A power asset management method and system based on an event-device association model

Abstract

This invention relates to the field of power asset management technology, and in particular to a power asset management method and system based on an event-equipment correlation model. It collects multi-source heterogeneous data from the power asset operating environment, generates a standardized event-equipment interaction dataset, constructs a dynamic causal correlation graph of events and equipment, employs a temporal causal hypergraph attention network combined with counterfactual interference minutiae, outputs a pure causal correlation weight matrix, and generates an event-equipment correlation model. This model deduces the equipment performance degradation trajectory and remaining useful life probability distribution under target event-driven conditions, quantifies the marginal benefits and risk exposure values ​​of assets at different operation and maintenance intervention nodes, solves for the optimal asset scheduling sequence through deep reinforcement learning, generates power asset management decision schemes, and achieves collaborative closed-loop evolution of the model and decisions based on online meta-learning. This invention effectively improves the intelligence level of power asset management and ensures the safe, stable, and economical operation of the power grid.

Description

Technical Field

[0001] This invention relates to the field of power asset management technology, specifically to a power asset management method and system based on an event-device association model. Background Technology

[0002] Equipment monitoring data, power grid operation event data, environmental meteorological data, and operation and maintenance history data generated during the operation of power assets are scattered across different business systems, lacking a unified data exchange standard between these systems. The spatiotemporal references of data from different sources are inconsistent, and unstructured power grid operation event logs cannot be effectively correlated with structured equipment monitoring data. Furthermore, a large amount of measurement noise and outliers introduced during data acquisition are difficult to effectively remove, resulting in inconsistent data quality and failing to provide reliable data support for subsequent analysis and modeling.

[0003] Existing technologies mostly employ correlation-based analysis methods to uncover the relationships between events and equipment. These methods can only discover statistical correlations between variables and cannot reveal the true driving mechanisms of events on equipment state changes. Confounding variables such as environmental and meteorological changes and power grid load fluctuations can generate a large number of spurious correlations, severely affecting the accuracy of correlation strength calculations. Furthermore, traditional binary graph models can only characterize the impact of a single event on a single device and cannot effectively depict the combined effects of multiple events or the multi-hop propagation effects of event influences along the power grid topology.

[0004] Existing methods for predicting the remaining useful life of equipment mostly rely on statistical models built from the equipment's historical aging data, neglecting the driving effect of external dynamic factors such as power grid operation events and environmental anomalies on equipment performance degradation. They cannot respond in real time to the impact of sudden power grid events on equipment status, nor can they accurately predict the performance degradation trajectory of equipment under the superposition of multiple events. The predicted results are mostly single lifetime values, lacking probability distribution and confidence interval information, making it difficult to quantify the uncertainty of the predicted results and failing to provide effective quantitative basis for operation and maintenance decisions.

[0005] Traditional power asset management often employs a model of periodic maintenance or reactive emergency repairs, which can easily lead to both over-maintenance and under-maintenance. Over-maintenance increases unnecessary operation and maintenance costs and may even introduce new equipment failures during maintenance. Under-maintenance, on the other hand, increases the risk of sudden equipment failures, causing widespread power outages. Furthermore, existing decision-making methods struggle to coordinate and optimize multiple decision dimensions, such as maintenance timing, spare parts allocation, and decommissioning / replacement, failing to achieve globally optimal allocation of operation and maintenance resources while meeting grid security constraints.

[0006] To address the aforementioned issues, there is an urgent need to propose a power asset management method and system that can achieve deep fusion of multi-source data, accurate causal relationship modeling, accurate equipment status prediction, global optimization of operation and maintenance decisions, and autonomous closed-loop evolution of the system, so as to comprehensively improve the overall management efficiency of power assets. Summary of the Invention

[0007] The purpose of this invention is to provide a power asset management method and system based on an event-device association model to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A power asset management method based on an event-device association model includes the following steps: S1: Collect multi-source heterogeneous data from the power asset operation environment. The multi-source heterogeneous data includes equipment monitoring time series data, power grid operation event logs, environmental meteorological data, and operation and maintenance history data. Based on a unified spatiotemporal benchmark, perform cross-modal alignment and feature denoising on the multi-source heterogeneous data to generate a standardized event-equipment interaction dataset containing timestamps, spatial topological coordinates, event semantic labels, and equipment operating parameters. S2: Based on a standardized event-device interaction dataset, event nodes and device entity nodes are extracted. Using the power grid physical topology as the framework, the event triggering timing and device state response lag relationship are introduced to construct a dynamic causal association graph of events and devices. A temporal causal hypergraph attention network is used to calculate the association strength of the dynamic causal association graph of events and devices. Specifically, this includes: aggregating the composite impact of multi-hop events on target devices through a multi-head hypergraph attention mechanism, and embedding the counterfactual interference quantifier of the structural causal model to remove environmental confounding variables, outputting a pure causal association weight matrix, and thus generating an event-device association model. S3: Input the real-time monitored target event stream into the event-device association model, couple the state space based on the causal association weight matrix and the initial health baseline of the device, and deduce the device performance degradation trajectory and remaining useful life probability distribution under the target event-driven model; map the probability distribution to the asset life cycle cost model to quantify the asset marginal benefits and risk exposure values ​​under different operation and maintenance intervention nodes; S4: With the goal of maximizing the overall efficiency of assets, and with risk exposure, asset marginal benefits, and grid security constraints as boundary conditions, an optimization space for power asset operation and maintenance strategies is constructed. A deep reinforcement learning strategy network is adopted, with a weighted combination of risk exposure and asset marginal benefits as the state reward signal. The optimal asset scheduling sequence is solved through strategy gradient iteration, generating a power asset management decision scheme that includes maintenance timing, spare parts allocation, and decommissioning and replacement plans. S5: Execute power asset management decision-making schemes and collect equipment status response data and economic indicator feedback after scheme execution to construct a strategy-effect comparison library; based on the strategy-effect comparison library, dynamically fine-tune the graph topology and weight parameters of the temporal causal hypergraph attention network through an online meta-learning mechanism to achieve the collaborative closed-loop evolution of the event-equipment association model and asset management decision-making schemes.

[0009] A power asset management system based on an event-device association model, wherein the system executes the above method, and the system includes: The data alignment and noise reduction module is used to collect multi-source heterogeneous data in the power asset operation environment. Based on a unified spatiotemporal benchmark, it performs cross-modal alignment and feature noise reduction on the multi-source heterogeneous data to generate a standardized event-device interaction dataset containing timestamps, spatial topological coordinates, event semantic labels and equipment operating parameters. The dynamic causal association modeling module, connected to the data alignment and noise reduction module, receives a standardized event-device interaction dataset, extracts event nodes and device entity nodes, and constructs a dynamic causal association graph of events and devices by introducing the event triggering time sequence and device state response lag relationship with the power grid physical topology as the skeleton. It then uses a temporal causal hypergraph attention network to calculate the association strength of the event-device dynamic causal association graph, aggregates the composite impact of multi-hop events on the target device through a multi-head hypergraph attention mechanism, and embeds the counterfactual interference quantifier of the structural causal model to remove environmental confounding variables, outputting a clean causal association weight matrix, and thus generating an event-device association model. The state deduction and benefit quantification module is connected to the dynamic causal relationship modeling module. It is used to receive the event-equipment relationship model, input the real-time monitored target event stream into the event-equipment relationship model, couple the state space based on the causal relationship weight matrix and the initial health baseline of the equipment, deduce the equipment performance degradation trajectory and the probability distribution of the remaining useful life under the target event, and map the probability distribution of the remaining useful life to the asset life cycle cost model to quantify the marginal benefits and dynamic risk exposure of the asset under different operation and maintenance intervention nodes. The strategy optimization and decision generation module is connected to the state deduction and benefit quantification module. It receives the marginal benefit of assets and the dynamic risk exposure value. Taking the maximization of comprehensive asset efficiency as the objective function, and using the dynamic risk exposure value, the marginal benefit of assets and the grid security constraints as boundary conditions, it constructs the power asset operation and maintenance strategy optimization space. It adopts a deep reinforcement learning strategy network to use the weighted combination of dynamic risk exposure value and asset marginal benefit as the state reward signal to solve the optimal asset scheduling sequence through strategy gradient iteration, and generates a power asset management decision scheme that includes maintenance timing, spare parts allocation and decommissioning replacement plan. The execution feedback and collaborative evolution module is connected to the strategy optimization and decision generation module for issuing and executing power asset management decision schemes, and collecting equipment status response data and economic indicator feedback after scheme execution to build a strategy-effect comparison library. Based on the strategy-effect comparison library, the graph topology and weight parameters of the temporal causal hypergraph attention network are dynamically fine-tuned through an online meta-learning mechanism, and the fine-tuned graph topology and weight parameters are fed back to the dynamic causal association modeling module to realize the collaborative closed-loop evolution of the event-equipment association model and the power asset management decision scheme.

[0010] As can be seen from the technical solution provided by the present invention above, the power asset management method and system based on the event-device association model provided by the present invention has the following beneficial effects: This invention establishes a cross-modal alignment and feature denoising system for multi-source heterogeneous data based on a unified spatiotemporal benchmark, achieving deep integration of equipment monitoring data, power grid operation event data, environmental meteorological data, and operation and maintenance history data. It solves the spatiotemporal inconsistency problem of data from different sources by using spatial coordinate projection transformation and time axis resampling alignment, achieves semantic structuring of unstructured event logs through a power industry professional knowledge graph, and effectively eliminates measurement noise and outliers through adaptive variational mode decomposition combined with the isolated forest algorithm, generating a standardized event-equipment interaction dataset with a unified structure and complete information, providing a high-quality data foundation for subsequent causal modeling and state inference. This invention proposes an event-device association modeling method based on temporal causal hypergraph attention network, which overcomes the technical bottleneck of traditional correlation analysis methods that cannot distinguish between causal associations and spurious associations. It constructs a dynamic causal association graph of events and devices with the physical topology of the power grid as the skeleton, transforms the joint effect of multiple events into hyperedge features through topological dimensionality enhancement, aggregates the composite influence of multi-hop events using a multi-head hypergraph attention mechanism, and embeds counterfactual interference quantifiers of structural causal model to remove environmental confounding variables, outputting a pure causal association weight matrix that only reflects the true driving effect of events, which significantly improves the accuracy and reliability of the assessment of the degree of impact of events on devices. This invention enables accurate extrapolation of event-driven equipment performance degradation trajectories and remaining useful life, quantifying the economic benefits and safety risks of different maintenance intervention nodes. A nonlinear state transition model is constructed based on causal correlation weights and the initial health baseline of the equipment. A rolling prediction time window mechanism is used to output the continuous time-domain trend of equipment performance changes. An adaptive kernel density estimation is used to generate a probability distribution of remaining useful life across multiple confidence intervals. Based on this, an asset lifecycle cost model is constructed to quantify the marginal benefits and dynamic risk exposure of assets when performing maintenance interventions at different time points, providing a comprehensive and quantitative assessment basis for maintenance decisions. This invention constructs a global optimization system for power asset operation and maintenance strategies under multiple constraints, realizing the coordinated optimization of maintenance timing, spare parts allocation, and decommissioning replacement. With maximizing overall asset efficiency as the objective function, it comprehensively considers risk exposure, asset benefits, and grid security constraints to construct an optimization space for operation and maintenance strategies. A deep reinforcement learning strategy network driven by both risk and benefit is used to solve for the optimal asset scheduling sequence. This method can achieve optimal allocation of operation and maintenance resources while meeting the requirements for safe grid operation, effectively reducing the total lifecycle operation and maintenance cost and improving equipment availability and regional grid power supply reliability. This invention establishes a closed-loop evolution mechanism for model and decision collaboration based on online meta-learning, endowing the system with long-term autonomous optimization capabilities. By collecting equipment status responses and economic indicator feedback after the execution of decision schemes, a strategy-effect comparison library is constructed. The online meta-learning mechanism is used to dynamically fine-tune the topology and weight parameters of the temporal causal hypergraph attention network. As the system's operating time increases, the strategy-effect comparison library continuously accumulates rich sample data, and the model's prediction accuracy and the scientific nature of decision-making will continue to improve. It can adapt to the dynamic changes in the power grid operating environment and realize the continuous iterative upgrade of power asset management capabilities. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the steps in a power asset management method based on an event-device association model according to the present invention. Figure 2 This is a schematic diagram of the structure of a power asset management system based on an event-device association model according to the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0013] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0014] like Figure 1-2 As shown, this embodiment of the invention provides a power asset management method based on an event-device association model, including the following steps: S1: Collect multi-source heterogeneous data from the power asset operation environment. The multi-source heterogeneous data includes equipment monitoring time series data, power grid operation event logs, environmental meteorological data, and operation and maintenance history data. Based on a unified spatiotemporal benchmark, perform cross-modal alignment and feature denoising on the multi-source heterogeneous data to generate a standardized event-equipment interaction dataset containing timestamps, spatial topological coordinates, event semantic labels, and equipment operating parameters. S2: Based on a standardized event-device interaction dataset, event nodes and device entity nodes are extracted. Using the power grid physical topology as the framework, the event triggering timing and device state response lag relationship are introduced to construct a dynamic causal association graph of events and devices. A temporal causal hypergraph attention network is used to calculate the association strength of the dynamic causal association graph of events and devices. Specifically, this includes: aggregating the composite impact of multi-hop events on target devices through a multi-head hypergraph attention mechanism, and embedding the counterfactual interference quantifier of the structural causal model to remove environmental confounding variables, outputting a pure causal association weight matrix, and thus generating an event-device association model. S3: Input the real-time monitored target event stream into the event-device association model, couple the state space based on the causal association weight matrix and the initial health baseline of the device, and deduce the device performance degradation trajectory and remaining useful life probability distribution under the target event-driven model; map the probability distribution to the asset life cycle cost model to quantify the asset marginal benefits and risk exposure values ​​under different operation and maintenance intervention nodes; S4: With the goal of maximizing the overall efficiency of assets, and with risk exposure, asset marginal benefits, and grid security constraints as boundary conditions, an optimization space for power asset operation and maintenance strategies is constructed. A deep reinforcement learning strategy network is adopted, with a weighted combination of risk exposure and asset marginal benefits as the state reward signal. The optimal asset scheduling sequence is solved through strategy gradient iteration, generating a power asset management decision scheme that includes maintenance timing, spare parts allocation, and decommissioning and replacement plans. S5: Execute power asset management decision-making schemes and collect equipment status response data and economic indicator feedback after scheme execution to construct a strategy-effect comparison library; based on the strategy-effect comparison library, dynamically fine-tune the graph topology and weight parameters of the temporal causal hypergraph attention network through an online meta-learning mechanism to achieve the collaborative closed-loop evolution of the event-equipment association model and asset management decision-making schemes.

[0015] In this embodiment, the core function of step S1 is to comprehensively collect heterogeneous data from different sources and with different structures in the power asset operation environment, achieve spatiotemporal and semantic alignment of multimodal data based on a unified spatiotemporal benchmark, remove measurement noise and environmental interference through an adaptive feature denoising algorithm, and finally generate a standardized event-device interaction dataset with a unified structure and complete spatiotemporal semantic information, providing a high-quality data foundation for the subsequent construction of event-device association models; the detailed steps are as follows: Step S1-1: Parallel acquisition of multi-source heterogeneous data and construction of the original multi-source data pool: It connects to the IoT sensor network, dispatch automation system, micro-meteorological monitoring terminal, and production operation and maintenance database respectively, and collects equipment monitoring time-series data, power grid operation event logs, environmental meteorological data, and operation and maintenance history data in parallel. The equipment monitoring time-series data includes continuous sampling sequences of equipment operating status parameters such as voltage, current, temperature, vibration, and partial discharge. The power grid operation event log includes records of abnormal power grid operation events such as switch actions, protection trips, load overruns, and equipment alarms. The environmental meteorological data includes environmental parameters that affect equipment operating status such as temperature, humidity, wind speed, rainfall, and lightning intensity. The operation and maintenance history data includes full lifecycle operation and maintenance information such as equipment commissioning time, maintenance records, spare parts replacement records, and fault handling records. The collected data undergoes communication protocol parsing, physical dimension normalization, and linear interpolation for missing breakpoints. Communication protocol parsing converts proprietary protocol data from different devices and systems into structured data in a unified format. Physical dimension normalization maps physical parameters of different dimensions and magnitudes to a unified numerical range, eliminating the impact of dimensional differences on subsequent model training. Linear interpolation for missing breakpoints fills in missing values ​​caused by communication interruptions or sensor malfunctions during data acquisition; the interpolation formula is as follows: ,in, For missing moments interpolated value, The most recent moment before the missing point The measured value, The most recent moment after the missing point The measured value; The pre-processed data is integrated and stored to build a raw multi-source data pool; Step S1-2: Spatiotemporal alignment processing under a unified spatiotemporal reference: Using the power grid GIS geographic information grid and the high-precision atomic clock timing signal as a unified spatiotemporal reference, spatial coordinate projection transformation and time axis resampling alignment are performed on the original multi-source data pool; Spatial coordinate projection transformation is used to convert the physical installation location of equipment in different coordinate systems into unified power grid GIS geographic information grid coordinates, and then map them into node-level spatial topology coordinates; each power device corresponds to a unique node-level spatial topology coordinate, which is used to characterize the positional relationship of the device in the physical topology of the power grid; Time axis resampling alignment is used to resample non-equal interval sampling sequences with different sampling frequencies and time bases to a unified clock beat; the resampling process uses linear interpolation to ensure the temporal continuity and accuracy of the data after resampling; all data frames after resampling carry a unified high-precision timestamp to achieve accurate alignment of data from different sources in the time dimension; Data that has undergone spatial coordinate projection transformation and time axis resampling alignment is associated and matched with spatial topological coordinates according to timestamps to generate spatiotemporally aligned data frames; Steps S1-3: Construction of cross-modal semantic alignment and multimodal interaction feature tensors: Cross-modal semantic alignment is performed based on spatiotemporally aligned data frames to achieve semantic association between unstructured event logs and structured monitoring data; A professional knowledge graph in the power field is used to perform named entity recognition and causal relationship extraction on power grid operation event logs. Named entity recognition is used to extract key entity information such as equipment name, event type, event level, and trigger time from unstructured event text. Causal relationship extraction is used to identify the sequential triggering relationship and causal association between different events. The extracted entity information and causal relationships are structured and encoded to output structured event semantic tags. Environmental meteorological time-series slices and equipment operating parameters are matched and stitched together at the feature level using a dynamic time warping algorithm. The dynamic time warping algorithm is used to solve the time lag problem between changes in environmental parameters and equipment state response, achieving optimal alignment of time-series data of different modalities in the time dimension. The aligned environmental meteorological features and equipment operating features are then stitched together dimensionally to construct a multimodal interactive feature tensor. The multimodal interactive feature tensor contains both equipment operating state information and external environmental influence information, which can comprehensively characterize the equipment's operating conditions. Steps S1-4: Feature Denoising and Standardization of Event-Device Interaction Dataset Generation: Adaptive variational mode decomposition combined with the isolated forest algorithm is performed on the multimodal interaction feature tensor for feature denoising and outlier removal; Adaptive variational mode decomposition is used to decompose multimodal interactive feature tensors into multiple intrinsic mode functions, separating high-frequency measurement noise, environmental drift interference and equipment intrinsic operating condition fluctuations; by adaptively selecting the number of mode decomposition layers, it ensures that the equipment intrinsic operating condition fluctuation information is completely preserved, while effectively filtering out high-frequency noise and low-frequency drift; The Isolation Forest algorithm is used to identify and remove outliers in multimodal interaction feature tensors. By constructing multiple isolated trees, the Isolation Forest algorithm can quickly isolate anomalous samples and efficiently handle the anomaly detection problem in high-dimensional data. After removing outliers, the effective feature data that reflects the true operating status of the equipment is retained. The denoised timestamps, spatial topological coordinates, event semantic labels, and equipment operating parameters are serialized and encapsulated according to a preset associative data structure; each data sample contains a unique timestamp and spatial topological coordinates, as well as a corresponding event semantic label and equipment operating parameter vector; all encapsulated data samples are arranged in chronological order to generate a standardized event-device interaction dataset.

[0016] In this embodiment, the core function of step S2 is to extract event and device entity nodes from the standardized event-device interaction dataset, introduce the relationship between event triggering timing and device state response lag using the power grid physical topology as the framework, construct a causal association graph that reflects the dynamic impact of events on devices, aggregate the composite impact of multi-hop events through a temporal causal hypergraph attention network, and embed the counterfactual interference quantifier of the structural causal model to remove environmental confounding variables, outputting a pure causal association weight matrix that only reflects the true driving effect of events, and finally generating an event-device association model that can be used for device state inference; the detailed steps are as follows: Step S2-1: Extraction of event and device nodes and construction of static power grid physical topology skeleton: The standardized event-device interaction dataset is parsed, and power grid operation events with independent event semantic labels and trigger timestamps are mapped to event nodes. Each event node carries event type, trigger intensity, impact range and time sequence characteristics. Power equipment carrying unique spatial topological coordinates and operating parameters is mapped to equipment entity nodes. Each equipment entity node carries equipment model, years of operation, health baseline and real-time operating parameters. The system utilizes a power grid electrical connection database and a geographic information grid to spatially cluster equipment entity nodes based on voltage level, electrical connection distance, and physical adjacency. Equipment with the same voltage level and an electrical connection distance less than a preset threshold is grouped into the same electrical cluster, while equipment physically adjacent and belonging to the same power supply area is grouped into the same geographic cluster. A static power grid physical topology skeleton is constructed based on the hierarchical relationship between electrical and geographic clusters, where edges represent electrical connection and physical adjacency relationships between equipment. Event nodes are projected onto the corresponding topology skeleton neighborhood according to their spatial influence radius, ensuring that each event node only has potential associations with equipment entity nodes within its spatial influence range. Step S2-2: Generation of temporal causal edges and construction of event-device dynamic causal relationship graph: Using the trigger timestamp of the event node as a reference, a preset observation window is slid backward along the time axis to calculate the deviation of the operating parameters of the equipment entity node within the window from the historical baseline and the response delay time; the formula for calculating the deviation of the operating parameters is: ,in, The deviation of operating parameters, For the dimensions of the operating condition parameters involved in the calculation, For the first Real-time values ​​of each operating parameter For the first Historical baseline mean of each operating parameter For the first Historical baseline standard deviation of each operating condition parameter; If the deviation exceeds the preset threshold and the response delay time falls within the device's inertial response range, a directed temporal causal edge is established between the corresponding event node and the device entity node. The direction of the directed temporal causal edge is from the event node to the device entity node, and the edge attributes include the event trigger strength, the device response hysteresis step size, and the spatial attenuation coefficient. The spatial attenuation coefficient decreases linearly as the electrical distance between the event node and the device entity node increases. A time-slicing sliding mechanism is used to stack and concatenate continuously generated directed temporal causal edges according to discrete time steps to form a multi-time-segment state snapshot sequence; the frequency and confidence scores of repeated causal edges in the snapshot sequence are statistically analyzed, and the confidence score is calculated using the following formula: ,in, For the causal edge confidence level, This represents the number of times the causal edge appears in all snapshots. This represents the total number of snapshot sequences. Occasional noise edges with confidence levels below a preset threshold are removed, while high-frequency stable causal links are retained. Finally, the static power grid physical topology skeleton and dynamic temporal causal edges are merged to generate an event-device dynamic causal relationship graph. Step S2-3: Topological dimensionality upgrade from binary graph to hypergraph: Multi-node interaction paths in the event-device dynamic causal relationship graph are identified as potential hyperedges; multi-hop backtracking is performed along the directed temporal causal edge with the target device entity node as the center, and event node clusters that are co-triggered and semantically related are merged into hypergraph hyperedges; each hyperedge connects a target device entity node and multiple event nodes with synergistic influence, which can characterize the joint effect of multiple events on the same device. The joint temporal distribution features and spatial topological adjacency matrix of event clusters are bound to each hyperedge. The joint temporal distribution features are used to describe the triggering temporal relationship and intensity superposition law of each event in the event cluster. The spatial topological adjacency matrix is ​​used to describe the spatial positional relationship and mutual influence degree between event nodes in the event cluster. This completes the topological dimensionality upgrade from binary graph to hypergraph and generates a temporal causal hypergraph. Step S2-4: Aggregation of the multi-dimensional composite effects of multi-head hypergraph attention mechanisms: A multi-head hypergraph attention network is initialized, and the hypergraph and hyperedge features and device node features are mapped to multiple independent semantic subspaces respectively. Each semantic subspace corresponds to an attention head, and different attention heads focus on different influencing factors such as electrical coupling, environmental stress and operation and maintenance history. Calculate the feature similarity between event nodes within hyperedges and target device nodes in each subspace to generate a spatial-temporal joint attention distribution; the attention distribution calculation formula is: ,in, For the first The attention weight of each event node to the target device node. Let W be the attention vector and W be the feature mapping matrix. For the first Feature vectors of each event node For the feature vector of the target device node, For feature splicing operations, This represents the number of event nodes contained within the hyperedge. It is a linear unit activation function with leakage correction, used to introduce nonlinear transformation, solve the problem of gradient vanishing in the negative region of the standard ReLU function, and improve the network's ability to capture weakly correlated features; By fusing the outputs of each head through a weighted summation mechanism, multi-dimensional composite influences are captured, and an initial associated feature vector is generated. Step S2-5: Removal of confounding variables and generation of pure causal association weight matrix in structural causal model: A structural causal model interference identification module is constructed, and environmental meteorological data and power grid load fluctuations that affect equipment status are extracted as a set of confounding variables. Confounding variables refer to variables that simultaneously affect event triggering and equipment status changes, which may lead to false causal associations. Counterfactual intervention is performed based on the initial association feature vector. The state variables of the target device node are fixed, and a virtual zeroing operation is applied to the set of confusion variables to generate a counterfactual state sequence. The difference in feature responses between the observed state sequence and the counterfactual state sequence is compared, and the spurious association bias caused by the confusion variables is calculated. The formula for calculating the spurious association bias is as follows: ,in, This is a spurious association bias. The time step of the state sequence. for The observed state feature vector at time t. for The counterfactual state feature vector at time step; The spurious association bias is removed from the initial association feature vector to obtain the deconfused causal effect features; the causal effect features are normalized and smoothed across time steps to eliminate time-series fluctuation noise; the output is a pure causal association weight matrix that only reflects the true driving force of the event. Step S2-6: Event-device association model solidification: The pure causal association weight matrix is ​​parameter-bound to the event-device dynamic causal association graph, so that the weight of each causal edge in the graph corresponds to an element in the pure causal association weight matrix; the inference rules of the temporal causal hypergraph attention network are solidified, including hyperedge construction rules, attention calculation rules, and counterfactual intervention rules; finally, an event-device association model containing a complete topology and causal association parameters is generated. This model can quickly retrieve the corresponding causal association weights based on the input event stream, and achieve accurate prediction of device state changes.

[0017] In this embodiment, the core function of step S3 is to input the real-time monitored target event stream into the trained event-device association model, construct an event-driven state transition system based on the pure causal association weights and the initial health baseline of the equipment, accurately deduce the equipment performance degradation trajectory and the probability distribution of remaining useful life, and then map the life prediction results to the asset lifecycle cost model to quantify the marginal benefits and dynamic risk exposure of the asset when performing operation and maintenance interventions at different time points, providing a quantitative decision-making basis for subsequent operation and maintenance strategy optimization; the detailed steps are as follows: Step S3-1: Real-time target event stream feature extraction and causal association weight matching: The system receives real-time monitored target event streams and analyzes the event type, trigger intensity, duration, and time sequence characteristics frame by frame. Event types include four main categories: power grid fault events, environmental anomaly events, load fluctuation events, and operation and maintenance events. Trigger intensity uses a normalized value to characterize the potential impact of the event on the equipment, with a value range of 0 to 1. Time sequence characteristics are used to describe the sequential triggering order and time interval relationship of multiple events. The extracted event features are input into the event-device association model. Based on the event type, spatial impact range and trigger timestamp, the corresponding causal association weight matrix fragment is retrieved. The retrieval process adopts a multi-level indexing mechanism. First, the power grid topology region where the target device is located is located by spatial topological coordinates. Then, the corresponding event cluster is matched by event semantic tags. Finally, the exclusive causal association weight matrix fragment of the event cluster to the target device is extracted. Step S3-2: Construction of the multi-dimensional health status space of the equipment: The system calls the initial health baseline library of the equipment and parses the multi-dimensional health indicator reference vector of the target equipment under standard commissioning conditions. The initial health baseline library of the equipment stores the benchmark performance parameters of various power equipment at the time of factory acceptance and first commissioning, covering three major categories of core indicators: electrical performance, mechanical performance and thermal performance. The equipment state space is constructed based on a multidimensional health indicator reference vector. The equipment state space is a three-dimensional orthogonal vector space, with the three dimensions being electrical insulation aging degree, mechanical fatigue accumulation degree, and thermal stress sensitivity. Electrical insulation aging degree characterizes the degree of deterioration of the equipment's insulating medium, mechanical fatigue accumulation degree characterizes the degree of wear and tear on the equipment's rotating parts and structural components, and thermal stress sensitivity characterizes the equipment's tolerance to temperature changes and thermal shock. The value of each dimension ranges from 0 to 1, with larger values ​​indicating a worse health condition in the corresponding aspect of the equipment. Step S3-3: Establishment of event-driven nonlinear state transition model: The matched causal association weight matrix fragments are used as state space driving factors and coupled with the device state space in a multidimensional feature tensor. The tensor coupling process uses the Hadamard product operation to multiply each element in the causal association weight matrix with the state value of the corresponding dimension in the device state space, thereby realizing the differentiated driving of events on different health dimensions of the device. An event-driven nonlinear state transition model is established based on the coupled feature tensor, and the state transition equation is: ,in, for The device state vector at time t, for Causal correlation weight matrix at time step, For Hadamard product operations, for The operating condition drift noise vector at any given time; The rolling prediction time window mechanism is used to iteratively solve the state transition model frame by frame along the future time domain. The length of the rolling prediction time window is set to 720 hours and the sliding step size is set to 24 hours. Each time step is slid, the causal correlation weight matrix and the device state vector are updated once. The degradation accumulation effect and operating condition drift compensation under the time sequence superposition of multi-source events are superimposed to output the device performance degradation trajectory in the continuous time domain. Step S3-4: Generation of the remaining useful lifetime probability distribution: A preset critical threshold envelope for equipment failure is defined. This envelope is a closed surface in three-dimensional space, corresponding to the combination of failure thresholds in the three dimensions of the equipment state space. When the equipment state vector exceeds this envelope, the equipment is deemed to have experienced a functional failure. The time-domain penetration point search is performed between the equipment performance degradation trajectory and the failure critical threshold envelope; the distance between the equipment state vector and the failure critical threshold envelope is calculated point by point along the time axis, and when the distance is less than or equal to 0, the time point is recorded as the failure penetration point; since the equipment performance degradation is random, multiple iterations of solving the state transition model will yield multiple different failure penetration points, forming a failure penetration point sequence; Based on the failure penetration point sequence, an adaptive kernel density estimation algorithm is used to fit the time probability density, generating the remaining useful lifetime probability distribution; the adaptive kernel density estimation formula is: ,in, Let be the probability density function of the remaining useful lifetime. The number of failed penetration points. For adaptive bandwidth, For Gaussian kernel function, For the first The time value of each failure penetration point; The remaining useful life intervals corresponding to different confidence intervals are calculated based on the probability density function, and a remaining useful life probability distribution containing 90%, 95% and 99% confidence intervals is generated. Step S3-5: Generation of Operation and Maintenance Intervention Node Sequence and Construction of Full Lifecycle Cost Model: Based on the predicted time axis of the probability distribution of remaining useful life, a multi-dimensional operation and maintenance intervention node sequence is generated by discretizing and cutting according to the preset operation and maintenance strategy granularity. The operation and maintenance strategy granularity is set according to the importance level of the equipment and the operation and maintenance management requirements. The operation and maintenance strategy granularity of important equipment is set to 24 hours, and the operation and maintenance strategy granularity of general equipment is set to 72 hours. Each operation and maintenance intervention node corresponds to a potential operation and maintenance operation execution time point. A full lifecycle cost model for assets is constructed, integrating basic equipment operating energy consumption costs, tiered preventative maintenance costs, emergency repair penalty costs for sudden failures, and asset residual value discount curves, forming a comprehensive cost evolution function that dynamically evolves with the forecast time axis; the comprehensive cost evolution function is as follows: ,in, for The total cost of an asset throughout its entire lifecycle at any given moment. For the device from the current moment to Cumulative operating energy cost at any given time For the device from the current moment to The cumulative cost of preventative maintenance over time. For the device from the current moment to The anticipated cost of emergency repairs in case of failure. For the equipment in The residual value of an asset at any given moment; Step S3-6: Quantifying the marginal benefit of assets and dynamic risk exposure: For each maintenance intervention node, calculate the failure probability cutoff rate and expected life extension effect after performing a specified maintenance intervention based on the remaining useful life probability distribution. The failure probability cutoff rate is the reduction ratio of the equipment failure probability after performing maintenance intervention; the expected life extension effect is the average increase in the remaining useful life of the equipment after performing maintenance intervention. Substituting the expected life extension effect and the failure probability cutoff rate into the comprehensive cost evolution function, the net present value of the asset's entire life cycle is calculated for the two cases of implementing operation and maintenance intervention and not implementing operation and maintenance intervention. The difference ratio of the net present value of the entire life cycle before and after intervention is solved, and the marginal benefit of the asset at the corresponding node is obtained after standardization. A positive marginal benefit of the asset indicates that implementing the operation and maintenance intervention can reduce the asset's entire life cycle cost, and the larger the value, the higher the economic benefit. Simultaneously extract the conditional failure probability density corresponding to each operation and maintenance intervention node from the probability distribution of remaining useful life; construct the risk consequence quantification function by calling the power grid topology vulnerability assessment matrix and the regional unit capacity power outage economic loss coefficient; perform spatiotemporal convolution mapping on the conditional failure probability density and the risk consequence quantification function to output the dynamic risk exposure value corresponding to each operation and maintenance intervention node; the dynamic risk exposure value represents the maximum economic loss and power grid safety risk that equipment failure may cause when the node does not perform operation and maintenance intervention. The marginal benefits and dynamic risk exposure values ​​of assets corresponding to each operation and maintenance intervention node are structured and encapsulated according to time sequence to generate an operation and maintenance intervention decision evaluation vector set.

[0018] In this embodiment, the core function of step S4 is to construct a high-dimensional power asset operation and maintenance strategy optimization space with the objective function of maximizing the overall efficiency of assets and the boundary conditions of dynamic risk exposure, asset marginal benefits, and grid security constraints. A deep reinforcement learning strategy network is then used to achieve global strategy optimization, solving for the optimal asset scheduling sequence that balances safety and economy. Finally, a power asset management decision-making scheme that can be directly implemented, including maintenance timing, spare parts allocation, and decommissioning / replacement plans, is generated. The detailed steps are as follows: Step S4-1: Extraction of multidimensional decision variable domain and construction of comprehensive performance evaluation function: Receive the set of operation and maintenance intervention decision evaluation vectors and reorganize them into multidimensional tensors according to asset category, spatial topology location and intervention time sequence; asset category is classified according to equipment type such as transformer, circuit breaker, switch cabinet, etc.; spatial topology location is classified according to the hierarchy of power grid GIS geographic information grid; intervention time sequence is sorted according to the timestamp of the operation and maintenance intervention node sequence. A multidimensional decision variable domain containing maintenance timing variables, spare parts inter-regional allocation variables, and decommissioning replacement trigger variables is extracted from the reorganized multidimensional tensor. The maintenance timing variable is a discrete variable, with the value being the timestamp of each maintenance intervention node. The spare parts inter-regional allocation variable is a continuous variable, with the value being the number of spare parts allocated from each spare parts warehouse to the target equipment. The decommissioning replacement trigger variable is a binary variable, with the value being 0 or 1, representing not performing decommissioning replacement and performing decommissioning replacement, respectively. With maximizing overall asset efficiency as the core guiding principle, a comprehensive efficiency evaluation function is constructed. This function is a weighted aggregation of the equipment availability maintenance index, the full-cycle operation and maintenance cost savings rate, and the regional power grid reliability. The formula is as follows: ,in, This represents the overall asset efficiency value. For the equipment availability maintenance index, To achieve cost savings throughout the entire operation and maintenance cycle, To ensure the reliability of power supply to the regional power grid, , , These are the weight coefficients of the corresponding indicators, and they satisfy... ; The equipment availability maintenance index characterizes the ability of equipment to maintain normal operation during the execution of the operation and maintenance strategy; the full-cycle operation and maintenance cost saving rate characterizes the proportion of cost reduction of the operation and maintenance strategy relative to the baseline strategy; the regional power grid power supply reliability characterizes the continuous power supply capability of the regional power grid after the execution of the operation and maintenance strategy; Step S4-2: Construction of Multidimensional Policy Constraint Boundaries and Cutting of Feasible Domain: Extract the dynamic risk exposure value and asset marginal benefit from the operation and maintenance intervention decision evaluation vector set, and construct a multi-dimensional strategy constraint boundary by combining the power flow transmission constraints, N-1 safety verification criteria and on-site operation resource limits in the power grid dispatch operation procedure; The multi-dimensional strategy constraint boundaries include dynamic tolerance thresholds for risk exposure, non-negativity constraints on asset marginal benefits, safety level constraints on spare parts inventory, constraints on parallel operations by maintenance teams, and power transmission limits at critical grid sections. The dynamic tolerance thresholds for risk exposure are dynamically adjusted based on the importance level of the equipment, with lower tolerance thresholds for critical equipment compared to general equipment. The non-negativity constraint on asset marginal benefits requires that all operational interventions must generate positive economic benefits. The safety level constraint on spare parts inventory requires that the inventory quantity in each spare parts warehouse must not fall below a preset minimum value. The constraints on parallel operations by maintenance teams require that the number of maintenance tasks performed simultaneously in the same area must not exceed the number of available maintenance teams. The power transmission limit at critical grid sections requires that operational operations must not cause the power transmission at critical grid sections to exceed their rated capacity. The multidimensional decision variable domain is mapped and coupled with the comprehensive performance evaluation function, and the multidimensional strategy constraint boundary is used as the feasible domain cutting operator to perform boundary filtering and invalid strategy pruning on the multidimensional decision variable domain; strategy combinations that violate the bottom line of power grid safe operation and economic bottom line are eliminated, and a power asset operation and maintenance strategy optimization space composed of the effective strategy path set, state transition probability distribution and resource allocation constraint matrix is ​​generated. Step S4-3: Deep reinforcement learning interactive environment mapping and state-action space definition: The optimization space of power asset operation and maintenance strategy is mapped to a deep reinforcement learning interactive environment, defining the environment state vector, strategy action space and strategy state transition rules. The environmental state vector is aggregated in real time with the target equipment health baseline offset, pending event flow characteristics, current spare parts inventory dynamic level, and power grid operation mode time slice. The target equipment health baseline offset represents the degree of deviation between the current equipment status and the initial health baseline. The pending event flow characteristics represent the type and intensity of currently unprocessed power grid operation events. The current spare parts inventory dynamic level represents the real-time inventory quantity of each spare parts warehouse. The power grid operation mode time slice represents the current power grid topology and power flow distribution. The strategy action space consists of a set of discrete-continuous mixed actions, including maintenance window selection instructions, spare parts cross-regional transfer instructions, and asset retirement threshold trigger instructions. Maintenance window selection instructions are discrete actions, corresponding to different maintenance timing nodes. Spare parts cross-regional transfer instructions are continuous actions, corresponding to different spare parts transfer quantities. Asset retirement threshold trigger instructions are discrete actions, corresponding to whether to perform asset retirement replacement. The strategy state transition rule is defined as the probability distribution of the environment state transitioning from the current state to the next state after performing a certain action. This probability distribution is determined by the state transition probability distribution matrix in the power asset operation and maintenance strategy optimization space. Step S4-4: Constructing the reward function for the risk-benefit dual-driven strategy: A risk-benefit dual-driven strategy reward function is constructed, which dynamically weights the dynamic risk exposure value output by the power asset operation and maintenance strategy optimization space with the asset marginal benefit. The dynamic weight is automatically adjusted according to the current environmental state. When there are high-risk events in the environment, the weight of the risk exposure value is automatically increased. When the environment is in a safe and stable state, the weight of the asset marginal benefit is automatically increased. When a strategy action brings the risk exposure value to within the tolerance threshold range and the marginal benefit of the asset continues to grow positively, a tiered positive reward is given; the lower the risk exposure value and the higher the marginal benefit of the asset, the greater the positive reward is; when an action triggers an over-limit risk or causes redundant operation and maintenance costs, a negative penalty is added; the higher the level of over-limit risk, the greater the redundant operation and maintenance costs, and the more severe the negative penalty is imposed. The formula for the strategy reward function is: ,in, This is the single-step reward value. This represents the dynamic risk exposure value after the current action is executed. This represents the risk exposure tolerance threshold. The marginal benefit of the asset after the current action is performed. , These are the dynamic weighting coefficients for risk and benefit, respectively; Generate a state reward signal sequence to guide the network gradient update. This sequence records the reward value obtained after each action is performed. Step S4-5: Training the deep reinforcement learning policy network and solving for the optimal asset scheduling sequence: Initialize a deep reinforcement learning policy network, which includes a temporal state feature encoding subnet, a cross-asset collaborative attention module, and a policy probability output head. The temporal state feature encoding subnet is used to map high-dimensional environmental state vectors to low-dimensional hidden layer representations. The cross-asset collaborative attention module is used to capture the topological coupling effect and resource competition relationship under the parallel operation and maintenance of multiple devices. The policy probability output head is used to output the execution probability distribution of each action. The environmental state vector is input into the temporal state feature encoding subnet to extract high-dimensional hidden layer representations. The cross-asset collaborative attention module captures the collaborative effects between multiple devices. The policy probability output head performs policy gradient iterative optimization based on the state reward signal sequence, outputs the action probability distribution, and generates candidate asset scheduling sequences by sampling according to the distribution. During the policy gradient iteration process, a trajectory experience replay mechanism is used to cache historical state-action-reward interaction segments; batch samples are randomly selected from the experience replay cache for network training to improve training stability and sample utilization; the network parameters are updated by backpropagation by combining advantage function estimation and gradient pruning techniques to avoid gradient vanishing and gradient exploding problems. Training is terminated when the convergence deviation of the comprehensive performance evaluation function is less than the set tolerance and the policy action distribution enters a steady state within a consecutive preset iteration period; the candidate sequence with the highest cumulative reward is extracted as the optimal asset scheduling sequence. Step S4-6: Physical Operation and Maintenance Dimension Analysis and Decision Scheme Generation: The optimal asset scheduling sequence is reverse-analyzed to the physical operation and maintenance execution dimension, and the corresponding equipment entity nodes, standardized maintenance process packages, spare parts specification and model library and capital budget items are aligned one by one according to the time axis. Align the physical nodes of the equipment to identify the specific equipment that needs to be overhauled, replaced with spare parts, or decommissioned; align the standardized maintenance process packages to identify the maintenance procedures and technical requirements for each piece of equipment; align the spare parts specification and model library to identify the required spare parts models and quantities; align the budget items to identify the budget for each maintenance operation. The system encapsulates and generates a power asset management decision-making scheme that includes precise maintenance timing windows, dynamic spare parts allocation paths, and asset decommissioning and replacement plans. Precise maintenance timing windows specify the optimal start and end times for maintenance of each piece of equipment. Dynamic spare parts allocation paths specify the transportation routes and time arrangements for each spare part from the warehouse to the target equipment. The asset decommissioning and replacement plan specifies the list of equipment to be decommissioned, the decommissioning time, and the installation and commissioning plan for the new equipment.

[0019] In this embodiment, the core function of step S5 is to issue and execute the generated power asset management decision scheme, simultaneously collect actual feedback data on equipment status response and economic indicators, construct a comparison system between strategy execution effect and expected results, and dynamically and adaptively fine-tune the topology and weight parameters of the temporal causal hypergraph attention network through an online meta-learning mechanism to achieve continuous self-evolution of the event-equipment association model and iterative optimization of the power asset management decision scheme, ultimately forming a complete closed loop of data collection, model inference, decision generation, and execution feedback. The detailed steps are as follows: Step S5-1: Issuance and Implementation of Decision-Making Plan and Collection of Multi-Dimensional Feedback Data: The power asset management decision plan is issued to the on-site operation and maintenance execution system and the material dispatch system, and the full-process monitoring of the plan execution is initiated simultaneously; the maintenance team carries out equipment maintenance work according to the precise maintenance window, the material department completes the transportation and delivery of spare parts according to the dynamic spare parts allocation path, and the asset department organizes the dismantling of old equipment and the installation and commissioning of new equipment according to the asset retirement and replacement plan. The system synchronously connects to the on-site equipment monitoring terminal and the operation and maintenance settlement system to collect real-time data on equipment operating condition response time series and economic indicator feedback data during the execution cycle of the plan. The equipment operating condition response time series data includes the change curves of core operating parameters such as equipment voltage, current, temperature, and vibration before and after maintenance. The economic indicator feedback data includes the actual settlement values ​​of maintenance labor costs, spare parts procurement costs, transportation costs, fault repair costs, and power outage loss costs. Using the maintenance window, spare parts allocation path, and decommissioning and replacement plan in the power asset management decision-making scheme as spatiotemporal anchors, the equipment operating condition response time sequence data and economic indicator feedback data are bidirectionally aligned and matched according to asset identification and execution time axis; ensuring that the execution time of each operation and maintenance operation corresponds one-to-one with the corresponding equipment status change and cost consumption data, and generating a strategy-effect comparison sample set; Step S5-2: Strategy-Effect Residual Calculation and Strategy-Effect Comparison Library Construction: The strategy-effect comparison sample set was analyzed to extract the expected risk exposure value and asset marginal benefit prediction sequence based on the event-equipment correlation model before the strategy was implemented, as well as the equipment condition decay deviation and cost settlement difference obtained by actual measurement after the strategy was implemented. Calculate the dynamic residual sequence between the expected sequence and the actual response. The residual calculation formula is as follows: ,in, for The dynamic residual value at time t. for Predicted value at time, for The actual measured value at that moment; The equipment state decay residual sequence and cost settlement residual sequence are calculated separately. The dynamic residual sequence, the corresponding event trigger feature vector, and the power grid operation condition snapshot are fused using tensors. The fused tensor data is versioned and archived according to execution time and asset category. Each version records the execution parameters and actual effects of all operation and maintenance strategies within the corresponding period. A strategy-effect comparison library containing multi-strategy trajectory mapping relationships is constructed for online adaptive learning of subsequent models. Step S5-3: Construction of Online Meta-Learning Task Set and Design of Meta-Gradient Calculation Path: Based on the strategy-effect comparison library, an online meta-learning task set is constructed according to the time evolution granularity and asset spatial clusters. The time evolution granularity is set to 7 days, and the asset spatial clusters are divided according to the substation level of the power grid GIS geographic information grid. Each meta-learning task corresponds to the execution effect data of all operation and maintenance strategies of a substation within a time period. Each meta-learning task is split into a support sample subset and a query sample subset; the support sample subset contains the top 80% of the policy-effect control samples in the task, which are used for rapid model adaptation; the query sample subset contains the bottom 20% of the policy-effect control samples in the task, which are used for verification and supervision of model adaptation. Using the decision execution characteristics of the supporting sample subset as the inner loop adaptation input and the actual state response of the query sample subset as the outer loop supervision signal, a meta-gradient calculation path for fast adaptation of graph network parameters is constructed. The inner loop quickly adjusts the local parameters of the model through a small number of samples to adapt to changes in specific scenarios. The outer loop updates the global meta-parameters of the model through the accumulated experience of multiple tasks, improving the model's generalization ability to different scenarios. Step S5-4: Dynamic topology and parameter fine-tuning of the temporal causal hypergraph attention network: The online meta-learning mechanism is driven to perform inner-layer task adaptation and outer-layer baseline update along the meta-gradient computation path, and joint dynamic fine-tuning of topology and weight parameters is implemented for temporal causal hypergraph attention networks; First, the graph topology is dynamically updated; then, redundant superedges that have failed verification in the strategy-effect comparison library are removed using a topology confidence pruning algorithm; the topology confidence calculation formula is: ,in, The topological confidence of the hyperedge. This represents the number of times the causal relationship corresponding to the superedge has been verified as correct during the execution of the historical strategy. This represents the total number of times the causal relationship corresponding to the hyperedge has been verified. Redundant hyperedges with topological confidence below a preset threshold are removed, the hypergraph connection relationships of high-gain event clusters are reorganized, and event nodes that have a significant impact on device status and are accurately verified are merged into new hyperedges, thus updating the topology of the temporal causal hypergraph. Simultaneously, a second-order gradient meta-optimizer is used to jointly iteratively calibrate the confusion bias parameters of the multi-head hypergraph attention weight matrix and the counterfactual interference quantifier; the second-order gradient meta-optimizer can utilize historical gradient information to accelerate parameter convergence and achieve rapid adaptation of model parameters under a small number of samples; the output is a time-series causal hypergraph attention network with calibrated parameters; Step S5-5: Two-way collaborative closed-loop evolution of model and decision: The parameter-calibrated temporal causal hypergraph attention network is replaced in the event-device association model to update the causal association reasoning logic and feature representation capability of the model; the device performance degradation trajectory and remaining useful life probability distribution of the next monitoring cycle are re-inferred to generate updated device status prediction results. The iteratively optimized pure causal correlation weight matrix is ​​injected into the reward function construction module of the deep reinforcement learning policy network to update the calculation parameters of the risk-benefit dual-driven policy reward function; this enables the policy network to evaluate the risks and benefits of different operation and maintenance actions based on more accurate causal correlation weights, and generate a better asset scheduling sequence. The system enables the self-evolution of feature representations in the event-device association model and the iterative optimization of strategies for power asset management decision-making schemes, achieving a two-way collaborative closed-loop evolution of model inference and operation and maintenance decisions. As the system's operating time increases, the strategy-effect comparison library continuously accumulates more sample data, and the model's prediction accuracy and the scientific nature of decision-making will continue to improve, ultimately realizing intelligent management of the entire life cycle of power assets.

[0020] A power asset management system based on an event-device correlation model is disclosed. This system implements the aforementioned power asset management method based on the event-device correlation model, adopts a modular and layered architecture design, and the modules achieve data interaction and functional collaboration through a standardized data bus. The system takes multi-source heterogeneous power operation data as input, takes dynamic causal correlation of events and devices as the core logic, and takes maximizing the comprehensive efficiency of assets as the ultimate goal. It realizes fully automated and intelligent management of the entire process from data acquisition and preprocessing, causal knowledge modeling, equipment status inference, operation and maintenance strategy optimization to execution feedback closed loop. The system is divided into five core functional modules, namely, data alignment and noise reduction module, dynamic causal correlation modeling module, status inference and benefit quantification module, strategy optimization and decision generation module, and execution feedback and collaborative evolution module. Data alignment and noise reduction module: This module serves as the system's data entry unit, responsible for the parallel acquisition, standardized preprocessing, and feature purification of multi-source heterogeneous data across all dimensions, providing high-quality data input in a unified format for all subsequent modules. The core of the module includes a multi-source data access unit, a spatiotemporal alignment processing unit, a cross-modal semantic alignment unit, and a feature denoising unit. The multi-source data access unit establishes dedicated data channels with the Internet of Things sensor network, the dispatch automation system, the micro-meteorological monitoring terminal, and the production and operation database, respectively, to realize the real-time parallel acquisition of equipment body monitoring time-series data, power grid operation event logs, environmental meteorological data, and operation and maintenance history data; and performs communication protocol parsing, physical dimension normalization, and linear interpolation of missing breakpoints on the acquired raw data to construct the original multi-source data pool. The spatiotemporal alignment processing unit uses the power grid GIS geographic information grid and the high-precision atomic clock timing signal as a unified spatiotemporal reference. It performs spatial coordinate projection transformation and time axis resampling alignment on the original multi-source data pool. It maps the physical installation location of the equipment under different coordinate systems to node-level spatial topological coordinates, and resamples non-equal interval sampling sequences with different sampling frequencies to a unified clock beat, generating a spatiotemporal aligned data frame carrying a unified timestamp and spatial coordinates. The cross-modal semantic alignment unit uses a power industry professional knowledge graph to perform named entity recognition and causal relationship extraction on the power grid operation event log, and outputs structured event semantic tags; it uses a dynamic time warping algorithm to perform feature-level matching and splicing of environmental meteorological time series slices and equipment body operation parameters to construct a multimodal interactive feature tensor that simultaneously includes the equipment's own state and the influence of the external environment. The feature denoising unit uses adaptive variational mode decomposition combined with the isolated forest algorithm to separate high-frequency measurement noise, environmental drift interference and equipment intrinsic operating condition fluctuations, and identify and remove outliers. The denoised timestamps, spatial topological coordinates, event semantic labels and equipment operating condition parameters are serialized and encapsulated according to a preset relational data structure to generate a standardized event-equipment interaction dataset, which is then transmitted to the dynamic causal association modeling module through the data bus. Dynamic causal relationship modeling module: This module is the core modeling unit of the system. It is responsible for mining the dynamic causal relationship between events and devices from standardized data, constructing a dynamic causal relationship graph of events and devices, and generating an event-device relationship model that can be used for state inference, realizing the transformation from raw data to causal knowledge. The core of the module includes a node extraction and topology construction unit, a temporal causal edge generation unit, a temporal causal hypergraph attention network unit, and a model solidification unit. The standard event-device interaction dataset is parsed by node extraction and topology building unit analysis. Power grid operation events with independent semantic labels and trigger timestamps are mapped as event nodes, and power equipment carrying unique spatial topological coordinates and operating parameters are mapped as device entity nodes. The power grid electrical connection relationship library and geographic information grid are called to spatially cluster the device entity nodes according to voltage level, electrical connection distance and physical adjacency relationship to construct a static power grid physical topology skeleton. Event nodes are then projected to the corresponding topology skeleton neighborhood according to the spatial influence radius. The temporal causal edge generation unit uses the trigger timestamp of the event node as a reference, slides a preset observation window backward along the time axis, and calculates the deviation of the equipment operating parameters relative to the historical baseline and the response delay time within the window. When the deviation exceeds the preset confidence threshold and the response delay time falls within the equipment inertial response range, a directed temporal causal edge is established between the corresponding event node and the equipment entity node, and the trigger strength, hysteresis step size and spatial decay coefficient are written as edge attributes. A time slice sliding mechanism is used to generate a multi-time period state snapshot sequence, and the frequency statistics and confidence accumulation of repeated causal edges are performed to remove occasional noise edges. The static topology and dynamic causal edges are fused to generate an event-equipment dynamic causal relationship graph. The temporal causal hypergraph attention network unit first completes the topological dimensionality upgrade from a binary graph to a hypergraph. It performs multi-hop backtracking centered on the target device entity node, merging co-occurring event node clusters with semantic associations into hypergraph hyperedges. Each hyperedge is bound to the joint temporal distribution features of the event clusters and the spatial topological adjacency matrix. Then, through a multi-head hypergraph attention mechanism, it aggregates the multi-dimensional composite impact of multi-hop events on the target device to generate an initial association feature vector. Finally, it embeds the counterfactual interference quantifier of the structural causal model, extracts environmental meteorological data and power grid load fluctuations as a set of confounding variables, calculates the spurious association bias and removes it from the initial association feature vector, and outputs a pure causal association weight matrix that only reflects the true driving effect of the events. The model solidification unit binds the pure causal association weight matrix with the event-device dynamic causal association graph, solidifies the hyperedge construction rules, attention calculation rules and counterfactual intervention rules, generates a complete event-device association model, and transmits it to the state inference and benefit quantification module. State simulation and benefit quantification module: This module is the system's prediction and evaluation unit. It is responsible for extrapolating the trend of equipment performance changes under real-time event-driven conditions based on the event-device correlation model, predicting the remaining useful life, and quantifying the economic benefits and safety risks of different operation and maintenance intervention nodes, providing a quantitative basis for subsequent decision generation. The core of the module includes a real-time event stream processing unit, an equipment state space construction unit, a performance degradation extrapolation unit, a remaining useful life prediction unit, and a benefit and risk quantification unit. The real-time event stream processing unit receives the target event stream monitored in real time, and analyzes the event type, trigger intensity, duration and time sequence characteristics frame by frame; it retrieves the event-device association model through a multi-level indexing mechanism and matches the causal association weight matrix segment corresponding to the target device. The equipment state space construction unit calls the equipment initial health baseline library, parses the multi-dimensional health indicator reference vector of the target equipment under standard operation conditions, and constructs a three-dimensional equipment state space covering electrical insulation aging degree, mechanical fatigue accumulation degree and thermal stress sensitivity. The performance degradation inference unit uses the matched causal correlation weight matrix fragment as the state space driving factor and couples it with the equipment state space in a multi-dimensional feature tensor to establish an event-driven nonlinear state transition model. It adopts a rolling prediction time window mechanism to iteratively solve the problem frame by frame along the future time domain, superimposes the degradation accumulation effect under the superposition of multiple source event time series and the operating condition drift compensation, and outputs the equipment performance degradation trajectory in the continuous time domain. The remaining useful life prediction unit presets the equipment failure critical threshold envelope, performs time-domain penetration point search on the equipment performance degradation trajectory and the failure critical threshold envelope to obtain the failure penetration point sequence; based on the failure penetration point sequence, an adaptive kernel density estimation algorithm is used to fit the time probability density to generate a remaining useful life probability distribution containing multiple confidence intervals; The benefit-risk quantification unit uses the predicted time axis of the remaining useful life probability distribution as a benchmark, and discretizes it according to the preset operation and maintenance strategy granularity to generate a multi-dimensional operation and maintenance intervention node sequence; it constructs an asset life cycle cost model that integrates the basic operating energy consumption cost of equipment, the cost of graded preventive maintenance, the cost of emergency repair for sudden failures, and the asset residual value discount curve; it traverses each operation and maintenance intervention node, calculates the asset marginal benefit and dynamic risk exposure value after the operation and maintenance intervention is performed at that node; it encapsulates the asset marginal benefit and dynamic risk exposure value corresponding to each node in a time sequence structure, generates an operation and maintenance intervention decision evaluation vector set, and transmits it to the strategy optimization and decision generation module; Strategy optimization and decision generation module: This module is the system's decision output unit, responsible for constructing the power asset operation and maintenance strategy optimization space. It uses a deep reinforcement learning algorithm to achieve global strategy optimization, solves for the optimal asset scheduling sequence that balances safety and economy, and generates a power asset management decision scheme that can be directly implemented. The core of the module includes an optimization space construction unit, a deep reinforcement learning interactive environment unit, a strategy network training unit, and a decision scheme generation unit. The optimization spatial construction unit receives the operation and maintenance intervention decision evaluation vector set, and reorganizes it into a multidimensional tensor according to asset category, spatial topology location, and intervention time sequence to extract a multidimensional decision variable domain containing maintenance timing variables, spare parts cross-regional allocation variables, and decommissioning and replacement trigger variables. With the maximization of comprehensive asset efficiency as the core orientation, a comprehensive efficiency evaluation function is constructed, which is a weighted aggregation of equipment availability maintenance index, full-cycle operation and maintenance cost saving rate, and regional power grid power supply reliability. Combining dynamic risk exposure value, asset marginal benefits, and power grid security constraints, a multidimensional strategy constraint boundary is constructed. The multidimensional decision variable domain is mapped and coupled with the comprehensive efficiency evaluation function, and the multidimensional strategy constraint boundary is used as the feasible domain cutting operator to eliminate invalid strategy combinations and generate the power asset operation and maintenance strategy optimization space. The deep reinforcement learning interactive environment unit maps the power asset operation and maintenance strategy optimization space to the deep reinforcement learning interactive environment, defining the environment state vector, strategy action space, and strategy state transition rules; the environment state vector aggregates in real time the target equipment health baseline offset, pending event flow characteristics, current spare parts inventory dynamic level, and grid operation mode time slice; the strategy action space consists of a discrete-continuous hybrid action set composed of maintenance window selection instructions, spare parts cross-regional transfer instructions, and asset decommissioning threshold trigger instructions. The policy network training unit constructs a risk-benefit dual-driven policy reward function, dynamically assigning weights to dynamic risk exposure values ​​and asset marginal benefits, and assigning tiered positive rewards or negative penalties based on the performance of actions. A deep reinforcement learning policy network is initialized, comprising a temporal state feature encoding subnet, a cross-asset collaborative attention module, and a policy probability output head. The environment state vector is input into the network, and policy gradient iterative optimization is performed based on the state reward signal sequence. Trajectory experience replay, advantage function estimation, and gradient pruning techniques are used to improve training stability. When the comprehensive performance evaluation function converges and the policy action distribution enters a steady state within a consecutive preset iteration period, the candidate sequence with the highest cumulative reward is extracted as the optimal asset scheduling sequence. The decision-making scheme generation unit reverse-parses the optimal asset scheduling sequence to the physical operation and maintenance execution dimension, unfolds it along the time axis and aligns it one by one with the corresponding equipment entity nodes, standardized maintenance process packages, spare parts specification and model library and capital budget items; it encapsulates and generates a power asset management decision-making scheme that includes precise maintenance timing windows, dynamic spare parts allocation paths and asset retirement and replacement plans, and sends it to the execution feedback and collaborative evolution module; Execution Feedback and Co-evolution Module: This module is the closed-loop optimization unit of the system, responsible for monitoring the execution of decision-making schemes and collecting feedback data. It constructs a comparison system between strategy execution effects and expected results, and achieves continuous iterative optimization of the event-equipment correlation model and asset management decision-making schemes through an online meta-learning mechanism, forming a complete self-evolving closed loop. The core of the module includes a scheme execution monitoring unit, a feedback data processing unit, a strategy-effect comparison library unit, an online meta-learning unit, and a model parameter update unit. The scheme execution monitoring unit distributes the power asset management decision-making scheme to the on-site operation and maintenance execution system and the material dispatching system, and simultaneously starts full-process execution monitoring; it tracks the progress of maintenance work, the status of spare parts transportation and the process of equipment decommissioning and replacement in real time to ensure that all operations are carried out according to plan. The feedback data processing unit synchronously connects to the equipment field monitoring terminal and the operation and maintenance settlement system, and collects equipment operating condition response time-series data and economic indicator feedback data in real time during the execution cycle of the plan. Taking the maintenance time window, spare parts allocation path and decommissioning and replacement plan in the decision-making plan as spatiotemporal anchors, the equipment operating condition response data and cost settlement data are bidirectionally aligned and matched according to asset identification and execution time axis to generate a strategy-effect comparison sample set. The strategy-effect comparison library unit analyzes the strategy-effect comparison sample set, extracts the expected risk exposure value and asset marginal benefit prediction sequence before strategy execution, and the actual equipment condition decay deviation and cost settlement difference after strategy execution, and calculates the dynamic residual sequence between the expected sequence and the actual response; tensor fusion and versioned archiving are performed on the dynamic residual sequence, the corresponding event trigger feature vector and the power grid operation condition snapshot to construct a strategy-effect comparison library containing multi-strategy trajectory mapping relationships; The online meta-learning unit is based on a policy-effect comparison library. It constructs an online meta-learning task set according to the temporal evolution granularity and asset space clusters. Each task is split into a support sample subset and a query sample subset. The decision execution characteristics of the support sample subset are used as the inner loop adaptation input, and the actual state response of the query sample subset is used as the outer loop supervision signal to construct a meta-gradient calculation path for fast adaptation of graph network parameters. This drives the online meta-learning mechanism to perform inner-layer task adaptation and outer-layer baseline update. The topology confidence pruning algorithm is used to remove redundant hyperedges that have failed verification. The hypergraph connection relationship of high-gain event clusters is reorganized to update the graph topology. At the same time, a second-order gradient meta-optimizer is used to jointly iteratively calibrate the confusion bias parameters of the multi-head hypergraph attention weight matrix and the counterfactual interference quantifier, and outputs a parameter-calibrated temporal causal hypergraph attention network. The model parameter update unit replaces the parameter-calibrated temporal causal hypergraph attention network into the event-device association model, updating the model's causal association reasoning logic and feature representation capabilities; it injects the iteratively optimized pure causal association weight matrix into the reward function construction module of the deep reinforcement learning policy network, updating the calculation parameters of the policy reward function; it realizes the self-evolution of feature representation of the event-device association model and the iterative optimization of the power asset management decision scheme, completing the two-way collaborative closed-loop evolution of model reasoning and operation and maintenance decision-making.

[0021] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A power asset management method based on an event-device association model, characterized in that: Includes the following steps: S1: Collect multi-source heterogeneous data from the power asset operation environment. The multi-source heterogeneous data includes equipment monitoring time series data, power grid operation event logs, environmental meteorological data, and operation and maintenance history data. Based on a unified spatiotemporal benchmark, perform cross-modal alignment and feature denoising on the multi-source heterogeneous data to generate a standardized event-equipment interaction dataset containing timestamps, spatial topological coordinates, event semantic labels, and equipment operating parameters. S2: Based on a standardized event-device interaction dataset, event nodes and device entity nodes are extracted. Using the power grid physical topology as the framework, the event triggering timing and device state response lag relationship are introduced to construct a dynamic causal association graph of events and devices. A temporal causal hypergraph attention network is used to calculate the association strength of the dynamic causal association graph of events and devices. Specifically, this includes: aggregating the composite impact of multi-hop events on target devices through a multi-head hypergraph attention mechanism, and embedding the counterfactual interference quantifier of the structural causal model to remove environmental confounding variables, outputting a pure causal association weight matrix, and thus generating an event-device association model. S3: Input the real-time monitored target event stream into the event-device association model, couple the state space based on the causal association weight matrix and the initial health baseline of the device, and deduce the device performance degradation trajectory and remaining useful life probability distribution under the target event-driven model; map the probability distribution to the asset life cycle cost model to quantify the asset marginal benefits and risk exposure values ​​under different operation and maintenance intervention nodes; S4: With the goal of maximizing the overall efficiency of assets, and with risk exposure, asset marginal benefits, and grid security constraints as boundary conditions, an optimization space for power asset operation and maintenance strategies is constructed. A deep reinforcement learning strategy network is adopted, with a weighted combination of risk exposure and asset marginal benefits as the state reward signal. The optimal asset scheduling sequence is solved through strategy gradient iteration, generating a power asset management decision scheme that includes maintenance timing, spare parts allocation, and decommissioning and replacement plans. S5: Execute power asset management decision-making schemes and collect equipment status response data and economic indicator feedback after scheme execution to construct a strategy-effect comparison library; based on the strategy-effect comparison library, dynamically fine-tune the graph topology and weight parameters of the temporal causal hypergraph attention network through an online meta-learning mechanism to achieve the collaborative closed-loop evolution of the event-equipment association model and asset management decision-making schemes.

2. The power asset management method based on the event-device association model according to claim 1, characterized in that: Multi-source heterogeneous data from the power asset operation environment are collected. Based on a unified spatiotemporal benchmark, cross-modal alignment and feature denoising are performed on the multi-source heterogeneous data to generate a standardized event-device interaction dataset containing timestamps, spatial topological coordinates, event semantic labels, and equipment operating parameters. Specifically, it includes: The system is connected to the Internet of Things sensor network, the dispatch automation system, the micro-meteorological monitoring terminal and the production and maintenance database respectively, and collects equipment monitoring time-series data, power grid operation event logs, environmental meteorological data and maintenance history data in parallel. The collected data is parsed for communication protocols, normalized for physical dimensions and linearly interpolated for missing breakpoints to construct the original multi-source data pool. Using the power grid GIS geographic information grid and high-precision atomic clock timing signal as a unified spatiotemporal reference, spatial coordinate projection transformation and time axis resampling alignment are performed on the original multi-source data pool to map the physical installation location of the equipment into node-level spatial topological coordinates, and non-equal interval sampling sequences are resampled to a unified clock beat to generate spatiotemporally aligned data frames. Cross-modal semantic alignment is performed based on spatiotemporal aligned data frames. Power domain professional knowledge graphs are used to perform named entity recognition and causal relationship extraction on power grid operation event logs, and output structured event semantic tags. Environmental meteorological time series slices and equipment body operation parameters are matched and stitched together at the feature level through dynamic time warping algorithm to construct multimodal interactive feature tensors. Adaptive variational mode decomposition combined with the isolated forest algorithm is performed on the multimodal interaction feature tensor to perform feature denoising and outlier removal, separating high-frequency measurement noise, environmental drift interference and equipment intrinsic operating condition fluctuations; the denoised timestamps, spatial topological coordinates, event semantic labels and equipment operating condition parameters are serialized and encapsulated according to a preset relational data structure to generate a standardized event-equipment interaction dataset.

3. The power asset management method based on the event-device association model according to claim 1, characterized in that: Based on a standardized event-device interaction dataset, event nodes and device entity nodes are extracted. Using the power grid physical topology as the framework, the relationship between event triggering timing and device state response lag is introduced to construct a dynamic causal relationship graph of events and devices, which specifically includes: The standardized event-device interaction dataset is parsed, and power grid operation events with independent event semantic labels and trigger timestamps are mapped to event nodes, and power equipment carrying unique spatial topological coordinates and operating parameters are mapped to device entity nodes. By calling the power grid electrical connection relationship database and geographic information grid, the equipment entity nodes are spatially clustered according to voltage level, electrical connection distance and physical adjacency relationship to construct a static power grid physical topology skeleton, and the event nodes are projected to the corresponding topology skeleton neighborhood according to the spatial influence radius; Based on the trigger timestamp of the event node, the preset observation window is slid backward along the time axis to calculate the deviation of the operating parameters of the device entity node within the window from the historical baseline and the response delay time. If the deviation exceeds the preset threshold and the response delay time falls within the inertial response range of the device, a directed temporal causal edge is established between the corresponding event node and the device entity node, and the trigger strength, hysteresis step size and spatial decay coefficient are written into the edge attribute. A time-slicing sliding mechanism is adopted to stack and splice continuously generated directed temporal causal edges in discrete time steps to form a multi-time state snapshot sequence. Frequency statistics and confidence accumulation are performed on the causal edges that appear repeatedly in the snapshot sequence to remove occasional noise edges and retain high-frequency stable causal links. Finally, the static power grid physical topology skeleton and dynamic temporal causal edges are integrated to generate an event-device dynamic causal relationship graph.

4. The power asset management method based on the event-device association model according to claim 3, characterized in that: A temporal causal hypergraph attention network is used to calculate the association strength of the event-device dynamic causal association graph. Specifically, this includes: aggregating the combined impact of multi-hop events on the target device through a multi-head hypergraph attention mechanism, embedding a counterfactual interference quantifier from a structural causal model to remove environmental confounding variables, outputting a clean causal association weight matrix, and then generating an event-device association model, specifically including: The multi-node interaction paths in the event-device dynamic causal relationship graph are identified as potential hyperedges. Multi-hop backtracking is performed along the directed temporal causal edge with the target device entity node as the center. Event node clusters that are co-triggered and semantically related are merged into hypergraph hyperedges. The joint temporal distribution features of the event cluster and the spatial topological adjacency matrix are bound to each hyperedge to complete the topological dimensionality upgrade from binary graph to hypergraph. Initialize a multi-head hypergraph attention network, map hypergraph and hyperedge features and device node features to multiple independent semantic subspaces respectively, calculate the feature similarity between event nodes and target device nodes within each subspace to generate a spatial-temporal joint attention distribution, and fuse the outputs of each head through a weighted summation mechanism to capture the multi-dimensional composite influence of electrical coupling, environmental stress and operation and maintenance history, and generate an initial associated feature vector; A structural causal model interference identification module is constructed. Environmental meteorological data and power grid load fluctuations affecting equipment status are extracted as a set of confounding variables. Counterfactual intervention is performed based on the initial correlation feature vector. The state variables of the target equipment node are fixed and a virtual zeroing operation is applied to the set of confounding variables to generate a counterfactual state sequence. The difference in feature response between the observed state sequence and the counterfactual state sequence is compared to calculate the spurious correlation bias caused by the confounding variables. The false association bias is removed from the initial association feature vector to obtain the deconfused causal effect features. The causal effect features are normalized and smoothed across time steps to output a pure causal association weight matrix that only reflects the true driving effect of the event. The pure causal association weight matrix is ​​parameter-bound to the event-device dynamic causal association graph and the network inference rules are solidified to generate the event-device association model.

5. The power asset management method based on the event-device association model according to claim 1, characterized in that: The real-time monitored target event stream is input into the event-device association model. Based on the causal association weight matrix and the initial health baseline of the device, state-space coupling is performed to deduce the device performance degradation trajectory and remaining useful life probability distribution driven by the target event. Specifically, this includes: Receive the target event stream under real-time monitoring, extract event type, trigger intensity and time sequence features, and retrieve and match the corresponding causal association weight matrix fragments through the event-device association model; The system calls upon the initial health baseline library of the equipment, parses the multi-dimensional health indicator reference vector of the target equipment under standard operation conditions, and constructs an equipment state space covering electrical insulation aging degree, mechanical fatigue accumulation degree, and thermal stress sensitivity. The causal correlation weight matrix fragments obtained by matching are used as state space driving factors and coupled with the equipment state space in a multi-dimensional feature tensor to establish an event-driven nonlinear state transition model. The rolling prediction time window mechanism is used to iteratively solve the state transition model frame by frame along the future time domain. The degradation accumulation effect and operating condition drift compensation under the superposition of multi-source event time sequence are superimposed to output the equipment performance degradation trajectory in the continuous time domain. A preset critical threshold envelope for equipment failure is established. The equipment performance degradation trajectory is compared with the critical threshold envelope for failure in the time domain to search for penetration points. Based on the penetration point sequence, an adaptive kernel density estimation algorithm is used to fit the time probability density, generating a probability distribution of remaining useful life containing multiple confidence intervals.

6. The power asset management method based on the event-device association model according to claim 5, characterized in that: Mapping probability distributions to an asset lifecycle cost model quantifies the marginal benefits and risk exposure of assets at different operational intervention points, specifically including: Based on the predicted time axis of the remaining useful life probability distribution, a multi-dimensional operation and maintenance intervention node sequence is generated by discretizing and segmenting according to the preset operation and maintenance strategy granularity. Construct an asset lifecycle cost model that integrates basic equipment operating energy consumption costs, tiered preventive maintenance costs, emergency repair penalty costs for sudden failures, and asset residual value discount curves to form a comprehensive cost evolution function that dynamically evolves with the forecast time axis. For each maintenance intervention node, calculate the failure probability cutoff rate and expected life extension effect after the node performs the specified maintenance intervention based on the probability distribution of the remaining useful life. Substitute the expected life extension effect and failure probability cutoff rate into the comprehensive cost evolution function to solve the difference ratio of the net present value of the whole life cycle before and after the intervention. After standardization, obtain the asset marginal benefit of the corresponding node. Simultaneously extract the conditional failure probability density corresponding to each operation and maintenance intervention node in the probability distribution of remaining useful life, call the power grid topology vulnerability assessment matrix and the regional unit capacity outage economic loss coefficient to construct the risk consequence quantification function, perform spatiotemporal convolution mapping on the conditional failure probability density and the risk consequence quantification function, and output the dynamic risk exposure value corresponding to each operation and maintenance intervention node. The marginal benefits and dynamic risk exposure values ​​of assets corresponding to each operation and maintenance intervention node are structured and encapsulated according to time sequence to generate an operation and maintenance intervention decision evaluation vector set.

7. The power asset management method based on the event-device association model according to claim 1, characterized in that: Using the maximization of overall asset efficiency as the objective function, and with risk exposure, asset marginal benefits, and grid security constraints as boundary conditions, an optimization space for power asset operation and maintenance strategies is constructed, specifically including: Receive the operation and maintenance intervention decision evaluation vector set, and reorganize it into a multidimensional tensor according to asset category, spatial topology location and intervention time sequence. Extract and form a multidimensional decision variable domain containing maintenance timing variables, spare parts cross-regional allocation variables and decommissioning replacement trigger variables. With the maximization of comprehensive asset efficiency as the core guiding principle, a comprehensive efficiency evaluation function is constructed. The comprehensive efficiency evaluation function is a weighted aggregation of equipment availability maintenance index, full-cycle operation and maintenance cost saving rate and regional power grid power supply reliability, which is used to quantify the overall asset operation level under different strategy combinations. The dynamic risk exposure value and asset marginal benefit are extracted from the evaluation vector set of operation and maintenance intervention decision. Combined with the power flow transmission constraints, N-1 safety verification criteria and the upper limit of on-site operation resources in the power grid dispatch operation procedure, a multi-dimensional strategy constraint boundary is constructed. The multi-dimensional strategy constraint boundary includes the dynamic tolerance threshold of risk exposure, the non-negativity constraint of asset marginal benefit, the safety level constraint of spare parts inventory, the parallel operation constraint of maintenance teams, and the power transmission limit of key sections of the power grid. The multidimensional decision variable domain is mapped and coupled with the comprehensive performance evaluation function. The multidimensional strategy constraint boundary is used as the feasible domain cutting operator to perform boundary filtering and invalid strategy pruning on the multidimensional decision variable domain. Strategy combinations that violate the bottom line of power grid safe operation and economic bottom line are eliminated, generating a power asset operation and maintenance strategy optimization space composed of the effective strategy path set, state transition probability distribution and resource allocation constraint matrix.

8. The power asset management method based on the event-device association model according to claim 7, characterized in that: A deep reinforcement learning policy network is employed, using a weighted combination of risk exposure and asset marginal benefit as the state reward signal. The optimal asset scheduling sequence is solved through policy gradient iteration, generating a power asset management decision scheme that includes maintenance timing, spare parts allocation, and decommissioning / replacement plans. Specifically, this includes: The optimization space of power asset operation and maintenance strategy is mapped to a deep reinforcement learning interactive environment, defining the environment state vector, strategy action space and strategy state transition rules; the environment state vector is aggregated in real time with the target equipment health baseline offset, pending event flow characteristics, current spare parts inventory dynamic level and grid operation mode time slice; the strategy action space is a set of discrete-continuous hybrid actions consisting of maintenance window selection instructions, spare parts cross-regional transfer instructions and asset retirement threshold trigger instructions. A risk-benefit dual-driven strategy reward function is constructed, which dynamically weights the dynamic risk exposure value of the power asset operation and maintenance strategy optimization space output with the asset marginal benefit. When the strategy action makes the risk exposure value converge to the tolerance threshold range and the asset marginal benefit maintains positive growth, a tiered positive reward is given. When the action triggers the risk of exceeding the limit or causes redundant operation and maintenance costs, a negative penalty term is superimposed to generate a state reward signal sequence to guide the network gradient update. A deep reinforcement learning policy network is initialized, which includes a temporal state feature encoding subnet, a cross-asset collaborative attention module, and a policy probability output head. The environmental state vector is input into the temporal state feature encoding subnet to extract high-dimensional hidden layer representations. The cross-asset collaborative attention module captures the topological coupling effect and resource competition relationship under the parallel operation and maintenance of multiple devices. The policy probability output head performs policy gradient iterative optimization based on the state reward signal sequence, outputs the action probability distribution, and generates candidate asset scheduling sequences by sampling according to the distribution. During the policy gradient iteration process, a trajectory experience replay mechanism is used to cache historical state-action-reward interaction segments. The network parameters are updated by backpropagation in combination with advantage function estimation and gradient pruning techniques. Training is terminated when the convergence deviation of the comprehensive performance evaluation function is less than the set tolerance and the policy action distribution enters a steady state within a continuous preset iteration period. The candidate sequence with the highest cumulative reward is extracted as the optimal asset scheduling sequence. The optimal asset scheduling sequence is reverse-analyzed to the physical operation and maintenance execution dimension. The corresponding equipment entity nodes, standardized maintenance process packages, spare parts specification and model library and capital budget items are aligned one by one according to the time axis. The resulting power asset management decision scheme includes precise maintenance timing windows, dynamic spare parts allocation paths and asset retirement and replacement plans.

9. The power asset management method based on the event-device association model according to claim 1, characterized in that: Implement power asset management decision-making schemes and collect equipment status response data and economic indicator feedback after scheme implementation to construct a strategy-effect comparison library. Based on the strategy-effect comparison library, dynamically fine-tune the graph topology and weight parameters of the temporal causal hypergraph attention network through an online meta-learning mechanism to achieve the coordinated closed-loop evolution of the event-equipment association model and asset management decision-making schemes. Specifically, this includes: The power asset management decision plan is issued and implemented, and the equipment field monitoring terminal and operation and maintenance settlement system are connected simultaneously to collect equipment operating condition response time series data and economic indicator feedback data in real time during the execution cycle of the plan. Using the maintenance time window, spare parts allocation path and decommissioning and replacement plan in the power asset management decision plan as spatiotemporal anchors, the equipment operating condition response time series data and economic indicator feedback data are bidirectionally aligned and matched according to asset identification and execution time axis to generate a strategy-effect comparison sample set. The strategy-effect comparison sample set is analyzed, and the expected risk exposure value and asset marginal benefit prediction sequence before strategy execution are extracted, as well as the actual state decay deviation and cost settlement difference after strategy execution. The dynamic residual sequence between the expected sequence and the actual response is calculated. The dynamic residual sequence, the corresponding event trigger feature vector and the power grid operation condition snapshot are tensor fused and versioned for archiving to construct a strategy-effect comparison library containing multi-strategy trajectory mapping relationships. Based on the policy-effect comparison library, an online meta-learning task set is constructed according to the temporal evolution granularity and asset space clusters. The online meta-learning task set is split into a support sample subset and a query sample subset. The decision execution characteristics of the support sample subset are used as the inner loop adaptation input, and the actual state response of the query sample subset is used as the outer loop supervision signal to construct a meta-gradient calculation path for fast adaptation of graph network parameters. The online meta-learning mechanism performs inner-layer task adaptation and outer-layer baseline update along the meta-gradient computation path, and implements dynamic fine-tuning for the temporal causal hypergraph attention network: redundant hyperedges that have failed verification in the policy-effect control library are removed by the topology confidence pruning algorithm, and the hypergraph connection relationship of high-gain event clusters is reorganized to update the graph topology; at the same time, a second-order gradient meta-optimizer is used to jointly iteratively calibrate the confusion bias parameters of the multi-head hypergraph attention weight matrix and the counterfactual interference quantifier, and outputs the parameter-calibrated temporal causal hypergraph attention network. The parameter-calibrated temporal causal hypergraph attention network is replaced in the event-device association model to re-infer the equipment performance degradation trajectory and remaining useful life probability distribution in the next monitoring cycle. The iteratively optimized pure causal association weight matrix is ​​injected into the reward function construction module of the deep reinforcement learning policy network to realize the self-evolution of the feature representation of the event-device association model and the iterative optimization of the power asset management decision scheme, thus completing the two-way collaborative closed-loop evolution of model inference and operation and maintenance decision.

10. A power asset management system based on an event-device association model, wherein the system executes the method according to any one of claims 1-9, characterized in that: The system includes: The data alignment and noise reduction module is used to collect multi-source heterogeneous data in the power asset operation environment. Based on a unified spatiotemporal benchmark, it performs cross-modal alignment and feature noise reduction on the multi-source heterogeneous data to generate a standardized event-device interaction dataset containing timestamps, spatial topological coordinates, event semantic labels and equipment operating parameters. The dynamic causal association modeling module, connected to the data alignment and noise reduction module, receives a standardized event-device interaction dataset, extracts event nodes and device entity nodes, and constructs a dynamic causal association graph of events and devices by introducing the event triggering time sequence and device state response lag relationship with the power grid physical topology as the skeleton. It then uses a temporal causal hypergraph attention network to calculate the association strength of the event-device dynamic causal association graph, aggregates the composite impact of multi-hop events on the target device through a multi-head hypergraph attention mechanism, and embeds the counterfactual interference quantifier of the structural causal model to remove environmental confounding variables, outputting a clean causal association weight matrix, and thus generating an event-device association model. The state deduction and benefit quantification module is connected to the dynamic causal relationship modeling module. It is used to receive the event-equipment relationship model, input the real-time monitored target event stream into the event-equipment relationship model, couple the state space based on the causal relationship weight matrix and the initial health baseline of the equipment, deduce the equipment performance degradation trajectory and the probability distribution of the remaining useful life under the target event, and map the probability distribution of the remaining useful life to the asset life cycle cost model to quantify the marginal benefits and dynamic risk exposure of the asset under different operation and maintenance intervention nodes. The strategy optimization and decision generation module is connected to the state deduction and benefit quantification module. It receives the marginal benefit of assets and the dynamic risk exposure value. Taking the maximization of comprehensive asset efficiency as the objective function, and using the dynamic risk exposure value, the marginal benefit of assets and the grid security constraints as boundary conditions, it constructs the power asset operation and maintenance strategy optimization space. It adopts a deep reinforcement learning strategy network to use the weighted combination of dynamic risk exposure value and asset marginal benefit as the state reward signal to solve the optimal asset scheduling sequence through strategy gradient iteration, and generates a power asset management decision scheme that includes maintenance timing, spare parts allocation and decommissioning replacement plan. The execution feedback and collaborative evolution module is connected to the strategy optimization and decision generation module for issuing and executing power asset management decision schemes, and collecting equipment status response data and economic indicator feedback after scheme execution to build a strategy-effect comparison library. Based on the strategy-effect comparison library, the graph topology and weight parameters of the temporal causal hypergraph attention network are dynamically fine-tuned through an online meta-learning mechanism, and the fine-tuned graph topology and weight parameters are fed back to the dynamic causal association modeling module to realize the collaborative closed-loop evolution of the event-equipment association model and the power asset management decision scheme.

Images