A Multi-Energy Outlier Collaborative Detection Method and System
By constructing a multi-energy knowledge graph and dynamic modal decomposition, combined with a causal reasoning model, the problems of data silos and delayed early warning in multi-energy systems are solved, enabling proactive early warning and accurate location of anomalies, and improving the system's situational awareness and operational efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA SOUTHERN POWER GRID COMPANY
- Filing Date
- 2025-11-12
- Publication Date
- 2026-06-30
AI Technical Summary
Existing technologies suffer from data silos, delayed early warnings, and difficulties in tracing the source of anomalies in multi-energy systems. They are unable to effectively identify the dynamic process of anomaly propagation and early weak characteristics between energy systems, resulting in delayed early warnings and difficulty in accurately locating the root cause of anomalies.
By constructing a multi-energy knowledge graph and combining dynamic mode decomposition and causal reasoning models, we can achieve real-time panoramic perception and forward-looking early warning of the operating status of multi-energy systems. We can use coupled dynamic law sets to track state trajectories and combine belief propagation algorithms and causal reasoning models to locate and trace anomalies.
It enables proactive early warning for multi-energy systems, with warning time 50% earlier, anomaly location accuracy improved by 30%, false alarm rate reduced by 25%, and operation and maintenance efficiency significantly improved.
Smart Images
Figure CN121705937B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent monitoring and fault diagnosis technology of energy systems, specifically involving a collaborative detection method and system for multiple energy outliers. Background Technology
[0002] With the rapid development of the energy internet and integrated energy systems, various energy systems such as electricity, heat, and gas are deeply coupled at both the physical and information levels, forming a complex multi-source heterogeneous energy network. Against this backdrop, the system's operational status exhibits highly nonlinear, strongly coupled, and dynamic propagation characteristics. Traditional energy management systems typically model and monitor single energy forms, with their detection methods built upon independent and closed data systems. This siloed management model leads to severe data barriers and information silos, making it impossible to perceive the mutual influence and energy interaction across energy categories from a holistic system perspective, and hindering the response to new operational risks arising from multi-energy coupling.
[0003] In existing technical solutions, the detection of anomalies in energy systems largely relies on alarms based on statistical thresholds or single signal processing algorithms. While these methods have some effectiveness within the energy system itself, they are essentially retrospective, meaning that alarms are only triggered after an anomaly has occurred and has a significant impact on the system, lacking foresight. More importantly, these methods generally lack the ability to model the system's inherent dynamic mechanisms and topological relationships. They treat data sequences as independent signals, ignoring the physical laws governing energy flow along the pipeline network topology. Therefore, they cannot capture the dynamic process of anomalies propagating between different energy systems, nor can they identify the early, subtle characteristics of the system's evolution from a steady state to a critical state. This results in severely delayed early warnings and is highly susceptible to frequent false alarms due to normal dynamic fluctuations in the system.
[0004] When anomalies occur, existing technologies often face significant challenges in pinpointing the root cause. Most methods stop at anomaly detection rather than anomaly tracing. While clustering or classification algorithms may identify anomalous data points, they cannot explain why the anomaly occurred or where it originated. Due to the lack of modeling the implicit causal logic within the system, when anomalies propagate through coupling paths between electrical, thermal, and gas systems, maintenance personnel struggle to sift through the complex alarm information to uncover the causal chain, ultimately relying on manual experience for troubleshooting—a process that is inefficient and lacks accuracy. Summary of the Invention
[0005] To overcome the problems of data silos, delayed early warning, and difficulty in tracing the source of problems in existing technologies, this invention provides a collaborative detection method, system, and storage medium for multi-energy outliers. By constructing a knowledge graph that integrates multi-source heterogeneous data, and combining it with a technical solution of dynamic mode decomposition and causal reasoning model, it achieves real-time panoramic perception of the operating status of multi-energy systems, forward-looking early warning from the stable domain to the critical domain, and accurate location and handling of the root cause of anomalies.
[0006] The specific technical solution of this application is as follows:
[0007] According to one aspect of this application, a multi-energy outlier cooperative detection method is provided, comprising:
[0008] Collect real-time operation data and equipment topology data of multi-energy systems, and construct a knowledge graph that includes energy equipment entities, pipeline topology relationships, and energy flow status.
[0009] Based on knowledge graphs, dynamic mode decomposition is used to analyze multi-energy data streams along the topological paths defined by the knowledge graphs, extracting the dominant dynamic modes and oscillation characteristics of energy systems at different time scales; through modal energy transfer analysis, the coupling dynamic laws of the dominant dynamic modes and oscillation characteristics across energy categories are identified, forming a set of coupling dynamic laws characterizing the overall operating state of the system.
[0010] Construct a high-dimensional state space for the energy system; use a set of coupled dynamic laws to track the trajectory of the energy system state in the high-dimensional state space; set three operating states for the energy system: stable domain, transition domain, and critical domain, and trigger a forward-looking warning when the state trajectory enters the transition domain.
[0011] If a forward-looking warning is triggered, a spatiotemporal evidence chain is constructed from monitoring data, status assessment results, and topological correlation information from different energy systems; a confidence propagation algorithm based on knowledge graph topology is adopted, combined with a spatiotemporal decay factor, to achieve distributed calculation of anomaly confidence.
[0012] Based on the distribution of anomaly confidence and the spatiotemporal evidence chain, a causal reasoning model is established; through intervention analysis and counterfactual reasoning, the root cause of the anomaly is located, and a complete solution including anomaly location, impact assessment and disposal recommendations is generated.
[0013] As a further optimization of this method, the knowledge graph is represented in a structured manner using entity-relationship-attribute as the basic unit;
[0014] The entities include energy equipment entities, the relationships include at least one of physical connection, energy flow direction, logical affiliation, and spatial co-location, and the attributes include real-time operating data and historical status data; the knowledge graph is stored in a graph database and formally represented as... , where: node set Represents various energy equipment entities; edge collection This represents the energy flow paths and topological connections between devices. A collection representing dynamic attributes.
[0015] As a further optimization of this method, the dominant dynamic mode and oscillation feature extraction steps include:
[0016] Based on the topological paths defined in the knowledge graph, the status data of devices along the paths are extracted and organized into a data matrix. ;
[0017] Data matrix Split into time offset matrix and ,in yes State vectors of all devices along the time path;
[0018] right Perform singular value decomposition: ,in and It is a unitary matrix. It is a singular value matrix;
[0019] Constructing the best-fit matrix in low-dimensional approximation : ;
[0020] For matrix Perform eigenvalue decomposition: Where Λ is a diagonal matrix, and its diagonal elements are... These are the eigenvalues of the dynamic mode. The column vectors of the eigenvector matrix W define the dynamic mode in the reduced-dimensional space.
[0021] As a further optimization of this method, after extracting the dominant dynamic mode, it is filtered by modal energy; for path The first There are 3 modes, and their modal energies are defined as follows: ,in It is the Frobenius norm; select the mode with the highest energy. Each mode is designated as the dominant dynamic mode, and the eigenvalues of the dominant mode are recorded. And the modal vector Φᵢ.
[0022] As a further optimization of this method, the step of forming the coupled dynamic law set includes:
[0023] Calculate the modal correlation among the dominant modes of different energy category pathways; the formula for calculating modal correlation is: ;in, It is the first The and the first Modal correlation coefficient between modes and No. The and the first The modal vector of each mode;
[0024] For correlated mode pairs, quantize the coupling strength coefficient. and transmission delay And recorded in the form of mode pair-coupling coefficient-transmission delay;
[0025] All identified cross-energy coupling patterns are summarized to form a set of coupling dynamic patterns. Its form is ,in and It is a mode vector with a coupling relationship.
[0026] As a further optimization of this method, the state space The dimensions correspond to the core state variables of key equipment entities in the knowledge graph;
[0027] The motion trajectory Tracing can be done through data-driven or model-driven methods, among which... At any moment The state.
[0028] As a further optimization of this method, the stability region, transition region and critical region are defined based on historical normal operation data, system physical constraints and coupled dynamic law set, and delineated by machine learning algorithms or optimization methods; wherein, the boundary of the stability region is defined as the high-density contour line of historical normal data points, the inner boundary of the transition region is the boundary of the stability region, and the outer boundary is determined by analyzing the common evolution characteristics of the state trajectory before instability.
[0029] As a further optimization of this method, the distributed computation of the belief propagation algorithm based on the knowledge graph topology includes:
[0030] For each section in the knowledge graph Initialize local anomaly confidence ;
[0031] The belief propagation algorithm is iteratively run on the knowledge graph topology, where each node... to neighboring nodes Send message Message computation takes into account the causal strength between nodes. ;
[0032] node Based on the received information and partial evidence Update anomaly confidence The update rule is as follows: ;in, From node Passed to node The news, It is a node Local evidence, It is a normalization constant. Represents a node The set of neighboring nodes;
[0033] Introducing the spatiotemporal decay factor Attenuate messages and evidence to ensure spatiotemporal relevance.
[0034] As a further optimization of this method, when conducting intervention analysis and counterfactual reasoning using a causal reasoning model, the following are included:
[0035] Constructing structural causal models or Bayesian networks based on knowledge graphs;
[0036] Intervene with candidate root cause nodes and set their status to normal.
[0037] Derive the state of other nodes in the system after intervention and compare it with the actual observed state;
[0038] If the system returns to normal after intervention, the node is determined to be the root cause, and a structured solution containing anomaly location, impact assessment, and handling recommendations is generated.
[0039] This application also provides a multi-energy outlier collaborative detection system, the system comprising:
[0040] The data acquisition and graph construction module is used to collect real-time operation data and equipment topology data of multi-energy systems, and to construct a knowledge graph that includes energy equipment entities, pipeline topology relationships and energy flow status.
[0041] The Coupled Dynamics Analysis module, connected to the Data Acquisition and Graph Construction module, is used to analyze multi-energy data streams based on knowledge graphs, employing dynamic mode decomposition along the topological paths defined by the knowledge graphs, extracting the dominant dynamic modes and oscillation characteristics of the energy system at different time scales; and identifying the coupled dynamic laws of the dominant dynamic modes and oscillation characteristics across energy categories through modal energy transfer analysis, forming a set of coupled dynamic laws characterizing the overall operating state of the system;
[0042] The state tracking and early warning module, connected to the coupled dynamic analysis module, is used to construct a high-dimensional state space of the energy system; it uses the coupled dynamic law set to track the motion trajectory of the energy system state in the high-dimensional state space; and based on the three energy system operating states of the set stable domain, transition domain, and critical domain, it triggers a forward-looking early warning when the state trajectory enters the transition domain.
[0043] The evidence chain and confidence calculation module, connected to the state tracking and early warning module, is used to construct a spatiotemporal evidence chain from monitoring data, state assessment results and topological association information from different energy systems when a forward-looking early warning is received; and adopts a confidence propagation algorithm based on knowledge graph topology structure, combined with spatiotemporal decay factor to realize the distributed calculation of abnormal confidence.
[0044] The causal reasoning and handling module, connected to the evidence chain and confidence calculation module, is used to establish a causal reasoning model based on the distribution of anomaly confidence and spatiotemporal evidence chains; through intervention analysis and counterfactual reasoning, it locates the root cause of the anomaly and generates a complete solution including anomaly location, impact assessment and handling recommendations.
[0045] The beneficial effects of this application are as follows:
[0046] This technical solution achieves proactive early warning of anomalies in multi-energy systems by constructing a multi-source heterogeneous knowledge graph and dynamic mode decomposition. Specifically, through state-space trajectory tracking and multi-domain partitioning, the system's operational status can be monitored in real time based on distance indicators, triggering an early warning when the system enters a transition domain. Compared to traditional a posteriori detection methods, the warning time is advanced by approximately 50%, effectively avoiding operational risks caused by delayed warnings. Simultaneously, by extracting cross-energy coupling patterns using dynamic mode decomposition, the dynamic characteristics of inter-energy interactions across multiple time scales can be captured. Experiments show that this method successfully identified over 85% of cross-energy coupling patterns in simulated data, significantly improving the overall situational awareness capability of the system.
[0047] This technical solution also achieves precise location of anomaly root causes and reduces false alarm rates through belief propagation algorithms and causal inference models. Utilizing spatiotemporal evidence chains and distributed belief propagation computation, it comprehensively considers topological structure and time factors, iteratively updating anomaly confidence levels. In test cases, anomaly location accuracy improved by over 30%, while the false alarm rate decreased by approximately 25%. Based on causal inference models for intervention analysis and counterfactual verification, it can quickly pinpoint the root cause from the candidate root cause set and generate structured solutions. In practical applications, this reduces the average troubleshooting time from several hours to minutes, significantly improving operational efficiency. Attached Figure Description
[0048] Figure 1 A schematic diagram of the overall process of the multi-energy outlier collaborative detection method;
[0049] Figure 2 Detailed flowchart of the S100 steps of the multi-energy outlier collaborative detection method;
[0050] Figure 3 Detailed flowchart of the S200 steps of the multi-energy outlier collaborative detection method;
[0051] Figure 4 Detailed flowchart of the S300 method for collaborative detection of multi-energy outliers;
[0052] Figure 5 Detailed flowchart of the S400 method for collaborative detection of multi-energy outliers;
[0053] Figure 6 Detailed flowchart of the S500 method for collaborative detection of multi-energy outliers; Detailed Implementation
[0054] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0055] As the physical carrier of the energy internet, integrated energy systems face challenges to safe and stable operation, including complex multi-energy flow coupling, highly time-varying dynamic processes, and concealed anomaly propagation paths. This invention achieves semantic unification and relational expression of multi-source heterogeneous data by constructing a multi-energy knowledge graph, extracts the essential dynamic features of the system along topological paths using dynamic mode decomposition, and achieves proactive early warning through trajectory tracking in the state space. Finally, it combines confidence propagation and causal reasoning to achieve precise anomaly localization and source tracing.
[0056] The multi-energy system operates as a time-varying dynamic system subject to topological constraints. Its state evolution process can be described as follows: ;in, Represents the system at time. The state vector is composed of the key state variables of all energy subsystems. It is the time-varying state matrix of the system, reflecting the dynamic characteristics of the system and the coupling relationship between energy devices; It is the input matrix; Represents external input. Time-varying state matrix. It contains the most important dynamic information and coupling laws of the energy system, but it is difficult to observe and obtain directly.
[0057] To address this issue, the present invention introduces a knowledge graph. The system is represented in a structured manner. This includes the node set. Representing various energy equipment entities, edge collection This represents the energy flow paths and topological connections between devices. A collection representing dynamic attributes.
[0058] Based on this, dynamic mode decomposition is applied to multi-energy time-series data streams along the topological paths defined in the knowledge graph. For a specific energy transport path in the graph... The device status data on it constitutes a data sequence. For data sequences The core of dynamic mode decomposition is finding the best-fitting linear operator. , making ,in and These are the data matrix and its time offset matrix, respectively. By analyzing... Eigenvalue decomposition is performed to obtain a series of dynamic mode vectors along the path. and their corresponding eigenvalues Eigenvalues The amplitude and phase reflect the growth rate and oscillation frequency of the mode, while the eigenvectors characterize the spatial structure of the mode.
[0059] Definition of the first The modal energy of each mode is ,in It is a dynamic mode vector The norm of [the energy type]. By analyzing the energy transfer relationships of dominant modes along different energy pathways, the cross-energy coupling strength is quantified, forming a set of coupling dynamic laws. Its mathematical expression is a set containing mode pairs, coupling coefficients, and propagation delays: .
[0060] To perform state assessment and early warning, the system's state space is constructed. Its dimensions are spanned by all key state variables in the knowledge graph. The system in The trajectory of motion Obtained through numerical integration or data-driven methods. Based on coupled dynamic law sets. And historical normal operation data, in the state space Three operational state domains are defined:
[0061] Stability region The area where the system can operate safely for a long time, where all state variables are within the safe threshold when the trajectory moves within this area.
[0062] Critical region This is a region where the system faces the risk of instability; the trajectory entering this region indicates that an anomaly is about to occur or has already occurred.
[0063] Transition domain The region between the stable region and the critical region is an early warning zone for the system's evolution from normal to abnormal.
[0064] Define system state points To the center of the stable domain The weighted Euclidean distance is used as an early warning indicator. : ;in, It is the first The weights of each state variable are determined by their relationship to the set of coupled dynamic laws. Participation level determines; and These are the current state and the stable center at the [number]th position, respectively. The value of the dimension. When the motion trajectory Entering the transition domain triggers a forward-looking warning.
[0065] To pinpoint the root cause of the anomaly, a causal reasoning model was constructed. This model is based on a knowledge graph. Its adjacency matrix is defined as If node Abnormal state may directly lead to node If the state is abnormal, then Otherwise, it is 0. This causal relationship can be determined through historical data analysis, physical principles, or expert knowledge. After receiving the warning, a spatiotemporal chain of evidence is constructed, and for each node... Calculate its anomaly confidence level A belief propagation algorithm based on knowledge graph topology is adopted, and its update rule is as follows: ;in, From node Passed to node The news, It is a node Local evidence, It is a normalization constant. Represents a node The set of neighboring nodes. Spatiotemporal decay factor. Introduced, among which It is the topological distance. The time difference is used to mitigate the impact of distant and outdated evidence. Through iterative propagation, the steady-state anomaly confidence distribution of all nodes is finally obtained. Based on this distribution, nodes with high confidence are selected as candidate root causes and verified through counterfactual reasoning: if the node is normal, will the current anomaly be eliminated? If the answer is yes, then the node is identified as the root cause.
[0066] The above theoretical framework provides a solid mathematical foundation for this invention, ensuring the synergy of multi-energy outlier detection, the foresight of early warning, and the interpretability of source tracing. The specific implementation methods of this invention will be described in detail below.
[0067] Please see Figure 1 This illustrates a multi-energy outlier collaborative detection method provided by an embodiment of the present invention, the method comprising:
[0068] S100: Collects real-time operation and topology data of multiple energy systems to construct a multi-source heterogeneous knowledge graph that integrates equipment, pipelines, and energy flows.
[0069] S200: Based on the knowledge graph topology path, dynamic mode decomposition is used to extract the dominant dynamic modes and cross-energy coupling laws under multiple time scales.
[0070] S300: Constructs a state space, tracks motion trajectories based on coupled dynamic laws, and divides stable, transitional, and critical domains to achieve forward-looking early warning.
[0071] S400: After the early warning is triggered, it integrates multi-source monitoring data and topological information to construct a spatiotemporal evidence chain, and calculates the anomaly confidence level in a distributed manner through the confidence propagation algorithm.
[0072] S500: Combining anomalous confidence distribution with spatiotemporal evidence chains, it uses causal reasoning models to locate the root cause and generate solutions that include assessment and action recommendations.
[0073] The specific plan is as follows:
[0074] In a collaborative detection method for multi-energy outliers, S100 achieves unified modeling and semantic association of heterogeneous energy data from multiple sources such as electricity, heat, and gas, providing a structured data foundation for subsequent collaborative analysis. The knowledge graph integrates physical devices, topological connections, and real-time operational data to form a panoramic digital mapping of the integrated energy system.
[0075] Please refer to Figure 2 The diagram illustrates a flowchart of an exemplary multi-energy outlier cooperative detection method S100 of this application, which includes:
[0076] S110: Multi-source heterogeneous energy data acquisition.
[0077] In terms of multi-energy data acquisition, this invention adopts a unified data access architecture based on edge gateways, which is deployed in the control centers or key stations of various energy systems to achieve cross-system data aggregation.
[0078] Specifically, the acquisition of multi-source heterogeneous energy data includes: power system data, thermal system data, gas system data, and equipment topology and asset data.
[0079] In one alternative implementation, data acquisition employs a combination of Industrial Internet of Things (IIoT) protocols and energy industry standard protocols to ensure the real-time nature and reliability of the data.
[0080] S120: Knowledge Graph Construction.
[0081] Knowledge graph construction involves transforming collected multi-source data into a structured knowledge network with entities, relationships, and attributes as basic units.
[0082] In one possible implementation, knowledge graph construction includes the following steps:
[0083] Energy equipment entities are extracted from equipment asset data and topology data, and each entity is assigned a globally unique identifier. The same entity described in different data sources is then aligned and merged.
[0084] Based on physical connections and energy flow direction, define and establish relationships between entities. The main relationship types include: Indicates physical connection, Indicates the direction of energy flow. Indicates logical attribution. Indicates spatial co-location.
[0085] Real-time operational data and historical status data are dynamically associated with the corresponding device entities, with each data point having a timestamp.
[0086] The constructed knowledge graph is stored in a graph database.
[0087] Completed knowledge graph It can be formally represented as Among them: node set Representing various energy equipment entities, edge set This represents the energy flow paths and topological connections between devices. Represents a node and There exists a type of Relationship; Represents a collection of dynamic attributes, with each node or edge having multiple attribute-value pairs.
[0088] For example, a gas turbine entity has attributes such as current power and efficiency, and is... The relationship is connected to the distribution network entity, and simultaneously through The relationship is connected to the waste heat boiler entity.
[0089] In a collaborative detection method for outliers in multiple energy sources, S200 uses a knowledge graph as a guide to perform dynamic mode decomposition on multiple energy data streams. This breaks through the limitations of traditional methods that only analyze single data sequences, thereby enabling the extraction of dynamic features and coupling laws that reflect the essence of system topology and energy flow.
[0090] Please refer to Figure 3 The diagram illustrates a flowchart of an exemplary multi-energy outlier cooperative detection method S200 of this application, the contents of which include:
[0091] S210: Data flow organization based on topology path.
[0092] Based on the relationships between entities defined in the knowledge graph, such as Identify the main energy transfer paths in the system by maintaining the same energy flow relationship. For each path Extract the status data of all devices located on the path within a specific time window, and organize them into a data matrix according to the order on the path. This ensures that the dynamic analysis is performed on a physically consistent topology.
[0093] S220: Path dynamic mode decomposition.
[0094] For each path Corresponding data matrix Perform dynamic mode decomposition to extract the dominant dynamic modes on the path.
[0095] In one possible implementation, the specific steps of path dynamic mode decomposition are as follows:
[0096] Data matrix Split into two time offset matrices: and ,in yes The state vectors of all devices on the time path.
[0097] For matrix Perform singular value decomposition: ,in and It is a unitary matrix. It is a singular value matrix.
[0098] Constructing the best-fit matrix in low-dimensional approximation : .
[0099] For matrix Perform eigenvalue decomposition: Where Λ is a diagonal matrix, and its diagonal elements are... These are the eigenvalues of the dynamic modes. The column vectors of the eigenvector matrix W define the dynamic modes in the reduced-dimensional space.
[0100] Mapping the modes in the reduced-dimensional space back to the original state space yields the dynamic modes in the original space. . Each column represents a dynamic mode, and its elements correspond to the path. The degree of participation and oscillation phase of each device in this mode.
[0101] S230: Dominant mode selection and feature extraction.
[0102] This step filters based on modal energy. For path The first There are 3 modes, and their modal energies are defined as follows: ,in It is the Frobenius norm. Select the mode with the highest energy. Each mode is designated as the dominant dynamic mode of the path. The eigenvalues of the dominant mode are recorded. And the modal vector Φᵢ.
[0103] S240: Forming a set of coupling rules.
[0104] In one possible implementation, the specific steps for forming the coupled dynamic law set include:
[0105] For the dominant modes from different energy category pathways, the modal correlation between the dominant modes is calculated to determine whether there is a statistical association between them. The formula for calculating the modal correlation between the dominant modes is as follows: ;in, and It is the first The and the first The mode vector of each mode.
[0106] For correlated mode pairs, analyze the energy transfer relationship, i.e., from the mode... To mode Coupling strength coefficient One feasible approach is to examine the cross-spectral density of the energy time series of the two modes or to quantize it using information theory tools such as transfer entropy.
[0107] The identified coupling relationships are recorded in the form of mode pairs-coupling coefficients-transmission delays, which is as follows: .in and It is a mode vector with coupling relationship. It is a mode vector To mode vector The coupling strength coefficient, It is a transmission delay.
[0108] All identified cross-energy coupling patterns are summarized to form a set of coupling dynamic patterns. Its form is .
[0109] In a multi-energy outlier collaborative detection method, S300 constructs a state space and tracks motion trajectories, linking the modes extracted by S200 with the macroscopic operating state, and achieves true forward-looking early warning through multi-domain partitioning.
[0110] Please refer to Figure 4 The diagram illustrates a flowchart of an exemplary multi-energy outlier cooperative detection method S300 of this application, the contents of which include:
[0111] S310: Construct the state space and perform system state trajectory tracking.
[0112] System state space Each dimension corresponds to a core state variable of a key device entity in the knowledge graph.
[0113] As the system operates over time, in the state space A trajectory is formed in the middle ,in, At any moment The state.
[0114] In one possible implementation, trajectory tracking is achieved through data-driven methods, that is, by directly utilizing the acquired real-time data sequence in the state space. Draw a line connecting the points in the middle.
[0115] In another possible implementation, trajectory tracking is achieved through model-driven methods, i.e., utilizing the set of coupled dynamic laws obtained from S200. A dynamic model of the system is established, and the state evolution is deduced through numerical integration, thereby generating the trajectory.
[0116] S320: Multi-operational state domain delimitation.
[0117] Based on historical normal operation data, system physical constraints, and coupled dynamic law sets In the state space The three regions are defined using machine learning algorithms or optimization methods:
[0118] Stability region This region contains the vast majority of the system's state points during normal operation. Its boundary is defined as a high-density contour line of historical normal data points.
[0119] Critical region This region is typically close to the system's physical limits or the state region where historical failures occurred.
[0120] Transition domain This region is a buffer zone between the stable region and the critical region. Its inner boundary is the boundary of the stable region, and its outer boundary is determined by analyzing the common evolution characteristics of the state trajectory before instability. For example, the average time required for the trajectory to reach the critical region after leaving the stable region is calculated, and the state region corresponding to half of this time is designated as the transition region.
[0121] S330: Forward-looking early warning triggered.
[0122] Real-time calculation of the current state point Distance index relative to the stability region Continuously monitor distance metrics. Once the movement trajectory is detected Upon entering the transition domain, a forward-looking warning is immediately triggered.
[0123] In a collaborative detection method for multi-energy outliers, the S400 initiates a refined anomaly confidence assessment process after receiving an early warning. By constructing a spatiotemporal evidence chain and utilizing a confidence propagation algorithm on a graph structure, efficient and distributed computation of anomaly source confidence on a knowledge graph is achieved.
[0124] Please refer to Figure 5 The diagram illustrates a flowchart of an exemplary multi-energy outlier cooperative detection method S400 of this application, the contents of which include:
[0125] S410: Construction of the spatiotemporal evidence chain.
[0126] Once an alert is triggered, the system not only focuses on the data at the current moment, but also traces back the historical data from a period of time before the alert and integrates the topological information of the knowledge graph to form a spatiotemporal evidence chain.
[0127] In one possible implementation, the spatiotemporal evidence chain includes: temporal evidence, spatial evidence, and state assessment evidence.
[0128] Specifically, temporal evidence refers to the degree of abnormal deviation of each device's state variables within the warning time window. Spatial evidence refers to the inter-device connectivity and energy flow relationships recorded in the knowledge graph. State assessment evidence refers to the state-space distance index calculated in S300. .
[0129] S420: Initialize local anomaly confidence.
[0130] For each node in the knowledge graph Determine an initial local anomaly confidence. .
[0131] In one possible implementation, each node Initial local anomaly confidence It is determined based on the degree to which its state variables deviate from the normal range.
[0132] S430: Distributed computing based on the belief propagation algorithm.
[0133] Distributed computing based on the belief propagation algorithm will run iteratively on the topology defined by the knowledge graph until the abnormal confidence of all nodes converges.
[0134] In one possible implementation, the confidence propagation algorithm iteratively runs on the knowledge graph topology, including the following steps:
[0135] In each iteration, each node It will send to all its neighboring nodes Send a message . Represents a node Based on one's current confidence level and from other neighbors (except...) The message received is for the node. Opinions or support levels regarding abnormal states. Message computation considers the causal strength between nodes. .
[0136] Each node After receiving messages from all the neighbors, I will combine them with my own partial evidence. Update its own abnormal confidence level The update rule is as follows: ;in, From node Passed to node The news, It is a node Local evidence, It is a normalization constant. Represents a node The set of neighboring nodes.
[0137] Introducing a spatiotemporal decay factor during message passing and confidence updates. For messages transmitted between nodes with large topological distances, their influence is multiplied by a spatiotemporal decay factor less than 1. Similarly, for evidence separated by a long time interval, its spatiotemporal decay factor... It will also be attenuated.
[0138] S440: Steady-state anomaly confidence distribution output.
[0139] After several iterations, the anomaly confidence scores of all nodes no longer change significantly, and the algorithm converges. The final steady-state anomaly confidence score distribution is output. It visually demonstrates the likelihood of each device being identified as an anomaly source within the entire integrated energy system.
[0140] In a multi-energy outlier collaborative detection method, S500, based on the anomaly confidence distribution calculated by S400 and the constructed spatiotemporal evidence chain, establishes a dynamic causal reasoning model to conduct rigorous intervention analysis and counterfactual verification on candidate anomaly sources, thereby accurately locating the root cause and generating a complete and operable solution.
[0141] Please refer to Figure 6 The diagram illustrates a flowchart of an exemplary multi-energy outlier cooperative detection method S500 of this application, the contents of which include:
[0142] S510: Determination of candidate root cause set based on confidence level and evidence chain.
[0143] This step involves selecting the most likely key suspects from a large pool of devices to identify those most likely to cause the current system anomaly. Its inputs are the steady-state anomaly confidence distribution output by S440 and the spatiotemporal evidence chain constructed by S410.
[0144] By combining confidence level, chronological order, and topological position, a candidate root cause set is finally formed.
[0145] S520: Construct a causal reasoning model and conduct intervention analysis and counterfactual reasoning through the causal reasoning model.
[0146] Based on the causal relationship edges already defined in the knowledge graph, a causal graph model is constructed. The causal graph model clarifies the causal direction and mechanism of action between node states.
[0147] In one possible implementation, the causal graphical model is a structural causal model or a Bayesian network.
[0148] For each candidate root cause node in the candidate root cause set, perform the following counterfactual reasoning:
[0149] In a causal model, it is assumed that an intervention is made on a certain node, that is, its state is forcibly set to the normal value.
[0150] Under the conditions of this intervention, the states of other nodes in the system are derived using a causal model.
[0151] The system state obtained through reasoning is compared with the actual observed state.
[0152] If, after intervention, the inferred system state is very close to the normal state, and the currently observed large-scale anomalies also disappear, then it proves that the node is the root cause of the current anomaly.
[0153] Once the root cause is identified, the system will automatically generate a structured diagnostic report and response plan, including: anomaly location, impact assessment, and response recommendations.
[0154] Example 2
[0155] According to Embodiment 1 of this application, a multi-energy outlier cooperative detection system is provided, the system comprising:
[0156] The data acquisition and graph construction module is used to collect real-time operation data and equipment topology data of multi-energy systems, and to construct a knowledge graph that includes energy equipment entities, pipeline topology relationships and energy flow status.
[0157] The Coupled Dynamics Analysis module, connected to the Data Acquisition and Graph Construction module, is used to analyze multi-energy data streams based on knowledge graphs, employing dynamic mode decomposition along the topological paths defined by the knowledge graphs, extracting the dominant dynamic modes and oscillation characteristics of the energy system at different time scales; and identifying the coupled dynamic laws of the dominant dynamic modes and oscillation characteristics across energy categories through modal energy transfer analysis, forming a set of coupled dynamic laws characterizing the overall operating state of the system;
[0158] The state tracking and early warning module, connected to the coupled dynamic analysis module, is used to construct a high-dimensional state space of the energy system; it uses the coupled dynamic law set to track the motion trajectory of the energy system state in the high-dimensional state space; and based on the three energy system operating states of the set stable domain, transition domain, and critical domain, it triggers a forward-looking early warning when the state trajectory enters the transition domain.
[0159] The evidence chain and confidence calculation module, connected to the state tracking and early warning module, is used to construct a spatiotemporal evidence chain from monitoring data, state assessment results and topological association information from different energy systems when a forward-looking early warning is received; and adopts a confidence propagation algorithm based on knowledge graph topology structure, combined with spatiotemporal decay factor to realize the distributed calculation of abnormal confidence.
[0160] The causal reasoning and handling module, connected to the evidence chain and confidence calculation module, is used to establish a causal reasoning model based on the distribution of anomaly confidence and spatiotemporal evidence chains; through intervention analysis and counterfactual reasoning, it locates the root cause of the anomaly and generates a complete solution including anomaly location, impact assessment and handling recommendations.
[0161] Those skilled in the art will understand that the embodiments of this application are provided as methods, systems, or computer program products. Therefore, this application takes the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application takes the form of a computer program product implemented on one or more computer storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer program code. The solutions in the embodiments of this application are implemented using various computer languages, exemplified by the object-oriented programming language Java and the interpreted scripting language JavaScript.
[0162] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. 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, are implemented by computer program instructions. These computer program instructions are 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 process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0163] These computer program instructions are also stored in a computer read-memory memory (CROM) that can direct a computer or other programmed data processing device to operate in a specific manner, such that the instructions stored in the CROM 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.
[0164] These computer program instructions are also loaded onto a computer or other programmed data processing device, causing a series of operational steps to be performed on the computer or other programmed device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmed device 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.
[0165] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0166] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A multi-energy outlier collaborative detection method, characterized in that, include: Collect real-time operation data and equipment topology data of multi-energy systems, and construct a knowledge graph that includes energy equipment entities, pipeline topology relationships and energy flow status; Based on knowledge graphs, dynamic mode decomposition is used to analyze multi-energy data streams along the topological paths defined by the knowledge graphs, extracting the dominant dynamic modes and oscillation characteristics of energy systems at different time scales; through modal energy transfer analysis, the coupling dynamic laws of the dominant dynamic modes and oscillation characteristics across energy categories are identified, forming a set of coupling dynamic laws characterizing the overall operating state of the system; Constructing a high-dimensional state space for the energy system; By utilizing coupled dynamic law sets, the motion trajectory of the energy system state in a high-dimensional state space can be tracked; Three operating states of the energy system are defined: stable region, transition region, and critical region. A forward-looking warning is triggered when the state trajectory enters the transition region. If a forward-looking warning is triggered, a spatiotemporal evidence chain will be constructed from monitoring data, status assessment results, and topological correlation information from different energy systems. A belief propagation algorithm based on knowledge graph topology is adopted, combined with a spatiotemporal decay factor to realize the distributed calculation of anomaly confidence. Based on the distribution of anomaly confidence and the spatiotemporal evidence chain, a causal reasoning model is established; through intervention analysis and counterfactual reasoning, the root cause of the anomaly is located, and a complete solution including anomaly location, impact assessment and disposal recommendations is generated. The steps for extracting dominant dynamic modes and oscillation features include: Based on the topological paths defined in the knowledge graph, the status data of devices along the paths are extracted and organized into a data matrix. ; Data matrix Split into time offset matrix and ,in yes State vectors of all devices along the time path; right Perform singular value decomposition: ,in and It is a unitary matrix. It is a singular value matrix; Constructing the best-fit matrix in low-dimensional approximation : ; For matrix Perform eigenvalue decomposition: Where Λ is a diagonal matrix, and its diagonal elements are... These are the eigenvalues of the dynamic mode. The column vectors of the eigenvector matrix W define the dynamic mode in the reduced-dimensional space. Distributed computation of the belief propagation algorithm based on knowledge graph topology includes: For each section in the knowledge graph Initialize local anomaly confidence ; The belief propagation algorithm is iteratively run on the knowledge graph topology, where each node... to neighboring nodes Send message Message computation takes into account the causal strength between nodes. ; node Based on the received information and partial evidence Update anomaly confidence The update rule is as follows: ;in, From node Passed to node The news, It is a node Local evidence, It is a normalization constant. Represents a node The set of neighboring nodes; Introducing the spatiotemporal decay factor Attenuate messages and evidence to ensure spatiotemporal relevance.
2. The multi-energy outlier collaborative detection method according to claim 1, characterized in that, The knowledge graph is represented in a structured manner using entities, relations, and attributes as basic units; The entities include energy equipment entities, the relationships include at least one of physical connection, energy flow direction, logical affiliation, and spatial co-location, and the attributes include real-time operating data and historical status data; the knowledge graph is stored in a graph database and formally represented as... , where: node set Represents various energy equipment entities; edge collection This represents the energy flow paths and topological connections between devices. A collection representing dynamic attributes.
3. The multi-energy outlier collaborative detection method according to claim 1, characterized in that, After extracting the dominant dynamic modes, they are filtered by modal energy; for path The first There are 3 modes, and their modal energies are defined as follows: ,in It is the Frobenius norm; select the mode with the highest energy. Each mode is designated as the dominant dynamic mode, and the eigenvalues of the dominant mode are recorded. And the modal vector Φᵢ.
4. The multi-energy outlier collaborative detection method according to claim 3, characterized in that, The steps for forming the coupled dynamic law set include: Calculate the modal correlation among the dominant modes of different energy category pathways; the formula for calculating modal correlation is: ;in, It is the first The and the first Modal correlation coefficient between modes and No. The and the first The modal vector of each mode; For correlated mode pairs, quantize the coupling strength coefficient. and transmission delay And recorded in the form of mode pair-coupling coefficient-transmission delay; All identified cross-energy coupling patterns are summarized to form a set of coupling dynamic patterns. Its form is ,in and It is a mode vector with a coupling relationship.
5. The multi-energy outlier collaborative detection method according to claim 1, characterized in that, The state space The dimensions correspond to the core state variables of key equipment entities in the knowledge graph; The motion trajectory Tracing can be done through data-driven or model-driven methods, among which... At any moment The state.
6. The multi-energy outlier collaborative detection method according to claim 1, characterized in that, The stability region, transition region, and critical region are defined based on historical normal operation data, system physical constraints, and a set of coupled dynamic laws, using machine learning algorithms or optimization methods. The boundary of the stability region is defined as a high-density contour line of historical normal data points, the inner boundary of the transition region is the boundary of the stability region, and the outer boundary is determined by analyzing the common evolution characteristics of the state trajectory before instability.
7. The multi-energy outlier collaborative detection method according to claim 1, characterized in that, When conducting intervention analysis and counterfactual reasoning using causal reasoning models, the following are included: Constructing structural causal models or Bayesian networks based on knowledge graphs; Intervene with candidate root cause nodes and set their status to normal. Derive the state of other nodes in the system after intervention and compare it with the actual observed state; If the system returns to normal after intervention, the node is determined to be the root cause, and a structured solution containing anomaly location, impact assessment, and handling recommendations is generated.
8. A multi-energy outlier collaborative detection system, characterized in that, The system for implementing the multi-energy outlier collaborative detection method as described in any one of claims 1 to 7, the system comprising: The data acquisition and graph construction module is used to collect real-time operation data and equipment topology data of multi-energy systems, and to construct a knowledge graph that includes energy equipment entities, pipeline topology relationships and energy flow status. The Coupled Dynamics Analysis module, connected to the Data Acquisition and Graph Construction module, is used to analyze multi-energy data streams based on knowledge graphs, employing dynamic mode decomposition along the topological paths defined by the knowledge graphs, extracting the dominant dynamic modes and oscillation characteristics of the energy system at different time scales; and identifying the coupled dynamic laws of the dominant dynamic modes and oscillation characteristics across energy categories through modal energy transfer analysis, forming a set of coupled dynamic laws characterizing the overall operating state of the system; The state tracking and early warning module, connected to the coupled dynamic analysis module, is used to construct a high-dimensional state space of the energy system; it uses the coupled dynamic law set to track the motion trajectory of the energy system state in the high-dimensional state space; and based on the three energy system operating states of the set stable domain, transition domain, and critical domain, it triggers a forward-looking early warning when the state trajectory enters the transition domain. The evidence chain and confidence calculation module, connected to the state tracking and early warning module, is used to construct a spatiotemporal evidence chain from monitoring data, state assessment results and topological association information from different energy systems when a forward-looking early warning is received; and adopts a confidence propagation algorithm based on knowledge graph topology structure, combined with spatiotemporal decay factor to realize the distributed calculation of abnormal confidence. The causal reasoning and handling module, connected to the evidence chain and confidence calculation module, is used to establish a causal reasoning model based on the distribution of anomaly confidence and spatiotemporal evidence chains; through intervention analysis and counterfactual reasoning, it locates the root cause of the anomaly and generates a complete solution including anomaly location, impact assessment and handling recommendations.