Power grid equipment BIM model parameter optimization system, electronic device and storage medium

Abstract

The present application relates to a kind of power grid equipment BIM model parameter optimization system, electronic equipment and storage medium, belong to data processing technical field, solve the problem that the existing BIM model optimization relies on artificial analysis and leads to long iteration cycle.It includes: component data processing module, for collecting the multi-source data of each component, respectively according to the multi-modal feature vector of each component in analysis time point extracted from each modal neural network model;Dynamic risk prediction module, for obtaining the multi-modal feature vector of each component in analysis time point output by component data processing module, and import into risk prediction model to obtain risk prediction result;Model parameter optimization module, for identifying the component to be optimized and its priority according to the risk prediction result output by dynamic risk prediction module;Based on parameter adjustment rule base, obtain the BIM model parameter adjustment rule of the component to be optimized, and form optimization scheme according to priority.It realizes the optimization of dynamic risk driven BIM model parameter.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a power grid equipment BIM model parameter optimization system, an electronic device and a storage medium. BACKGROUND

[0002] In the digital operation of power grid engineering, the system based on building information model (BIM) has become the core carrier, and the model data covers geometry, electricity, operation and maintenance, and environment and other multi-source heterogeneous information. However, the existing BIM platform and related system have obvious limitations in data processing and decision support.

[0003] At the data fusion level, the existing system lacks effective multi-source heterogeneous data correlation analysis capability. The traditional system usually uses rule engines or simple statistical analysis modules to process data, and these modules cannot deeply mine the complex and nonlinear implicit correlations between the BIM geometric structure, the electrical parameter time series curve, the unstructured operation text and the environmental monitoring data. For example, the system cannot automatically establish a deep relationship model between the "regional continuous heavy rain" data and the "foundation settlement risk" in the BIM model of the power transmission tower, resulting in that the data value is buried, forming a "data island", and the overall perception ability of the system is limited.

[0004] At the risk warning level, the built-in warning function of the existing system is mostly based on a static threshold judgment module. Such a module usually presets fixed rules (such as "temperature exceeding limit means alarm"), and cannot reflect the dynamic evolution trend of the risk, so the warning is seriously lagging. At the same time, the system lacks core components for time series modeling and analysis of historical and real-time data, and cannot capture the time series dependency between data, resulting in low prediction accuracy and insufficient foresight of the system.

[0005] At the model optimization level, the model updating and maintenance function of the existing BIM system is highly dependent on manual intervention. When receiving the feedback that "some components frequently fail and need to be adjusted in design", the engineer must manually find and modify the related component parameters in the BIM platform. The whole system lacks a core algorithm that can automatically mine the model optimization direction from multi-source data, resulting in a long BIM model iteration cycle, and the digital twin model cannot quickly adapt to the dynamic changes of the actual running state, so the effectiveness of the digital twin model is greatly reduced. SUMMARY

[0006] In view of the above analysis, the embodiments of the present application aim to provide a power grid equipment BIM model parameter optimization system, an electronic device and a storage medium, to solve the problem of long iteration cycle caused by the dependence of the existing BIM model optimization on manual analysis.

[0007] In one aspect, the embodiments of the present application provide a power grid equipment BIM model parameter optimization system, comprising:

[0008] The component data processing module is used to collect multi-source data of each component and extract the multi-modal feature vectors of each component at the analysis time according to the neural network model of each modality.

[0009] The dynamic risk prediction module is used to obtain the multimodal feature vectors of each component at the analysis time output by the component data processing module, and input them into the risk prediction model to obtain the risk prediction results. The risk prediction results include: component risk probability, component risk level, component failure type, equipment risk index, and equipment risk type.

[0010] The model parameter optimization module is used to identify components to be optimized and their priorities based on the risk prediction results output by the dynamic risk prediction module; obtain the BIM model parameter adjustment rules for the components to be optimized based on the parameter adjustment rule library, and form an optimization scheme according to the priority of the components to be optimized; identifying the components to be optimized includes: calculating the basic risk value of the component based on the component risk probability and component risk level by weighted summation; calculating the dynamic risk threshold of the component based on the equipment risk index, high risk threshold, and basic risk threshold; selecting components whose basic risk value is greater than their dynamic risk threshold as the first component; selecting neighboring components whose component risk probability is greater than the neighbor screening threshold and do not belong to the first component from the neighboring components of each first component as the second component; the neighbor screening threshold is less than the basic risk threshold; and merging the first and second components as the component to be optimized.

[0011] Based on further improvements to the above system, the parameter adjustment rule base is obtained through the following steps:

[0012] Natural language processing technology is used to parse power grid engineering design standards, technical specifications, and expert experience documents to construct a knowledge graph;

[0013] Based on historical operation and maintenance records and knowledge graphs, the mapping relationship between fault types and BIM model parameters is extracted, and a fault-parameter association table is established.

[0014] Using the fault-parameter association table and knowledge graph as the knowledge context, the large language model is guided to generate preliminary adjustment rules through prompt word templates. After being reviewed and corrected by domain experts, these rules are stored in the parameter adjustment rule library.

[0015] Based on further improvements to the above system, the basic risk value of a component is calculated using the following formula:

[0016] ,

[0017] in, Representation of components The basic risk value, and Representation of components a component risk probability and a component risk level, and respectively represent a weight of a component risk probability and a component risk level, represents a normalization function.

[0018] Based on the further improvement of the above system, the dynamic risk threshold is calculated by the following formula:

[0019] ,

[0020] wherein, represents a dynamic risk threshold of a component ; and respectively represent a preset high risk threshold and a basic risk threshold, ; represents a weight coefficient; represents a device risk index of a device to which a component belongs.

[0021] Based on the further improvement of the above system, the priority of a first component in the component to be optimized is obtained by calculating the product of the basic risk value of the first component and the topological centrality weight of the first component; the priority of a second component is obtained by calculating the product of the basic risk value of the second component, the edge weight between the second component and the first component associated with the second component, and the topological centrality weight of the second component.

[0022] Based on the further improvement of the above system, the topological centrality weight is obtained based on at least one of the degree centrality, the betweenness centrality and the eigenvector centrality of the corresponding component.

[0023] Based on the further improvement of the above system, the optimization scheme is formed according to the priority of the component to be optimized, comprising:

[0024] for a first component in the component to be optimized, arranging the first component in descending order of priority to form a main sequence;

[0025] for each first component in the main sequence, arranging the second components associated with the first component in descending order of priority to form a subsequence of the first component;

[0026] according to the order of the main sequence and each subsequence, sequentially outputting the corresponding parameter adjustment rule to obtain the optimization scheme.

[0027] Based on the further improvement of the above system, the parameter adjustment rule meets three constraints simultaneously: adjustment range constraint, coordination constraint and amplitude constraint; wherein the adjustment range constraint is used to ensure that the adjusted parameter value is within the upper and lower limits specified by the standard; the coordination constraint is used to ensure that the parameter adjustment between the associated components is matched; the amplitude constraint is used to limit the amplitude of single parameter adjustment to not exceed the preset safety threshold.

[0028] Based on the further improvement of the above system, the multi-source data of each component includes: BIM geometric data, topological relationship data, electrical parameter data, fault data and environmental parameter data; the multi-modal feature vector of each component at the analysis time is extracted according to the neural network model of each modal, including:

[0029] For each component, based on the BIM geometric data and topological relationship data, the topological features of the component are extracted by using the graph attention network; based on the fault data at the analysis time and before the analysis time, the fault text features of the component at the analysis time are extracted by using the text embedding model; based on the electrical parameter data, the electrical parameter features of the component at the analysis time are extracted by using the electrical time sequence feature extraction model; based on the environmental parameter data, the environmental parameter features of the analysis time are extracted by using the environmental time sequence feature extraction model, and the components of the power grid equipment in the same region share the same environmental parameter features.

[0030] The topological features, fault text features, electrical parameter features and environmental parameter features of each component at the analysis time constitute the multi-modal feature vector of each component at the analysis time.

[0031] Based on the further improvement of the above system, the topological features of the component are extracted by using the graph attention network, including:

[0032] The graph structure data is constructed by taking the component as the node, the initial feature of the BIM geometric data of the component as the node, and the relationship between the components as the edge; the relationship includes: electrical connection relationship, mechanical linkage relationship and spatial proximity relationship;

[0033] The graph structure data is input into the multi-layer multi-channel graph attention network, and the node embedding representation is output to obtain the topological features of each component; wherein each channel corresponds to the edge of a kind of relationship, and is used to independently calculate the attention score and normalize the edge of the kind of relationship.

[0034] Based on the further improvement of the above system, after the electrical parameter data and the environmental data are time-aligned according to the unified time axis, the electrical time sequence data and the environmental time sequence data of three time scales are extracted from the data that have been time-aligned before the analysis time, wherein different time scales are resampled by using different sampling frequencies.

[0035] Based on the further improvement of the above system, the model structures of the electrical time sequence feature extraction model and the environmental time sequence feature extraction model are the same, and each sequentially includes: a parameter-aware time sequence encoding module, a gated fusion module and a dynamic attention pooling module; wherein the parameter-aware time sequence encoding module is used to encode the time sequence data of each parameter of each time scale to obtain the feature representation of each time scale; the gated fusion module is used to fuse the feature representations of the same parameter at different time scales to obtain the fusion feature representation of the parameter; and the dynamic attention pooling module is used to aggregate the fusion feature representations of multiple parameters of the same component to obtain the electrical parameter feature or the environmental parameter feature corresponding to the component.

[0036] Based on the further improvement of the above system, the parameter-aware time sequence encoding module sequentially includes:

[0037] a parameter type encoding unit, configured to generate an embedding vector for each parameter type;

[0038] an input mapping unit, configured to concatenate the embedding vector of the parameter type and the current time position encoding for each time step in different time scales, and then map the same through a linear layer to obtain an input sequence;

[0039] a plurality of improved LSTM units, respectively encoding the input sequence of one time scale, wherein the input gate, the forget gate and the output gate of each improved LSTM unit all fuse the embedding vector of the parameter type, and the attention weights of all time steps are calculated through the attention mechanism to obtain the feature representation of the current time scale.

[0040] Based on the further improvement of the above system, the risk prediction model sequentially includes a multi-modal correlation model and a hierarchical prediction model; the multi-modal correlation model is used to output a global context representation, a correlation analysis result and a multi-modal fusion representation of each component according to the multi-modal feature vector of the component at the analysis moment, and is transmitted to the hierarchical prediction model; and the hierarchical prediction model is used to output the risk prediction result of each component and the device to which it belongs in the future period.

[0041] Based on the further improvement of the above system, the multi-modal correlation model performs the following steps:

[0042] projecting the multi-modal feature vectors of each component at the analysis moment to a unified dimension and constructing an input sequence, and adding the mixed position encoding of each modality to the projected feature vector in the input sequence through element-by-element addition;

[0043] inputting the input sequence with mixed position encoding into a Transformer encoding layer, and performing cross-correlation between modalities according to a unified multi-head self-attention layer and a hierarchical attention mask matrix, to output a multi-modal fusion representation of each component and a global context representation in each sample.

[0044] The correlation analysis result is extracted from the attention weights of the Transformer encoding layer.

[0045] Based on the further improvement of the above system, the positions of mutual attention in the hierarchical attention mask matrix are set to 0, and the positions of prohibited attention are set to negative infinity, wherein the positions of mutual attention include:

[0046] The position of the global token is mutually attentive with the positions of all other vectors in the input sequence;

[0047] The positions of the feature vectors of different modalities within the same component are mutually attentive;

[0048] The positions of the feature vectors of the same modality of different components are mutually attentive only when the corresponding components are neighbors in the topological relationship.

[0049] Based on the further improvement of the above system, the correlation analysis result includes an inter-modal correlation matrix and a cross-component modal correlation matrix; the inter-modal correlation matrix is obtained by averaging the attention weights between different modal position sets in all Transformer layers and attention heads; and the cross-component modal correlation matrix is obtained by averaging the attention weights of the same modality between neighbor components in the topological relationship in all Transformer layers and attention heads.

[0050] Based on the further improvement of the above system, the hierarchical prediction model successively includes a component-level risk prediction module, a device-level risk prediction module and a prediction enhancement module; the component-level risk prediction module is used to output the component risk probability, the component risk level and the component failure type by using the multilayer perceptron according to the multi-modal fusion representation of each component; the device-level risk prediction module is used to output the device risk index and the device risk type after respectively aggregating the global context representation and the multi-modal fusion representation of each component according to the device; and the prediction enhancement module is used to enhance the component risk probability and the device risk index according to the correlation analysis result, and output the final component risk probability and device risk index.

[0051] In another aspect, the embodiment of the present application provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor executes the functions of the power grid equipment BIM model parameter optimization system of any one of the above aspects when executing the computer program.

[0052] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the functions of the power grid equipment BIM model parameter optimization system of any one of the above aspects are realized when the computer program is executed by the processor.

[0053] Compared with the prior art, the present application can at least achieve one of the following beneficial effects:

[0054] 1. From data preprocessing, correlation mining, risk prediction to the formation of parameter optimization scheme, a complete closed-loop system is realized, effectively solving the problems of data island, early warning lag and long BIM model parameter optimization iteration cycle in the prior art; according to the actual operation state and risk prediction result, autonomous, closed-loop and continuous iteration is carried out to promote parameter optimization, the operation demand of power grid is quickly adapted to the changing operation demand of power grid, the dependence on manual analysis is reduced, human errors are reduced, and the safe, stable and efficient operation of power grid is greatly valuable.

[0055] 2. Through the unified and standardized processing of the component data processing module, the system can map the heterogeneous multi-modal data such as BIM geometric data, topological relationship data, electrical parameter data, fault text data and environmental parameter data to a unified semantic space; instead of considering various data in isolation, the system actively discovers and quantifies the complex and nonlinear implicit correlations between them, greatly excavates the value of multi-source data, and provides a comprehensive data view for accurate decision-making.

[0056] 3. The dynamic risk prediction module uses the fused deep features to capture the time evolution law through an improved parameter-aware LSTM, and combines multi-task learning and context enhancement to perform hierarchical (component level / equipment level) and multi-dimensional (probability / level / type) risk prediction, so that the prediction is more in line with the physical reality, the precision is high and the false alarm rate is low; the traditional fixed threshold-based lag alarm mode is changed, so that the operation and maintenance is changed from passive response to active intervention, greatly reducing the probability of unplanned shutdown and accident risk.

[0057] In the present application, the above technical solutions can be combined with each other to realize more preferred combination schemes. Other features and advantages of the present application will be described in the subsequent specification, and some advantages will become apparent from the specification, or will be understood by implementing the present application. The purpose and other advantages of the present application can be achieved and obtained from the contents specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0058] The accompanying drawings are included to provide a further understanding of the embodiments, and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and, together with the description, serve to explain the principles of the application; and in which:

[0059] Figure 1 Figure 1 is a schematic diagram of a power grid equipment BIM model parameter optimization system structure in Embodiment 1 of the present application. DETAILED DESCRIPTION

[0060] Preferred embodiments of the present application will be described in detail below with reference to the drawings, in which the same or similar components are denoted by the same reference numerals, for the purpose of illustrating the principles of the present application. It is to be understood that the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0061] Those skilled in the art should know that the following specific embodiments or specific implementation manners are a series of optimized settings listed by the present application to further explain the specific inventive content, and these settings can be used in combination or association with each other, unless the present application explicitly indicates that certain or a specific embodiment or implementation manner cannot be associated with other embodiments or implementation manners or cannot be used together. At the same time, the following specific embodiments or implementation manners are only the most optimized settings, and not as an understanding of the protection scope of the present application.

[0062] Embodiment 1

[0063] One specific embodiment of the present application discloses a power grid equipment BIM model parameter optimization system, which generates optimization decisions of BIM model parameters by dynamic risk perception driving of components in power grid equipment through deep data fusion and intelligent algorithm.

[0064] As shown in Figure 1 The system of the present embodiment includes three functional modules: component data processing module 100, dynamic risk prediction module 200, and model parameter optimization module 300. These modules are connected in turn and iteratively executed to form a closed-loop system from data perception to decision execution.

[0065] The component data processing module 100 is used to collect multi-source data of each component, and extract multi-modal feature vectors of each component at the analysis time according to the neural network model of each modality;

[0066] The dynamic risk prediction module 200 is used to obtain the multi-modal feature vectors of each component at the analysis time output by the component data processing module, and input them into the risk prediction model to obtain the risk prediction result;

[0067] The model parameter optimization module 300 is used to identify the components to be optimized and their priorities according to the risk prediction result output by the dynamic risk prediction module; based on the parameter adjustment rule library, obtain the BIM model parameter adjustment rule of the components to be optimized, and form an optimization scheme according to the priorities of the components to be optimized.

[0068] Specifically, the component data processing module 100 is the foundation of the system, and its goal is to build a unified and standardized "component-level multi-modal feature map". The specific implementation steps of the component data processing module include S101 and S102.

[0069] S101, collect and pre-process multi-source data of each component:

[0070] It should be noted that the power grid equipment is divided into 5 categories according to application categories: transmission lines, substations (converter stations), communication equipment, monitoring devices and public facilities. Each power grid equipment contains many components. According to actual needs, the components that need to be optimized are selected, and multi-source data is collected, including but not limited to: BIM geometric data, topological relationship data, electrical parameter data, fault data and environmental parameter data.

[0071] Among them, the BIM geometric data and the topological relationship data are obtained by analyzing the BIM model file to obtain the geometric data of each component in the power grid equipment and the relationship between the components; the electrical parameter data and the environmental parameter data are obtained by numerically completing and time sequence aligning the electrical parameters of each component and the meteorological data of the area where the power grid equipment is located based on a unified time axis; the fault data is obtained by analyzing the operation and maintenance records to obtain the fault data of each component.

[0072] ① Collection and preprocessing of BIM geometric data.

[0073] It should be noted that the BIM geometric data comes from the BIM model file. By calling the interface provided by the BIM platform, the BIM model file is parsed, and the BIM geometric data of each component is extracted, including: device type, component type, coordinates, size and component properties.

[0074] Among them, the device type and the component type are the classification of the device and the component in the actual application scenario, such as the device type is transformer, and the component type is radiator; the coordinates are used to represent the relative position of the component in the device; the size is the length, width and height of the bounding box of each component; the component properties include basic geometric properties and optimizable parameter properties.

[0075] For example, for the radiator of the transformer, the "radiator area" and the "radiator density" are the key optimizable parameters that affect the heat dissipation efficiency; for the high-voltage bushing, the "insulation material" and the "dry arc distance" are the core properties that determine its insulation performance. By targeted extraction and normalization of these key properties, on the one hand, it is convenient to distinguish components and reflect the unique features of each component, and on the other hand, it is convenient to compare the changes before and after adjusting the parameter property values.

[0076] In order to convert the above BIM geometric data into model processable numerical features, the following preprocessing is needed:

[0077] One-hot encoding is used to convert the category variables such as equipment type and component type into binary vectors;

[0078] The coordinates of all components are scaled to the range of [0, 1] based on the unified BIM model coordinate system using the min-max normalization method to eliminate the influence of the absolute coordinate value dimension;

[0079] The dimensions of each component are standardized to eliminate the influence of the dimension;

[0080] The numerical values of the component attributes are normalized.

[0081] Finally, the processed equipment type, component type, coordinates, dimensions, and component attributes are spliced together to form the geometric feature sequence of each component. This series of processing effectively solves the problem of mixed qualitative and quantitative data and different dimensions in BIM geometric data.

[0082] ②Collection and preprocessing of topological relationship data.

[0083] It should be noted that the BIM model file also includes relationship information between components. The analysis of the BIM model file obtains the relationship between the core components, including electrical connection relationship, mechanical linkage relationship, and spatial proximity relationship.

[0084] The collected relationship between components is represented in the form of an adjacency list or adjacency matrix. If there is a relationship between components, it is set to 1, otherwise it is set to 0; The component association relationship implicitly contained in the BIM model is made explicit and structured, providing key information for understanding the propagation path of risks within the equipment.

[0085] ③Collection and preprocessing of electrical parameter data.

[0086] It should be noted that one component may contain multiple electrical parameters. Electrical parameter data is usually obtained from SCADA systems or other monitoring files, and the format is usually CSV. In this embodiment, the Pandas library is used to parse the CSV data to obtain the electrical parameter time series data of each component, such as sleeve current, radiator inlet and outlet oil temperature, etc.

[0087] For cases where only device-level data can be obtained (such as total current of transformer), the data is copied to all components as shared electrical parameter data.

[0088] Since electrical parameter data is time series data, after collecting the electrical parameters at a specific analysis time, historical data before the analysis time is still needed to analyze the trend of the data. Therefore, the data at the historical time is also stored for analysis.

[0089] The preprocessing of electrical parameter data includes data cleaning, time series alignment and padding, and standardization.

[0090] Among them, data cleaning is to eliminate obvious outliers, such as illegal data with current greater than 500A.

[0091] Preferably, an outlier detection method based on dynamic threshold is adopted, for each parameter of each component, the moving average and standard deviation of its historical data are calculated, and the data points obviously exceeding (such as ±5 times the standard deviation) the range are regarded as outliers and are eliminated. This method can adapt to the normal fluctuation range of different devices and different parameters, and the cleaning is more accurate.

[0092] Time series alignment and padding is to unify data of different sampling frequencies to the same time granularity (such as 5 minutes), and for missing values, linear interpolation method is used for padding to ensure the continuity of time series.

[0093] Standardization is to perform Z-Score standardization on continuous electrical parameters.

[0094] The above preprocessing effectively solves the common noise, missing and dimension problems in real-time data, and lays a solid foundation for accurate time series pattern mining.

[0095] ④ Collection and preprocessing of fault data.

[0096] Fault data comes from recent device fault records and inspection reports, in the form of unstructured Word or PDF documents.

[0097] This embodiment extracts text from the document, uses Python NLP tools (such as jieba segmentation) and pre-trained models to automatically analyze and encode the text, and its preprocessing includes text cleaning, entity recognition and text vectorization.

[0098] Among them, text cleaning includes removing special characters, stop words, and unifying professional terms to reduce text noise.

[0099] Entity recognition is to extract device and / or component name, fault type, fault description, severity, etc. from the text, so as to convert unstructured text into structured knowledge.

[0100] Preferably, a two-level entity linking strategy is adopted in entity recognition: the first level uses a pre-defined device and component dictionary for accurate matching. The second level, for fault descriptions that cannot be accurately matched, uses a semantic similarity model (such as Sentence-BERT) to calculate the similarity between the fault description and the function description of all components, and associates the fault description to the component with the highest similarity.

[0101] In addition, the fault time is extracted from the text based on the date and time regularization rules, and is bound with the fault record, which facilitates the association with the electrical parameter data and the environmental parameter data according to the time, so as to accurately label the risk label of the corresponding time.

[0102] It should be noted that if the extracted is the fault data of a specific component, a higher association confidence is set; if the extracted is the device-level fault and cannot be clearly associated with a specific component, the fault data is associated with all core components of the device, and a lower association confidence is set. Through this intelligent association strategy, the accuracy of information utilization and association is considered.

[0103] Finally, the preprocessed fault data includes: fault time, device name and / or component name, fault type, fault description, severity and association confidence.

[0104] ⑤Collection and preprocessing of environmental parameter data.

[0105] The environmental parameter data is derived from the meteorological data of the area where the power grid equipment is located, and the format is JSON. By calling the meteorological data interface, the time series data of environmental temperature, environmental humidity and rainfall accumulation is obtained.

[0106] When preprocessing the environmental parameter data, the same time axis as the electrical parameter data is used, and linear interpolation or spline interpolation is used to fill in the data and align the time sequence.

[0107] Based on the above preprocessing of multi-source data, the multi-source data of each component at the analysis time is obtained, including: BIM geometric data and topological relationship data, different types or quantities of electrical parameter data for each component, different fault data, and the same environmental parameter data for components in the same area.

[0108] S102, respectively according to the neural network model of each modality, the multi-modal feature vector of each component at the analysis time is extracted.

[0109] For each component, the following processing steps are used to obtain the topological feature, fault text feature, electrical parameter feature and environmental parameter feature at the analysis time, to form the multi-modal feature vector of each component at the analysis time.

[0110] ①Based on the BIM geometric data and topological relationship data, the topological feature of the component is extracted by using the graph attention network.

[0111] Specifically, the component is taken as a node, the BIM geometric data of the component is taken as the initial feature of the node, and the relationship between the components is taken as the edge to construct the graph structure data; the relationship includes: electrical connection relationship, mechanical linkage relationship and spatial proximity relationship.

[0112] Preferably, the edge weight is set according to the relationship type of the edge, such as the weight of the electrical connection relationship being 1.0, the weight of the mechanical linkage relationship being 0.8, and the weight of the spatial proximity relationship being 0.5.

[0113] It should be noted that the graph attention network used in this step is a multi-layer multi-channel graph attention network, which adopts a separate pre-training strategy, learns the topological structure representation of the power grid components through a graph reconstruction task, and calculates the difference between the reconstructed adjacency matrix of the node embedding and the original adjacency matrix as a loss function. After pre-training, end-to-end fine-tuning is performed with the risk prediction model in the dynamic risk prediction module 200 of the embodiment.

[0114] In implementation, the graph structure data is input into the trained multi-layer multi-channel graph attention network, and the node embedding representation is output to obtain the topological features of each component; each channel corresponds to an edge of a certain relationship, and is used to independently calculate the attention score and normalize the edge of the certain relationship. This multi-channel attention mechanism enables the model to learn the differences in the influence of different types of neighbor relationships on the center node.

[0115] Preferably, a two-layer multi-channel graph attention network is used, which first performs linear transformation on the initial features of all nodes, and then shares in all channels, and then calculates the attention score of each channel through the following formula: Taking the channel as an example, the attention score is calculated by the following formula:

[0116] ,

[0117] wherein, represents the unnormalized attention score between the node and the node in the relationship of the layer, represents the learnable attention vector of the layer in the relationship , and represent the linearly transformed features of the node and the node in the layer, represents the learnable scaling factor, represents the weight (edge weight) of the relationship , represents vector splicing, T represents the transpose operation; represents the activation function.

[0118] The neighbors of the node are all the nodes in the relationship , and the softmax normalization is as follows:

[0119] ,

[0120] wherein, denotes the normalized attention weight, denotes the node in the neighbor set under the relation . denotes the exponential function.

[0121] Further, under the channel of the relation , the neighbor information is aggregated to obtain the aggregation result of the relation , and the formula is as follows:

[0122] ,

[0123] wherein, denotes the aggregation result of the node at the layer under the relation .

[0124] The attention weight is calculated independently for each relation type, and the neighbor information is aggregated to obtain the node representation of each channel; the outputs of all channels are spliced and linearly transformed to generate the node feature of this layer, that is, the first-order message of the node (i.e., the output of the first-layer multi-channel graph attention network), and the above process is repeated to aggregate the second-order neighbor information, and the final embedding representation of the node is output. The representation not only contains the attributes of the node itself, but also integrates the structural context information in the topological network, providing a feature basis rich in spatial and functional relationships for subsequent multi-modal fusion.

[0125] Through the multi-channel attention mechanism model, it can be learned that when evaluating the risk of a "breaker arc chamber", the load state of its electrical connection components (such as current transformers) should be focused on, and the state of its mechanical linkage components (such as operating mechanisms) should be relatively weakened. This ability of neighbor screening and attention focusing makes the generated topological features no longer a simple average of neighbor features, but a refined structured representation containing network functional logic, laying a solid foundation for risk propagation prediction in topological networks.

[0126] ②Based on the fault data at the analysis time and before the analysis time, the fault text features of the components at the analysis time are extracted by using a text embedding model.

[0127] Specifically, based on a unified time axis, the fault data up to the analysis time is extracted, and the text embedding model is trained to convert it into a fixed-dimensional feature vector, obtaining the fault text feature representation of each component at the analysis time.

[0128] Exemplarily, the text embedding model adopts a Bert model and is trained by a conventional method, which will not be described herein.

[0129] Preferably, the correlation confidence of the fault data is taken as a weight, and the fault text feature representation is obtained by weighted average of the fault data up to the analysis time.

[0130] Through the text vectorization processing, the abstract text fault data is converted into a numerical vector with semantic information, so that the text information can participate in the fusion processing of the multi-modal data together with other numerical modal data.

[0131] ③Based on the electrical parameter data, the electrical time sequence feature extraction model is used to extract the electrical parameter features of the components at the analysis time; based on the environmental parameter data, the environmental time sequence feature extraction model is used to extract the environmental parameter features at the analysis time, and the components of the power grid equipment in the same region share the same environmental parameter features.

[0132] The electrical parameters and the environmental parameters of the power grid are typical time sequence data, which contain evolution rules. The embodiment effectively captures the short-term fluctuations, medium-term and long-term trends in these data through three time scales (time windows).

[0133] Specifically, for the electrical parameter data and the environmental parameter data, the electrical time sequence data and the environmental time sequence data of three time scales are extracted from the data that have been time sequence aligned before the analysis time and the analysis time, wherein different time scales are resampled at different sampling frequencies.

[0134] Exemplarily, based on the 7-day time sequence data before each time, the three time scales are 3 hours, 24 hours and 7 days, respectively, and the resampling is performed according to the sampling frequencies of 5 minutes, 1 hour and 12 hours, that is, each time corresponds to three time scale data sequences. If the electrical parameter data and the environmental parameter data are time sequence aligned at a frequency of 5 minutes, when resampling is performed according to a sampling frequency of 1 hour, all data points at 5-minute intervals within each hour are collected and aggregated (for example, the average is taken) to obtain a data point representing 1 hour.

[0135] The model structures of the electrical time sequence feature extraction model and the environmental time sequence feature extraction model are the same, and both include a parameter-aware time sequence encoding module, a gating fusion module and a dynamic attention pooling module in sequence; wherein the parameter-aware time sequence encoding module is used to encode the time sequence data of each parameter of each time scale to obtain the feature representation of each time scale; the gating fusion module is used to fuse the feature representations of the same parameter at different time scales to obtain the fusion feature representation of the parameter; and the dynamic attention pooling module is used to aggregate the fusion feature representations of multiple parameters of the same component to obtain the electrical parameter feature or the environmental parameter feature corresponding to the component.

[0136] Specifically, the parameter-aware temporal encoding module comprises, in sequence:

[0137] a parameter type encoding unit configured to generate an embedding vector for each parameter type;

[0138] an input mapping unit configured to concatenate the embedding vector of the parameter type and the current time position encoding after the numerical value of each time step in different time scales, and map the same through a linear layer to obtain an input sequence;

[0139] a plurality of improved LSTM units, each of which encodes the input sequence of one time scale, wherein the input gate, the forget gate and the output gate of each improved LSTM unit all fuse the embedding vector of the parameter type, and the hidden states of all time steps are weighted and summed through an attention mechanism to obtain the feature representation of the time scale corresponding to the current branch.

[0140] It should be noted that the implementation of the parameter type encoding unit is through a trainable embedding layer (Embedding Layer) to map discrete parameter types (such as “current”, “temperature”) into a dense embedding vector.

[0141] At the same time, a standard position encoding (such as sine-cosine encoding) is used to capture the time sequence information to generate the current time position encoding for different time scales.

[0142] In the input mapping unit, the embedding vector of the parameter type and the current time position encoding are concatenated after the numerical value of each time step in different time scales, and are mapped into an input sequence of the same dimension through a linear layer. This way, the model can inherently consider the physical nature of the parameter when memorizing or forgetting information. For example, for the “temperature” parameter, the model may learn a slowly changing, inertial memory pattern; while for the “current” parameter, it may learn a pattern more sensitive to instantaneous spikes.

[0143] Further, the number of improved LSTM units is the same as the number of time scales, and in this embodiment, there are three improved LSTM units to process short-term, medium-term and long-term time sequences. The improvement is that the input gate, the forget gate and the output gate of the LSTM unit all fuse the embedding vector of the parameter type, and the formula is as follows:

[0144] ,

[0145] wherein, , and represent the input gate, the forget gate and the output gate vectors at time , , and denote the weight matrices of the input gate, forget gate and output gate, respectively, for projecting the concatenated vector to the hidden layer dimension; denotes the hidden state at the previous time step, which contains the history information up to time ; denotes the Sigmoid activation function, which compresses the output to the range of (0, 1); denotes the input data at time ; denotes the embedding vector of the parameter type; , and denote the bias vectors of the input gate, forget gate and output gate, respectively.

[0146] Further, after the LSTM output, a temporal attention is added to fully capture the temporal pattern. The hidden states of all time steps of the LSTM are taken as the input, and a trainable network is used to calculate the attention score of each time step and perform Softmax normalization to obtain the weight. The final feature vector is the weighted sum of all weights and the corresponding hidden states. This mechanism enables the model to automatically focus on key time segments, such as abnormal fluctuation periods, rather than treating all time steps equally.

[0147] Specifically, the gating fusion module receives the feature representations of three time scales output by the temporal feature extraction model for each parameter, concatenates the feature representations of different time scales, and calculates the gating weight vectors (three-dimensional vectors) of different time scales through a fully connected layer and a Sigmoid activation function, as shown in the following formula:

[0148] ,

[0149] wherein, denotes the gating weight vector, including the weights of short-term, medium-term and long-term three time scales , and ; and denote the weight matrix and bias vector of the fully connected layer, respectively, , and denote the feature of short-term, medium-term and long-term three time scales, respectively.

[0150] Further, the feature representations of the corresponding time scales are weighted and summed according to the gating weight vectors of different time scales to obtain the fused feature representation of the parameter.

[0151] It should be noted that the gating vector of the gating fusion module is dynamic. For predicting short-term sudden failures, The weight of the component that is likely to be automatically increased; for predicting long-term aging failure, The weight of the component that is likely to be automatically increased; for predicting long-term aging failure,

[0152] Since a component usually has multiple electrical parameters (variable in number), the dynamic attention pooling module is used to aggregate the fusion feature representations of variable parameters into a fixed-dimensional component feature representation, including:

[0153] The parameter representation enhancement unit is used to splice the fusion feature representation of each parameter with the embedding vector of the parameter type and the importance prior weight to obtain the enhanced feature representation of each parameter; wherein the importance prior weight is preset based on domain knowledge, and is used to distinguish the importance of different electrical parameters and environmental parameters to risk prediction;

[0154] The attention pooling unit uses a learnable global query vector to calculate the attention weight of each parameter, and aggregates the enhanced feature representation of each parameter according to the attention weight, and outputs the electrical parameter feature or environmental parameter feature corresponding to the component.

[0155] Specifically, the learnable global query vector Q is a global context shared by all components, and the enhanced feature representation of each parameter is mapped into a key vector K and a value vector V respectively through an MLP; further, the attention score of the query vector and all key vectors is calculated, and after normalization, it is used as the attention weight to weight sum all value vectors V, to obtain the electrical parameter feature or environmental parameter feature corresponding to the component.

[0156] It should be noted that the two time series feature extraction models are independently trained in advance using autoencoder reconstruction loss or contrastive learning loss, or can be trained end-to-end with the model of the subsequent steps of the embodiment. No independent loss function is needed, but the model parameters are learned from the final risk prediction loss through gradient backpropagation.

[0157] When independently training, if using the autoencoder reconstruction loss, the time series feature extraction model is used as the encoder, and another reverse LSTM or fully connected layer is designed as the decoder to reconstruct the original sequence from the output of the encoder. The loss function is the mean square error of the reconstructed sequence and the original sequence. If using the contrastive learning loss, two time enhancements are performed on the same time series, and features are extracted using the time series feature extraction model. The loss function is to make the two features as close as possible in the representation space, and as far away as possible from the features of other time series.

[0158] It should be noted that the two time sequence feature extraction models have the same structure, but use independent model parameters, that is, the time sequence of multiple time scales of electrical parameters and the time sequence of multiple time scales of environmental parameters are used to train the respective time sequence feature extraction models.

[0159] The electrical time sequence data of three time scales extracted at the analysis moment of each component is transmitted to the trained electrical time sequence feature extraction model to obtain the electrical parameter feature at the analysis moment; the environmental time sequence data of three time scales extracted at the analysis moment of each component is transmitted to the trained environmental time sequence feature extraction model to obtain the environmental parameter feature at the analysis moment.

[0160] It should be noted that the risk prediction model in the dynamic risk prediction model 200 sequentially includes a multi-modal correlation model and a hierarchical prediction model, and is obtained by end-to-end training through construction of a sample set, and the specific implementation steps include S201-S204.

[0161] S201, constructing a sample set.

[0162] The multi-source data of each component at each moment in history is collected by using the component data processing module 100, and then the multi-modal feature vector of each component at each moment in history is obtained.

[0163] Further, the risk label is labeled using the fault data at each moment in history. The risk label includes a component level label and a device level label.

[0164] Specifically, the component level label includes:

[0165] The component risk probability label is set to 1 for the component that fails within the subsequent T days based on the historical moment, and otherwise 0; for example, T is 7 days.

[0166] The component risk level label is set according to the severity in the fault data.

[0167] For example, levels 1 to 5 are set, including: a slight anomaly, a risk level label of 1, which does not affect operation; a general defect, a risk level label of 2, which requires planned maintenance; a significant anomaly, a risk level label of 3, which requires near-term treatment; a serious defect, a risk level label of 4, which requires emergency treatment; and an immediate failure, a risk level label of 5, which requires immediate shutdown.

[0168] The component failure type label is one-hot encoded according to the failure type in the fault data. For example, the component failure type includes: overheating failure, insulation failure, mechanical failure, and electrical failure.

[0169] Specifically, the device level label includes:

[0170] The device risk index label is a comprehensive continuous value calculated according to the component risk level and the device failure coefficient corresponding to each moment in history, and the formula is as follows:

[0171] ,

[0172] Among them, indicates the device risk index label, and respectively indicate the preset ratio; indicates the function of taking the minimum value; indicates the high-risk component ratio, which is based on the component risk level in the fault data to obtain the ratio of the number of high-risk components in the device to the total number of components in the device; indicates the historical device failure coefficient, which is calculated according to the number of failures of the device before each moment in history, for example, if no failure has occurred, the historical device failure coefficient is 0, if the number of failures is within 3, the historical device failure coefficient is 0.5; if the number of failures is more than 3, the historical device failure coefficient is 1.

[0173] The device risk type label is a one-hot encoding of the risk type of the device as a whole, including insulation aging, overload risk, environmental impact and comprehensive factors.

[0174] All components are sorted, and the topological features, fault text features, electrical parameter features and environmental parameter features of all components at each moment are taken as a sample, which is put into the sample set together with the corresponding risk label. That is, each sample is a snapshot data at a time point, including the feature vectors of all components in all modalities.

[0175] S202, constructing a multi-modal correlation model.

[0176] The multi-modal correlation model is used to mine and quantify the complex and non-linear implicit correlation between different modal data, thereby breaking the "data island" and providing feature representation rich in context information for accurate risk prediction.

[0177] The multi-modal correlation model is a Transformer model fused with hierarchical attention mask, which realizes deep and controllable multi-modal fusion through fine position encoding, hierarchical attention mask and cross attention mechanism, while maintaining physical rationality and interpretability. The following steps are used to output the multi-modal fusion representation, global context representation and correlation analysis result of each sample:

[0178] ①Project the feature vector of each modality of each component in each sample to a unified dimension and construct an input sequence. The projected feature vectors of each modality in the input sequence are stacked with their own mixed position encoding through element-wise addition.

[0179] Specifically, the feature vector of each modality of each component in each sample is projected to a unified dimension using a linear layer respectively ; According to the component index and the fixed modality order, splice them together, add a special CLS token of the same dimension at the beginning of the sequence as a global token, which is used to aggregate global information, and get the input sequence.

[0180] It should be noted that the global token is initially a learnable vector. Each position in the input sequence corresponds to a dimension vector, and each component has 4 modalities, so the total number of positions in the input sequence is , where represents the number of components, and the 4 modalities are arranged in the order of topology, fault text, electrical parameters and environmental parameters. Position 0 corresponds to the feature vector of the CLS token, positions 1 to correspond to the topology feature vectors of the respective components, positions to correspond to the fault text feature vectors of the respective components, positions to correspond to the electrical parameter feature vectors of the respective components, and positions to correspond to the environmental parameter feature vectors of the respective components.

[0181] Further, the feature vectors of each modality in the input sequence are stacked with their respective mixed position encodings by element-wise addition to help the model distinguish different modalities and different components. The dimension of the mixed position encoding is the same as that of the feature vector after projection of each modality, and the dimension of the feature vector of each modality does not change after stacking. The mixed position encoding includes:

[0182] Intra-modality position encoding is a dimensional vector generated by assigning an index to each component according to its functional importance in the device and using the sine and cosine functions;

[0183] Intra-component position encoding is a dimensional vector generated by assigning a fixed position offset to each modality in the input sequence;

[0184] Inter-modality position encoding is a dimensional vector obtained from the embedding vector of the modality type.

[0185] It should be noted that the mixed position encoding is also calculated for the global token, the intra-modality position encoding is calculated by assigning an index of 0, the intra-component position encoding is calculated by assigning a position offset of -1, and the inter-modality position encoding is obtained by obtaining the embedding vector of "CLS".

[0186] Finally, the input sequence with mixed position encoding is obtained, and the input is input into the Transformer encoding layer.

[0187] In the self-attention mechanism of the Transformer, a pre-defined hierarchical attention mask matrix is applied to inject domain knowledge, control the path of information flow, and enhance the physical rationality and interpretability of the model.

[0188] It should be noted that the dimension of the hierarchical attention mask matrix is the same as the length of the input sequence, which is a matrix used to control the attention permission between different positions in the input sequence.

[0189] Specifically, the positions that need to be mutually attended in the hierarchical attention mask matrix are set to 0, and the positions that need to be prohibited from attention are set to negative infinity. The positions that need to be mutually attended include:

[0190] The position of the global token is mutually attended with the positions of all other vectors in the input sequence to aggregate the global context.

[0191] The positions of the feature vectors of different modalities within the same component are mutually attended, allowing them to freely calculate attention.

[0192] The positions of the feature vectors of the same modality of different components are only mutually attended when the corresponding components are neighbors in the topological relationship (such as electrical direct connection, belonging to the same device, spatial proximity).

[0193] ②The input sequence with mixed position encoding is input into the Transformer encoding layer, and cross-correlation between modalities is performed according to the unified multi-head self-attention layer and the hierarchical attention mask matrix, outputting the multi-modal fusion representation and the global context representation of each sample.

[0194] It should be noted that the Transformer encoding layer includes multiple Transformer encoders, and each Transformer encoder includes a multi-head self-attention mechanism, a feedforward network, a residual connection, and layer normalization.

[0195] Specifically, in the self-attention mechanism of the Transformer, query vectors, key vectors and value vectors are calculated according to the input sequence with mixed position encoding; then, the original attention scores are calculated according to the query vectors and the key vectors, and the final attention scores are obtained by adding the hierarchical attention mask matrix; and then the attention weights are obtained by softmax normalization. At this time, since the attention scores of the prohibited positions are negative infinity, the values of these positions are still negative infinity after adding the original attention scores, and are 0 after softmax normalization, thereby realizing the set attention mechanism.

[0196] The embodiment uniformly calculates query vectors, key vectors and value vectors according to the input sequence, without explicitly specifying the modal sources of the query, the key and the value, each attention head projects the input sequence into different feature subspaces through a respective independent parameter matrix, and naturally learns different modal cross modes, that is, rich and diverse multi-modal cross fusion can be realized.

[0197] Through this hierarchical and controlled attention mechanism, the model is forced to learn the interaction mode conforming to the physical law in the training process. For example, the electrical parameter feature of a component cannot directly pay attention to the text description of another non-neighbor component, but can indirectly affect it through a shared global token or intra-component full connection. The “operation and maintenance text” feature of a component (such as describing “abnormal noise”) will first be deeply fused with its own “electrical feature” (such as increased current harmonic) and “topological feature” (such as located at a vibration-sensitive mechanical connection point). Then, this fused information affects the judgment of other components through the CLS token or the topological neighbor relationship. This enables the model to learn complex interactions while avoiding illogical associations, improving the reliability of the results. This process simulates the reasoning chain of human experts when performing fault diagnosis, and realizes guided and interpretable deep fusion.

[0198] After the cross correlation of the Transformer encoding layer, the output vectors of multiple modalities of each component are aggregated to generate a multi-modal fusion representation, and the output of the global token is used as a global context representation.

[0199] ③Extract the correlation analysis result from the attention weight of the Transformer encoding layer.

[0200] The correlation analysis results include: an inter-modal correlation matrix and a cross-component modal correlation matrix. The inter-modal correlation matrix is used to quantify the average correlation strength between different modalities; for example, the correlation between the topology modality and the text modality. The cross-component modal correlation matrix is used to quantify the correlation strength of the same modality between neighboring components; for example, the correlation between the topology modality of component A and the topology modality of neighboring component B. These correlation matrices are not only key to interpretability, but also provide important contextual features for subsequent risk prediction.

[0201] It should be noted that the correlation analysis results are obtained by extracting and analyzing the attention weight matrices of each Transformer layer and each attention head, and then aggregating the weights by modality and component.

[0202] Specifically, the inter-modal correlation matrix is obtained by averaging the attention weights between different modality position sets in all Transformer layers and attention heads.

[0203] The cross-component modal correlation matrix is obtained by averaging the attention weights of the same modality between neighboring components in the topology relationship in all Transformer layers and attention heads.

[0204] Compared with the prior art, the multi-modal correlation model guides the model to learn data correlation patterns consistent with the physical reality of the power grid by mixing position encoding and hierarchical attention masks; it not only ensures sufficient fusion of different modal features within the same component, but also avoids non-neighbor components from having direct influence that is not logical, achieving controllable and physically consistent multi-modal information interaction; and it extracts inter-modal correlation strength and cross-component modal correlation strength from the model attention weights, forming quantitative correlation analysis results for use in hierarchical prediction models, achieving intelligent perception and fusion of context.

[0205] S203, constructing a hierarchical prediction model.

[0206] The hierarchical prediction model of the embodiment simultaneously performs component-level fine-grained prediction and device-level macro-situation assessment. The model sequentially includes: a component-level risk prediction module, a device-level risk prediction module, and a prediction enhancement module.

[0207] It should be noted that the component-level risk prediction module is used to output component risk probability, component risk level, and component fault type using a multi-layer perceptron based on the multi-modal fusion representation of each component in each sample, to achieve component-level fine-grained risk identification.

[0208] Specifically, the multi-layer perceptron extracts abstract features highly related to risk from the multi-modal fusion representation through layer-by-layer nonlinear transformation. Subsequently, the branches are three independent outputs, achieving multi-task learning:

[0209] The component risk probability is a scalar between 0 and 1, representing the probability of the component failing in the future time period;

[0210] The component risk level outputs a multi-dimensional vector corresponding to the component risk level label, for example, outputs a 5-dimensional vector, representing the probability distribution of the component belonging to "level 1" to "level 5" risk, corresponding to slight, general, significant, severe, immediate failure respectively, and this classification output facilitates quick classification management by operation and maintenance personnel;

[0211] The component failure type outputs a multi-dimensional vector corresponding to the component failure type label, identifying the possible failure type, and this classification output facilitates revealing the main risk inducement.

[0212] It should be noted that the device-level risk prediction module is used to aggregate the global context representation of each sample and the multi-modal fusion representation of each component, and then output the device risk index and the device risk type, realizing the device-level macro situation assessment.

[0213] Specifically, the device-level risk prediction module first aggregates the global context representation of each sample and the multi-modal fusion representation of each component according to the device, and then encodes and decodes through a fully connected network to output the device risk index and the device risk type.

[0214] Among them, the global context representation of each sample and the multi-modal fusion representation of each component are aggregated according to the device, including:

[0215] According to the mapping relationship from the component to the device, the multi-modal fusion representations of all components of the same device are aggregated within the device to obtain a device feature representation; the intra-device aggregation includes maximum pooling, average pooling and risk attention pooling operations;

[0216] The global context representation and the device feature representation are spliced to obtain a comprehensive feature representation.

[0217] It should be noted that the risk attention pooling is to take the predicted component risk probability as the attention weight, and to obtain the weighted sum of the multi-modal fusion representations of each component within the same device.

[0218] The comprehensive feature representation is encoded and decoded through a fully connected network, and finally outputs the device risk index and the device risk type. The device risk index is a scalar between 0 and 1, used to quantify the overall risk level of the device; the device risk type outputs a multi-dimensional vector corresponding to the device failure type label, identifying the possible device dominant risk type.

[0219] It should be noted that the prediction enhancement module is configured to enhance the component risk probability and the equipment risk index according to the correlation analysis result, and output the final component risk probability and the equipment risk index.

[0220] Specifically, the inter-modal correlation matrix in the correlation analysis result is flattened, the cross-component modal correlation matrix in the correlation analysis result is subjected to max-pooling and average-pooling, and the context vectors are obtained by splicing the matrices, respectively. The context vectors are mapped to the enhancement coefficients by a multilayer perception machine.

[0221] Further, the component risk probability and the equipment risk index are enhanced by the following formulas according to the enhancement coefficients, respectively.

[0222] ,

[0223] wherein, represents the enhanced component risk probability or the enhanced equipment risk index, is a learnable parameter, represents the component risk probability output by the component-level risk prediction module or the equipment risk index output by the equipment-level risk prediction module, represents the enhancement coefficient.

[0224] The prediction enhancement module adopts a learnable scaling mechanism. When the correlation context strongly indicates the risk (for example, the state of a component is acceptable, but its multiple key neighbors and the environment in which it is located all show abnormalities), the model will moderately increase the risk value, so as to make it more consistent with the actual physical situation.

[0225] S204, end-to-end training is performed by using the sample set.

[0226] The multi-modal correlation model and the hierarchical prediction model are optimized by end-to-end training by using the sample set, and the overall loss function is The overall loss function is obtained by calculating the weighted sum of the component-level joint loss , the equipment-level joint loss and the correlation-driven loss , and the formula is as follows:

[0227] ,

[0228] wherein, , and represent the weights of the component-level joint loss, the equipment-level joint loss and the correlation-driven loss, respectively.

[0229] It should be noted that the component-level joint loss is a weighted sum of the component risk probability loss, the component risk level loss, and the component failure type loss. Specifically, the component risk probability loss is obtained by calculating the binary cross-entropy loss between the component-level risk probability prediction and the true label; the component risk level loss is obtained by calculating the cross-entropy loss between the component-level risk level classification prediction and the true label; and the component failure type loss is obtained by calculating the multi-label binary cross-entropy loss between the component failure type multi-label prediction and the true label.

[0230] It should be noted that the equipment-level joint loss is a weighted sum of the equipment risk index loss and the equipment risk type loss; the equipment risk index loss is obtained by calculating the mean squared error loss between the predicted equipment risk index value and the true label; the equipment risk type loss is obtained by calculating the cross-entropy loss between the predicted equipment risk type classification and the true label.

[0231] It should be noted that the correlation-driven loss is calculated based on the predicted value of the equipment risk index and the inter-modal correlation matrix in the correlation analysis results. This means that when the model finds strong correlations between certain modes, its prediction of the equipment risk index should be more reliable (probability value close to 0 or 1).

[0232] Specifically, the formula is as follows:

[0233] ,

[0234] in, This indicates the number of samples in the training batch. Indicates the first Number of devices in each sample Indicates the first In the nth sample Predicted risk index value for each device Indicates the first The modal correlation matrix of each sample. This represents the smoothing constant, a very small value used to prevent mathematical errors in numerical calculations.

[0235] During implementation, the component data processing module 100 outputs the multimodal feature vectors of each component at the analysis time and transmits them to the dynamic risk prediction module 200. The trained multimodal association model outputs the global context representation, association analysis results, and multimodal fusion representation of each component based on the multimodal feature vectors of the components at the analysis time, and transmits them to the trained hierarchical prediction model. The hierarchical prediction model outputs the risk prediction results of each component and its associated equipment in the future time period, including: component risk probability, component risk level, component failure type, equipment risk index, and equipment risk type.

[0236] Preferably, the various neural network models involved in each module of the system of the present embodiment are encapsulated as Restful API interfaces and integrated into the background service of a domestic BIM platform (such as Guanglianda GLD-BIM), which is automatically called periodically to obtain the risk prediction results.

[0237] The risk prediction results are visually rendered in the three-dimensional model of the BIM platform. The color of the corresponding component changes dynamically according to its risk level. The operation and maintenance personnel click on the highlighted component to view the corresponding multi-source data information and risk prediction results.

[0238] Further, the model parameter optimization module 300 receives the risk prediction results from the dynamic risk prediction module 200, and if it identifies that there are components to be optimized, it intelligently provides adjustment schemes for the BIM model parameters of the components to be optimized based on a pre-constructed parameter adjustment rule library, thereby forming a complete closed loop of "perception-decision-optimization". The specific implementation steps of this module include S301-S303.

[0239] S301, construct a parameter adjustment rule library.

[0240] Parameter adjustment is not arbitrary and must strictly follow grid design standards and engineering experience. For this purpose, natural language processing technology is introduced to assist in constructing an intelligent and structured parameter adjustment rule library.

[0241] ① Use natural language processing technology to analyze grid engineering design standards, technical specifications, and expert experience documents to construct a knowledge graph.

[0242] Collect and input relevant grid engineering design standards, technical specifications, and expert experience documents. These documents include but are not limited to national standards and industry specifications such as "DL / T 1573-2016 Electrical Power Transformer Maintenance Guidelines", "GB 50150-2016 Electrical Equipment Interim Test Standards for Electrical Device Installation Engineering", etc.

[0243] Through natural language processing technology, the input documents are deeply semantically analyzed to extract key technical requirements, parameter constraints, safe operation boundaries, and recommended engineering practices. The analysis process focuses on identifying the following: the allowed range of technical parameters of equipment and components, parameter adjustment suggestions under different working conditions, threshold boundary conditions for safe operation, and specific technical measures for fault prevention.

[0244] Based on the information obtained from the analysis, entities, entity attributes, and relationships between entities are constructed to form a structured knowledge graph.

[0245] ② Based on historical operation and maintenance records and the knowledge graph, the mapping relationship between fault types and BIM model parameters is extracted to establish a fault-parameter association table.

[0246] The fault data in the component data processing module is collected by the method, and each fault type (such as insulation breakdown, overheating aging, mechanical fatigue, etc.) and parameter in the historical operation and maintenance record is obtained;

[0247] By comparing the similarity of entity vectors in the knowledge graph, the names of the fault types and parameters are standardized, and the associated parameters are obtained according to the knowledge graph to form a fault-parameter association table.

[0248] For example, "insulation failure" is mapped to "insulation material grade", "insulation thickness", etc. "Overheating failure" is mapped to "heat dissipation area", "cooling efficiency", etc. "Mechanical failure" is mapped to "structural strength", "material toughness", etc.

[0249] Preferably, by calculating the correlation between the occurrence probability of the fault in the historical operation and maintenance record and the parameter, such as calculating the Pearson correlation coefficient, the sensitivity of different parameters to the fault type is obtained after normalization, which is used to quantify the relative importance of each parameter in risk mitigation.

[0250] ③The fault-parameter association table and the knowledge graph are used as knowledge context, and the big language model is guided to generate preliminary adjustment rules through prompt word templates, which are stored in the parameter adjustment rule library after being reviewed and corrected by field experts.

[0251] It should be noted that the prompt word template instructs the big language model to generate parameter adjustment rules based on the knowledge context, including fault type, parameter name, adjustment direction, recommended amplitude, basis clause and expected effect.

[0252] For example, the prompt word template is: "You are a senior power grid engineering expert and BIM model standard consultant, proficient in DL / T 1573-2016 Power Transformer Maintenance Guide and other power grid design and operation standards. Your task is to generate accurate and executable BIM model parameter adjustment rules based on the provided domain knowledge. Please generate rules based on the following two parts of knowledge context: \n 1. Fault-parameter association table\n[Enter fault-parameter association table here]\n 2. Related standard knowledge\n[Enter the knowledge fragment retrieved from the knowledge graph according to the fault type here]\n. For fault type [enter specific fault type, such as insulation failure], please generate specific parameter adjustment rules, including: 1. Which parameters need to be adjusted; 2. Adjustment direction and recommended amplitude; 3. Parameter constraints and standard clauses relied on; 4. Expected risk reduction effect.".

[0253] The initial adjustment rules generated by the large language model are sent to domain experts for manual review, including: verifying the rationality and feasibility of the adjustment range, supplementing special constraints in actual engineering, correcting rule logic that does not conform to engineering practice, and optimizing the application priority ranking of rules.

[0254] After optimization by domain experts, the final adjustment rules are formed, formatted, and stored in the parameter adjustment rule library. This rule library is then integrated into the BIM platform, and a version management mechanism is established to support continuous updates and improvements.

[0255] S302. Based on the risk prediction results output by the dynamic risk prediction module, identify the components to be optimized and their priorities.

[0256] It should be noted that the components to be optimized are identified through the following steps:

[0257] ①Calculate the basic risk value of the component by weighted summation based on the component risk probability and component risk level.

[0258] Specifically, the basic risk value of the component is calculated using the following formula:

[0259] ,

[0260] in, Representation of components The basic risk value, and Representation of components The component risk probability and component risk level, and These represent the weights of the component risk probability and the component risk level, respectively. This represents the normalization function; for example, a minimum-maximum normalization function may be used.

[0261] ②Calculate the dynamic risk threshold of the component based on the equipment risk index, high risk threshold, and basic risk threshold.

[0262] This embodiment employs a dynamic threshold mechanism, combined with an equipment risk index for enhanced identification. For equipment with a high overall risk, the risk threshold for its components is appropriately relaxed. The dynamic risk threshold for each component is calculated using the following formula:

[0263] ,

[0264] in, Representation of components Dynamic risk threshold; and These represent the preset high-risk threshold and basic risk threshold, respectively. ; Indicates the weighting coefficient; Representation of components The equipment risk index of the equipment.

[0265] ③ Select components whose basic risk value is greater than their dynamic risk threshold as the first component; select neighbor components whose component risk probability is greater than the neighbor screening threshold and do not belong to the first component from the neighbor components of each first component as the second component; the neighbor screening threshold is less than the basic risk threshold.

[0266] This step includes related components whose propagation risk exceeds a lower neighbor screening threshold as components to be optimized, ensuring synergy in parameter optimization. For example, the neighbor screening threshold is set to 0.3.

[0267] ④ The first and second components are combined as the component to be optimized.

[0268] For example, if the component risk probability is 0.65, the component risk level is 4 (minimum level is 1, maximum level is 5, normalized to 0.75), and the weights are 0.7 and 0.3 respectively, then the basic risk value is 0.68. The high-risk threshold is set to 0.7, the basic risk threshold is set to 0.5, and the weighting coefficients are... With a base risk value of 0.3 and an equipment risk index of 0.8, the calculated dynamic risk threshold is approximately 0.642. Since the base risk value 0.68 > 0.642, this component is classified as a first-tier component. If selection were based solely on a high-risk threshold, it would not be selected as a first-tier component because 0.68 < 0.7. This demonstrates how a dynamic threshold mechanism appropriately relaxes the selection criteria.

[0269] Furthermore, the components to be optimized are processed in descending order of priority to ensure that key risks are resolved first.

[0270] It should be noted that the first component is the direct source of risk and has a higher priority than the second component. Therefore, the priority of the first component is obtained by multiplying its basic risk value by its topological centrality weight; the priority of the second component is obtained by multiplying its basic risk value, the edge weights between the second component and its associated first component, and its topological centrality weight.

[0271] The topological centrality weight is calculated based on at least one of the degree centrality, betweenness centrality and eigenvector centrality of the component, wherein the shortest path used to calculate betweenness centrality is determined based on the electrical function path.

[0272] It should be noted that the topological centrality weight can also be preset according to the importance of the component in the topological network, for example, the weight of a key node is 1.2 and that of an ordinary node is 1.0.

[0273] S303, based on the parameter adjustment rule library, obtain the BIM model parameter adjustment rule of the to-be-optimized component, and form an optimization scheme according to the priority of the to-be-optimized component.

[0274] According to the component failure type of each to-be-optimized component, the corresponding parameter adjustment rule is queried from the parameter adjustment rule library, and an optimization scheme is formed according to the priority of the to-be-optimized component.

[0275] Specifically, forming an optimization scheme according to the priority of the to-be-optimized component comprises:

[0276] Arranging the first components in the to-be-optimized component in descending order of priority to form a main sequence;

[0277] Arranging the second components associated with each first component in the main sequence in descending order of priority to form a sub-sequence of each first component;

[0278] According to the order of the main sequence and each sub-sequence, the corresponding parameter adjustment rule is output in sequence to obtain an optimization scheme.

[0279] This two-level sorting method facilitates simultaneous adjustment of the parameters of direct risk sources and coordinated adjustment of the components associated with the highest risk.

[0280] When applying the adjustment rule to adjust the BIM model parameters of the component, the following three constraints are strictly followed:

[0281] Adjustment range constraint: Ensure that the new parameter value is within the upper and lower limits specified by the standard.

[0282] Coordination constraint: Check and ensure that the parameter adjustment between associated components is matched (for example, if the sleeve insulation level is improved, the insulation level of the connected conductor also needs to be considered accordingly).

[0283] Amplitude constraint: Limit the single adjustment amplitude not to be too large (such as ≤30%), to avoid overdesign and severe impact on the existing system.

[0284] Preferably, one or more BIM model parameters are first adjusted in the BIM model of the test environment, and new risk prediction results are obtained based on the latest real-time data. If the component risk probability is not reduced, the adjustment amplitude is appropriately adjusted or an alternative parameter is selected for a new round of calculation until a scheme that meets all constraints and effectively reduces the risk is found.

[0285] Through the parameter synchronization interface, the finally determined parameter adjustment scheme is automatically written into the parameter database of the localization BIM platform. The BIM platform updates the geometric and attribute information of the three-dimensional model in sequence, realizing the real-time evolution of the digital twin model.

[0286] Preferably, a log of adjustment of each parameter is recorded, facilitating comparison and analysis before and after adjustment, and tracing back to any model version in history to provide data track for long-term optimization strategy analysis.

[0287] Compared with the prior art, the power grid equipment BIM model parameter optimization system provided by the embodiment realizes a complete closed loop system from data preprocessing, correlation mining, risk prediction to the formation of a parameter optimization scheme, effectively solves the problems of data island, early warning lag and long BIM model parameter optimization iteration period in the prior art, realizes autonomous, closed loop and continuous iteration to promote parameter optimization according to the actual operation state and the risk prediction result, quickly adapts to the changing operation demand of the power grid, reduces the dependence on manual analysis and human errors, and has great value for guaranteeing the safe, stable and efficient operation of the power grid. Through unified and standardized processing of the component data processing module, the system can map heterogeneous multi-modal data such as BIM geometric data, topological relationship data, electrical parameter data, fault text data and environmental parameter data to a unified semantic space; instead of considering various data in isolation, the system actively discovers and quantifies the complex and nonlinear implicit correlations between them, greatly mines the value of multi-source data, and provides a comprehensive data view for accurate decision-making. The dynamic risk prediction module uses the fused deep features to capture the time evolution law through an improved parameter-aware LSTM, and combines multi-task learning and context enhancement to perform hierarchical (component level / equipment level) and multi-dimensional (probability / level / type) risk prediction, so that the prediction is more in line with the physical reality, has high precision and low false alarm rate; the traditional lag alarm mode based on fixed threshold is changed, so that the operation and maintenance is changed from passive response to active intervention, greatly reducing the probability of unplanned shutdown and accident risk.

[0288] Embodiment 2

[0289] Another embodiment of the present application discloses an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor.

[0290] The electronic device of the embodiment can be a server, a terminal or any other electronic device with necessary computing and / or processing capability.

[0291] In one specific embodiment, the electronic device includes a processor, a memory and a communication interface connected through a system bus. The processor of the electronic device provides necessary computing, processing and / or control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory, the non-volatile storage medium stores an operating system, a computer program and serves as a working area for runtime, and temporarily stores intermediate results and final results; the internal memory provides a running environment for the operating system and the computer program in the non-volatile storage medium; the communication interface of the electronic device is used for network connection and communication with external devices; and the computer program is executed by the processor to implement the power grid equipment BIM model parameter optimization system of embodiment 1.

[0292] Embodiment 3

[0293] In another embodiment of the present application, a computer readable storage medium is disclosed, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the power grid equipment BIM model parameter optimization system of embodiment 1.

[0294] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the computer program can be stored in a computer readable storage medium. The computer readable storage medium is a disk, an optical disk, a read-only memory or a random access memory, etc.

[0295] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A power grid equipment BIM model parameter optimization system, characterized in that, The method comprises the following steps: a component data processing module is used to collect multi-source data of each component, and multi-modal feature vectors of each component at an analysis time are extracted according to a neural network model of each modality; a dynamic risk prediction module is used to obtain the multi-modal feature vectors of each component at the analysis time output by the component data processing module, and the multi-modal feature vectors are input into a risk prediction model to obtain a risk prediction result; the risk prediction result comprises a component risk probability, a component risk level, a component failure type, a device risk index and a device risk type; a model parameter optimization module is used to identify a component to be optimized and a priority of the component to be optimized according to the risk prediction result output by the dynamic risk prediction module, obtain a BIM model parameter adjustment rule of the component to be optimized based on a parameter adjustment rule library, and form an optimization scheme according to the priority of the component to be optimized; the identification of the component to be optimized comprises: a basic risk value of the component is calculated by weighted summation according to the component risk probability and the component risk level; a dynamic risk threshold of the component is calculated according to the device risk index, a high-risk threshold and a basic risk threshold; a first component is obtained by taking the component whose basic risk value is greater than its dynamic risk threshold; a second component is obtained by taking, from neighbor components of each first component, a component whose component risk probability is greater than a neighbor screening threshold and which does not belong to the first component; the neighbor screening threshold is less than the basic risk threshold; and the first component and the second component are combined as the component to be optimized.

2. The power grid equipment BIM model parameter optimization system of claim 1, wherein, The parameter adjustment rule library is obtained by the following steps: natural language processing technology is used to analyze power grid engineering design standards, technical specifications and expert experience documents, and a knowledge graph is constructed; a mapping relationship between a fault type and a BIM model parameter is extracted based on historical operation and maintenance records and the knowledge graph, and a fault-parameter association table is established; the fault-parameter association table and the knowledge graph are used as a knowledge context, a large language model is guided to generate preliminary adjustment rules through a prompt word template, and the preliminary adjustment rules are stored in the parameter adjustment rule library after being reviewed and corrected by a field expert.

3. The power grid equipment BIM model parameter optimization system of claim 1, wherein, The basic risk value of the component is calculated by the following formula: ， wherein represents a component of a basic risk value, and represents a component of a component risk probability and a component risk level, and represent a weight of the component risk probability and the component risk level, respectively, represents a normalization function.

4. The power grid equipment BIM model parameter optimization system of claim 1, wherein, The dynamic risk threshold is calculated by the following formula: ， wherein, represents a component of a dynamic risk threshold; and respectively represent a preset high risk threshold and a base risk threshold, ; represents a weight coefficient; represents a component of a device risk index of the device.

5. The power grid equipment BIM model parameter optimization system of claim 1, wherein, The priority of the first component in the component to be optimized is obtained by multiplying the basic risk value of the first component by a topological centrality weight of the first component; The priority of the second component is obtained by multiplying the basic risk value of the second component, an edge weight between the second component and an associated first component of the second component, and a topological centrality weight of the second component.

6. The power grid equipment BIM model parameter optimization system of claim 5, wherein, The topological centrality weight is calculated based on at least one of a degree centrality, a betweenness centrality and a feature vector centrality of the corresponding component.

7. The power grid equipment BIM model parameter optimization system of claim 5, wherein, The optimization scheme is formed according to the priority of the component to be optimized, which comprises: the first components in the component to be optimized are arranged in descending order of priority to form a main sequence; for each first component in the main sequence, the second components associated with the first component are arranged in descending order of priority to form a subsequence of the first component; the corresponding parameter adjustment rules are output in sequence according to the order of the main sequence and the subsequences to obtain the optimization scheme.

8. The power grid equipment BIM model parameter optimization system of claim 7, wherein, The parameter adjustment rule satisfies three constraints simultaneously: an adjustment range constraint, a coordination constraint and a magnitude constraint; the adjustment range constraint is used to ensure that the adjusted parameter value is within the upper and lower limits specified by the standard; the coordination constraint is used to ensure that the parameter adjustment between the associated components is matched; and the magnitude constraint is used to limit the magnitude of a single parameter adjustment to a preset safety threshold.

9. The power grid equipment BIM model parameter optimization system of claim 1, wherein, The multi-source data of each component includes BIM geometric data, topological relationship data, electrical parameter data, fault data and environmental parameter data; the multi-modal feature vector of each component at the analysis time is extracted according to the neural network model of each modality, including: For each component, the topological features of the component are extracted based on the BIM geometric data and the topological relationship data using a graph attention network; the fault text features of the component at the analysis time are extracted based on the fault data at the analysis time and before the analysis time using a text embedding model; the electrical parameter features of the component at the analysis time are extracted based on the electrical parameter data using an electrical time series feature extraction model; and the environmental parameter features at the analysis time are extracted based on the environmental parameter data using an environmental time series feature extraction model, and the components of the power grid equipment in the same region share the same environmental parameter features. The topological features, the fault text features, the electrical parameter features and the environmental parameter features of each component at the analysis time constitute the multi-modal feature vector of each component at the analysis time.

10. The power grid equipment BIM model parameter optimization system of claim 9, wherein, The topological features of the component are extracted using a graph attention network, including: The graph structure data is constructed with the components as nodes, the initial features of the nodes being the BIM geometric data of the components, and the edges between the nodes being the relationships between the components; the relationships include electrical connection relationships, mechanical linkage relationships and spatial proximity relationships; The graph structure data is input into a multi-layer multi-channel graph attention network to output node embedding representations, thereby obtaining the topological features of each component; each channel corresponds to an edge of a certain relationship and is used to independently calculate and normalize the attention score of the edge of the certain relationship.

11. The power grid equipment BIM model parameter optimization system of claim 9, wherein, After the electrical parameter data and the environmental data are time series aligned according to a unified time axis, the electrical time series data and the environmental time series data of three time scales are extracted from the data that have been time series aligned at the analysis time and before the analysis time, wherein different time scales are resampled at different sampling frequencies.

12. The power grid equipment BIM model parameter optimization system of claim 11, wherein, The model structures of the electrical time series feature extraction model and the environmental time series feature extraction model are the same, and each model structure includes, in sequence, a parameter-aware time series encoding module, a gated fusion module and a dynamic attention pooling module; the parameter-aware time series encoding module is used to encode the time series data of each parameter of each time scale to obtain the feature representation of each time scale; the gated fusion module is used to fuse the feature representations of the same parameter at different time scales to obtain the fused feature representation of the parameter; and the dynamic attention pooling module is used to aggregate the fused feature representations of multiple parameters of the same component to obtain the electrical parameter features or the environmental parameter features corresponding to the component.

13. The power grid equipment BIM model parameter optimization system of claim 12, wherein, The parameter-aware time series encoding module includes, in sequence: a parameter type encoding unit configured to generate an embedding vector for each parameter type; An input mapping unit is configured to map the embedded vectors of the numerical concatenation parameters of each time step in different time scales and the current time position encoding to an input sequence through a linear layer after encoding. A plurality of improved LSTM units are configured to encode the input sequence of one time scale, wherein the input gate, the forget gate and the output gate of each improved LSTM unit are fused with the embedded vectors of the parameter types, and the attention weights of the hidden states of all time steps are calculated through the attention mechanism to obtain the feature representation of the current time scale.

14. The power grid equipment BIM model parameter optimization system of claim 1, wherein, The risk prediction model comprises a multi-modal correlation model and a hierarchical prediction model in sequence; the multi-modal correlation model is configured to output a global context representation, a correlation analysis result and a multi-modal fusion representation of each component according to the multi-modal feature vector of the component at the analysis moment, and transmit the multi-modal fusion representation to the hierarchical prediction model; the hierarchical prediction model is configured to output a risk prediction result of each component and the device to which the component belongs in a future period.

15. The power grid equipment BIM model parameter optimization system of claim 14, wherein, The multi-modal correlation model performs the following steps: Projecting the multi-modal feature vector of each component at the analysis moment to a unified dimension and constructing an input sequence, and adding the mixed position encoding of each modality projected feature vector in the input sequence through element-by-element addition; Inputting the input sequence with mixed position encoding into a Transformer encoding layer, and performing cross-correlation between modalities according to the unified multi-head self-attention layer and the hierarchical attention mask matrix, to output a multi-modal fusion representation of each component in each sample and a global context representation; Extracting a correlation analysis result from the attention weights of the Transformer encoding layer.

16. The power grid equipment BIM model parameter optimization system of claim 15, wherein, The positions of mutual attention in the hierarchical attention mask matrix are set to 0, and the positions of prohibited attention are set to negative infinity, wherein the positions of mutual attention include: The position of the global token and the positions of all other vectors in the input sequence mutually attend to each other; The positions of the feature vectors of different modalities in the same component mutually attend to each other; The positions of the feature vectors of the same modality of different components mutually attend to each other only when the corresponding components are neighbors in the topological relationship.

17. The power grid equipment BIM model parameter optimization system of claim 15, wherein, The correlation analysis result includes an inter-modal correlation matrix and a cross-component modal correlation matrix; the inter-modal correlation matrix is obtained by averaging the attention weights between different modality position sets in all Transformer layers and attention heads; and the cross-component modal correlation matrix is obtained by averaging the attention weights of the same modality between neighbor components in the topological relationship in all Transformer layers and attention heads.

18. The power grid equipment BIM model parameter optimization system of claim 14, wherein, The hierarchical prediction model comprises, in sequence, a component-level risk prediction module, a device-level risk prediction module and a prediction enhancement module; the component-level risk prediction module is configured to output a component risk probability, a component risk level and a component failure type by using a multi-layer perception according to a multi-modal fusion representation of each component; the device-level risk prediction module is configured to output a device risk index and a device risk type after performing aggregation on each component according to a global context representation and a multi-modal fusion representation of each component; and the prediction enhancement module is configured to enhance the component risk probability and the device risk index according to an association analysis result, and output a final component risk probability and a final device risk index.

19. An electronic device, comprising: The electronic device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to perform the functions of the power grid equipment BIM model parameter optimization system according to any one of claims 1-18.

20. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the functions of the power grid equipment BIM model parameter optimization system according to any one of claims 1-18.

Images