CN-121765605-B - Intelligent analysis method and system for industrial electricity utilization abnormal movement based on multi-mode large model

CN121765605BCN 121765605 BCN121765605 BCN 121765605BCN-121765605-B

Abstract

The invention relates to an industrial electricity utilization abnormal intelligent analysis method and system based on a multi-mode large model, comprising the following steps of S1, acquiring multi-source data of an electric power system, preprocessing to obtain a standardized time sequence data set, S2, carrying out feature extraction based on the standardized time sequence data set to construct a feature data set, S3, constructing an industrial electricity utilization abnormal knowledge graph based on the feature data set and combining a multi-level association relation network, S4, constructing a multi-mode abnormal detection model integrating a time sequence Transformer, a graph neural network and a knowledge embedding module, training based on the feature data set and the electric power knowledge graph, S5, deploying the trained multi-mode model on a distributed stream processing platform to realize real-time abnormal detection and multi-dimensional intelligent analysis, and S6, carrying out visual processing on real-time abnormal detection and multi-dimensional intelligent analysis results. The invention can face the actual demands of complex industries and realize the omnibearing intelligent monitoring of abnormal electricity consumption of the industries.

Inventors

GAO SONG
XU YI
LIAO CHENGJIE
YANG JUNYI
XIA CHAOPENG
GAO ZHIYE
XUE HEYU
GUO XUEQING
ZHU JIANLI
Huang Pingfa
XIE HONGYU

Assignees

国网江苏省电力有限公司
国网信通亿力科技有限责任公司

Dates

Publication Date: 20260508
Application Date: 20260303

Claims (8)

1. An intelligent analysis method for industrial electricity utilization abnormal state based on a multi-mode large model is characterized by comprising the following steps: S1, acquiring multi-source data of a power system, and preprocessing to obtain a cross-mode standardized time sequence data set; s2, carrying out feature extraction based on a cross-modal standardized time sequence data set to construct a feature data set; S3, based on the characteristic data set, combining a multi-level association relation network to construct an industry electricity utilization abnormal knowledge graph; S4, constructing a multi-mode abnormal motion detection model integrating a time sequence transducer, a graph neural network and a knowledge embedding module, and training based on a characteristic data set and an electric power knowledge graph; S5, deploying the trained multi-mode model on a distributed stream processing platform to realize real-time abnormal movement detection and multi-dimensional intelligent analysis; S6, carrying out visual processing on real-time abnormal detection and multidimensional intelligent analysis results; The multi-modal large model comprises an input layer, a multi-modal processing layer, a multi-modal fusion layer and a task specific decoding layer, and concretely comprises an input layer input characteristic data set X T ={x 1 ,x 2 ,...,x t ,...,x T }, wherein X t is time T input characteristic data, T is the total number of time sequences, multi-level association relation network characteristics G= (V, E, T V ,T E , X), wherein V is all entity relations, E is node type, TE is side type, X is node characteristic matrix, knowledge graph characteristics KG= { (hs, r, ts) }, wherein hs is head entity, r is relation, ts is tail entity, the multi-modal processing layer acquires corresponding characterization characteristics through three-way coding of a time sequence Transformer, a graph neural network and a knowledge embedding module, fuses the three-mode characterization, and the task specific decoding layer is used for designing the multi-modal wharf based on the fusion characterization and comprises an electric quantity prediction wharf, an anomaly detection wharf solution wharf and a behavior analysis wharf.
2. The intelligent analysis method for industrial electricity utilization variation based on the multi-mode large model of claim 1, wherein the method is characterized in that the method comprises the steps of obtaining multi-source data of an electric power system, preprocessing the multi-source data to obtain a standardized time sequence data set, wherein the multi-source data of the electric power system comprise electric power service data, production operation data, external environment data and unstructured information, the electric power service data comprise electricity consumption of industrial users, load curves, voltage current records and power factor high-frequency sampling data, the production operation data obtain output, production plan, man-hour, process parameters and equipment operation states of an enterprise MES system, the external environment data comprise climate factors, time factors and price mechanisms, and the unstructured information comprises industrial news, public opinion data, environment policy files, enterprise bulletin texts, equipment monitoring images and infrared image visual information; The method comprises the steps of carrying out unified cleaning and space-time alignment on multi-source data, removing noise by adopting a dynamic window anomaly detection algorithm, reconstructing and repairing incomplete data by utilizing a missing value, carrying out unified coding on text data after word segmentation, entity identification and semantic correction, carrying out normalization, target detection and feature extraction on image data, and finally outputting a cross-modal standardized time sequence data set in a unified format by means of Z-score standardization.
3. The intelligent analysis method for the electricity utilization variation of the industry based on the multi-mode large model according to claim 2 is characterized in that the characteristic data set is combined with a multi-level association relation network to construct a structured electricity knowledge graph, and specifically comprises the steps of firstly constructing a multi-level association relation network between equipment and a user, wherein the multi-level association relation network comprises a user-equipment relation, an equipment-line topological relation and a geographic position relation, modeling in a relation graph form, and meanwhile, constructing a domain knowledge ontology by combining professional knowledge, operation and maintenance experience and expert rules of the electricity industry to form the structured electricity knowledge graph, wherein the method comprises the following steps of: The multi-level association relation network between the construction equipment and the user comprises a user-equipment relation, an equipment-line topological relation and a geographic position relation, and is modeled in a relation diagram form, and the multi-level association relation network is specifically as follows: The user-device relationship modeling maps the association between the user entity u= { U 1 ,u 2 ,...,u i ,...,u n } and the device entity d= { D 1 ,d 2 ,...,d j ,...,d m } into a bipartite graph structure G UD : G UD =(U∪D,E UD ); Wherein u i is the ith user entity, n is the number of user entities, d j is the j-th device entity, m is the number of device entities, E UD represents the edge set between the user and the device, each edge Carry weight information w (u i ,d j ): ; Wherein f is a weight weighting calculation function; Using adjacency matrices Wherein the association of user u i with device entity d j is : ; Modeling the device-line topological relation, constructing a power network physical topological graph G T =(V T ,E T ), wherein a vertex set V T comprises all power devices and nodes, and an edge set E T represents the physical connection relation: The geographic position relationship modeling is used for constructing a geographic space relationship matrix D geo based on equipment geographic coordinates (lon, lat): ; Wherein, the The geographic coordinates of the equipment a and the equipment b are respectively, lon is latitude, and lat is longitude; is the geospatial relationship of devices a and b; Is the earth radius; And (3) combining the modeling to obtain a multi-level association relation network feature G= (V, E, T V ,T E and X), wherein V is all nodes, E is all entity relations, T V is node type, T E is edge type, and X is node feature matrix.
4. The industrial electricity utilization abnormal intelligent analysis method based on the multi-mode large model according to claim 3 is characterized in that the technical knowledge, operation and maintenance experience and expert rules of the electric power industry are combined to construct a domain knowledge ontology to form a structured electric power knowledge graph, specifically, a concept set C, a relation set R and an attribute set P are defined by summarizing node types T V and side types T E of a multi-level association relation network, the concept set C comprises all clear entities and node types, the attribute set P comprises attributes carried by each type of nodes or sides, the relation set R comprises connection relations among nodes, all entity relations E are traversed to construct (entity, attribute and attribute value) triples for the attributes of each node of all nodes V, and a rule template corresponding to a high-frequency abnormality mode is generated based on historical fault and operation and maintenance large data and operation and maintenance, monitoring and fault response rules of the electric power industry are introduced to obtain the structured electric power knowledge graph.
5. The intelligent analysis method for industrial electricity utilization abnormal movement based on the multi-mode large model according to claim 1, wherein the three paths of codes are performed through a time sequence Transformer, a graph neural network and a knowledge embedding module to obtain corresponding characterization characteristics, and the method is characterized in that the time sequence Transformer adds position codes PE t into input characteristic data x t to obtain input characteristics : ; Layer i multi-headed self-attention and feed forward: ; The system comprises an MHA, an FFN, an LN, a feedback neural network, a multi-head self-attention module, a feedback neural network and a feedback neural network, wherein the MHA is a multi-head self-attention module and models global dependence on an input sequence; The intermediate state after the multi-head attention and residual error are connected and the layers are normalized; the hidden state is output at the time t of the first layer; the hidden state set is a hidden state set of all nodes at all moments of the layer 1; The final global time sequence feature obtains a time sequence feature representation vector by a weighting mode : ; Wherein, the Attention weight for time t; is the hidden state at the time t of the L T th layer; the graph neural network inputs multi-level association relation network characteristics, and for each layer l ', the characteristics of the node i ' are updated through information weighting of the node i ' and neighbors: ; Wherein, the A vector representation for a layer i' node of the first layer; A weight matrix for the first layer; is the feature vector of the node i' of the upper layer; A neighbor node set of the node i'; The attention weight of the node i 'of the first layer to the neighbor node j' is given, and sigma is an activation function; finally, the structural feature z graph is obtained by global pooling: ; Wherein V is the set of all nodes in the graph, The knowledge embedding module is used for embedding the knowledge graph through RotatE to obtain an embedded vector, setting E rel as an entity directly related to the current analysis object, E KG [ E ] as an embedded vector of an element E, and weighting and polymerizing E rel into a knowledge feature z knowledge : ; Where γ e is the weighting factor of element e.
6. The industrial electricity consumption abnormal intelligent analysis method based on the multi-mode large model according to claim 1 is characterized in that the trained multi-mode model is deployed on a distributed stream processing platform to achieve real-time abnormal detection and multi-dimensional intelligent analysis, and specifically comprises the steps of utilizing Kafka/to construct a stream data channel, receiving electricity consumption index stream in real time and calculating increment characteristics, carrying out online reasoning on new data by the multi-mode model, outputting abnormal grades, types and potential reasons in real time, carrying out multi-hop reasoning by combining a knowledge graph structure, tracking possible causes of abnormal, evaluating the influence range of events on adjacent enterprises or industries by a graph propagation model, and carrying out future electricity consumption trend and risk prediction output by combining a time sequence prediction module to generate abnormal early warning signals in advance.
7. The industrial electricity consumption abnormal intelligent analysis method based on the multi-mode large model is characterized by comprising the steps of carrying out visualization processing on real-time abnormal detection and multi-dimensional intelligent analysis results, specifically, displaying histories and predicted abnormal tracks by adopting a real-time fluctuation curve and a periodic trend graph, displaying abnormal distribution thermodynamic diagrams based on a GIS system, comparing abnormal conditions of different industrial clusters by an industry dimension through Sang Ji graphs or radar graphs, generating abnormal cause link views by utilizing a knowledge graph structure, representing causal paths from external factors to electricity consumption abnormal, automatically generating periodic abnormal reports and treatment suggestions, wherein the reports comprise abnormal description, influence ranges, cause diagnosis and optimization strategies, and generating intelligent reports by combining natural languages of the large model.
8. An industrial electricity consumption abnormal intelligent analysis system based on a multi-mode large model is characterized by comprising a processor, a memory and a computer program stored on the memory, wherein the steps in the industrial electricity consumption abnormal intelligent analysis method based on the multi-mode large model are specifically executed when the processor executes the computer program.

Description

Intelligent analysis method and system for industrial electricity utilization abnormal movement based on multi-mode large model Technical Field The invention relates to the field of electric power monitoring, in particular to an industrial electricity consumption abnormal intelligent analysis method and system based on a multi-mode large model. Background With the rapid development of industrial modernization and digital economy, the industry electricity consumption behavior is increasingly complex and multiple. A large number of novel production equipment, automatic production lines and intelligent manufacturing systems are continuously connected into a power network, and enterprise power loads show stronger dynamic performance and uncertainty. Meanwhile, multiple external factors such as policy regulation, energy saving and carbon reduction, climate fluctuation, market fluctuation and the like also continuously influence the power demand and the power utilization structure, so that industry power utilization abnormal events frequently occur. The traditional electricity consumption monitoring and anomaly detection method mainly uses single time sequence analysis and static threshold judgment, is difficult to capture implicit complex association between multi-source data, and is also difficult to timely and accurately identify and explain essential reasons of electricity consumption fluctuation. Disclosure of Invention In order to solve the problems, the invention aims to provide the industrial electricity consumption abnormal intelligent analysis method and the system based on the multi-mode large model, which can face to the actual demands of the complex industry and realize the omnibearing intelligent monitoring of the industrial electricity consumption abnormal. In order to achieve the above purpose, the present invention adopts the following technical scheme: An intelligent analysis method for industrial electricity utilization abnormal state based on a multi-mode large model comprises the following steps: S1, acquiring multi-source data of a power system, and preprocessing to obtain a cross-mode standardized time sequence data set; s2, carrying out feature extraction based on a cross-modal standardized time sequence data set to construct a feature data set; S3, based on the characteristic data set, combining a multi-level association relation network to construct an industry electricity utilization abnormal knowledge graph; S4, constructing a multi-mode abnormal motion detection model integrating a time sequence transducer, a graph neural network and a knowledge embedding module, and training based on a characteristic data set and an electric power knowledge graph; S5, deploying the trained multi-mode model on a distributed stream processing platform to realize real-time abnormal movement detection and multi-dimensional intelligent analysis; And S6, carrying out visualization processing on the real-time abnormal detection and multidimensional intelligent analysis results. Further, multi-source data of the power system are obtained and preprocessed to obtain a standardized time sequence data set, wherein the multi-source data of the power system comprise power service data, production operation data, external environment data and unstructured information, the power service data comprise power consumption of industry users, load curves, voltage and current records and power factor high-frequency sampling data, the production operation data obtain output, production plans, working hours, process parameters and equipment operation states of an enterprise MES system, the external environment data comprise climate factors, time factors and price mechanisms, and the unstructured information comprises industry news, public opinion data, environment policy files, enterprise bulletin texts, equipment monitoring images and infrared image visual information; The method comprises the steps of carrying out unified cleaning and space-time alignment on multi-source data, removing noise by adopting a dynamic window anomaly detection algorithm, reconstructing and repairing incomplete data by utilizing a missing value, carrying out unified coding on text data after word segmentation, entity identification and semantic correction, carrying out normalization, target detection and feature extraction on image data, and finally outputting a cross-modal standardized time sequence data set in a unified format by means of Z-score standardization. Further, based on the characteristic data set, a multi-level association relation network is combined to construct a structured power knowledge graph, which is specifically characterized in that firstly, the multi-level association relation network between equipment and a user is constructed, wherein the multi-level association relation network comprises a user-equipment relation, an equipment-line topological relation and a geographic position relation, and modeling is performed in a