extracted from relational semantic constraints, which includes allowed edges, prohibited edges, and preferred edges and their weight information.

[0126] Among them, multi-environment invariance screening is a process of testing whether the relationship between variables and risks remains stable under different operating scenarios / environment slices, and thereby screening out unrobust spurious correlation variables.

[0127] Among them, the core candidate factor set is a set of candidate factor variables that have been retained after screening for multi-environment invariance and have stable explanatory power for risk across scenarios.

[0128] Among them, constraint structure search is the process of searching and solving the causal graph structure under the constraints of the set of structural causal constraints to determine the causal connection relationship between variables.

[0129] Among them, the structural causal model structure is the causal graph structure obtained by the constraint structure search, which describes the causal edges and directions between the core candidate factors and risk outcomes, as well as between the factors themselves.

[0130] Among them, constraint parameter estimation is the process of fitting and estimating the parameters of each structural equation based on a given cause-effect graph structure and in combination with business and physical constraints.

[0131] Among them, the structural causal model parameters are the set of structural equations or conditional distribution parameters obtained from the constraint parameter estimation, which are used to perform intervention deduction and effect calculation.

[0132] Among them, the calculation of causal effect is a process of applying do intervention to variables based on the parameters of the structural causal model and quantifying their impact on risk outcomes in order to obtain a causal contribution measure.

[0133] Specifically, the path elements in the candidate driving factor path set are structurally decomposed. Node attributes (such as equipment health, load level, customer sensitivity, resource arrival delay, etc.), edge relationship parameters (such as topological connection strength, geographical adjacency weight, historical co-occurrence weight, propagation direction marker, etc.), and event elements (such as fault / power outage event intensity, work order arrival rate, complaint frequency, public opinion heat, weather level, etc.) are uniformly mapped into computable candidate factor variables. Each variable is assigned a clear value domain, temporal granularity, and spatial / topological anchor point. Triggering risk indicators (such as risk probability, risk intensity, or risk increment) are defined as outcome variables. Variables at different stages of the path are organized according to their temporal sequence into an input-mediation-output variable sequence, ultimately forming the candidate factor variable set and its corresponding sample dataset.

[0134] Based on the spatiotemporal knowledge graph data of power supply services, semantic constraints that can serve as priors for causal structures are extracted from the candidate factor variable set. These constraints include, but are not limited to: topological direction constraints (the direction of influence of upstream equipment on downstream power supply units), physical reachability constraints (causal edges are prohibited if there is no physical connection or if the equipment is unreachable), geographical adjacency constraints (the influence of spatially nearest neighbors takes precedence over distant unrelated ones), service affiliation constraints (the priority of service events within the same power supply station / area on the risk of customers in the same domain), temporal order constraints (the time of occurrence of the cause must be earlier than the time of the result), and domain rule constraints (such as the direct impact of maintenance plans on power outage risk, the upstream impact of weather on failure rate, etc.). The above constraints are formalized into a computable set of "allowed edges / prohibited edges / priority edges" and their confidence weights to obtain the set of structural causal constraints.

[0135] The candidate factor variable set is divided into multiple environments and subjected to invariance tests. Historical samples are sliced ​​into multiple "environments" according to different operating environments or scenarios (e.g., by season, weather level, load level range, holidays / weekdays, different transformer substation types, different proportions of sensitive customer groups, different operation and maintenance strategy stages, etc.). Within each environment, the conditional association between candidate factors and outcome variables is estimated, and their stability across environments is tested (e.g., through invariant risk minimization, stable regression / invariance tests, or robustness tests against key condition distribution drift). Variables that are only effective in individual environments or are significantly affected by confounding or selection bias are removed, while variables that have stable explanatory power for the results in most environments are retained, resulting in the core candidate factor set. This completes the "cross-scenario robustness" factor dimensionality reduction and spurious elimination before structure learning.

[0136] Under the constraints of "allowed edges / prohibited edges / preferred edges" in the structural causal constraint set, the search space is narrowed by utilizing the constraints of the structural causal constraint set (prohibiting edges that do not conform to topological / temporal / physical rules). Meanwhile, preferred edges from the structural causal constraint set are added to the structural score as penalty terms or prior scores. Candidate structures are obtained by searching the core candidate factor set through score search (e.g., directed acyclic graph scoring based on BIC / MDL) or constraint search (e.g., skeleton learning based on conditional independence tests). The feasibility of the candidate structures is verified by combining the domain constraints of the structural causal constraint set (e.g., prohibiting the formation of loops that violate temporal causal directions). Finally, a structural causal model structure that satisfies the constraints and can optimally explain the data is output, where the structural causal model structure clearly defines the causal connections between key factors and between key factors and risk outcome variables.

[0137] Given a fixed structure, for each structural causal model structure, an appropriate parameterization form is selected and fitted to the causal equation (the generation mechanism of each node under the conditions of its parent node). For continuous variables, linear / nonlinear regression, generalized linear models, or neural parameterized structural equations can be used, while for discrete variables, logistic regression or conditional probability tables can be used. Regularization and constraints (such as monotonicity constraints, nonnegativity constraints, amplitude boundaries, and sparsity constraints) are introduced to conform to the power grid business mechanism and improve generalization. The parameters of each equation and their uncertainty measures are obtained through maximum likelihood, Bayesian estimation, or variational inference, ultimately forming the structural causal model parameters.

[0138] Based on the parameters of the structural causal model, a do-intervention is constructed for each candidate factor (the factor is set to different values ​​or adjusted according to the perturbation amplitude). The remaining variables are then integrated / sampled according to the conditional distribution in the model to obtain the expected difference or quantile difference of the risk outcome before and after the intervention, forming the average treatment effect, conditional treatment effect, or path decomposition effect. Factors are screened and ranked based on effect size, significance, and robustness (cross-environment consistency), retaining the factors that contribute the most to the increase in risk as the set of key driving factors, and outputting the corresponding causal effect values, thus completing the convergence and quantification from "candidate paths" to "key causal drivers."

[0139] In this embodiment, by mapping the candidate driving factor path set to path element variables to form a candidate factor variable set, the graph path evidence can be transformed into a modelable and quantifiable causal analysis input, improving the computability of source tracing analysis. Furthermore, based on the spatiotemporal knowledge graph of power supply services, relational semantic constraints are extracted to form a structural causal constraint set. This allows the topological physical laws of the power grid, temporal sequence, and business semantic priors to be explicitly injected into the modeling process, significantly reducing the structural search space and decreasing the generation of unreasonable causal edges. On this basis, multi-environment invariance screening is performed to obtain a core candidate factor set, eliminating pseudo-correlated variables that are only effective in local scenarios and improving the robustness of causal conclusions under different seasons, loads, weather, and customer structures. Subsequently, under the constraints of the structural causal constraint set, a structural search is performed on the core candidate factors, and model parameters are estimated, resulting in a structural causal model that conforms to domain mechanisms and is interpretable. Finally, based on the model parameters, causal effect calculation is performed, and a key driving factor set and causal effect values ​​are output. This converges "related clues" into "causal contributions" and achieves quantitative ranking, thereby improving the accuracy, interpretability, and anti-confounding ability of identifying the causes of risk escalation.

[0140] In an exemplary embodiment, based on a set of key driving factors and causal effect values, counterfactual effects are extrapolated for at least two candidate actions targeting the target power grid to obtain the expected risk reduction effect, including steps 902 to 912. Wherein:

[0141] Step 902: Perform action controllability mapping on the driving factors in the key driving factor set to obtain the corresponding relationship data of factor disposal actions; Step 904: Decompose the data on the correspondence between factor treatment actions into interventionable sites to obtain a set of intervention sites. Step 906: Based on the intervention site set data, perform comparative learning counterfactual intervention coding on at least two candidate treatment actions to obtain candidate treatment action intervention vector data; Step 908: Based on the candidate intervention vector data and causal effect values, perform path-oriented causal message passing inference on the power supply service spatiotemporal knowledge graph data to obtain path-level risk response sequence data. Step 910: Perform particle-based counterfactual simulation aggregation on the path-level risk response sequence data to obtain risk change trajectory data; Step 912: Perform a multi-objective constraint assessment on the risk change trajectory data to obtain the expected risk reduction effect.

[0142] Among them, driving factors are key influencing variables that contribute causally to the increase in power supply service risk, such as equipment status, load fluctuation, response delay, and coverage.

[0143] Among them, action controllability mapping is a process of determining the driving factors according to whether they are direct / indirect / uncontrollable, and mapping them to a set of actions that can affect them.

[0144] Among them, the factor-action correspondence data is a structured correspondence data describing which actions can intervene in each driving factor, the direction of intervention, and the intervention boundaries (magnitude / cost / delay, etc.).

[0145] Intervention site decomposition is a process that breaks down the role of a treatment action in the graph into specific actionable locations and objects (nodes / edges / events).

[0146] The intervention site set data consists of the set of interventionable locations and their constraint information obtained from site decomposition, including site identifiers, range of action, effective time window, and amplitude boundaries.

[0147] Among them, contrastive learning counterfactual intervention coding is a coding process that uses contrastive learning of positive and negative sample pairs to encode candidate actions into counterfactual intervention vectors, making the representation of effective actions more aggregated and the representation of ineffective actions more separated.

[0148] Among them, the candidate treatment action intervention vector data is an intervention vector representation formed by at least two candidate treatment actions, which includes intervention site, intervention direction, intervention magnitude and timing information.

[0149] Among them, path-oriented causal message transmission inference involves the targeted dissemination and updating of causal impact messages along the risk propagation direction on the knowledge graph, in order to infer the process of how the response actions affect the evolution of the risk.

[0150] Among them, the path-level risk response sequence data is a time-varying risk response sequence derived from causal message passing, which depicts the dynamic changes of risk along the propagation path within the prediction window.

[0151] Among them, particle-based counterfactual simulation aggregation is a simulation method that generates multiple counterfactual risk evolution trajectories through multiple random samplings and performs statistical aggregation to evaluate robustness under uncertainty.

[0152] Among them, the risk change trajectory data is the risk change trajectory over time and its statistical characterization (mean, quantiles, worst case and dispersion, etc.) output by particle simulation aggregation.

[0153] Among them, multi-objective constraint assessment is an evaluation process that comprehensively weighs the benefits of risk reduction against constraints such as costs, resource consumption, time limits, and compliance, and outputs an evaluation of the effectiveness of the actions.

[0154] Specifically, each driving factor in the set of key driving factors is classified as directly controllable, indirectly controllable, and uncontrollable. Factor-action mapping rules are established by combining the available resources and business measures of the target power grid (such as pre-positioning of emergency repair resources, power transfer / uninterrupted operation, maintenance window adjustment, proactive notification and segmented outreach, customer-side load management, service scripts and channel strategies, etc.). In actual operation, each action is abstracted into a control operator that affects the factor's value or distribution (e.g., reducing "response delay," increasing "power transfer success rate," increasing "reach coverage rate," and reducing "fault incidence rate," etc.). The direction of action, upper limit of effect intensity, execution cost, and delay constraints are recorded, ultimately forming factor-action correspondence data.

[0155] The data on the correspondence between factor handling actions is decomposed into "intervention sites." This means breaking down the impact of each action in the data on the correspondence between factor handling actions into its specific position and target on the spatiotemporal knowledge graph of power supply services. Intervention sites include at least: the node site where the target factor is located (such as a device node in a certain transformer area, a customer group node, or a resource node), the associated edge site (such as a propagation strength edge, an adjacency weight edge, or a service affiliation edge), and the event site (such as the work order arrival process, the emergency repair dispatch process, or the proactive notification event). At the same time, the scope and granularity of the intervention sites are limited (e.g., limited to a specific transformer area, a specific time period, or a specific customer group), and the controllable boundaries of the action (maximum amplitude, minimum effective unit, and executable delay) are bound to the site to obtain the set of intervention site data.

[0156] Based on the set of intervention sites, each candidate treatment action is represented as an intervention vector applied to multiple intervention sites (including components such as site ID, intervention direction, intervention amplitude, and temporal effective window). To avoid incomparability or semantic drift between different action codes, a contrastive learning mechanism is introduced. This involves constructing positive sample pairs for "actions-scenarios that have historically significantly reduced risk" and negative sample pairs for "actions-scenarios that have poor effects or have increased risk." Contrastive loss ensures that similar effective actions are closer in the representation space, while ineffective actions are further removed. Simultaneously, site consistency constraints (different amplitudes at the same site should form a monotonic geometric relationship) and cross-scenario alignment constraints (the same action should be consistently encoded in similar scenarios) can be added to obtain candidate treatment action intervention vector data.

[0157] Candidate action intervention vector data is used as the "do" intervention input for key driving factors and their propagation edges into the power supply service spatiotemporal knowledge graph data inference engine, enabling targeted causal message transmission along the risk propagation path. Specifically, using the intervention points of the candidate action intervention vector data as the source, "risk impact messages" are iteratively propagated in the power supply service spatiotemporal knowledge graph data according to the propagation direction (upstream → downstream, disturbance source → recipient). At each propagation, the message strength is updated based on the corresponding causal effect value, and the temporal state of nodes / edges (e.g., load periods, resource response delays, event arrival processes) is integrated to form a time-based recursion. Simultaneously, paths irrelevant to the task or gated are attenuated or truncated to ensure the inference focuses on key transmission channels, ultimately outputting path-level risk response sequence data that evolves over time.

[0158] Uncertainties in path-level risk response sequence data (such as weather disturbances, randomness of work order arrival, equipment status noise, fluctuations in contact achievement success rate, and fluctuations in arrival delay) are modeled as random variables. Multiple "particle trajectories" are generated using particle filtering / Monte Carlo methods, with each particle trajectory corresponding to a risk response sequence sampled under a set of uncertainties. All particle trajectories are aggregated and statistically analyzed to output risk change trajectory data. This risk change trajectory data includes at least the risk mean trajectory, quantile trajectories (such as P10 / P50 / P90), worst-case trajectory, and trajectory dispersion, characterizing the dynamic evolution and robustness differences of candidate actions under uncertainty conditions.

[0159] The evaluation incorporates both "risk reduction" and "feasibility" from risk change trajectory data. Risk objectives include the magnitude of risk reduction within ΔT, the probability of achieving the target, and the scope of spillover mitigation. Constraint objectives include handling costs, resource consumption, execution delay, number of affected users, and compliance and security constraints. At least two candidate actions are compared using weighted scoring, constraint optimization, or Pareto ranking to calculate the expected risk reduction effect relative to the baseline scenario. Corresponding feasibility scores and priority rankings are output, ultimately yielding the expected risk reduction effect.

[0160] In this embodiment, by mapping the controllability of key driving factors and establishing the factor-action correspondence, the decision-making process can be transformed from experience-based matching into a structured, computable problem of "factor controllability - action attainability," avoiding blind intervention in uncontrollable factors. Further decomposition of the correspondence into intervention points allows abstract actions to be mapped to specific nodes / edges / events within the knowledge graph, improving the feasibility and accuracy of the deduction. Based on this, comparative learning-based counterfactual intervention encoding is used to generate action intervention vectors, enabling comparability of different actions within a unified representation space and strengthening... "Effective actions" indicate that the reliability of comparing response plans is improved; combining causal effect values ​​with path-oriented causal message transmission simulation can quantify the link-level impact of actions on risk evolution along the risk propagation channel, enhancing interpretability and reducing interference from irrelevant information; furthermore, by aggregating particle-based counterfactual simulation, the trajectory of risk change is obtained, which can assess the robust range of response effects under uncertainty conditions and avoid the randomness of single-point simulation; finally, multi-objective constraint assessment outputs the expected effect of risk reduction, which can achieve a comprehensive trade-off between risk and benefit and cost, time, and resource constraints, improving the optimality, robustness, and feasibility of response action selection.

[0161] Based on the same inventive concept, this application also provides a power supply service digital map linkage monitoring, control, and early warning device for implementing the aforementioned power supply service digital map linkage monitoring, control, and early warning method. For example... Figure 3 As shown, it includes: a spatiotemporal anchor module, a knowledge fusion module, a risk prediction module, and a result deduction module. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more power supply service digital map linkage monitoring, control and early warning device embodiments provided below can be found in the limitations of a power supply service digital map linkage monitoring, control and early warning method described above, and will not be repeated here.

[0162] The various modules in the aforementioned power supply service digital-map linkage monitoring, control, and early warning device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0163] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. This computer device includes a processor, memory, input / output interfaces (I / O), and communication interfaces. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0164] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0165] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0166] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0167] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0168] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0169] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0170] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for monitoring, controlling, and issuing early warnings based on a digital map of power supply services, characterized in that, The method includes: Spatiotemporal anchoring processing is performed on the multi-source heterogeneous power supply service data corresponding to the target power grid to obtain spatiotemporal anchor event data; Based on the spatiotemporal anchor point event data, spatiotemporal knowledge fusion is performed on the multi-source correlation of the target power grid to obtain spatiotemporal knowledge graph data of power supply service; Based on the power supply service spatiotemporal knowledge graph data, multi-task risk prediction is performed on the power supply service risk of the target power grid within a preset prediction time window to obtain dynamic early warning triggering results. Based on the dynamic early warning triggering results, a risk propagation path backtracking analysis is performed on the power supply service spatiotemporal knowledge graph data to obtain a set of candidate driving factor paths; The set of candidate driving factor paths is fitted with a structural causal model constraint to obtain the set of key driving factors and causal effect values. Based on the set of key driving factors and the causal effect value, counterfactual effects are extrapolated for at least two candidate actions for the target power grid to obtain the expected risk reduction effect. The expected effect of risk reduction is transformed into power grid early warning rules to obtain interpretable early warning results; the interpretable early warning results are used to orchestrate the behavior of the target power grid.

2. The method according to claim 1, characterized in that, The step of performing multi-task risk prediction on the power supply service risk of the target power grid within a preset prediction time window based on the power supply service spatiotemporal knowledge graph data, and obtaining dynamic early warning triggering results, includes: Based on the power supply service spatiotemporal knowledge graph data, the historical risk result data and risk tolerance parameters of the target power grid are quantized to obtain dynamic early warning triggering gating parameters. Multi-task risk prediction is performed on the power supply service spatiotemporal knowledge graph data to obtain the risk prediction results within the preset prediction time window; Based on the risk prediction results and the dynamic early warning triggering gating parameters, the power supply service spatiotemporal knowledge graph data is subjected to counterfactual perturbation inference to obtain the dynamic early warning triggering results.

3. The method according to claim 2, characterized in that, The step of performing multi-task risk prediction on the spatiotemporal knowledge graph data of the power supply service to obtain the risk prediction results within the preset prediction time window includes: Self-supervised representation learning is performed on the spatiotemporal knowledge graph data of the power supply service to obtain general spatiotemporal graph representation data; Based on the general representation data of the spatiotemporal graph, task conditional subgraph retrieval is performed on each risk task within the preset prediction time window to obtain the targeted propagation subgraph representation data. Multi-task spatiotemporal joint prediction is performed on the directional propagation subgraph representation data to obtain the multi-task risk prediction value within the preset prediction time window; The multi-task risk prediction values ​​are fused and corrected based on statistical features to obtain the risk prediction results within the preset prediction time window.

4. The method according to claim 3, characterized in that, The step of performing multi-task spatiotemporal joint prediction on the directed propagation subgraph representation data to obtain the multi-task risk prediction value within the preset prediction time window includes: Task causal gating structure analysis is performed on the task identification data and propagation direction data in the directional propagation subgraph representation data to obtain task coupling gating graph data. Based on the task coupling gating graph data, gating graph diffusion encoding is performed on the directed propagation subgraph representation data to obtain task coupling diffusion representation data. The task coupling diffusion characterization data is folded into multi-scale time windows to obtain multi-scale spatiotemporal sequence characterization data. The multi-scale spatiotemporal sequence characterization data is subjected to confidence-modulated hybrid decoding to obtain candidate multi-task risk prediction values; Supervised parameterization mapping is performed on the candidate multi-task risk prediction values ​​to obtain the multi-task risk prediction values ​​within the preset prediction time window.

5. The method according to claim 2, characterized in that, The step of performing counterfactual perturbation deduction on the power supply service spatiotemporal knowledge graph data based on the risk prediction results and the dynamic early warning triggering gating parameters to obtain the dynamic early warning triggering results includes: Based on the risk prediction results and the dynamic early warning triggering gating parameters, risk propagation path is extracted from the power supply service spatiotemporal knowledge graph data to obtain targeted intervention subgraph data. The minimum counterfactual perturbation solution is applied to the targeted intervention subgraph data to obtain the set of perturbation variables and perturbation amplitude data; Based on the set of disturbance variables and the disturbance amplitude data, causal intervention sampling and inference are performed on the power supply service spatiotemporal knowledge graph data to obtain multi-counterfact scenario graph data; Risk envelope estimation is performed on the multi-counterfact scenario map data to obtain robust risk prediction distribution data; The robust risk prediction distribution data and the dynamic early warning triggering gating parameters are subjected to gating discrimination processing to obtain the dynamic early warning triggering result.

6. The method according to claim 1, characterized in that, The step of performing structural causal model constraint fitting on the candidate driver path set to obtain the key driver set and causal effect values ​​includes: The candidate driving factor path set is mapped by path element variableization to obtain the candidate factor variable set; Based on the spatiotemporal knowledge graph data of the power supply service, the set of candidate factor variables is subjected to relational semantic constraints to obtain a set of structural causal constraints. The candidate factor variable set is subjected to multi-environment invariance screening to obtain the core candidate factor set; Based on the set of structural causal constraints, a constraint structure search is performed on the set of core candidate factors to obtain the structure of the structural causal model. The constraint parameters of the structural causal model are estimated to obtain the structural causal model parameters. The causal effect is calculated based on the parameters of the structural causal model to obtain the set of key driving factors and the causal effect value.

7. The method according to claim 1, characterized in that, The step involves performing counterfactual effect deductions on at least two candidate actions for the target power grid based on the set of key driving factors and the causal effect values ​​to obtain the expected risk reduction effect, including: The action controllability mapping is performed on the driving factors in the set of key driving factors to obtain the corresponding relationship data of factor disposal actions; The data on the correspondence between the factor treatment actions are decomposed into intervention sites to obtain a set of intervention sites. Based on the intervention site set data, comparative learning counterfactual intervention coding is performed on at least two candidate treatment actions to obtain candidate treatment action intervention vector data; Based on the candidate intervention vector data and the causal effect value, path-oriented causal message passing inference is performed on the power supply service spatiotemporal knowledge graph data to obtain path-level risk response sequence data. The path-level risk response sequence data is aggregated using particle-based counterfactual simulation to obtain risk change trajectory data; The risk change trajectory data is evaluated under multi-objective constraints to obtain the expected risk reduction effect.

8. A power supply service digital map linkage monitoring, control and early warning device, characterized in that, The device includes: The spatiotemporal anchor module is used to perform spatiotemporal anchoring processing on the multi-source heterogeneous power supply service data corresponding to the target power grid to obtain spatiotemporal anchor event data. The knowledge fusion module is used to perform spatiotemporal knowledge fusion on the multi-source correlation of the target power grid based on the spatiotemporal anchor point event data to obtain power supply service spatiotemporal knowledge graph data; The risk prediction module is used to perform multi-task risk prediction on the power supply service risk of the target power grid within a preset prediction time window based on the power supply service spatiotemporal knowledge graph data, and obtain dynamic early warning triggering results. The result inference module is used to perform risk propagation path backtracking analysis on the power supply service spatiotemporal knowledge graph data based on the dynamic early warning triggering result, and obtain a set of candidate driving factor paths; The set of candidate driving factor paths is fitted with a structural causal model constraint to obtain the set of key driving factors and causal effect values. Based on the set of key driving factors and the causal effect value, counterfactual effects are extrapolated for at least two candidate actions for the target power grid to obtain the expected risk reduction effect. The expected effect of risk reduction is transformed into power grid early warning rules to obtain interpretable early warning results; the interpretable early warning results are used to orchestrate the behavior of the target power grid.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

Images