Integrated electrical system reliability assessment method fusing multi-source data

Abstract

This application relates to the field of power system reliability assessment technology, and in particular to a comprehensive electrical system reliability assessment method that integrates multi-source data. The method includes deploying intelligent acquisition units at multiple system levels to simultaneously acquire five-dimensional data on operation, environment, equipment, users, and communication; decoupling system vulnerability and disturbance sensitivity through dual-channel feature construction; generating a comprehensive reliability index using a three-level fusion process by a regional collaborative assessment center, and issuing collaborative control strategies; and continuously evolving assessment capabilities by updating model parameters and historical event databases online based on operational feedback. This application enhances the comprehensiveness, dynamism, and adaptability of the assessment.

Description

Technical Field

[0001] This invention belongs to the field of power system reliability analysis and evaluation technology, specifically a comprehensive electrical system reliability evaluation method that integrates multi-source data. Background Technology

[0002] As new power systems rapidly evolve towards higher proportions of renewable energy, multi-energy complementarity, and intelligence, the structure of integrated electrical systems is becoming increasingly complex. Their operational status is affected by multiple factors, including weather, equipment aging, load fluctuations, and topology changes. Traditional reliability assessment methods are often based on single data sources or static models, making it difficult to comprehensively reflect the true reliability level of the system under dynamic environments. Therefore, there is an urgent need for a new method that can integrate multi-source heterogeneous data to achieve comprehensive reliability assessment of integrated electrical systems across all dimensions, supporting the safe, efficient, and intelligent operation of the power grid.

[0003] A search revealed a method and apparatus for assessing the reliability of a distribution network, with publication number CN115313511B, published on October 24, 2025. This patent assesses distribution network reliability by constructing a day-ahead optimization scheduling model and an islanding model, combining remaining energy storage capacity and island power supply range, and recording the number of power outages and cumulative outage time. However, this approach primarily relies on preset electrical parameters and historical outage statistics in a simulation system, failing to effectively integrate multi-source heterogeneous data such as real-time operational data, environmental monitoring information, and equipment status perception. Furthermore, its assessment dimensions are limited to outage indicators, lacking a comprehensive consideration of system vulnerability, recovery capabilities, and the coupling effects of multiple faults, making it difficult to adapt to the dynamic reliability characteristics of distribution networks with high penetration rates of distributed energy access.

[0004] A search revealed a power distribution network reliability assessment system and method, publication number CN119005531B, published on February 28, 2025. This patent calculates dynamic failure rates based on electrical parameter data and introduces a fault impact transmission matrix to assess the correlation between nodes, improving the responsiveness to equipment aging and environmental factors. However, this method still relies on single electrical operation data and does not integrate external multi-source information such as meteorological, geographical, user-side response, and communication status. Furthermore, while its failure probability model considers dynamism, it lacks a cross-domain data fusion mechanism, failing to achieve collaborative modeling of the coupled effects of multiple physical fields such as electricity, heat, cold, and information in a comprehensive electrical system. This results in insufficient generalization ability of the assessment results in complex scenarios.

[0005] The aforementioned problems indicate that existing technologies suffer from limitations in data sources, lack of comprehensiveness in evaluation dimensions, and difficulty in supporting the deep fusion and dynamic updating of multi-source heterogeneous data in their model architecture. Therefore, this invention proposes a "Comprehensive Electrical System Reliability Assessment Method Integrating Multi-Source Data." This method aims to construct a reliability assessment framework that integrates data from multiple dimensions, including operational data, environmental data, equipment status, user behavior, and network communication, thereby achieving accurate, dynamic, and comprehensive quantification of the reliability level of integrated electrical systems. This provides scientific decision support for the planning, scheduling, and operation and maintenance of new power systems. Summary of the Invention

[0006] The technical problem to be solved by this invention is: how to address the issues of single data sources, one-sided evaluation dimensions, and lack of dynamic fusion capabilities in existing technologies, and provides a comprehensive electrical system reliability assessment method that integrates multi-source data. This invention solves the above technical problem through the following technical solution, comprising the following steps: S1: Deploying intelligent acquisition units with edge sensing and preprocessing capabilities at substations, feeder nodes, distributed energy access points, and user-side metering terminals of the comprehensive electrical system, simultaneously acquiring operational layer data, environmental layer data, equipment status layer data, user behavior layer data, and communication network status data; wherein, operational layer data includes voltage, current, power, frequency, and harmonic content; environmental layer data includes temperature, humidity, wind speed, rainfall, and light intensity; and equipment status layer data includes insulation resistance, partial discharge, mechanical vibration, and temperature rise curves. User behavior layer data includes load response delay, frequency of electricity consumption mode switching, and demand-side regulation participation; communication network status data includes latency jitter, packet loss rate, and link availability. All raw data is aligned to a unified timestamp and then processed for missing value imputation, outlier removal, and dimensional normalization, forming a multi-source heterogeneous data stream. S2: Each intelligent acquisition unit performs dual-channel feature construction based on a local embedded inference engine. The first channel spatiotemporally aligns the operational layer data and device status layer data, then inputs it into a graph neural network guided by physical constraints to generate a structural vulnerability vector representing the current vulnerability of the system. The second channel integrates environmental layer data, user behavior layer data, and... Communication network status data is concatenated into an external disturbance feature sequence, segmented by a sliding window, and then input into a lightweight timing encoder to output a disturbance sensitivity feature vector. S3: The intelligent acquisition unit concatenates the structural vulnerability vector and the disturbance sensitivity feature vector into a local reliability feature package, which is then bound to a current system topology snapshot. When the preliminary reliability score calculated locally is lower than a set threshold, the feature package, along with the topology snapshot, is uploaded to the regional collaborative evaluation center. S4: After receiving data uploaded by multiple intelligent acquisition units, the regional collaborative evaluation center initiates a three-level fusion evaluation process: First, it uses a historical event database to perform similar scenario matching on the disturbance sensitivity feature vector. First, the system generates scenario adaptation coefficients. Second, it inputs structural vulnerability vectors, disturbance sensitivity feature vectors, and topology snapshots into a cross-domain fusion model based on a multi-head graph attention mechanism, outputting a system-level dynamic reliability probability. Finally, it combines the scenario adaptation coefficients and dynamic reliability probability to generate a comprehensive reliability index through a differentiable weighted gating mechanism. S5: Based on the comprehensive reliability index and predefined reliability level classification rules, the regional collaborative assessment center determines the reliability level of the current system and generates a collaborative control strategy that includes topology reconstruction suggestions, resource scheduling instructions, and early warning area identifiers. This strategy is then distributed to relevant intelligent acquisition units and scheduling decision terminals.S6: After receiving the coordinated control strategy, the intelligent acquisition unit analyzes the topology reconstruction suggestions and triggers local relay protection logic adjustments. Simultaneously, it updates the edge weight parameters in the graph neural network to reflect the vulnerability distribution under the new topology. The scheduling decision terminal generates an operation ticket based on the warning area identifier and resource scheduling instructions, and pushes it to the handheld terminal of the maintenance personnel. After the maintenance personnel execute the operation, they transmit the actual operation record and on-site verification data back to the regional collaborative evaluation center. S7: Based on the transmitted operation record and verification data, the regional collaborative evaluation center fine-tunes the attention weights of the cross-domain fusion model using an online gradient update method. It also assigns a time-decrease factor to similar scenario samples in the historical operation event database based on the effectiveness of the operation, achieving continuous evolution of the evaluation model.

[0007] Furthermore, in step S2, the process of generating the structural vulnerability vector is as follows: S211: Construct a directed graph with electrical nodes as vertices and branches as edges. The node features include voltage magnitude, phase angle, and equipment health index, while the edge features include impedance magnitude, power flow direction, and protection action history; S212: Input the graph into a two-layer graph convolutional network, followed by residual connections and layer normalization at each layer, and output the vulnerability embedding of each node; S213: Perform global average pooling on all node embeddings to obtain the structural vulnerability vector.

[0008] Furthermore, in step S2, the generation process of the perturbation sensitivity feature vector is as follows: S221: Group the external perturbation feature sequences within the sliding window, with the environmental layer data as the first subsequence and the user behavior layer and communication network state data merged into the second subsequence; S222: Extract local patterns from the input one-dimensional convolutional layer of the two subsequences respectively, then capture long-term dependencies through a bidirectional gated recurrent unit, and finally concatenate the final hidden states of the two subsequences to form the perturbation sensitivity feature vector.

[0009] Furthermore, in step S4, the scenario adaptation coefficient is calculated as follows: the regional collaborative evaluation center maintains a historical operational event database, each event containing a disturbance sensitivity feature vector, the meteorological code at the time of occurrence, the load type distribution, and the final reliability level; for the received disturbance sensitivity feature vector, the K nearest neighbor events are retrieved in the event database using Euclidean distance, and the weighted average of their reliability levels is calculated, with the weight being the reciprocal of the distance. This weighted average is then mapped by Sigmoid and used as the scenario adaptation coefficient.

[0010] Furthermore, in step S4, the structure of the cross-domain fusion model includes: the input layer receives the structural vulnerability vector, the disturbance sensitivity feature vector, and the adjacency matrix corresponding to the topology snapshot; the intermediate layer consists of three parallel branches, which respectively process electrical structure information, external disturbance information, and topology dynamics; the fusion layer adopts a multi-head graph attention mechanism to interactively weight the three types of information at the node level, and finally generates the system-level dynamic reliability probability through a global readout function.

[0011] Furthermore, in step S4, the specific implementation of the differentiable weighted gating mechanism is as follows: Let the scene adaptation coefficient be α and the dynamic reliability probability be β, then the comprehensive reliability index is... Where σ is the Softmax function, and These are learnable parameters that are updated after each evaluation via backpropagation.

[0012] Furthermore, before step S5, the method includes: calculating the rate of change of the comprehensive reliability index of the current system over the past N evaluation periods, which is obtained by the slope of the linear regression of the first-order difference sequence; and constructing a two-dimensional reliability state plane with the comprehensive reliability index as the horizontal axis and the rate of change as the vertical axis, dividing it into four sub-regions: "high reliability stable region", "medium reliability fluctuating region", "low reliability deterioration region" and "emergency instability region".

[0013] Furthermore, in step S5, the method for generating topology reconfiguration suggestions is as follows: based on the current topology snapshot and structural vulnerability vector, identify the M branches with the highest vulnerability, combine distributed energy output prediction and load distribution, solve the islanding scheme under the minimum load shedding, and encode the scheme as a switching operation sequence as a topology reconfiguration suggestion.

[0014] Furthermore, in step S6, the intelligent acquisition unit updates the edge weights of the graph neural network by setting the initial weights of the edges corresponding to newly added or disconnected branches in the new topology to 0.1 or 0, and dynamically adjusting them in subsequent evaluation cycles based on the actual power flow and fault isolation effect. The adjustment formula is as follows: Where η is the learning rate, For the measured current, To predict the current for the model, This is the rated current.

[0015] Furthermore, in step S7, the specific process of online gradient update is as follows: construct the operation record and verification data into a supervision signal, calculate the cross-entropy loss between the cross-domain fusion model output and the actual reliability result, and perform only one step of stochastic gradient descent on the attention weight matrix; the update rule of the time decay factor is: if a historical event is successfully matched and the operation is valid, its decay factor is multiplied by 1.05, otherwise it is multiplied by 0.95. The decay factor is used to scale the event feature vector in subsequent retrievals.

[0016] The beneficial effects of this invention are as follows: This comprehensive electrical system reliability assessment method, which integrates multi-source data, simultaneously acquires five-dimensional data on operation, environment, equipment, users, and communication by deploying intelligent acquisition units at multiple levels of the system. It constructs a dual-channel feature extraction mechanism of a physically constrained graph neural network and a time-series encoder, achieving decoupled modeling of system vulnerability and sensitivity to external disturbances. The regional collaborative assessment center adopts a three-level fusion assessment process, combining historical scenario matching and a cross-domain graph attention model to generate a comprehensive reliability index with physical interpretability. The collaborative control strategy includes topology reconfiguration, resource scheduling, and early warning indicators. Based on this, the intelligent acquisition units dynamically adjust local model parameters and protection logic, and the scheduling terminal forms a closed-loop operation feedback. The regional center continuously optimizes the assessment model through online gradient updates and time-decrease mechanisms, enabling the reliability assessment results to adaptively adjust with system evolution. This overcomes the shortcomings of traditional methods that rely on a single data source, static model, and local indicators, supporting the safe, efficient, and intelligent operation of new power systems in highly uncertain environments. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall architecture of the integrated electrical system reliability assessment method that integrates multi-source data according to the present invention, showing the data flow and control flow interaction relationship between the intelligent acquisition unit, the regional collaborative assessment center and the scheduling decision terminal.

[0018] Figure 2 This is a schematic diagram of the three-level fusion evaluation process of the regional collaborative evaluation center in this invention, which sequentially shows the process of scene adaptation coefficient calculation, cross-domain fusion model reasoning and comprehensive reliability index generation.

[0019] Figure 3 This is a flowchart of the overall electrical system reliability assessment method of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Specific implementation examples are given below.

[0022] Specific implementation methods of the integrated electrical system reliability assessment method that integrates multi-source data of the present invention, combined with Figures 1 to 3 The structure shown will be described in detail; for example Figure 1 As shown, the entire system consists of a three-layer collaborative architecture comprised of intelligent acquisition units deployed at key nodes of the integrated electrical system, a regional collaborative assessment center, and a dispatch decision terminal. The intelligent acquisition units are installed at substation busbar outgoing lines, feeder branch nodes, distributed photovoltaic / energy storage access points, and user-side smart meters, establishing a bidirectional data channel with the regional collaborative assessment center via wired or wireless communication links. The regional collaborative assessment center connects to the dispatch decision terminal via dedicated fiber optic or 5G slicing networks, forming a closed-loop feedback control path. During operation, the intelligent acquisition units simultaneously collect five types of raw data, align them with a unified timestamp, and then preprocess them to generate multi-source heterogeneous data streams. When the local preliminary assessment result is below a threshold, a local reliability feature packet containing structural vulnerability vectors and disturbance sensitivity feature vectors is uploaded to the regional collaborative assessment center. After performing a three-level fusion assessment, the regional collaborative assessment center generates a collaborative control strategy and distributes it to the relevant intelligent acquisition units and the dispatch decision terminal. The dispatch decision terminal generates an operation ticket based on the strategy and pushes it to the handheld device of the maintenance personnel. After performing the operation, the maintenance personnel return the operation record and verification data to the regional collaborative assessment center, completing a full assessment-control-feedback cycle. Inside the intelligent acquisition unit, its embedded hardware platform is equipped with a dual-channel feature construction module. The first channel receives data from the operation layer and the device status layer, including real-time electrical quantities (voltage, current, power, frequency, harmonic content) from voltage transformers and current transformers, as well as device status parameters (insulation resistance, partial discharge quantity, mechanical vibration amplitude, temperature rise curve) output from insulation monitoring devices, partial discharge sensors, vibration accelerometers, and infrared temperature measurement modules. These data are first spatiotemporally aligned in the local FPGA or ARM Cortex-M7 core to ensure that electrical quantities and device status quantities within the same time window correspond to the same physical moment. Subsequently, a directed graph is constructed with electrical nodes as vertices and branches as edges: the vertex feature vector consists of the voltage amplitude, phase angle, and health score calculated based on the device health index, where the device health index is obtained by weighted normalization of parameters such as insulation resistance and partial discharge quantity; the edge feature vector includes the branch impedance magnitude (read from the line parameter database), real-time power flow direction (determined by power flow direction), and the value of the branch impedance over the past 30 days. The number of times the branch-related protection device operates; the graph structure is input to a physically constrained graph neural network, which consists of two graph convolutional layers. Each layer uses Chebyshev polynomials to approximate spectral domain convolution, and the convolution kernel parameters are constrained by Kirchhoff's laws to ensure that the node injection power and the branch power flow satisfy the physical conservation relationship; each graph convolution is followed by a residual connection and layer normalization module, which outputs a 128-dimensional vulnerability embedding for each node; finally, through global average pooling, the average value of all node embedding vectors is calculated according to their dimensions to obtain a 128-dimensional structural vulnerability vector. The second channel receives environmental layer, user behavior layer, and communication network status data, including temperature, humidity, wind speed, rainfall, and light intensity provided by the weather station; load response delay time and power consumption mode switching frequency (such as air conditioner start-stop times) recorded by user-side smart meters; demand-side response participation records; and latency jitter, packet loss rate, and link availability indicators reported by the communication network management system. This data is concatenated at 1-minute sampling intervals into an external disturbance feature sequence of length T=60, segmented by a sliding window (window length 30, step size 10), and then input into a lightweight timing encoder. This encoder first divides the sequence into two sub-sequences based on the data source: the first sub-sequence is 5-dimensional environmental data, and the second sub-sequence is 7-dimensional user and communication data. Each sub-sequence is then passed through a one-dimensional convolutional layer (convolutional...). The kernel (3 cores, 32 output channels) extracts local timing patterns and then feeds them into a bidirectional gated recurrent unit (GRU, 64-dimensional hidden unit) to capture long-term dependencies. The final forward and backward hidden states of the two GRUs are concatenated to form a 128-dimensional disturbance sensitivity feature vector. The structural vulnerability vector and the disturbance sensitivity feature vector are concatenated in the main control chip of the intelligent acquisition unit to form a 256-dimensional local reliability feature package, which is then bound and stored with the current system topology snapshot (the switch state matrix synchronously acquired by the SCADA system). The local embedded inference engine uses a lightweight fully connected network to calculate the initial reliability score. If the score is lower than the 0.7 threshold, the upload mechanism is triggered, and the local reliability feature package and the topology snapshot are encrypted and uploaded to the regional collaborative evaluation center via the MQTT protocol. The regional collaborative assessment center is deployed in the municipal-level power dispatch building and equipped with a GPU-accelerated server cluster. Its three-level fusion assessment process first calls upon the historical operational event database, which stores a quadruple of all major operational events from the past three years: {disturbance sensitivity feature vector, meteorological code (e.g., heavy rain, high temperature codes), load type distribution (industrial / commercial / residential percentage), and final reliability level (0-1 continuous value)}. For the received disturbance sensitivity feature vector, a KD-Tree index is used to retrieve the K=10 events with the closest Euclidean distance from the event database, and the weighted average of their reliability levels is calculated, with weights of [weight missing]. To prevent division by zero, the weighted average is mapped to the (0,1) interval using the Sigmoid function and used as the scene adaptation coefficient. Subsequently, the structural vulnerability vector, disturbance sensitivity feature vector, and topology snapshot (converted to adjacency matrix form) are input into the cross-domain fusion model. The input layer of this model feeds the three types of data into three parallel branches: the electrical structure branch uses a two-layer graph attention network to process the structural vulnerability vector and adjacency matrix, with 4 attention heads per layer, outputting a node-level vulnerability enhancement representation; the external disturbance branch processes the disturbance sensitivity feature vector through a two-layer Transformer encoder; the topology dynamic branch flattens the adjacency matrix and compresses it into a topology embedding through a fully connected layer; the fusion layer uses a multi-head graph attention mechanism, with the output of the electrical structure branch as the query and the outputs of the other two branches as keys and values, performing cross-domain information interaction weighting at the node level, and finally generating a dynamic reliability probability between 0 and 1 through global max pooling and a fully connected layer; a differentiable weighted gating mechanism receives the scene adaptation coefficient and dynamic reliability probability, and calculates the comprehensive reliability index. Where σ is the Softmax function, and The parameter is a learnable parameter, initially set to 0.5, and updated after each evaluation via backpropagation; After generating the comprehensive reliability index, the regional collaborative assessment center calculates its first-order difference sequence over the past N=10 assessment periods (5 minutes per period) and performs linear regression on the sequence; the slope is the rate of change. A two-dimensional reliability state plane is constructed with the comprehensive reliability index as the horizontal axis and the rate of change as the vertical axis, dividing the region into four sub-regions: high reliability stable region (index ≥ 0.85 and rate of change ≥ -0.01), medium reliability fluctuating region (0.7 ≤ index < 0.85 or rate of change < -0.01 but ≥ -0.03), low reliability deterioration region (0.5 ≤ index < 0.7 and rate of change < -0.03), and emergency instability region (index < 0.5 or rate of change < -0.05). The reliability level is determined based on the current location within the region. If the region is in the low reliability deterioration region or the emergency instability region, the collaborative control strategy generation module is activated. Based on the current topology snapshot and structural vulnerability vector, the branches with the highest vulnerability embedding value (M=5) are identified. Combining short-term load forecasting (for the next 15 minutes) and distributed energy output forecasting (from a new energy cloud platform), a mixed-integer linear programming model is constructed. The objective function is to minimize the load shedding amount, with constraints including power flow balance, voltage limits, islanded power balance, and black-start capability. The optimal islanding scheme is obtained after solving the model and encoded as a switching operation sequence (e.g., "open #123 switch, close #456 tie switch") as a topology reconfiguration suggestion. Simultaneously, based on user behavior data in the disturbance sensitivity feature vector, demand-side resource scheduling instructions (e.g., calling interruptible loads) are generated. Furthermore, based on the spatial distribution of the structural vulnerability vector, feeder segments with vulnerability values ​​exceeding the threshold are identified as early warning areas. This coordinated control strategy is distributed to relevant intelligent acquisition units and scheduling decision terminals via the IEC 61850-7-420 standard protocol. After receiving the coordinated control strategy, the intelligent acquisition unit analyzes the topology reconfiguration suggestions and adjusts the overcurrent protection settings and operating time limits in the local relay protection device. For example, it shortens the protection time limit of the island boundary branch by 20% to accelerate fault isolation. Simultaneously, it updates the edge weight parameters in the physically constrained graph neural network: for newly added branches in the strategy (such as branches formed by the closing of tie switches), the initial weight of the corresponding edge in its adjacency matrix is ​​set to 0.1; for disconnected branches, the weight is set to 0. In subsequent evaluation cycles, the weight is adjusted based on the measured current. Predicting current with graph neural networks The deviation is dynamically adjusted by weighting, and the adjustment formula is as follows: Where η = 0.01 is the learning rate. The rated current of the branch is 10 ... After receiving the operation records and verification data, the regional collaborative evaluation center constructs a monitoring signal: if the system does not experience a failure and the reliability index recovers after the operation, it is marked as a positive sample; otherwise, it is marked as a negative sample. The cross-entropy loss between the cross-domain fusion model output and the actual result is calculated, and only one step of stochastic gradient descent is performed on the attention weight matrix in the multi-head graph attention mechanism, with a learning rate set to 10. -4 Simultaneously, for event samples in the historical event database that are successfully matched (i.e., the K events returned contain similar events) and whose corresponding operations are valid, their time-decrease factor is multiplied by 1.05; if the operation is invalid or the match fails, the decrease factor is multiplied by 0.95. This decrease factor is used to scale the event feature vector in the subsequent scenario adaptation coefficient calculation, that is, the decrease factor × the original feature vector is used in the distance calculation during retrieval, so that the model pays more attention to recent valid events. Through the above mechanism, the entire system realizes a complete closed loop from data acquisition, feature extraction, fusion evaluation, collaborative control to model self-evolution, supporting the dynamic reliability assessment and active defense of integrated electrical systems in complex and ever-changing environments. In order to enable those skilled in the art to fully understand and implement the present invention, the following further supplements the specific implementation principle of the present invention in conjunction with a specific application scenario. In the integrated electrical system of an industrial park in a coastal city, distributed photovoltaic clusters, electrochemical energy storage stations, combined cooling, heating and power (CCHP) units, and a large number of adjustable industrial loads are deployed. The system operation is affected by multiple disturbances, including heavy rainfall during typhoon season, high temperature and humidity, accelerated equipment aging, and fluctuations in user-side response. To address the reliability assessment needs under this complex scenario, the method of this invention is implemented according to the following steps: First, intelligent data acquisition units were deployed at the 110kV substation busbar outgoing terminals, 10kV feeder branch nodes, rooftop photovoltaic grid connection points, energy storage PCS interfaces, and smart meters of key enterprise users in the park. Each intelligent data acquisition unit simultaneously collected data from the operation layer and equipment status layer, including the three-phase voltage output from the voltage transformer (accuracy class 0.2), the line current measured by the current transformer, the power factor, and the 5th harmonic distortion rate. Simultaneously, it acquired the partial discharge pulse count at the cable terminal recorded by the partial discharge sensor, the temperature rise curve of the ring main unit fed back by the infrared temperature measurement module, and the mechanical vibration amplitude of the transformer monitored by the vibration accelerometer. These data were spatiotemporally aligned in the ARM Cortex-M7 core with a 10ms window to construct a directed graph containing 32 electrical nodes and 48 branches. The vertex feature vector consists of the node voltage amplitude (per unit), voltage phase angle (radians), and equipment health index, which is calculated using the formula... calculate, To measure the insulation resistance, This refers to the partial discharge quantity. The temperature rise value is used; the edge feature vector integrates the impedance magnitude, real-time power flow direction indicator (+1 indicates forward, -1 indicates reverse) and the number of times the corresponding circuit breaker has operated in the past 30 days from the line impedance parameter library; this graph structure is input to a graph neural network guided by physical constraints, and its first layer graph convolution kernel is expanded to K=3 using Chebyshev polynomials, with the convolution weight matrix... During the training phase, Kirchhoff's current law constraint is introduced using the Lagrange multiplier method. This ensures that the embedding space satisfies physical conservation. After two layers of graph convolution and residual connection, a 128-dimensional vulnerability embedding for each node is output. Then, a structural vulnerability vector is generated through global average pooling. The larger the value of this vector, the more significant the structural weakness in the system topology. Meanwhile, the second channel of the intelligent acquisition unit receives environmental layer, user behavior layer, and communication network status data: the meteorological micro-station provides minute-updated temperature (38.2℃), humidity (85%), wind speed (6.3m / s), rainfall intensity (12mm / h), and solar irradiance (210W / m²); smart meters of key enterprise users upload air conditioner start / stop frequency (up to 7 times in the past 10 minutes), demand response command response delay (4.2 seconds), and load fluctuation amplitude (ΔP=180kW); the communication network management system reports end-to-end latency jitter (standard deviation 18ms) and packet loss rate (0.7%) of the MQTT link; the above 12-dimensional data are spliced ​​into a time series of length T=60, divided into 4 segments by a sliding window (window length 30, step size 10), and each segment is input into a lightweight time encoder; the encoder sends the first 5-dimensional environmental data into a one-dimensional convolutional layer. The system extracts local patterns such as "sustained high temperature accompanied by high humidity"; the last 7 dimensions of user and communication data are processed by another convolutional layer to capture the composite disturbance features of "high-frequency load switching superimposed on communication degradation"; the two outputs are respectively fed into a bidirectional GRU (hidden_size=64), and its final forward hidden state... With backward hidden state The vector is concatenated into a 128-dimensional perturbation sensitivity feature vector, which reflects the system's dynamic response sensitivity to external perturbations. When the preliminary reliability score calculated by the local embedded inference engine (based on the output of the fully connected network of the 256-dimensional local reliability feature package) is 0.63, which is lower than the threshold of 0.7, the intelligent acquisition unit will upload the local reliability feature package along with the current topology snapshot (switch state matrix, dimension 32×32) synchronously acquired by SCADA to the regional collaborative evaluation center via MQTT over TLS encryption. The regional collaborative assessment center accesses a historical operational event database containing 217 operational event records from 2021 to 2024 during Typhoons "Muifa" and "Haikui". For the received disturbance sensitivity feature vector, KD-Tree retrieves the K=10 events with the closest Euclidean distance in the feature space (e.g., the high temperature and humidity + frequent industrial load switching event on September 15, 2023), calculates their reliability levels (e.g., 0.58, 0.61, etc.) as a weighted average, and then obtains the scenario adaptation coefficient (α=0.65) through Sigmoid mapping, reflecting the similarity between the current operating conditions and historical high-risk scenarios. Subsequently, the structural vulnerability vector, disturbance sensitivity feature vector, and topological snapshot (converted to an adjacency matrix) are input into the cross-domain fusion model: the electrical structure branch uses an adjacency matrix as the graph structure and performs two layers of GAT (head=4) on the structural vulnerability vector to enhance the vulnerability representation of key nodes; the external disturbance branch uses two layers... The modeling perturbations internal dependencies; the topological dynamic branch flattens the adjacency matrix into a 1024-dimensional vector, which is then compressed into a 64-dimensional topological embedding through an FC layer; the fusion layer uses the GAT output as the query and the outputs of the other two branches as the key / value pairs, and calculates cross-domain interaction weights on 32 nodes through a multi-head graph attention mechanism, finally outputting a dynamic reliability probability (β=0.59) through global max pooling and MLP; the differentiable weighted gating mechanism uses the initial... = =0.5 Calculate the overall reliability index: γ = σ(0.5×0.65 + 0.5×0.59) = 0.62; After the intelligent acquisition unit analyzes the strategy, the overcurrent protection time limit for branch #123 is adjusted from 0.3s to 0.24s; simultaneously, in the graph neural network guided by physical constraints, the weights of the adjacent matrix elements corresponding to branch #456 are set to 0.1, and those for branch #123 are set to 0; in the next evaluation cycle, if the measured current... =185A, model prediction =192A, rated current =200A, then the weight is updated to w new = 0.1 + 0.01×(1 - |185-192| / 200) = 0.10965, achieving adaptive calibration of the model for the new topology; After the scheduling decision terminal verifies that there is no risk of asynchronous loop merging in the operation sequence, it generates an operation ticket and pushes it to the handheld terminal of the operation and maintenance personnel. The operation and maintenance personnel execute the operation on-site and transmit the operation record and verification data (including switch position photos and clamp meter readings). Based on this, the regional collaborative evaluation center constructs a supervision label (if no failure occurs within 30 minutes after the operation, it is marked as a positive sample) and performs a one-step SGD update on the attention weight matrix of the cross-domain fusion model. At the same time, the time decay factor of the event on September 15, 2023 in the event database is updated from 0.88 to 0.88×1.05=0.924, which improves its priority in future scenario matching. Through the above process, this invention achieves deep fusion of multi-source heterogeneous data, collaborative modeling of physical laws and data-driven approaches, and closed-loop evolution of evaluation-regulation-learning, effectively supporting the dynamic reliability quantification and proactive defense of integrated electrical systems under high-penetration renewable energy access.

[0023] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A comprehensive electrical system reliability assessment method integrating multi-source data, characterized in that, Includes the following steps: S1: Intelligent data acquisition units are deployed at substations, feeder nodes, distributed energy access points, and user-side metering terminals of the integrated electrical system to synchronously acquire operational layer data, environmental layer data, equipment status layer data, user behavior layer data, and communication network status data. Operational layer data includes voltage, current, power, frequency, and harmonic content; environmental layer data includes temperature, humidity, wind speed, rainfall, and light intensity; equipment status layer data includes insulation resistance, partial discharge, mechanical vibration, and temperature rise curves; user behavior layer data includes load response delay, frequency of power consumption mode switching, and demand-side regulation participation; and communication network status data includes latency jitter, packet loss rate, and link availability. All raw data are processed according to a unified time... After interpolation alignment, missing value imputation, outlier removal, and dimensional normalization are performed to form a multi-source heterogeneous data stream; S2: Each intelligent acquisition unit performs dual-channel feature construction based on a local embedded inference engine. The first channel spatiotemporally aligns the runtime layer data and device status layer data, then inputs it into a graph neural network guided by physical constraints to generate a structural vulnerability vector; the second channel concatenates the environmental layer data, user behavior layer data, and communication network status data into an external disturbance feature sequence, segments it through a sliding window, and inputs it into a lightweight time encoder to output a disturbance sensitivity feature vector; S3: The intelligent acquisition unit concatenates the structural vulnerability vector and the disturbance sensitivity feature vector into a local reliability feature package and binds it to the current system topology snapshot. When the preliminary reliability score calculated locally is lower than a set threshold, the feature package along with the topology snapshot is uploaded to the regional collaborative evaluation center; S4: After receiving data uploaded by multiple intelligent acquisition units, the regional collaborative evaluation center initiates a three-level fusion evaluation process: First, it uses the historical event database to perform similar scenario matching on the disturbance sensitivity feature vector to generate scenario adaptation coefficients; Second, it inputs the structural vulnerability vector, disturbance sensitivity feature vector, and topology snapshot into a cross-domain fusion model based on a multi-head graph attention mechanism to output a system-level dynamic reliability probability; Finally, it combines the scenario adaptation coefficient and the dynamic reliability probability to generate a comprehensive reliability index through a differentiable weighted gating mechanism; S5: The regional collaborative evaluation center, based on the comprehensive reliability... The reliability index, combined with predefined reliability level classification rules, determines the current system's reliability level and generates a collaborative control strategy that includes topology reconfiguration suggestions, resource scheduling instructions, and early warning area identifiers. This strategy is then distributed to relevant intelligent acquisition units and scheduling decision terminals. S6: After receiving the collaborative control strategy, the intelligent acquisition unit parses the topology reconfiguration suggestions and triggers local relay protection logic adjustments. Simultaneously, it updates the edge weight parameters in the graph neural network to reflect the vulnerability distribution under the new topology. The scheduling decision terminal generates an operation ticket based on the early warning area identifier and resource scheduling instructions and pushes it to the handheld terminal of the maintenance personnel. After the maintenance personnel execute the operation, they transmit the actual operation record and on-site verification data back to the regional collaborative evaluation center.S7: Based on the returned operation records and verification data, the regional collaborative evaluation center fine-tunes the attention weights of the cross-domain fusion model using an online gradient update method, and assigns time-decrease factors to similar scenario samples in the historical operation event database according to the effectiveness of the operation, thereby achieving continuous evolution of the evaluation model.

2. The method for comprehensive electrical system reliability assessment based on multi-source data as described in claim 1, characterized in that, In step S2, the process of generating the structural vulnerability vector is as follows: S211: Construct a directed graph with electrical nodes as vertices and branches as edges. The node features include voltage magnitude, phase angle, and equipment health index, and the edge features include impedance magnitude, power flow direction, and protection action history; S212: Input the graph into a two-layer graph convolutional network, with residual connections and layer normalization after each layer, and output the vulnerability embedding of each node; S213: Perform global average pooling on all node embeddings to obtain the structural vulnerability vector.

3. The method for comprehensive electrical system reliability assessment based on multi-source data as described in claim 1, characterized in that, In step S2, the generation process of the perturbation sensitivity feature vector is as follows: S221: Group the external perturbation feature sequence within the sliding window, with the environmental layer data as the first sub-sequence and the user behavior layer and communication network state data merged into the second sub-sequence; S222: Extract local patterns from the input one-dimensional convolutional layer of the two sub-sequences respectively, then capture long-term dependencies through a bidirectional gated recurrent unit, and finally concatenate the final hidden states of the two sub-sequences to form the perturbation sensitivity feature vector.

4. The method for comprehensive electrical system reliability assessment by integrating multi-source data according to claim 1, characterized in that, In step S4, the scenario adaptation coefficient is calculated as follows: the regional collaborative evaluation center maintains a historical operation event database, and each event contains a disturbance sensitivity feature vector, the meteorological code at the time of occurrence, the load type distribution, and the final reliability level; for the received disturbance sensitivity feature vector, the K nearest neighbor events are retrieved in the event database using Euclidean distance, and the weighted average of their reliability levels is calculated, with the weight being the reciprocal of the distance. This weighted average is then mapped by Sigmoid and used as the scenario adaptation coefficient.

5. The method for comprehensive electrical system reliability assessment based on multi-source data as described in claim 1, characterized in that, In step S4, the structure of the cross-domain fusion model includes: the input layer receives the structural vulnerability vector, the disturbance sensitivity feature vector, and the adjacency matrix corresponding to the topology snapshot; the intermediate layer consists of three parallel branches, which respectively process electrical structure information, external disturbance information, and topology dynamics; the fusion layer adopts a multi-head graph attention mechanism to interactively weight the three types of information at the node level, and finally generates the system-level dynamic reliability probability through a global readout function.

6. The method for comprehensive electrical system reliability assessment based on multi-source data as described in claim 1, characterized in that, In step S4, the specific implementation of the differentiable weighted gating mechanism is as follows: Let the scene adaptation coefficient be α and the dynamic reliability probability be β, then the comprehensive reliability index... Where σ is the Softmax function, and These are learnable parameters that are updated after each evaluation via backpropagation.

7. The method for comprehensive electrical system reliability assessment based on multi-source data as described in claim 1, characterized in that, Before step S5, the method further includes: calculating the rate of change of the comprehensive reliability index of the current system over the past N evaluation periods, which is obtained by the linear regression slope of the first-order difference sequence; constructing a two-dimensional reliability state plane with the comprehensive reliability index as the horizontal axis and the rate of change as the vertical axis, dividing it into four sub-regions: "high reliability stable region", "medium reliability fluctuating region", "low reliability deterioration region" and "emergency instability region".

8. The method for comprehensive electrical system reliability assessment based on multi-source data as described in claim 1, characterized in that, In step S5, the method for generating topology reconfiguration suggestions is as follows: based on the current topology snapshot and structural vulnerability vector, identify the M branches with the highest vulnerability, combine distributed energy output prediction and load distribution, solve the islanding scheme under the minimum load shedding, and encode the scheme as a switching operation sequence as a topology reconfiguration suggestion.

9. The method for comprehensive electrical system reliability assessment based on multi-source data as described in claim 1, characterized in that, In step S6, the intelligent acquisition unit updates the edge weights of the graph neural network by setting the initial weights of the edges corresponding to newly added or disconnected branches in the new topology to 0.1 or 0, and dynamically adjusting them in subsequent evaluation cycles based on the actual power flow and fault isolation effect. The adjustment formula is as follows: Where η is the learning rate, For the measured current, To predict the current for the model, This is the rated current.

10. The method for comprehensive electrical system reliability assessment based on multi-source data according to claim 1, characterized in that, In step S7, the specific process of online gradient update is as follows: construct the operation record and verification data into a supervision signal, calculate the cross-entropy loss between the cross-domain fusion model output and the actual reliability result, and perform one step of stochastic gradient descent only on the attention weight matrix; the update rule of the time decay factor is: if a historical event is successfully matched and the operation is valid, its decay factor is multiplied by 1.05, otherwise it is multiplied by 0.

95. The decay factor is used to scale the event feature vector in subsequent retrievals.

Images