A Method and System for Power Supply Reliability Prediction and Evaluation Based on Multi-Source Fusion

By constructing a cross-system mapping mechanism and a graph attention hybrid prediction model, combined with hierarchical incremental learning, the problems of multi-source data alignment and model adaptability were solved, achieving high-precision power supply reliability prediction and real-time assessment.

CN123117482APending Publication Date: 2026-11-17XIANNING POWER SUPPLY COMPANY OF STATE GRID HUBEIELECTRIC POWER +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611543798.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-30
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for power supply reliability prediction suffer from problems such as high maintenance costs due to the strong reliance on manual mapping for spatiotemporal alignment of multi-source heterogeneous data, model concept drift, and high consumption of computing resources, and are unable to adapt to dynamic changes in the power grid in real time.

Method used

By constructing an automated cross-system mapping mechanism based on the correlation between the electrical topology of the distribution network and the voltage curve of the equipment, embedding a graph attention hybrid prediction model, and adopting a hierarchical incremental closed-loop learning mechanism, the automated alignment of multi-source data and dynamic updating of the model are achieved.

Benefits of technology

It significantly improves the accuracy of power supply reliability prediction, reduces maintenance costs, realizes real-time adaptability of the model and efficient online evaluation, and solves the problems of accuracy degradation and computing power consumption of traditional models in dynamic power grid environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117482A_ABST
    Figure CN123117482A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for predicting and evaluating power supply reliability based on multi-source fusion. The method includes the following steps: acquiring multi-source heterogeneous data from distribution automation, dispatch management, and electricity consumption information collection systems; performing cross-system equipment mapping based on the distribution network feeder topology and Pearson correlation of equipment voltage curves to generate a uniformly encoded spatiotemporal feature tensor; inputting the tensor into a graph attention-bidirectional long short-term memory hybrid prediction model, aggregating topological feature encoding of outage propagation paths and extracting temporal dependencies, and outputting prediction indicators; calculating actual values ​​based on real-time data and comparing them with prediction indicators, triggering a hierarchical incremental learning mechanism to independently update different levels of the model at different frequencies. This invention achieves automated spatiotemporal alignment of heterogeneous data, the model deeply matches the physical propagation mechanism of the power grid, and the hierarchical closed-loop mechanism balances prediction agility and computational overhead, significantly improving prediction accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of power system automation and distribution network reliability analysis technology, and specifically relates to a method and system for predicting and evaluating power supply reliability based on multi-source fusion. Background Technology

[0002] Power supply reliability is a key indicator for measuring the continuous power supply capability of a distribution network. Currently, power supply reliability management mainly relies on the following three types of systems: (1) Distribution Automation System (SCADA / DA): Real-time collection of data such as distribution network switch status, feeder current and voltage, and equipment operating parameters; (2) Dispatch Management System (OMS): Carrying out grid dispatch operation management business, including dispatch management information such as equipment ledger, maintenance plan, operation ticket, and fault record; (3) Electricity Information Collection System: Collecting data such as electricity consumption, voltage, and power outage events on the user side through smart meters.

[0003] To improve the intelligent analysis level of distribution networks, existing technologies include schemes that integrate multi-source data and utilize deep learning models for analysis. For example, Chinese invention patent application CN121682129A discloses an intelligent fault assessment system and method for distribution networks based on artificial intelligence and multi-source data fusion. This scheme models the pre-processed multi-source data of the distribution network as a dynamic heterogeneous graph structure, embeds feature vectors into nodes and edges, uses a graph attention network to calculate the probability of fault propagation between nodes, and employs a long short-term memory network for temporal fault mode identification. Chinese invention patent application CN122046226A discloses a method and system for distribution network fault prediction based on multi-source data fusion. This scheme also utilizes a graph neural network model to aggregate the spatial dimensions of the fused heterogeneous feature vectors and uses a long short-term memory network model for temporal dimension analysis. When identifying network nodes, this scheme pre-establishes a global asset index table containing all physical devices in the distribution network and assigns a fixed global index ID to each physical device to achieve data mapping.

[0004] Despite the progress made by the aforementioned existing technologies in combining graph neural networks and time-series models, the following significant shortcomings still exist in practical engineering applications for power supply reliability prediction: The SCADA system uses dispatch naming conventions as units, the OMS uses asset management codes, and the electricity information collection system uses marketing user numbers; these three coding systems are incompatible. Current technologies still require the pre-establishment and maintenance of a large and fixed global asset index table for node mapping. This matching method, heavily reliant on manual experience and static configuration, is extremely costly and error-prone when the distribution network undergoes frequent upgrades, expansions, or topology reconfigurations. Furthermore, it lacks automated cross-system device topology association methods based on the inherent electrical mechanisms of the distribution network.

[0005] Although existing technologies can extract spatiotemporal features using GAT+LSTM models, the operating environment of power distribution networks is highly dynamic. Traditional models can lead to conceptual drift and decreased prediction accuracy over time; the latter consumes enormous computing resources and cannot meet the requirements of real-time online evaluation. Summary of the Invention

[0006] This invention provides a method and system for predicting and evaluating power supply reliability based on multi-source fusion. By constructing an automated cross-system mapping mechanism based on the correlation between the electrical topology of the distribution network and the voltage curve of the equipment, a graph attention hybrid prediction model embedded in the physical spatial topology of the distribution network, and a hierarchical incremental closed-loop learning mechanism based on the sensitivity differences of network parameters, this invention aims to solve the problems in existing technologies, such as insufficient fusion due to the strong reliance on manual mapping for spatiotemporal alignment of multi-source heterogeneous data, limited prediction accuracy due to the lack of perception of the causal propagation mechanism of power physics in general artificial intelligence models, and the high computational cost of model retraining and inability to adapt to dynamic changes in the power grid in real time due to the separation between the current prediction and online evaluation systems.

[0007] To address the aforementioned technical problems, this invention proposes a power supply reliability prediction and evaluation method based on multi-source fusion, comprising the following steps: Acquire multi-source data from power distribution automation, dispatch management, and electricity consumption information collection systems; The multi-source data is mapped based on the distribution network feeder topology and the Pearson correlation of equipment voltage curves to generate a unified coded spatiotemporal feature tensor. The spatiotemporal feature tensor is input into the prediction model, and the power supply reliability prediction index is output. The prediction model includes a graph attention network and a bidirectional long short-term memory network, and integrates the distribution network topology adjacency matrix and the outage causal propagation chain.

[0008] Preferably, after outputting the predicted index, the method further includes online evaluation and closed-loop update steps: The actual value of power supply reliability is dynamically calculated based on the real-time access data stream and compared with the predicted index. When the prediction deviation exceeds a preset threshold, a hierarchical incremental learning mechanism is triggered. Based on the differences in the sensitivity of the model parameters to changes in the power grid state, the output layer, the time-series network layer, and the full parameters of the hybrid prediction model are updated independently.

[0009] Preferably, the independent update includes: The output layer parameters of the prediction model are periodically adjusted using the latest data at a first frequency, and the parameters of the graph attention network layer and the bidirectional long short-term memory network layer are frozen. The parameters of the bidirectional long short-term memory network layer and the temporal attention layer are periodically adjusted at a second frequency, and the parameters of the graph attention network layer are frozen. The prediction model is retrained periodically using historical global data at a third frequency; wherein the first frequency > the second frequency > the third frequency.

[0010] Preferably, the method for generating the spatiotemporal feature tensor is as follows: Based on the feeder topology of the power distribution automation system, a topology skeleton is constructed, and the set of equipment nodes and electrical connection edges are extracted. The device names in the scheduling management system are matched with the scheduling names of the device nodes using fuzzy string matching to establish a first mapping relationship; Extract the feeder identifier from the transformer area number in the electricity information collection system and connect it to the corresponding distribution transformer node. Calculate the Pearson correlation coefficient between the transformer area's total meter voltage curve and the low-voltage side voltage measurement curve of the distribution transformer. When the correlation coefficient is greater than a preset threshold, establish a second mapping relationship. Based on the first and second mapping relationships, a unified code is generated for each device node in the topology skeleton.

[0011] Preferably, the generation step of the spatiotemporal feature tensor further includes time dimension alignment processing: A unified time grid is established, and the continuous measurement data in the multi-source data are linearly interpolated and resampled on the unified time grid. For discrete event data in the multi-source data, the number of events and cumulative impact duration within each time grid interval are counted and transformed into continuous features of the corresponding time grid to construct a spatiotemporal feature tensor containing device nodes, time steps and multidimensional features.

[0012] Preferably, the graph attention network integrates the distribution network topology adjacency matrix and the outage causal propagation chain, specifically in the following manner: Based on the feeder topology of the power distribution network, an adjacency matrix representing the electrical connection relationships between devices is constructed; In the graph attention layer, the adjacency matrix is ​​used to calculate the attention coefficient between the target device node and its neighboring device nodes; The features of adjacent device nodes are weighted and aggregated using the attention coefficient, so that the feature representation of the target device node encodes the operating status of its upstream and downstream neighboring nodes and the impact of power outage propagation.

[0013] Preferably, the prediction model employs a multi-task learning architecture to simultaneously predict power supply reliability, average user outage time, and average user outage frequency, specifically including: The bidirectional long short-term memory network, combined with a temporal attention mechanism, outputs a temporal feature representation for each device node; The time-series features are represented as shared common features and input to three independent fully connected layers, which synchronously output the predicted values ​​of power supply reliability, average user outage time, and average user outage frequency.

[0014] A second aspect of the present invention also proposes a power supply reliability prediction and evaluation system based on multi-source fusion, the system being used to implement the prediction and evaluation method as described in the first aspect of the present invention, comprising: The data acquisition module is used to acquire multi-source data from power distribution automation, dispatch management, and electricity consumption information collection systems. The spatiotemporal feature mapping module is used to map the multi-source data based on the distribution network feeder topology and the Pearson correlation of equipment voltage curves to generate a uniformly encoded spatiotemporal feature tensor. The prediction module is used to input the spatiotemporal feature tensor into the prediction model and output a power supply reliability prediction index; the prediction model includes a graph attention network and a bidirectional long short-term memory network, and integrates the distribution network topology adjacency matrix and the outage causal propagation chain; and / or The online evaluation and closed-loop update module is used to dynamically calculate the actual value of power supply reliability based on the real-time access data stream and compare it with the power supply reliability prediction index. When the prediction deviation exceeds the preset threshold, the hierarchical incremental learning mechanism is triggered. According to the difference in the sensitivity of the model parameters to changes in the power grid state, the output layer, the time-series network layer and the full parameters of the hybrid prediction model are updated independently.

[0015] A third aspect of the present invention also provides an electronic device comprising: One or more processors; Memory, used to store one or more computer programs; One or more computer programs stored in the memory are executed by the one or more processors, causing the one or more processors to implement the prediction and evaluation method as described in the first aspect of the invention.

[0016] In a fourth aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the prediction and evaluation method as described in the first aspect of the present invention.

[0017] Compared with the prior art, the present invention has the following technical effects: 1. The prediction and evaluation method proposed in this invention integrates multi-source heterogeneous data acquisition, spatiotemporal automatic alignment, deep learning prediction, and online dynamic evaluation and incremental model update, breaking down the business barriers of traditional prediction and evaluation functions, and significantly improving the prediction accuracy of power supply reliability and the efficiency of lean operation and maintenance management of distribution networks under complex dynamic conditions.

[0018] 2. The prediction and evaluation method proposed in this invention uses the inherent electrical topology of the distribution network and the Pearson correlation of equipment voltage curves as physical anchors for cross-system association, realizing accurate physical-level alignment of heterogeneous data at the bottom layer of distribution automation, dispatch management and electricity information collection systems; this mechanism eliminates the dependence on manual maintenance of static mapping tables, can automatically adapt to the switching of power grid equipment and topology reconstruction, and solves the industry pain point of non-interoperability of underlying data encoding.

[0019] 3. The prediction and evaluation method proposed in this invention introduces a graph attention network into the prediction architecture, which explicitly encodes the real physical topology adjacency relationship of the distribution network into the model structure, enabling the model to perceive the downstream cascading propagation path of power outage events along the feeder topology. Compared with general time series prediction models, the hybrid model of this invention integrates both physical spatial topology and time series evolution prior knowledge, improving the prediction accuracy in power-specific scenarios.

[0020] 4. The prediction and evaluation method proposed in this invention proposes a hierarchical incremental learning mechanism. By setting high, medium and low independent update frequencies for the output layer, time series layer and full parameters, it not only ensures the model's rapid and agile response to recent changes in power grid operation, but also effectively avoids the huge computing power overhead and overfitting risk caused by high-frequency full model retraining. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the prediction and evaluation method described in this invention; Figure 2 This is a schematic diagram of automatic cross-system device association based on topology provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0023] Example 1 This embodiment provides a power supply reliability prediction and evaluation method based on multi-source fusion, such as... Figure 1 As shown, it includes the following steps one through three: Step 1: Acquire multi-source data from power distribution automation, dispatch management, and electricity consumption information collection systems.

[0024] Power supply reliability is comprehensively affected by the physical operating status of equipment, dispatching and management behavior, and actual power consumption on the user side; a single data source cannot truly reflect the overall picture of the power grid. Therefore, this embodiment independently collects heterogeneous data from the following three dimensions: First, data reflecting the real-time physical state of the distribution network is obtained from the Supervisory Control and Data Automation (SCADA / DA) system. This includes, but is not limited to: the opening and closing sequence of each switch, power flow data of the lines (such as three-phase current, three-phase voltage, active power, and reactive power), measurement data of distribution transformers (such as load factor), and the equipment topology connection data at the lower level of the distribution network. This topology connection refers to the physical spatial connection framework between switches, lines, and distribution transformers, which can typically be directly derived from the feeder topology model of the distribution automation system.

[0025] Secondly, data reflecting the power grid management status and equipment static attributes are obtained from the Operational Management System (OMS). This includes, but is not limited to, equipment ledger information, various maintenance plans, and historical fault records. Specifically, equipment ledger information includes at least static characteristic parameters such as equipment name, equipment model code, and commissioning date / years of service; maintenance plans include the number of planned maintenance items for the day and the number of households scheduled for power outages; and historical fault records include the number of sliding window fault trips in the past 30 days.

[0026] Secondly, data reflecting the actual perceived status on the user side should be obtained from the electricity consumption information collection system. This includes, but is not limited to: transformer area number, daily electricity consumption of users (total electricity consumption of transformer area meters), voltage qualification rate, user power outage event records, and the number of user repair work orders; among which, user power outage event records should include the start and end time of the power outage, the cumulative power outage duration, and the number of affected users.

[0027] At the beginning of the day-ahead prediction link in this embodiment, the system connects in parallel with the underlying distribution automation system (SCADA / DA) database, OMS database, and electricity consumption information collection system database through a multi-source data acquisition module. This completes the panoramic extraction of multi-source heterogeneous data from the underlying layers of the three major systems, thereby providing original underlying data support for subsequent cross-system device association, spatiotemporal alignment, and feature fusion.

[0028] To help those skilled in the art better understand the scale and specific objects of data acquisition in Step 1, a concrete example scenario is provided: acquiring panoramic data of a 10kV feeder in a certain region. First, the basic physical topology and equipment operation measurement data of the feeder are exported from the distribution automation system. This confirms that the feeder includes one substation outgoing switch, five sectionalizing switches, three tie switches, and twelve distribution transformers. The scheduling names of these nodes are obtained, such as "Sectionalizing Switch of 10kV Industrial Line #01 Pole in Substation A". Simultaneously, equipment ledgers and historical maintenance records strongly related to this specific feeder are extracted and imported from the OMS dispatch management system, for example, a total of 21 relevant records are obtained. At the same time, electricity consumption and outage event information of the transformer substations connected to this feeder are selectively extracted from the electricity consumption information acquisition system, for example, relevant data for 12 transformer substations are obtained. Through the above multi-source concurrent acquisition operations, the construction of the original basic dataset for this feeder in the three major business systems is completed.

[0029] Step 2: Based on the distribution network feeder topology and the Pearson correlation of equipment voltage curves, the multi-source data is mapped to generate a unified coded spatiotemporal feature tensor.

[0030] In this embodiment, the distribution automation system uses switches, lines, and transformers as equipment units and adopts the dispatch naming standard; the dispatch management system uses equipment ledgers as management units and adopts asset management codes; and the electricity consumption information collection system uses transformer area-user as collection units and adopts marketing user numbers. The underlying equipment coding systems of the three systems are incompatible and cannot be directly linked. This step utilizes the inherent electrical topology connection relationship of the distribution network as the anchor point for cross-system equipment association to achieve physical-level data fusion.

[0031] Furthermore, for the aforementioned spatiotemporal feature tensor, this embodiment provides a specific generation method, which specifically includes: A topology skeleton is constructed based on the feeder topology of the power distribution automation system, and the set of equipment nodes and electrical connection edges is extracted. Combined with... Figure 2 As shown, the feeder topology is first derived from the distribution automation system and represented as a directed graph G=(V,E). Here, V is the set of equipment nodes, including physical equipment such as substation outgoing switches, sectionalizing switches, tie switches, and distribution transformers, and each equipment node... It records its scheduling name and topology location in the distribution automation system, for example... Figure 2 The 10kV XX line outgoing switch and the #01 pole sectionalizing switch are mentioned; E is the set of electrical connection edges representing the actual physical connection between the above devices.

[0032] The device names in the dispatch management system are matched with the dispatch names of the device nodes using fuzzy string matching to establish a first mapping relationship. The feeder identifier in the transformer area number of the electricity information collection system is extracted and linked to the corresponding distribution transformer node. The Pearson correlation coefficient between the transformer area's total meter voltage curve and the low-voltage side voltage measurement curve of the distribution transformer is calculated. When the correlation coefficient is greater than a preset threshold, a second mapping relationship is established.

[0033] like Figure 2 The illustrated connection process for the electricity consumption information collection system involves first extracting the feeder identifier (e.g., the first 6 digits of the distribution transformer number) from the distribution transformer number for coarse matching, since the distribution transformer area corresponds to the distribution transformer in the distribution network in terms of physical attributes. This coarse matching connects the target distribution transformer area to the distribution transformer node on the same feeder in the topology skeleton. To resolve the association ambiguity caused by multiple distribution transformers on the same feeder, the system further extracts the voltage curve of the distribution transformer's total meter and the voltage measurement curve of the low-voltage side of the distribution transformer in the distribution automation system, and calculates the Pearson correlation coefficient between the two. In this embodiment, when the correlation coefficient is >0.85, it is considered that the electrical characteristics are highly consistent, and the matching is successful, thus establishing a second mapping relationship from the distribution transformer area to the distribution transformer node.

[0034] Based on the first and second mapping relationships, a unified code is generated for each device node in the topology skeleton. For example... Figure 2 As shown in the unified coding generation module, after the connection of the two major peripheral systems is completed, a unified code is generated for each device node in the topology skeleton. The coding rule can be set as: UCODE = Feeder ID + Topology Level + Device Type + Serial Number. This unified code will serve as a unique primary key, completely integrating and aggregating the data from the three systems. In addition, the coverage rate of the three systems' data can be calculated for each node. If the data missing rate of a node exceeds 30%, it is marked as having insufficient association quality and will be assigned a lower weight or subjected to neighbor data imputation in subsequent calculations.

[0035] To provide a more concrete understanding of the spatial mapping process for those skilled in the art, a scenario example from step one is used: Taking a 10kV feeder in a certain region as an example, its topology includes one outgoing switch, five sectionalizing switches, three tie switches, and 12 distribution transformers. 21 ledger records are imported from the OMS. For example, the edit distance between the OMS record "Industrial Line #01 Pole Sectional Switch" and the topology node "A Substation 10kV Industrial Line #01 Pole Sectional Switch" is 6, with a confidence level of 0.92, indicating a successful match. Twelve distribution areas are imported from the electricity consumption information collection system. Through feeder prefix filtering and the aforementioned voltage curve correlation coefficient calculation, a precise one-to-one match between the 12 distribution areas and the 12 distribution transformers is achieved. For example, a unified code "FEEDER_IND_01_SW" is ultimately generated for the #01 pole sectionalizing switch. Verification shows that the automated association accuracy of this mechanism reaches 94.7%.

[0036] After completing the spatial alignment of the physical topology, due to the frequency differences in multi-source data (e.g., distribution automation typically involves continuous measurements at the 5-minute level, while electricity consumption information acquisition systems generate hourly discrete and continuous data), the generation step of the spatiotemporal feature tensor in this embodiment further includes time dimension alignment processing. A unified time grid is established, and continuous measurement data from the multi-source data are linearly interpolated and resampled on the unified time grid. In specific implementation, time slots are divided with a basic time interval of 1 hour, and the entire day is divided into 24 grid nodes. For continuous measurement data (such as current and voltage at the second / minute level) acquired by the power distribution automation system and continuous data such as electricity consumption data from the electricity information collection system, linear interpolation or averaging operations are performed on the aforementioned unified 1-hour time grid.

[0037] For discrete event data in the multi-source data, the number of events and cumulative impact duration within each time grid interval are counted and transformed into continuous features of the corresponding time grid to construct a spatiotemporal feature tensor containing device nodes, time steps and multidimensional features.

[0038] For sporadic discrete event data such as OMS maintenance plans, fault trip events, and user power outage events in the electricity information collection system, the frequency of occurrence and cumulative power outage duration are statistically analyzed within their respective time slot intervals, and then transformed into continuous feature dimensions suitable for deep learning processing.

[0039] Through the above spatial alignment and temporal unification operations, a final result based on device nodes was generated. and time slot Three-dimensional spatiotemporal feature tensor with dual indexes .in, This represents the total number of device nodes in the topology skeleton. Representing a 24-hour node. This represents the total number of feature dimensions. For example, in this embodiment, a total of 27 features were extracted, including electrical measurements, static ledgers, and event statistics. This structured spatiotemporal feature tensor provides standardized data input for subsequent power supply reliability prediction models.

[0040] The 27-dimensional features used in this embodiment are shown in Table 1 below: Table 1. Specific Components of Input Features

[0041] Step 3: Input the spatiotemporal feature tensor into the prediction model and output the power supply reliability prediction index; the prediction model includes a graph attention network and a bidirectional long short-term memory network, and integrates the distribution network topology adjacency matrix and the outage causal propagation chain.

[0042] Since conventional general-purpose prediction models (such as conventional Attention-BiLSTM) only process time series data and cannot perceive the topology of the distribution network and the physical propagation path of power outage events, this step, by inputting the 27-dimensional spatiotemporal feature tensor generated in step two, embeds topological adjacency priors that are highly correlated with the operating states of adjacent devices, as well as outage causal propagation priors that allow power outage events to propagate upstream and downstream along the feeder topology, into the model. This significantly improves the prediction accuracy in power scenarios.

[0043] Specifically, the graph attention network integrates the distribution network topology adjacency matrix and the outage causal propagation chain. The method is as follows: an adjacency matrix representing the electrical connection relationship between devices is constructed based on the distribution network feeder topology; in the graph attention layer, the attention coefficient between the target device node and its adjacent device nodes is calculated using the adjacency matrix; the features of adjacent device nodes are weighted and aggregated using the attention coefficient, so that the feature representation of the target device node encodes the operating status of its upstream and downstream neighboring nodes and the impact of outage propagation.

[0044] In a physical sense, this graph attention mechanism enables the model to automatically learn and identify which neighboring devices have a greater impact on the power supply reliability of the target device. For example, the model can automatically learn the causal chain that upstream switch tripping inevitably leads to downstream transformer outages, thus assigning higher attention weights to the abnormal states of upstream devices during feature aggregation.

[0045] After aggregating spatial topological features, the model further processes the extraction of temporal features and outputs the final target. The prediction model employs a multi-task learning architecture to simultaneously predict power supply reliability, average user outage time, and average user outage frequency, specifically including: The bidirectional long short-term memory network, combined with a time attention mechanism, outputs a temporal feature representation for each device node. This temporal feature representation is then used as a shared common feature and input into three independent fully connected layers, simultaneously outputting the predicted values ​​for power supply reliability, average user outage time, and average user outage frequency.

[0046] Through the above process, the system can output the predicted values ​​of the core power supply reliability indicators for each hour from 0:00 to 24:00 the following day at a fixed time every day (e.g., 8:00 AM), i.e., output tensors. .

[0047] The following provides a unified and detailed description of the specific network structure and training strategy of the prediction model involved in step three above.

[0048] From the perspective of the overall physical architecture, the structure of this prediction model is arranged from top to bottom, with data flowing sequentially through: the input layer, the graph attention layer, the temporal attention BiLSTM layer, and the multi-task output layer. Its core processing logic lies in the fact that the model first uses GAT to connect the physical topology in the spatial dimension, and then inputs the results into the BiLSTM to mine historical patterns in the temporal dimension.

[0049] The original 3D data tensor first enters the GAT layer. The GAT layer is responsible for encoding the topology of the distribution network into the model, allowing each device node to preferentially aggregate the feature information of its neighboring nodes. The input to the graph attention layer is the spatiotemporal tensor generated in step two. For each time slot Construct the feature matrix of device nodes Construct an adjacency matrix based on the distribution network topology skeleton. ,in Indicates equipment With equipment There is an electrical connection. For the node... and its neighboring nodes Using a learnable weight matrix and attention vector Calculate the coefficients:

[0050] in, Indicates in time slot Next, target device node Its adjacent device nodes The unnormalized attention coefficients (raw relevance scores) between them; Represents a learnable attention vector The transpose of the matrix is ​​used to map the concatenated features to a scalar score; This represents the learnable weight matrix. This is used to perform linear transformations on input features to increase or reduce dimensionality. and Representing time slots Next, target device node with neighboring nodes The original input feature vector; This represents the vector concatenation operation, which combines nodes... and nodes The transformed features are connected together.

[0051] Then, Softmax normalization is performed to obtain... Utilizing the ELU (Exponential Linear Unit) activation function Perform feature weighted aggregation:

[0052] in, This represents the attention weight coefficients after normalization using the Softmax function, and represents the neighboring nodes. For the target node Weighting of the impact of power supply reliability; Indicates the device node in the distribution network topology. The set of all adjacent nodes; This represents the ELU activation function, used to introduce non-linear expressive power. The final output of this layer is a feature tensor after topological aggregation. Preferably, .

[0053] The BiLSTM layer takes the aggregated feature tensor output by the GAT layer as its direct input. Building upon the completed topological information fusion, the BiLSTM layer extracts the forward and backward dependencies in the time series for each device node.

[0054] Based on topology aggregation, for each device node The time-series input is extracted using a bidirectional LSTM to extract temporal dependencies. In this embodiment, the BiLSTM (Bidirectional Long Short-Term Memory) network is configured with two layers, with a hidden layer dimension of 128, including 64 dimensions for the forward pass and 64 dimensions for the backward pass. A temporal attention mechanism is introduced that focuses on time steps rather than feature dimensions: for each time step... Hidden state Calculate its importance weight relative to the entire sequence. Finally, the hidden states of all time steps are weighted and summed to output a temporal feature vector for each device node that incorporates spatiotemporal information. .

[0055] To simultaneously predict power supply reliability (RS-1), average user outage time (SAIDI), and average user outage frequency (SAIFI), the model shares the above... As a public characteristic:

[0056]

[0057]

[0058] In the formula, This represents the final output device node after processing by the GAT and BiLSTM layers. A comprehensive feature vector integrating spatiotemporal information, with dimensions of [missing information]. ; The learnable weight matrices in the fully connected layers corresponding to three independent prediction tasks, each with dimension [missing information]. ; These correspond to the bias terms of three independent prediction tasks; This represents the Sigmoid activation function, which maps the output value to the (0,1) interval, and then multiplies it by 100 to convert it into a percentage; The rectified linear unit is used to ensure that the output values ​​of the two predictive indices, SAIDI and SAIFI, are non-negative.

[0059] Finally, a regional-level indicator is obtained by weighting the predicted values ​​of all device nodes by the proportion of users.

[0060] The model employs a multi-task joint loss function. Joint optimization can be performed, and the weights can be configured based on the dimensions and importance of the indicators. , , Where L represents the total loss function value during model training; This represents the mean squared error, which is the expected value of the square of the difference between the predicted value and the true label value of the corresponding indicator. This represents the joint loss weight coefficients assigned to different prediction tasks, used to balance the uneven gradient updates caused by differences in the dimensions of various indicators and the importance of business operations.

[0061] Furthermore, to ensure the model correctly learns the physical mechanisms, this embodiment employs a two-stage freeze training strategy: first, the BiLSTM layer is frozen, and the GAT layer is trained separately for 10 epochs to allow the topological aggregation capability to converge preferentially; then, all parameters are unfrozen, and the model is jointly trained until it converges overall, with early stopping used to prevent overfitting. Ablation experiments demonstrate that after embedding topological information, this mechanism significantly reduces the relative prediction error by 37.8%.

[0062] In a preferred embodiment, after outputting the predicted index, the method further includes a fourth step of online evaluation and closed-loop update: The actual value of power supply reliability is dynamically calculated based on the real-time access data stream and compared with the predicted index.

[0063] During implementation, real-time measurement data from the power distribution automation system (updated every 5 minutes) and real-time power outage event data from the electricity consumption information collection system are continuously input through stream processing platforms (such as Apache Kafka and Apache Flink). Based on this streaming data, the system calculates the actual value of the current power supply reliability index at a relatively high frequency (such as every 15 minutes). This calculation process typically uses a sliding window mechanism to statistically analyze data from the past hour. Subsequently, for each device node... and the current moment The prediction deviation is calculated using a preset formula, for example. ,in It is a very small constant to prevent division by zero anomalies. In this embodiment, when the regional average deviation... When the threshold exceeds 5%, the system will trigger an early warning mechanism and a subsequent incremental learning loop.

[0064] When the prediction deviation exceeds a preset threshold, a hierarchical incremental learning mechanism is triggered. Based on the differences in the sensitivity of model parameters to changes in grid conditions, the output layer, temporal network layer, and all parameters of the hybrid prediction model are updated independently. In this mechanism, the graph attention network layer encodes the relatively static underlying topology of the distribution network, which changes the slowest; while the multi-task output layer directly maps to the final index and is particularly sensitive to recent short-term fluctuations. Therefore, setting a stepped update frequency for different layers enables the model to have both agile response and long-term stability.

[0065] In this embodiment, the independent update includes: periodically adjusting the output layer parameters of the prediction model using the latest data at a first frequency, and freezing the parameters of the graph attention network layer and the bidirectional long short-term memory network layer; periodically adjusting the parameters of the bidirectional long short-term memory network layer and the temporal attention layer at a second frequency, and freezing the parameters of the graph attention network layer; and periodically retraining all parameters of the prediction model using historical global data at a third frequency; wherein, the first frequency > the second frequency > the third frequency.

[0066] Specifically, the first frequency is high-frequency, such as once a day, fixed after the daily prediction execution, using the actual operating data of the previous day for calibration. In the first-level update, the system strictly freezes the parameters of the GAT layer and BiLSTM layer, and only performs rapid fitting optimization on the weights and bias parameters of the three fully connected output layers. Preferably, the optimization objective is to minimize the weighted sum of squares of the prediction bias over the past 7 days, and assigns higher weights to more recent samples. This adjustment can be completed within minutes, greatly improving the model's sensitivity to recent changes in operating conditions such as temporary maintenance and short-term load fluctuations.

[0067] The second frequency is a medium frequency, for example, executed once a week, using a sliding data window from the past four weeks for fine-tuning. In the second-level update, the system focuses on updating the core time series extraction module (BiLSTM and the time attention layer). The technical purpose is to enable the model to adapt to seasonal climate changes and cyclical changes in electricity consumption, such as the grid load switching from summer high-temperature cooling mode to winter heating mode. This process consumes moderate computing power and will not cause the risk of full overfitting.

[0068] The third frequency is an extremely low frequency, such as once per quarter, and is usually scheduled during the spring and autumn seasons when the power grid operation mode and load are relatively stable. The system will extract all historical fusion data from the past year and thoroughly retrain all network parameters, including the GAT layer, from scratch. This aims to enable the model to learn deeply and adapt to major physical evolutions at the bottom layer of the distribution network, such as the commissioning of new feeders, network structure reconfiguration, and the batch replacement of old equipment.

[0069] To enable those skilled in the art to clearly understand the significant effects of the above-mentioned layering mechanism in practical engineering applications, a test example of a running scenario is provided here: A regional power grid experienced a sudden and prolonged period of extreme high temperatures during the summer, leading to a surge in load and a sudden outbreak of equipment failures. This caused severe distortion in the day-ahead forecasting model, which was originally trained based on stable spring data, resulting in prediction biases as high as 8% to 12%. After three consecutive days of bias exceeding the limit (greater than 5%), the system seamlessly initiated a first-level high-frequency update, adjusting only the output layer parameters. After calibration, the bias quickly dropped to 5.2% on the fourth day and decreased to 4.1% on the fifth day. Entering the third week of the high-temperature period, the system automatically triggered a second-level weekly fine-tuning, updating the time-series parameters to keep the bias stable within 3% for an extended period. After the summer cycle ended, the system initiated a quarterly full retraining to fully absorb the potential impact of aging equipment. Comparative data shows that after six months of continuous operation, the mean absolute percentage error (MAPE) of this hierarchical closed-loop scheme remained within an excellent range of 2.5%. In contrast, the MAPE of the static model that was not updated deteriorated to 5.8% after three months. If the traditional daily full model retraining scheme is adopted, although the accuracy is maintained, it takes about 4 hours, and its computational resource consumption is more than 10 times that of the daily 15-minute high-frequency fine-tuning scheme in this embodiment.

[0070] Example 2 This embodiment provides a power supply reliability prediction and evaluation system based on multi-source fusion. The system is used to implement the prediction and evaluation method as described in Embodiment 1, including: The data acquisition module is used to acquire multi-source data from power distribution automation, dispatch management, and electricity consumption information collection systems.

[0071] The spatiotemporal feature mapping module is used to map the multi-source data based on the distribution network feeder topology and the Pearson correlation of equipment voltage curves to generate a uniformly encoded spatiotemporal feature tensor.

[0072] The prediction module is used to input the spatiotemporal feature tensor into the prediction model and output the power supply reliability prediction index; the prediction model includes a graph attention network and a bidirectional long short-term memory network, and integrates the distribution network topology adjacency matrix and the outage causal propagation chain.

[0073] Furthermore, in a preferred embodiment, in order to achieve a closed-loop model retraining for day-ahead prediction and real-time evaluation, the system further includes an online evaluation and closed-loop update module, specifically used for: dynamically calculating the actual value of power supply reliability based on the real-time access data stream, and comparing it with the power supply reliability prediction index; when the prediction deviation exceeds a preset threshold, triggering a hierarchical incremental learning mechanism, and independently updating the output layer, time-series network layer, and all parameters of the hybrid prediction model according to the differences in the sensitivity of model parameters to changes in grid state.

[0074] Furthermore, the system also includes a visualization module, which is used to present the power supply reliability prediction index output by the reliability prediction module, as well as the actual power supply reliability value and prediction deviation warning information output by the online evaluation and closed-loop update module, to the dispatching and operation personnel in a graphical interface to provide visual decision support.

[0075] Example 3 This embodiment provides an electronic device, including: One or more processors; Memory, used to store one or more computer programs; One or more computer programs stored in the memory are executed by the one or more processors, causing the one or more processors to implement the prediction and evaluation method as described in Embodiment 1.

[0076] Example 4 This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the prediction and evaluation method as described in Embodiment 1.

[0077] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A power supply reliability prediction and evaluation method based on multi-source fusion, characterized in that, Includes the following steps: Acquire multi-source data from power distribution automation, dispatch management, and electricity consumption information collection systems; The multi-source data is mapped based on the distribution network feeder topology and the Pearson correlation of equipment voltage curves to generate a unified coded spatiotemporal feature tensor. The spatiotemporal feature tensor is input into the prediction model, and the power supply reliability prediction index is output. The prediction model includes a graph attention network and a bidirectional long short-term memory network, and integrates the distribution network topology adjacency matrix and the outage causal propagation chain.

2. The method according to claim 1, characterized in that, After outputting the predicted indicators, the method also includes online evaluation and closed-loop update steps: The actual value of power supply reliability is dynamically calculated based on the real-time access data stream and compared with the predicted index. When the prediction deviation exceeds a preset threshold, a hierarchical incremental learning mechanism is triggered. Based on the differences in the sensitivity of the model parameters to changes in the power grid state, the output layer, the time-series network layer, and the full parameters of the hybrid prediction model are updated independently.

3. The method according to claim 2, characterized in that, The independent update includes: The output layer parameters of the prediction model are periodically adjusted using the latest data at a first frequency, and the parameters of the graph attention network layer and the bidirectional long short-term memory network layer are frozen. The parameters of the bidirectional long short-term memory network layer and the temporal attention layer are periodically adjusted at a second frequency, and the parameters of the graph attention network layer are frozen. The prediction model is retrained periodically using historical global data at a third frequency; wherein the first frequency > the second frequency > the third frequency.

4. The method according to claim 1 or 2, characterized in that, The method for generating the spatiotemporal feature tensor is as follows: Based on the feeder topology of the power distribution automation system, a topology skeleton is constructed, and the set of equipment nodes and electrical connection edges are extracted. The device names in the scheduling management system are matched with the scheduling names of the device nodes using fuzzy string matching to establish a first mapping relationship; Extract the feeder identifier from the transformer area number in the electricity information collection system and connect it to the corresponding distribution transformer node. Calculate the Pearson correlation coefficient between the transformer area's total meter voltage curve and the low-voltage side voltage measurement curve of the distribution transformer. When the correlation coefficient is greater than a preset threshold, establish a second mapping relationship. Based on the first and second mapping relationships, a unified code is generated for each device node in the topology skeleton.

5. The method according to claim 4, characterized in that, The generation step of the spatiotemporal feature tensor also includes time dimension alignment processing: A unified time grid is established, and the continuous measurement data in the multi-source data are linearly interpolated and resampled on the unified time grid. For discrete event data in the multi-source data, the number of events and cumulative impact duration within each time grid interval are counted and transformed into continuous features of the corresponding time grid to construct a spatiotemporal feature tensor containing device nodes, time steps and multidimensional features.

6. The method according to claim 1 or 2, characterized in that, The graph attention network integrates the distribution network topology adjacency matrix and the outage causal propagation chain, specifically in the following way: Based on the feeder topology of the power distribution network, an adjacency matrix representing the electrical connection relationships between devices is constructed; In the graph attention layer, the adjacency matrix is ​​used to calculate the attention coefficient between the target device node and its neighboring device nodes; The features of adjacent device nodes are weighted and aggregated using the attention coefficient, so that the feature representation of the target device node encodes the operating status of its upstream and downstream neighboring nodes and the impact of power outage propagation.

7. The method according to claim 1 or 2, characterized in that, The prediction model employs a multi-task learning architecture to simultaneously predict power supply reliability, average user outage time, and average user outage frequency, specifically including: The bidirectional long short-term memory network, combined with a temporal attention mechanism, outputs a temporal feature representation for each device node; The time-series features are represented as shared common features and input to three independent fully connected layers, which synchronously output the predicted values ​​of power supply reliability, average user outage time, and average user outage frequency.

8. A power supply reliability prediction and evaluation system based on multi-source fusion, characterized in that, The system is used to implement the prediction and evaluation method as described in any one of claims 1-7, including: The data acquisition module is used to acquire multi-source data from power distribution automation, dispatch management, and electricity consumption information collection systems. The spatiotemporal feature mapping module is used to map the multi-source data based on the distribution network feeder topology and the Pearson correlation of equipment voltage curves to generate a uniformly encoded spatiotemporal feature tensor. The prediction module is used to input the spatiotemporal feature tensor into the prediction model and output a power supply reliability prediction index; the prediction model includes a graph attention network and a bidirectional long short-term memory network, and integrates the distribution network topology adjacency matrix and the outage causal propagation chain; and / or The online evaluation and closed-loop update module is used to dynamically calculate the actual value of power supply reliability based on the real-time access data stream and compare it with the power supply reliability prediction index. When the prediction deviation exceeds the preset threshold, the hierarchical incremental learning mechanism is triggered. According to the difference in the sensitivity of the model parameters to changes in the power grid state, the output layer, the time-series network layer and the full parameters of the hybrid prediction model are updated independently.

9. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs; The feature is that one or more computer programs stored in the memory are executed by the one or more processors, causing the one or more processors to implement the prediction and evaluation method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the prediction and evaluation method as described in any one of claims 1-7.

Citation Information

Patent Citations

  • Intelligent research and judgment system and method for multi-source data fusion fault of power distribution network based on artificial intelligence

    CN121682129A

  • Fusion power distribution network fault prediction method and system based on multi-source data

    CN122046226A