Industry electricity abnormal change intelligent analysis method and system based on multi-modal large model

Abstract

The present application relates to a multi-modal large model-based industry electricity abnormality intelligent analysis method and system, comprising the following steps: S1: obtaining power system multi-source data and preprocessing to obtain standardized time series data set; S2: based on the standardized time series data set, feature extraction is carried out, and a feature data set is constructed; S3: based on the feature data set, an industry electricity abnormality knowledge graph is constructed in combination with a multi-level correlation relationship network; S4: a multi-modal abnormality detection model is constructed by fusing time series Transformer, graph neural network and knowledge embedding module, and is trained based on the feature data set and the power knowledge graph; S5: the trained multi-modal model is deployed on a distributed stream processing platform to realize real-time abnormality detection and multi-dimensional intelligent analysis; S6: the real-time abnormality detection and multi-dimensional intelligent analysis results are visualized. The present application can realize all-round intelligent monitoring of industry electricity abnormalities facing complex industry actual demands.

Description

Technical Field

[0001] This invention relates to the field of power monitoring, and in particular to an intelligent analysis method and system for industry power consumption anomalies based on a multimodal large model. Background Technology

[0002] With the rapid development of industrial modernization and the digital economy, industry electricity consumption behavior is becoming increasingly complex and diverse. A large number of new production equipment, automated production lines, and intelligent manufacturing systems are constantly being connected to the power grid, making enterprise electricity loads more dynamic and uncertain. At the same time, multiple external factors, such as policy regulation, energy conservation and carbon reduction, climate fluctuations, and market volatility, are constantly influencing electricity demand and consumption structure, leading to frequent abnormal electricity consumption events in industries. Traditional electricity monitoring and anomaly detection methods mainly rely on single time-series analysis and static threshold judgments, making it difficult to capture the complex correlations hidden between multi-source data and to identify and explain the underlying causes of electricity consumption anomalies in a timely and accurate manner. Summary of the Invention

[0003] To address the aforementioned issues, the present invention aims to provide an intelligent analysis method and system for industry power consumption anomalies based on a multimodal large model, which can meet complex industry real-world needs and achieve comprehensive intelligent monitoring of industry power consumption anomalies.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A method for intelligent analysis of industry power consumption anomalies based on a multimodal large model includes the following steps:

[0006] S1: Acquire multi-source data from the power system and preprocess it to obtain a cross-modal standardized time series dataset;

[0007] S2: Based on the cross-modal standardized time series dataset, perform feature extraction and construct a feature dataset;

[0008] S3: Based on the feature dataset and combined with a multi-level relational network, construct an industry electricity consumption anomaly knowledge graph;

[0009] S4: Construct a multimodal anomaly detection model that integrates temporal Transformer, graph neural network and knowledge embedding modules, and train it based on feature dataset and power knowledge graph;

[0010] S5: Deploy the trained multimodal model on a distributed stream processing platform to achieve real-time anomaly detection and multidimensional intelligent analysis;

[0011] S6: Visualize the results of real-time anomaly detection and multi-dimensional intelligent analysis.

[0012] Furthermore, multi-source data from the power system is acquired and preprocessed to obtain a standardized time-series dataset, specifically as follows: The multi-source data from the power system includes power business data, production and operation data, external environmental data, and unstructured information; the power business data includes electricity consumption, load curves, voltage and current records, and high-frequency sampling data of power factors from industry users; the production and operation data includes output, production plans, working hours, process parameters, and equipment operating status from the enterprise's MES system; the external environmental data includes climate factors, time factors, and pricing mechanisms; the unstructured information includes industry news, public opinion data, environmental policy documents, enterprise announcement texts, as well as equipment monitoring images and infrared visual information.

[0013] Multi-source data undergoes unified cleaning and spatiotemporal alignment. A dynamic window anomaly detection algorithm is used to remove noise, and missing values ​​are used to reconstruct and repair incomplete data. Text data is uniformly encoded after word segmentation, entity recognition, and semantic correction. Image data is normalized, target detection is performed, and feature extraction is conducted. Time-series data is standardized using Z-score, and the final output is a cross-modal standardized time-series dataset in a unified format.

[0014] Furthermore, based on the feature dataset and combined with a multi-level relational network, a structured power knowledge graph is constructed, as follows: First, a multi-level relational network between equipment and users is constructed, including user-equipment relationships, equipment-line topology relationships, and geographical location relationships, and modeled in the form of a relational graph; simultaneously, combining professional knowledge, operation and maintenance experience, and expert rules from the power industry, a domain knowledge ontology is constructed to form a structured power knowledge graph, as follows:

[0015] The constructed multi-level network of relationships between devices and users includes user-device relationships, device-line topology relationships, and geographical location relationships, and is modeled in the form of a relationship diagram, as follows:

[0016] The user-device relationship modeling defines the user entity U = {u1, u2, ..., u...} i ,...,u n} and the device entity D={d1,d2,...,d j ,...,d m The association mapping between} is a bipartite graph structure G. UD :

[0017] G UD =(U∪D,E UD );

[0018] Where u i Let d be the i-th user entity, n be the number of user entities, and d be the number of user entities. j Let M be the j-th device entity, and m be the number of device entities; E UDThis represents the set of edges between the user and the device, where each edge... Carrying weight information w(u) i ,d j ):

[0019] ;

[0020] Where f is the weighted calculation function;

[0021] Adjacency matrix Among them, user u i With device entity d j The association is :

[0022] ;

[0023] Equipment-line topology modeling, constructing the power network physical topology diagram G T =(V T E T ), where the vertex set V T Contains all power equipment and nodes, edge set E T Indicates physical connection relationships:

[0024] The geographic location relationship modeling is based on the device's geographic coordinates (lon, lat) to construct a geospatial relationship matrix D. geo :

[0025] ;

[0026] in, These are the geographical coordinates of devices a and b, respectively; lon is the latitude, and lat is the longitude. The geographic spatial relationship between devices a and b; The radius of the Earth;

[0027] Based on the above modeling, the feature G=(V,E,T) of the multi-level association network is obtained. V ,T E Let V be all nodes, E be all entity relations, and T be X, where V represents all nodes, E represents all entity relations, and T represents all entity relations. V For node type, T E Let X be the edge type and X be the node feature matrix.

[0028] Furthermore, by combining professional knowledge, operation and maintenance experience, and expert rules from the power industry, a domain knowledge ontology is constructed to form a structured power knowledge graph, specifically as follows: The node types T of the multi-level relational network are summarized. V With edge type T EDefine a concept set C, a relation set R, and an attribute set P. The concept set C includes all explicit entity and node types, the attribute set P includes the attributes carried by each type of node or edge, and the relation set R includes the connection relationships between nodes. Traverse all entity relations E and convert them into standard triples according to (source, relation, target). Traverse all entity relations and construct (entity, attribute, attribute value) triples for the attributes of each node of all nodes V. Based on historical faults and operation and maintenance big data, statistically analyze abnormal patterns and introduce the operation and maintenance, monitoring, and fault response rules of the power industry to generate rule templates for corresponding high-frequency abnormal patterns, thus obtaining a structured power knowledge graph.

[0029] Furthermore, the multimodal large model includes an input layer, a multimodal processing layer, a multimodal fusion layer, and a task-specific decoding layer, as follows: The input layer takes into account the feature dataset X. T ={x1,x2,...,x t ,...,x T}, where x t Input feature data for time t, where T is the total number of time series; Multi-level association network feature G=(V,E,T) V ,T E The knowledge graph features KG = {(hs, r, ts)}, where hs is the head entity, r is the relation, and ts is the tail entity; the multimodal processing layer performs three-way encoding through a temporal Transformer, a graph neural network, and a knowledge embedding module to obtain corresponding representation features; the multimodal fusion layer fuses the representations of the three modalities; the task-specific decoding layer designs a multi-task decoding head based on the fused representation, including a power prediction decoding head, an anomaly detection decoding head, and a behavior analysis decoding head.

[0030] Furthermore, the corresponding representation features are obtained through three-way encoding using a temporal Transformer, a graph neural network, and a knowledge embedding module, as detailed below: The temporal Transformer processes the input feature data x... t Add position encoding PE t , to obtain input features :

[0031] ;

[0032] Then the first l Multi-head self-attention and feedforward:

[0033] ;

[0034] Among them, MHA is a multi-head self-attention module that models the global dependency of the input sequence; FFN is a feedforward neural network, which includes two linear layers and an activation function; LN is a layer normalization network. This represents the intermediate state after multi-head attention, residual connections, and layer normalization. Let be the hidden output state at time t in the l-th layer; It is the set of hidden states of all nodes at all times in the (l-1)th layer;

[0035] Finally, the global temporal features are weighted to obtain the temporal feature representation vector. :

[0036] ;

[0037] in, Let be the attention weight at time t; For the Lth T The hidden state at time t in the layer;

[0038] The graph neural network takes multi-level relational network features as input, and for each layer... l The feature representation of node i' is updated using a weighted average of its own and its neighbors' information:

[0039] ;

[0040] in, Let i be the vector representation of node i' in layer l'; Let l′ be the weight matrix of the l-th layer; This is the feature vector of the previous node i′; Let i be the set of neighboring nodes of node i′; Let σ be the attention weight of node i′ in layer l′ to its neighbor node j′; σ is the activation function.

[0041] Finally, structural features are obtained through global graph pooling. z graph :

[0042] ;

[0043] Where V is the set of all nodes in the graph. This represents the final representation of each node i' at the top level L′; the knowledge embedding module embeds the knowledge graph using the RotatE method to obtain the embedding vector, let E... rel For entities directly related to the current object of analysis, E KG [e] is the embedding vector of element e, for E rel Weighted aggregation forms knowledge features z knowledge :

[0044]

[0045] Where γ e is the weighting coefficient for element e.

[0046] Furthermore, the trained multimodal model is deployed on a distributed stream processing platform to achieve real-time anomaly detection and multidimensional intelligent analysis, specifically as follows: A streaming data channel is constructed using Kafka to receive electricity consumption index streams in real time and calculate incremental features. The multimodal model performs online inference on the new data, outputting the anomaly level, type, and potential causes in real time. Multi-hop inference is performed using a knowledge graph structure to track possible triggers for anomalies, and the impact of the event on neighboring enterprises or industries is assessed through a graph propagation model. Finally, a time-series prediction module is used to predict future electricity consumption trends and risks, generating early warning signals for anomalies in advance.

[0047] Furthermore, the results of real-time anomaly detection and multi-dimensional intelligent analysis are visualized, specifically as follows: real-time fluctuation curves and periodic trend charts are used to display historical and predicted anomaly trajectories; anomaly distribution heatmaps are displayed based on a GIS system; at the industry level, anomalies in different industrial clusters are compared using Sankey diagrams or radar charts; anomaly cause link view is generated using a knowledge graph structure to represent the causal path from external factors to abnormal power consumption; periodic anomaly reports and handling suggestions are automatically generated, including anomaly descriptions, scope of impact, cause diagnosis, and optimization strategies, combined with natural language generation from a large model to generate intelligent reports.

[0048] An intelligent analysis system for industry power consumption anomalies based on a multimodal large model includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent analysis method for industry power consumption anomalies based on a multimodal large model as described above.

[0049] The present invention has the following beneficial effects:

[0050] 1. This invention innovatively combines temporal Transformer, graph neural network and knowledge embedding model to construct a knowledge-driven, multimodal fusion intelligent analysis model. By introducing industry power consumption anomaly knowledge graph and multi-level association network, the model not only has deep feature extraction and causal reasoning capabilities, but also performs well in terms of anomaly detection interpretability and adaptability, and can effectively cope with complex and ever-changing real-world application scenarios.

[0051] 2. This invention realizes end-to-end real-time anomaly detection and multi-dimensional analysis, supports the visualization of results and intelligent decision-making assistance. Through distributed stream processing and intelligent visualization technology, the system can quickly respond to and analyze power consumption anomalies in the industry from multiple perspectives, greatly improving the risk warning capability and operation and maintenance management efficiency of the power system, and providing efficient, intuitive and intelligent decision support for industry users and management departments. Attached Figure Description

[0052] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0053] The following is in conjunction with the appendix Figure 1 The present invention will be further described in detail with reference to specific embodiments:

[0054] refer to Figure 1 In this embodiment, an intelligent analysis method for industry power consumption anomalies based on a multimodal large model is provided, including the following steps:

[0055] S1: Acquire multi-source data from the power system and preprocess it to obtain a cross-modal standardized time series dataset;

[0056] S2: Based on the cross-modal standardized time series dataset, perform feature extraction and construct a feature dataset;

[0057] S3: Based on the feature dataset and combined with a multi-level relational network, construct an industry electricity consumption anomaly knowledge graph;

[0058] S4: Construct a multimodal anomaly detection model that integrates temporal Transformer, graph neural network and knowledge embedding modules, and train it based on feature dataset and power knowledge graph;

[0059] S5: Deploy the trained multimodal model on a distributed stream processing platform to achieve real-time anomaly detection and multidimensional intelligent analysis;

[0060] S6: Visualize the results of real-time anomaly detection and multi-dimensional intelligent analysis.

[0061] In this embodiment, multi-source data from the power system is acquired and preprocessed to obtain a standardized time-series dataset, as follows: The multi-source data from the power system includes power business data, production and operation data, external environmental data, and unstructured information; the power business data includes electricity consumption, load curves, voltage and current records, and high-frequency sampling data of power factor from industry users; the production and operation data includes output, production plan, working hours, process parameters, and equipment operating status from the enterprise's MES system; the external environmental data includes climate factors (temperature, humidity, wind speed), time factors (holidays, seasons), and pricing mechanisms (peak-valley electricity prices); the unstructured information includes industry news, public opinion data, environmental policy documents, enterprise announcement texts, as well as equipment monitoring images and infrared visual information;

[0062] The multi-source data is uniformly cleaned and spatiotemporally aligned. A dynamic window anomaly detection algorithm is used to remove noise, and incomplete data is repaired by missing value reconstruction (KNN, interpolation, and temporal imputation). Text data is uniformly encoded after word segmentation, entity recognition, and semantic correction. Image data is normalized, target detection is performed, and feature extraction is performed. Temporal data is standardized using Z-score, and finally, a cross-modal standardized temporal dataset in a unified format is output.

[0063] In this embodiment, a structured power knowledge graph is constructed based on the feature dataset, as follows: First, a multi-level relationship network between devices and users is constructed, including user-device relationships, device-line topology relationships, and geographical location relationships, and modeled in the form of a relationship graph; at the same time, domain knowledge ontology is constructed by combining professional knowledge, operation and maintenance experience and expert rules of the power industry, forming a structured power knowledge graph.

[0064] In this embodiment, a multi-level relationship network between devices and users is constructed, including user-device relationships, device-line topology relationships, and geographical location relationships, and is modeled in the form of a relationship diagram, as follows:

[0065] The user-device relationship modeling defines the user entity U = {u1, u2, ..., u...} i ,...,u n} and the device entity D={d1,d2,...,d j ,...,d m The association mapping between} is a bipartite graph structure G. UD :

[0066] G UD =(U∪D,E UD );

[0067] Where u i Let d be the i-th user entity, n be the number of user entities, and d be the number of user entities. j Let M be the j-th device entity, and m be the number of device entities; E UD This represents the set of edges between the user and the device, where each edge... Carrying weight information w(u) i ,d j ):

[0068] ;

[0069] Where f is the weighted calculation function;

[0070] Adjacency matrix Among them, user u i With device entity d j The association is :

[0071] ;

[0072] Equipment-line topology modeling, constructing the power network physical topology diagram G T =(V T E T ), where the vertex set V T Contains all power equipment and nodes, edge set E T Indicates physical connection relationships:

[0073] The geographic location relationship modeling is based on the device's geographic coordinates (lon, lat) to construct a geospatial relationship matrix D. geo :

[0074] ;

[0075] in, These are the geographical coordinates of devices a and b, respectively; lon is the latitude, and lat is the longitude. The geographic spatial relationship between devices a and b; The radius of the Earth;

[0076] Based on the above modeling, the feature G=(V,E,T) of the multi-level association network is obtained. V ,T E Let V be all nodes (users, devices, lines, stations, etc.), E be all entity relationships (user-device, device-line, adjacency, geographical proximity, etc.), and T be the number of nodes (X, V). V For node type, T E Let X be the edge type (e.g., "own", "connect", "geographically proximate"), and let X be the node feature matrix.

[0077] In this embodiment, by combining professional knowledge, operation and maintenance experience, and expert rules from the power industry, a domain knowledge ontology is constructed to form a structured power knowledge graph, as detailed below:

[0078] Summarize the node types TV and edge types TE of the multi-level relationship network, and define the concept set C, the relationship set R, and the attribute set P; the concept set C includes all explicit entity and node types, the attribute set P includes the attributes carried by each type of node or edge, and the relationship set R includes the connection relationships between nodes.

[0079] Specifically, C = {user, electrical equipment, line, station, protection device, switch, …}; P = {rated capacity, geographical coordinates, manufacturing date, current alarm threshold, status code, …}; R = {owned, deployed at, powered at, adjacent, abnormal trigger, geographical association, …};

[0080] Iterate through all entity relations E, convert them into standard triples according to (source, relation, target), and iterate through all entity relations to construct (entity, attribute, attribute value) triples for the attributes of each node of all nodes V.

[0081] It incorporates the operation, maintenance, monitoring, and fault response experience of the power industry. For example, when a certain type of equipment detects "three-phase imbalance," an abnormal warning needs to be sent to all connected users; when a protection device malfunctions, an associated work order is automatically generated and maintenance personnel are notified; data-driven rule summarization: based on historical faults and operation and maintenance big data, high-frequency abnormal patterns are statistically analyzed, and operation, maintenance, monitoring, and fault response rules of the power industry are introduced to generate rule templates for corresponding high-frequency abnormal patterns. For example, if multiple adjacent substations report short circuits simultaneously within a short period of time, it is inferred that there is a common fault in the upstream line; a structured power knowledge graph is formed.

[0082] In this embodiment, the multimodal large model includes an input layer, a multimodal processing layer, a multimodal fusion layer, and a task-specific decoding layer, as detailed below:

[0083] The input layer receives the feature dataset X. T ={x1,x2,...,x t ,...,x T}, where x t Input feature data at time t, where T is the total number of time series; multi-level association network feature G=(V,E,T) V ,T E Knowledge graph features KG={(hs,r,ts)}, where hs is the head entity, r is the relation, and ts is the tail entity;

[0084] The multimodal processing layer uses a temporal Transformer, a graph neural network, and a knowledge embedding module to perform three-way encoding to obtain corresponding representation features;

[0085] The multimodal fusion layer fuses the representations of the three modalities;

[0086] The task-specific decoding layer, based on fused representation, designs a multi-task decoding head, including a power prediction decoding head, an anomaly detection decoding head, and a behavior analysis decoding head.

[0087] In this embodiment, three-way encoding is performed using a temporal Transformer, a graph neural network, and a knowledge embedding module to obtain the corresponding representation features, as detailed below:

[0088] The temporal Transformer processes the input feature data x. t Add position encoding PE t , to obtain input features :

[0089] ;

[0090] Then the first l Multi-head self-attention and feedforward:

[0091] ;

[0092] Among them, MHA is a multi-head self-attention module that models the global dependency of the input sequence; FFN is a feedforward neural network, which includes two linear layers and an activation function; LN is a layer normalization network. This represents the intermediate state after multi-head attention, residual connections, and layer normalization. Let be the hidden output state at time t in the l-th layer; It is the set of hidden states of all nodes at all times in the (l-1)th layer;

[0093] Finally, the global temporal features are weighted to obtain the temporal feature representation vector. :

[0094] ;

[0095] in, Let be the attention weight at time t; For the Lth T The hidden state at time t in the layer;

[0096] The graph neural network takes multi-level relational network features as input, and for each layer... l The feature representation of node i' is updated using a weighted average of its own and its neighbors' information:

[0097] ;

[0098] in, Let i be the vector representation of node i' in layer l'; Here is the weight matrix for the l′ layer; This is the feature vector of the previous node i′; Let i be the set of neighboring nodes of node i′; Let σ be the attention weight of node i′ in layer l′ to its neighbor node j′; σ is the activation function.

[0099] Finally, structural features are obtained through global graph pooling. z graph :

[0100] ;

[0101] Where V is the set of all nodes in the graph. This represents the final representation of each node i' in the top-level L';

[0102] The knowledge embedding module embeds the knowledge graph using the RotatE method to obtain an embedding vector, where E is the embedding vector. rel For entities directly related to the current object of analysis, E KG [e] is the embedding vector of element e, for E rel Weighted aggregation forms knowledge features z knowledge :

[0103] ;

[0104] Where γ e is the weighting coefficient for element e.

[0105] In this embodiment, the trained multimodal model is deployed on a distributed stream processing platform to achieve real-time anomaly detection and multidimensional intelligent analysis. Specifically, the following steps are taken: a streaming data channel is built using Kafka to receive electricity consumption index streams in real time and calculate incremental features. The multimodal model performs online inference on the new data and outputs the anomaly level, type, and potential causes in real time. Multi-hop inference is performed using a knowledge graph structure to track possible causes of anomalies (such as policy adjustments or climate change), and the impact of the event on neighboring enterprises or industries is assessed using a graph propagation model. The TemporalDecoder module is used to predict future electricity consumption trends and risks, generating anomaly warning signals in advance.

[0106] In this embodiment, the results of real-time anomaly detection and multi-dimensional intelligent analysis are visualized, specifically as follows: real-time fluctuation curves and periodic trend charts are used to display historical and predicted anomaly trajectories; anomaly distribution heatmaps are displayed based on a GIS system; at the industry level, anomalies in different industrial clusters are compared using Sankey diagrams or radar charts; anomaly cause link view is generated using a knowledge graph structure to represent the causal path from external factors to abnormal power consumption; periodic anomaly reports and handling suggestions are automatically generated, including anomaly descriptions, impact range, cause diagnosis, and optimization strategies, and intelligent reports are generated using natural language from a large model.

[0107] A power multidimensional monitoring and analysis system based on a large power model includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent analysis method for industry power consumption anomalies based on a multimodal large model described above.

[0108] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0109] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0110] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0111] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent analysis of industry power consumption anomalies based on a multimodal large model, characterized in that, Includes the following steps: S1: Acquire multi-source data from the power system and preprocess it to obtain a cross-modal standardized time series dataset; S2: Based on the cross-modal standardized time series dataset, perform feature extraction and construct a feature dataset; S3: Based on the feature dataset and combined with a multi-level relational network, construct an industry electricity consumption anomaly knowledge graph; S4: Construct a multimodal anomaly detection model that integrates temporal Transformer, graph neural network and knowledge embedding modules, and train it based on feature dataset and power knowledge graph; S5: Deploy the trained multimodal model on a distributed stream processing platform to achieve real-time anomaly detection and multidimensional intelligent analysis; S6: Visualize the results of real-time anomaly detection and multi-dimensional intelligent analysis; The multimodal large model includes an input layer, a multimodal processing layer, a multimodal fusion layer, and a task-specific decoding layer, as follows: The input layer takes into account the feature dataset X. T ={x1,x2,...,x t ,...,x T }, where x t Input feature data at time t, where T is the total number of time series; multi-level association network feature G=(V,E,T) V ,T E ,X), where V represents all nodes, E represents all entity relations, TV represents node type, TE represents edge type, and X represents node feature matrix; The knowledge graph feature KG = {(hs, r, ts)}, where hs is the head entity, r is the relation, and ts is the tail entity; the multimodal processing layer uses a temporal Transformer, graph neural network, and knowledge embedding module to perform three-way encoding to obtain corresponding representation features; the multimodal fusion layer fuses the representations of the three modalities; the task-specific decoding layer designs a multi-task decoding head based on the fused representation, including a power prediction decoding head, an anomaly detection decoding head, and a behavior analysis decoding head.

2. The intelligent analysis method for industry power consumption anomalies based on a multimodal large model according to claim 1, characterized in that, The acquisition and preprocessing of multi-source power system data to obtain a standardized time-series dataset is as follows: The multi-source power system data includes power business data, production and operation data, external environmental data, and unstructured information; the power business data includes electricity consumption, load curves, voltage and current records, and high-frequency power factor sampling data of industry users; the production and operation data includes output, production plan, working hours, process parameters, and equipment operating status of the enterprise's MES system; the external environmental data includes climate factors, time factors, and pricing mechanisms; the unstructured information includes industry news, public opinion data, environmental policy documents, enterprise announcement texts, as well as equipment monitoring images and infrared visual information. Multi-source data is uniformly cleaned and spatiotemporally aligned, noise is removed using a dynamic window anomaly detection algorithm, and incomplete data is reconstructed and repaired using missing values; text data is uniformly encoded after word segmentation, entity recognition, and semantic correction. Image data is normalized, target detection is performed, and feature extraction is performed; time series data is standardized using Z-score, and the final output is a cross-modal standardized time series dataset in a unified format.

3. The intelligent analysis method for industry power consumption anomalies based on a multimodal large model according to claim 2, characterized in that, Based on the feature dataset and combined with a multi-level relational network, a structured power knowledge graph is constructed, as follows: First, a multi-level relational network between equipment and users is constructed, including user-equipment relationships, equipment-line topology relationships, and geographical location relationships, and modeled in the form of a relational graph; simultaneously, combining professional knowledge, operation and maintenance experience, and expert rules from the power industry, a domain knowledge ontology is constructed to form a structured power knowledge graph, as follows: The constructed multi-level network of relationships between devices and users includes user-device relationships, device-line topology relationships, and geographical location relationships, and is modeled in the form of a relationship diagram, as follows: The user-device relationship modeling defines the user entity U = {u1, u2, ..., u...} i ,...,u n } and the device entity D={d1,d2,...,d j ,...,d m The association mapping between} is a bipartite graph structure G. UD : G UD =(U∪D,E UD ); Where u i Let d be the i-th user entity, n be the number of user entities, and d be the number of user entities. j Let M be the j-th device entity, and m be the number of device entities; E UD This represents the set of edges between the user and the device, where each edge... Carrying weight information w(u) i ,d j ): ； Where f is the weighted calculation function; Adjacency matrix Among them, user u i With device entity d j The association is : ; Equipment-line topology modeling, constructing the power network physical topology diagram G T =(V T E T ), where the vertex set V T Contains all power equipment and nodes, edge set E T Indicates physical connection relationships: The geographic location relationship modeling is based on the device's geographic coordinates (lon, lat) to construct a geospatial relationship matrix D. geo : ; in, These are the geographical coordinates of devices a and b, respectively; lon is the latitude, and lat is the longitude. The geographic spatial relationship between devices a and b; The radius of the Earth; Based on the above modeling, the feature G=(V,E,T) of the multi-level association network is obtained. V ,T E Let V be all nodes, E be all entity relations, and T be X, where V represents all nodes, E represents all entity relations, and T represents all entity relations. V For node type, T E Let X be the edge type and X be the node feature matrix.

4. The intelligent analysis method for industry power consumption anomalies based on a multimodal large model according to claim 3, characterized in that, By combining professional knowledge, operation and maintenance experience, and expert rules from the power industry, a domain knowledge ontology is constructed, forming a structured power knowledge graph, specifically as follows: The node types T of the multi-level relational network are summarized. V With edge type T E Define a concept set C, a relation set R, and an attribute set P. The concept set C includes all explicit entity and node types, the attribute set P includes the attributes carried by each type of node or edge, and the relation set R includes the connection relationships between nodes. Traverse all entity relations E and convert them into standard triples according to (source, relation, target). Traverse all entity relations and construct (entity, attribute, attribute value) triples for the attributes of each node of all nodes V. Based on historical faults and operation and maintenance big data, statistically analyze abnormal patterns and introduce the operation and maintenance, monitoring, and fault response rules of the power industry to generate rule templates for corresponding high-frequency abnormal patterns, thus obtaining a structured power knowledge graph.

5. The intelligent analysis method for industry power consumption anomalies based on a multimodal large model according to claim 1, characterized in that, The process involves three-way encoding using a temporal Transformer, a graph neural network, and a knowledge embedding module to obtain corresponding representation features, as detailed below: The temporal Transformer processes the input feature data x... t Add position encoding PE t , to obtain input features : ； Then the first l Multi-head self-attention and feedforward: ； Among them, MHA is a multi-head self-attention module that models the global dependency of the input sequence; FFN is a feedforward neural network, which includes two linear layers and an activation function; LN is a layer normalization network. This represents the intermediate state after multi-head attention, residual connections, and layer normalization. Let be the hidden output state at time t in the l-th layer; It is the set of hidden states of all nodes at all times in the (l-1)th layer; Finally, the global temporal features are weighted to obtain the temporal feature representation vector. : ； in, Let be the attention weight at time t; For the Lth T The hidden state at time t in the layer; The graph neural network takes multi-level relational network features as input, and for each layer... l The feature representation of node i' is updated using a weighted average of its own and its neighbors' information: ； in, Let i be the vector representation of node i' in layer l'; Here is the weight matrix for the l′ layer; This is the feature vector of the previous node i′; Let i be the set of neighboring nodes of node i′; Let σ be the attention weight of node i′ in layer l′ to its neighbor node j′; σ is the activation function. Finally, structural features are obtained through global graph pooling. z graph : ； Where V is the set of all nodes in the graph. This represents the final representation of each node i' at the top level L′; the knowledge embedding module embeds the knowledge graph using the RotatE method to obtain the embedding vector, let E... rel For entities directly related to the current object of analysis, E KG [e] is the embedding vector of element e, for E rel Weighted aggregation forms knowledge features z knowledge : ； Where γ e is the weighting coefficient for element e.

6. The intelligent analysis method for industry power consumption anomalies based on a multimodal large model according to claim 1, characterized in that, The trained multimodal model is deployed on a distributed stream processing platform to achieve real-time anomaly detection and multidimensional intelligent analysis. Specifically, Kafka is used to build a streaming data channel to receive electricity consumption index streams in real time and calculate incremental features. The multimodal model performs online inference on the new data and outputs the anomaly level, type and potential cause in real time. By combining knowledge graph structure for multi-hop reasoning, the possible causes of anomalies can be tracked, and the impact of events on neighboring enterprises or industries can be assessed through graph propagation model. Combined with time series prediction module, future electricity consumption trends and risks can be predicted and output, and anomaly warning signals can be generated in advance.

7. The intelligent analysis method for industry power consumption anomalies based on a multimodal large model according to claim 1, characterized in that, The visualization processing of real-time anomaly detection and multi-dimensional intelligent analysis results is as follows: real-time fluctuation curves and periodic trend charts are used to display historical and predicted anomaly trajectories; anomaly distribution heatmaps are displayed based on a GIS system; anomaly situations in different industrial clusters are compared using Sankey diagrams or radar charts at the industry level; anomaly cause link view is generated using a knowledge graph structure to represent the causal path from external factors to abnormal power consumption; periodic anomaly reports and handling suggestions are automatically generated, and the reports include anomaly descriptions, scope of impact, cause diagnosis, and optimization strategies, combined with natural language generation of large models to generate intelligent reports.

8. An intelligent analysis system for industrial power consumption anomalies based on a multimodal large model, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent analysis method for industry power consumption anomalies based on a multimodal large model as described in any one of claims 1-7.

Images