Smart park power distribution equipment intelligent operation and maintenance method and system based on big data

Abstract

The invention provides a smart park power distribution equipment intelligent operation and maintenance method and system based on big data, and relates to the technical field of smart park operation and maintenance management.The method comprises the steps that firstly, a structured expression mode of operation behaviors of a multi-park power distribution system is constructed, and the relationship between common operation characteristics of parks and equipment behaviors is extracted; rapid modeling and migration prediction of the new park are realized; secondly, an intelligent layered response mechanism based on abnormal event triggering is constructed, and an adaptive analysis and diagnosis strategy can be dynamically selected according to the type and grade of an abnormal event; meanwhile, a closed-loop process from event triggering to fault tracing, equipment positioning and response strategy making is realized in combination with a data driving and structured reasoning method; in addition, various heterogeneous data sources are fused, an information barrier is broken through, the comprehensive judgment of the system is improved, the output of the AI model is verified by introducing electrical physical constraint conditions, and it is ensured that the result logic is reasonable.

Description

Technical Field

[0001] The present invention relates to the technical field of smart park operation and maintenance management, and in particular to a big data-based intelligent operation and maintenance method and system for smart park power distribution equipment. Background Art

[0002] Against the backdrop of the current acceleration of energy digitization and smart city development, smart campuses have become a key application scenario for power distribution systems. Campus power distribution systems typically feature clear boundaries and moderate scale, but complex equipment types and significantly differentiated user loads. This provides a solid technical foundation for intelligent operations and maintenance.

[0003] However, during actual deployment, the intelligent operation and maintenance system for campus power distribution equipment still faces a number of prominent problems. First, current operation and maintenance systems generally rely on local historical data for model training. This results in difficulty in starting the model due to data scarcity and insufficient labels during the initial construction of the campus or when new equipment is online, resulting in a significant cold start problem. Second, the significant differences in topology, equipment types, and power usage behavior between different campuses make it difficult to migrate trained models and their versatility is poor, limiting the rapid replication and promotion of intelligent operation and maintenance systems across multiple campuses. Third, existing systems typically use a unified model to respond to various anomalies, lacking the ability to identify the severity and type of abnormal events. This leads to confusion between the severity of the anomaly, redundant or untimely processing, and increased false alarm rates and system response delays. At the same time, many AI models output results in a black box manner, lacking explainability, making it difficult to trace the source or guide specific operation and maintenance measures. This is especially weak when it comes to chain failures or multi-device coupling issues. Furthermore, smart campuses contain numerous sensors and complex data types, including heterogeneous information such as voltage and current waveforms, equipment temperature, humidity, sound, and operation logs. Currently, there is a lack of effective integration and unified analysis mechanisms, resulting in information silos and further weakening intelligent judgment capabilities. More critically, some AI prediction methods ignore physical constraints, resulting in results that contradict electrical logic and undermine system reliability.

[0004] In summary, the existing power distribution equipment operation and maintenance solutions have significant shortcomings in data cold start, model migration, response strategy, explainability, multi-source fusion and physical consistency, and urgently need systematic innovative solutions. Summary of the Invention

[0005] In response to the above problems, the present invention proposes a smart operation and maintenance method and system for power distribution equipment in smart parks based on big data, focusing on solving practical pain points such as the inability to deploy models due to data scarcity in newly built parks, difficulty in transferring knowledge between multiple parks, low efficiency in abnormal response, and unexplainable operation and maintenance results.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] In a first aspect of the present invention, a method for intelligent operation and maintenance of power distribution equipment in a smart park based on big data is provided, comprising the following steps:

[0008] S1. Collect and pre-process multi-source data from the smart park, including electrical operation data, environmental monitoring data, structural information data, and operation and maintenance logs;

[0009] S2. Construct a cross-campus portable causal structure map based on the preprocessed data, divide the devices into several priority levels, calculate the dynamic granularity priority score for each node in the causal structure map, and perform the resampling operation on each node;

[0010] S3. In multiple deployed historical smart parks, a set of causal structure subgraphs with stability and migration capabilities are extracted from the pre-processed multi-source data, which are used as the structure for constructing new park models;

[0011] S4. In the actual deployment and operation of smart parks, based on the common causal structure graph and the online real-time data streams of multiple sensors, dynamic identification of multimodal abnormal data that may appear during operation is carried out, and it is encapsulated into structured event objects and their corresponding feature input tensors;

[0012] S5. When an abnormal event occurs, conduct abnormality tracking and locate the root cause device at the structural path level based on the common causal structure map and structured abnormality feature input;

[0013] S6. After completing the exception tracing, generate intelligent operation and maintenance response suggestions that can be directly implemented.

[0014] Preferably, in S1, the electrical operation data includes the three-phase current, voltage and frequency of electrical equipment A, B and C collected by the telemetry unit; the environmental monitoring data includes the temperature and humidity in the equipment cabin, infrared imaging temperature and noise signal strength; the structural information data includes the structural data of the primary wiring diagram parsed into a diagram, including the connection relationship between the devices.

[0015] Preferably, in S2, a hierarchical sampling score function for the structure graph nodes is introduced to divide the devices into several priority levels, and the dynamic granularity priority score of each node is:

[0016] ;

[0017] in, For nodes Total sampling score of For its structure diagram The centrality score in ; Representing a tensor Middle Dimension in time The value of is derived from the electrical or environmental sensor signal of the field device; is a unified number of time slices; A nonlinear difference function designed to enhance change detection, whose cubic term is used to emphasize sudden changes; is the change incentive factor; is the structural / dynamic equilibrium parameter.

[0018] More preferably, after the dynamic granularity priority score is calculated, all nodes are sorted according to the dynamic granularity priority score and divided into L layers, each layer corresponding to a different sampling window length. After performing the resampling operation on each node, the length is padded to obtain a resampled tensor set.

[0019] Preferably, in S3, the connection direction, occurrence frequency and variable correspondence of the edges of the causal structure subgraph with stability and migration ability are fixed, and the structural matching area with the causal structure subgraph is found in the new park topology structure, and the migration model is based on it.

[0020] Preferably, in S4, a structure-guided linkage offset function is designed, and its specific calculation formula is as follows:

[0021] ;

[0022] in, represents a causal edge in the structure graph, is the frequency of the edge appearing in all parks, represents the standard deviation normalized offset representation of the data, It represents the standard deviation normalized offset value of variable k at time t, which is used to measure the degree of deviation of its current observation value from the historical steady-state benchmark;

[0023] By setting a dynamically adjusted trigger threshold, when When the value of is greater than the trigger threshold, it is determined that the current state is a triggerable event state, and an abnormal event object is generated. The abnormal event object contains information such as the current timestamp, trigger path, abnormal level, etc., and the current data window is recorded at the same time, and the current data window is defined as a structured input.

[0024] Preferably, in S6, the intelligent operation and maintenance response suggestion includes:

[0025] Low-risk suggestion: Generate a dispatch patrol work order;

[0026] Medium-risk advice: Call on-duty personnel for offline inspection;

[0027] High-risk advice: Immediately issue isolation and power-off instructions, and arrange for personnel evacuation.

[0028] In a second aspect of the present invention, there is also provided a smart park power distribution equipment intelligent operation and maintenance system based on big data, comprising:

[0029] Data acquisition and processing module, used to collect multi-source data of smart parks and perform pre-processing;

[0030] A causal structure map construction module is used to construct a cross-campus portable causal structure map based on preprocessed data, divide the devices into several priority levels, calculate the dynamic granularity priority score of each node in the causal structure map, and perform the resampling operation on each node;

[0031] The new park structure migration module is used to extract causal structure subgraphs with stability and migration capabilities, which are used as the structure for constructing the new park model;

[0032] Anomaly recognition module, which is used to dynamically identify multimodal anomaly data that may occur during operation based on common causal structure graphs and online real-time data streams from multiple sensors;

[0033] The anomaly tracking and positioning module is used to track anomalies at the structural path level and locate the root cause device based on the common causal structure map and structured anomaly feature input when an abnormal event occurs;

[0034] The intelligent report generation module is used to generate intelligent operation and maintenance response suggestions that can be directly implemented.

[0035] The beneficial effects of the present invention are:

[0036] This invention first constructs a structured representation of the operating behavior of multi-park power distribution systems. By extracting common operating characteristics and equipment behavior relationships across parks, it enables rapid modeling and migration prediction for new parks, effectively overcoming the model cold start problem. Secondly, this invention constructs an intelligent hierarchical response mechanism based on abnormal event triggering. This mechanism dynamically selects appropriate analysis and diagnostic strategies based on the type and level of the abnormal event, improving system response speed and resource utilization efficiency.

[0037] At the same time, by combining data-driven and structured reasoning methods, a closed-loop process from event triggering to fault tracing, equipment location, and response strategy formulation is achieved, enhancing the system's interpretability and operational guidance capabilities. This method also integrates multiple heterogeneous data sources, breaks down information barriers, improves the system's comprehensive judgment, and verifies the output of the AI ​​model by introducing electrical and physical constraints to ensure that the results are logically reasonable and can be actually adopted. The overall solution builds a highly integrated intelligent operation and maintenance system from modeling, monitoring, analysis, interpretation, and response. It comprehensively solves the key obstacles in existing technologies and has strong adaptability, deployability, and engineering feasibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of the intelligent operation and maintenance method of smart park power distribution equipment based on big data of the present invention.

[0039] Figure 2 This is a block diagram of the intelligent operation and maintenance system of smart park power distribution equipment based on big data of the present invention. DETAILED DESCRIPTION

[0040] See also Figure 1 As shown, in a first aspect of the present invention, a method for intelligent operation and maintenance of power distribution equipment in a smart park based on big data is provided, comprising the following steps:

[0041] S1. Collect and pre-process multi-source data from the smart park, including electrical operation data, environmental monitoring data, structural information data, and operation and maintenance logs;

[0042] The goal of this step is to build a standardized data representation system that can be used for subsequent causal structure learning and model migration, to uniformly represent the status and structure of power distribution equipment from multiple historical smart parks. Due to the diverse types of power distribution equipment, widely varying operating environments, and inconsistent topological structures in smart parks, data cannot be directly spliced ​​together in the conventional way. Instead, the input must be reconstructed and enhanced to meet the needs of subsequent tasks. The innovation of this step lies in the fusion encoding of time series data, environmental monitoring data, and electrical structure information into feature tensors, and the introduction of a graph structure confidence screening mechanism combined with edge stability, thereby obtaining an input data structure that is more transferable and valuable for causal modeling.

[0043] Input is A collection of original distribution system data from historical parks , each It contains four types of data. The first type is electrical operation data, which usually comes from telemetry units (such as DTU and FTU) inside equipment such as switch cabinets, ring network cabinets, and transformers. The collected parameters include A, B, and C three-phase current, voltage, and frequency, etc., and are sampled at a granularity of one minute to form a time series tensor. The second category is environmental monitoring data, such as temperature and humidity in the equipment cabin, infrared imaging temperature, noise signal strength, etc., which are collected by temperature and humidity sensors, infrared cameras, and sound sensors, and then aligned with electrical data by timestamp to form a matrix. The third category is structural information, which is derived from a wiring diagram from a BIM or GIS system and parsed into a diagram structure. ,in Represents a device node, Indicates the connection relationship between devices. The fourth category is operation and maintenance logs , the format is structured text, which comes from the park energy management platform, but is not used in this step.

[0044] The lengths of all time series are unified by linear interpolation and equally spaced downsampling. .right and Normalize the standard deviation separately and splice them into a joint tensor according to the time dimension In order to improve the subsequent causal structure mapping effect, a structure-aware sorting mechanism is introduced to rearrange the feature dimensions according to the connectivity of the devices in the structure diagram. For example, the related variables of transformers with high node degrees will be placed at the front of the tensor to enhance the model's ability to learn the relationships between high-impact nodes. The tensor after splicing and sorting is recorded as:

[0045] ;

[0046] in, Indicates the Data collected from power monitoring equipment in the park dimensional electrical characteristic sequence; Indicates the data collected from environmental sensors such as temperature, humidity, noise, and infrared. dimensional environmental data series; is the sampling length; Indicates column-wise concatenation operation; Indicates reordering the concatenated variable dimensions according to the connectivity of each node in the structure graph.

[0047] Considering that different park structure diagrams may have import errors or inconsistent equipment modeling problems, in order to improve structural consistency, Then the edge confidence screening mechanism is introduced. Construct edge confidence , indicating the edge The frequency ratio in the historical park, if it is lower than the threshold (e.g. 0.5), it is considered to have no migration universality and is eliminated. The screening formula is as follows:

[0048] ;

[0049] in, For the edge exist The proportion of occurrence in is the co-occurrence threshold (it is recommended to be set to ); The new graph structure after confidence filtering only retains stable edges to improve the reliability of subsequent structure migration.

[0050] Ultimately, each park will receive Yes. Among them, is the sorted electrical and environmental joint feature tensor, The structure of the graph is to remove unstable edges. The variable order and graph structure have been standardized to facilitate subsequent causal modeling operations.

[0051] The output is two sets: (1) feature tensor set ; (2) Confidence graph structure set Both will serve as direct input to the causal structure graph learning step.

[0052] S2. Construct a cross-campus portable causal structure map based on the preprocessed data, divide the devices into several priority levels, calculate the dynamic granularity priority score for each node in the causal structure map, and perform the resampling operation on each node;

[0053] The goal of this step is to build a structured input data (feature tensor set) based on the structured input data constructed in the previous step S1. and structure diagram collection ), proposed a "dynamic granularity control and data sampling scheduling" mechanism for intelligent O&M of power distribution equipment. This mechanism not only addresses the information redundancy and resource waste issues faced by traditional unified sampling strategies in heterogeneous campus environments, but also further enhances the adaptability of the causal structure learning phase, ensuring sufficient temporal resolution, clear feature expression, and complete capture of abnormal behavior in key devices.

[0054] A prominent feature of smart campuses is the diverse composition of devices, with significant differences in state and behavior. Some edge devices remain in a low-activity state for extended periods, making them inefficient for real-time, high-frequency sampling. Therefore, sampling strategies must be structurally sensitive and dynamically adjustable. The innovation of this step lies in combining the node interaction strengths in the topology graph with the non-stationarity of the time series signal to jointly define a "sampling priority function." This function is then used to generate structured resampled tensor inputs, improving the efficiency and stability of overall structural modeling.

[0055] The input is the output of step 1 , indicating the A normalized tensor of parks, where are electrical dimensions (such as three-phase current, voltage, frequency, etc.), are environmental dimensions (such as infrared temperature, cabin humidity, noise intensity, etc.), is a unified number of time slices; This is the structural diagram corresponding to the park, where the node Numbered in sequence, strictly corresponding to the tensor column dimensions, that is, Variables represent nodes State. Graph structure After the co-occurrence edge confidence screening in step 1, it has good stability. Next, granular adaptive resampling will be performed based on this.

[0056] In this step, a "stratified sampling scoring function" is introduced for the nodes in the structure graph to classify devices into several priority levels. This function combines two pieces of information: structural centrality, which measures the node's criticality within the topology. For example, transformers and main switchgear often connect to multiple devices; and the intensity of the node's state change over time, which measures the dynamic complexity of its characteristics within the monitoring time window.

[0057] The dynamic granularity priority score of each node is defined as:

[0058] ;

[0059] in, For nodes Total sampling score of For its structure diagram centrality scores in (e.g. betweenness centrality or approximate eigencentrality, unit normalized); Representing a tensor Middle Dimension in time The value of is derived from the electrical or environmental sensor signal of the field device; , a nonlinear difference function designed to enhance change detection, whose cubic term is used to emphasize sudden changes; is the change incentive factor (usually set to ~ ); is the structural / dynamic equilibrium parameter (it is recommended to take The core of this formula's design is to capture sudden behaviors (such as short-term jump currents and noise spikes) through nonlinear enhancement terms rather than simple linear fluctuations, thereby increasing the sampling density of possible fault precursor behaviors.

[0060] After the sampling score is calculated, all nodes are Sort, divide into Layer (such as ), each layer corresponds to a different sampling window length (such as 1 minute, 3 minutes, 5 minutes). The sampling window uses the sliding window averaging method to compress high-frequency data. For example, for The nodes of the layer will have their time series sorted by length The sliding window calculates the local average to form a new sequence column vector. The window length is set by the developer or automatically configured through structural optimization search. The recommended solution is as follows: Layer 1 is the core node, ; Layer 2 is the secondary key node, ; Layer 3 is the edge node, .

[0061] After performing the resampling operation on each node, the entire tensor Refactored to , where the original column order is retained and the length is unified as , all missing points are filled by linear interpolation.

[0062] The final structure is shown below:

[0063] ;

[0064] in, Represents the process of adaptively setting the resampling window according to the sampling score level and updating the tensor; output tensor , maintain structural consistency and meet the input requirements of subsequent structure learning algorithms.

[0065] The output includes:

[0066] (1) Resampling tensor set ,Each tensor retains high-frequency information in key device dimensions and is compressed on edge nodes;

[0067] (2) Structure diagram collection , remains unchanged and is used for subsequent causal graph modeling and structural migration.

[0068] The innovations of this step are mainly reflected in three aspects: first, the introduction of a composite node sampling score function that combines structural centrality and state mutation degree to adapt to the actual scenario of multiple types of equipment coexisting in the park; second, the design of a nonlinear difference function for mutation detection , which can improve the sampling resolution for equipment with potential fault characteristics (such as a sudden increase in current before tripping); thirdly, a structure-oriented Figure 1 The consistent granularity and hierarchical scheduling mechanism ensures the dimensional consistency of tensors and improves the efficiency and effectiveness of time series modeling through unified sliding window reconstruction and interpolation alignment.

[0069] S3. In multiple deployed historical smart parks, a set of causal structure subgraphs with stability and migration capabilities are extracted from the pre-processed multi-source data, which are used as the structure for constructing new park models;

[0070] The goal of this step is to use the device status feature tensor to analyze the device status feature tensors in multiple deployed historical smart parks. and structure diagram collection , extracting a set of causal structure subgraphs with stability and migration capabilities to serve as structural priors for constructing new campus models. This step goes beyond general causal structure learning and instead incorporates customized optimization based on the specific structural characteristics of smart campus power distribution systems (such as "short inter-device connection paths," "stable backbone paths," and "weak causal relationships between edge devices"). In particular, the "co-occurrence stability" and "physical dependency" of structural edges are simultaneously modeled to form a causal graph tailored for O&M migration.

[0071] The input of this step is the output of step 1, including the feature tensor set and confidence structure graph collection .in, Indicates the park The joint state sequence tensor of is the total feature dimension), each The nodes are sorted by their connectivity and normalized. It is a park Electrical structure diagram after edge confidence screening, including node device topology.

[0072] In traditional causal graph modeling, independent regression prediction is often used to construct a structural equation model. However, considering the special characteristics of the one-to-one correspondence between the structural graph and the state data in the smart park (for example, a node represents a transformer with a certain number, and its state variable can be derived from its connected equipment), a structural coupling regularization term is introduced in the regression model. , used to convert the connection relationship of devices into a priori bias of causal edges. The following optimization objectives are constructed:

[0073] ;

[0074] in For the The causal coefficient matrix of each park, For the variables in the park The first term is the reconstruction error, and the second term is Sparse regularization term, controlling the number of edges in the causal structure, the third term The structural graph prior coupling term innovatively introduced by the present invention is used to guide the causal graph construction to be biased towards the path with strong electrical connection logic. Specifically, , which means that a greater penalty is imposed on candidate causal edges that are not in the edge set of the structure graph. It can be set to a constant or related to the connection distance (for example, the longer the path hops, the greater the penalty).

[0075] The variables are described as follows: Indicates the From the variables to variables the strength of the causal relationship; yes Middle Columns, corresponding to the time series of a single variable; is the sparsity regularization term; is the structure graph guided regularization term, used to couple the topological prior; are two regularization hyperparameters.

[0076] Through the above regularized mapping process, a causal graph coupled with its structural graph G′ can be obtained in each park. ,This graph is not only a set of causal edges driven by the statistical correlation of variables, but also reflects the physical connection logic and operation dependency path of electrical equipment. Perform "edge co-occurrence rate" analysis, that is, count each causal edge The frequency of occurrence in all parks. Let the edge occurrence index be , define the common structure diagram for:

[0077] ;

[0078] The common structure diagram This is the structural prior for subsequent migration to the new park, which has two important characteristics: (1) its edge sets come from high-frequency causal paths in multiple parks and have strong stability; (2) its construction process integrates the physical prior of the structural graph G′ to ensure that the modeling direction is consistent with the actual dependency logic of the distribution equipment, thereby improving the interpretability and safety of the prediction after migration.

[0079] Output is a common causal structure map , where the connection direction, occurrence frequency and variable correspondence of the edges are fixed. The subsequent steps will be to find the structural matching area of ​​the subgraph in the new campus topology and migrate the modeling module accordingly.

[0080] This step is innovative in two key aspects: First, instead of using independent causal modeling for a single campus, it introduces a structural aggregation method called "cross-campus edge co-occurrence rate" to form the causal commonality required for structural migration. Second, for the first time, the structural graph is embedded as an explicit regularizer in the causal structure learning objective function, so that edge selection is driven not only by the statistical relationships between variables but also reflects the preferences of the electrical physical structure. This joint structure-state mapping approach has clear scenario value in the field of smart campus power distribution equipment operation and maintenance, and significantly improves the controllability and rationality of structural migration.

[0081] S4. In the actual deployment and operation of smart parks, based on the common causal structure graph and the online real-time data streams of multiple sensors, dynamic identification of multimodal abnormal data that may appear during operation is carried out, and it is encapsulated into structured event objects and their corresponding feature input tensors;

[0082] The goal of this step is to use the common causal structure map output in step S3 in the actual deployment and operation of the smart park. , combining online real-time data streams from multiple sensors , dynamically identify multimodal anomalies that may occur during operation and encapsulate them into structured event objects and its corresponding feature input tensor .

[0083] The input consists of two parts: the common causal structure map output in step S3 , and sliding window multimodal monitoring data collected in real time during deployment .in Indicates the total number of sensor channels (including electrical and environmental), Indicates the length of the current sliding window (e.g. 5 minutes). Data collection sources include:

[0084] Electrical quantities (current, voltage, power factor): collected by electrical transformers and acquisition modules on on-site circuit breakers, transformers and other equipment;

[0085] Thermal quantities (equipment housing temperature, connection point temperature rise): Infrared thermal imaging equipment is used to capture images and extract values ​​through average filtering;

[0086] Acoustic quantities (partial discharge noise, arc discharge sound): high-frequency sampling by a directional microphone array;

[0087] Environmental variables (temperature, humidity, and air dust): acquired through conventional meteorological sensors.

[0088] These data are uploaded in real time through the edge computing unit to form a sliding data window under a unified time axis. , and the system builds a set of steady-state reference benchmarks from the initialization phase , which represents the expected state of each variable in the non-abnormal stage. In order to judge the current degree of variable deviation, the standard deviation normalized deviation representation is introduced:

[0089] ;

[0090] in is a variable At the moment The observed value of and are the mean and standard deviation of the variable in the historical stable period. Normalization becomes a dimensionless value, allowing for uniform comparison of different modal variables. In particular, in smart parks, monitored variables come from different physical systems and have large unit spans (such as current [A] and temperature [°C]). Failure to normalize this value can lead to significant imbalances in structural scores, making this operation necessary at the system level.

[0091] In order to mine "whether structural anomalies are occurring" from multimodal offsets, the system designs a structure-guided linkage offset function :

[0092] ;

[0093] in Representation structure diagram A causal edge in is the frequency of the edge appearing in all parks (e.g., if it appears in 70% of parks, ), as the stability weight of the edge, This function represents the normalized standard deviation of variable k at time t, measuring the degree of deviation of its current observation from the historical steady-state baseline. This function assesses whether related variables in the structure graph path are synchronously shifted at the same time. If multiple critical paths are abnormal, it indicates a global coordination risk in the system.

[0094] The system sets a dynamically adjusted trigger threshold (Initially set based on sample standard deviation, and later dynamically adjusted based on false alarm rate feedback). When the system determines that the current state is a triggerable event state, it generates an abnormal event object , including the current timestamp, trigger path, abnormal level and other information, while recording the current data window For structured input , for subsequent model scoring and path tracing calls.

[0095] The variables are described as follows: For the current moment variable sampling values; The mean and standard deviation during its historical period constitute the steady-state benchmark; It is a Z-score type offset indicator, which eliminates the physical dimension difference; is the stability weight of the edge in the causal structure graph; is the structural consistency offset score; Is a structured exception event object; is the multimodal state input corresponding to the time window.

[0096] The output is two structures: one is the event object , including fields such as abnormal variable set, trigger time, structure path and abnormal level; the second is the state tensor , is the window data of all monitoring variables at the triggering moment, and the input dimension is , which will be used as input in the model scoring phase.

[0097] S5. When an abnormal event occurs, conduct abnormality tracking and locate the root cause device at the structural path level based on the common causal structure map and structured abnormality feature input;

[0098] The goal of this step is to Based on the new park causal structure map output in step 3, , combined with the structured anomaly feature input output from step 4 , conduct anomaly tracking and root cause device location at the structural path level. The core function of this step is to expand the anomaly detection results from "whether it is abnormal" to "why the anomaly occurs", and infer the potential propagation chain and the initial inducing node based on the causal path, thereby providing an accurate basis for subsequent response control.

[0099] Input includes causal structure graph , output from step 3, is a causal dependency structure represented by a set of directed edges, each edge Representing variables An abnormal offset may cause the variable Also includes abnormal events , output from step 4, including the time point when the exception occurred and the set of variables involved And the abnormal feature input tensor , which represents the sliding window data when the event is triggered.

[0100] In order to evaluate the credibility of the abnormal propagation path, this step introduces a path scoring mechanism with offset activation weights. Suppose any causal path , the comprehensive activation of the path is defined as follows:

[0101] ;

[0102] in, is an edge in the causal structure graph The weight is derived from the normalized value of the frequency of each edge in multiple parks in step 3 (e.g., the frequency is but ), represents the stability of the edge; Representing variables The offset vector in the event window (taken from the normalized vector in step 4) The corresponding variable dimension is ); It is the time series cosine similarity function, which is used to measure whether the offset patterns of the variables at both ends of the edge are similar. between.

[0103] This scoring formula measures whether all causal edges in a path are simultaneously in the "high confidence edge + high synchronization offset" activation state. The higher the value, the more likely the path is an abnormal propagation chain. The abnormal variable in is the end point of the path, and all reachable paths are traversed in reverse to filter out The largest paths as candidate propagation chains.

[0104] Then, define each node The root cause score of is the cumulative weight sum of its appearance in all candidate paths:

[0105] ;

[0106] This indicator is used to evaluate the node Whether it is a potential anomaly origin, the higher the score, the greater the possibility of it being the source of transmission. Finally, the system generates a ranked list of device root causes. , as the tracing result of this exception, and output the selected key path set.

[0107] The variables are described as follows:

[0108] : A directed path in the structure graph;

[0109] : comprehensive activation of the pathway;

[0110] : The stability weight of the edge in the causal structure graph comes from the edge frequency normalization in step 3;

[0111] :variable A sequence of normalized offsets in the event window, of length , from Middle OK;

[0112] : cosine similarity function;

[0113] :node The root cause score of is the sum of the scores of all high-activation paths containing this node;

[0114] : sorted list of root cause nodes;

[0115] : The multimodal input tensor at the abnormal triggering moment output in step 4, with a dimension of .

[0116] The output includes two structures: (1) a sorted list of root cause devices , a set of variable numbers arranged from high to low likelihood; (2) a set of key causal paths , which are the directed paths that are most likely to cause anomaly propagation in the current event. Both will be directly used for the next response and dispatch decision.

[0117] S6. After completing the anomaly tracing, generate intelligent operation and maintenance response suggestions that can be directly implemented;

[0118] The goal of this step is to: After completing the abnormality tracing, sort the root cause device list based on the output of step 5. and key causal path set , combined with the abnormal trigger input tensor output in step 4 , generate intelligent operation and maintenance response suggestions that can be directly implemented The input consists of three parts: (1) a sorted list of root causes , output from step five, each Number the device nodes that may cause the current exception and sort them from high to low according to the possibility of the exception propagation source; (2) The set of key causal paths , which contains a high-activation anomaly propagation chain starting from the root cause device, used to identify potential propagation risks in the system; (3) Abnormal input tensor , which is output from step 4, contains the observed values ​​of all monitoring variables that trigger abnormal events within the sliding window, with a dimension of , used to evaluate the current state.

[0119] The goal of the system is to automatically generate hierarchical response instructions based on this, which mainly include three levels: (1) Dispatching inspection (low response level, only generate work orders); (2) Manual confirmation (medium level, requires on-duty personnel to make offline judgments); (3) Emergency treatment (high level, immediately issue isolation, power off and other operation instructions). Each response object is a candidate root cause device , its response level is determined by factors such as its anomaly intensity, transmission risk location, and operating status. To this end, the following multidimensional scoring function is constructed:

[0120] ;

[0121] in, Score the root cause output from step 5 to reflect the device the credibility of being traced to the source of the anomaly; For devices The depth in the abnormal path increases with the risk closer to the starting point; Representation device The intensity of fluctuations within the anomaly triggering time window can be calculated using the standard deviation or Z-score variance; The weight coefficient is set by experience and is used to adjust the influence of the three factors. The default value is as follows: .

[0122] Afterwards, the system The value range automatically matches the response template:

[0123] like , generate inspection work orders;

[0124] like , generate manual confirmation tasks;

[0125] like , triggering emergency commands, including remote switching, control scheduling or priority alarm reporting. The above threshold It can be adjusted dynamically according to the park level and equipment importance.

[0126] : Response classification threshold, empirically set, such as .

[0127] The output is the final executable response instruction set , each instruction is in the form of a triple ,in is the device number to be responded to, is the recommended response level (e.g. 1 for inspection, 2 for manual review, 3 for emergency action), The output structure is directly fed into the campus energy dispatch system or EAM platform to implement work order distribution and remote control.

[0128] See also Figure 2 As shown, in a second aspect of the present invention, there is also provided a smart park power distribution equipment intelligent operation and maintenance system based on big data, comprising:

[0129] Data acquisition and processing module, used to collect multi-source data of smart parks and perform pre-processing;

[0130] A causal structure map construction module is used to construct a cross-campus portable causal structure map based on preprocessed data, divide the devices into several priority levels, calculate the dynamic granularity priority score of each node in the causal structure map, and perform the resampling operation on each node;

[0131] The new park structure migration module is used to extract causal structure subgraphs with stability and migration capabilities, which are used as the structure for constructing the new park model;

[0132] Anomaly recognition module, which is used to dynamically identify multimodal anomaly data that may occur during operation based on common causal structure graphs and online real-time data streams from multiple sensors;

[0133] The anomaly tracking and positioning module is used to track anomalies at the structural path level and locate the root cause device based on the common causal structure map and structured anomaly feature input when an abnormal event occurs;

[0134] The intelligent report generation module is used to generate intelligent operation and maintenance response suggestions that can be directly implemented.

[0135] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. The intelligent operation and maintenance method of power distribution equipment in smart parks based on big data is characterized by: The following steps are involved: S1. Collect and pre-process multi-source data from the smart park, including electrical operation data, environmental monitoring data, structural information data, and operation and maintenance logs; S2. Construct a cross-campus portable causal structure map based on the preprocessed data, divide the devices into several priority levels, calculate the dynamic granularity priority score for each node in the causal structure map, and perform the resampling operation on each node; S3. In multiple deployed historical smart parks, a set of causal structure subgraphs with stability and migration capabilities are extracted from the pre-processed multi-source data, which are used as the structure for constructing new park models; S4. In the actual deployment and operation of smart parks, based on the common causal structure graph and the online real-time data streams of multiple sensors, dynamic identification of multimodal abnormal data that may appear during operation is carried out, and it is encapsulated into structured event objects and their corresponding feature input tensors; S5. When an abnormal event occurs, conduct abnormality tracking and locate the root cause device at the structural path level based on the common causal structure map and structured abnormality feature input; S6. After completing the exception tracing, generate intelligent operation and maintenance response suggestions that can be directly implemented.

2. The intelligent operation and maintenance method for power distribution equipment in a smart park based on big data according to claim 1 is characterized in that: In S1, the electrical operation data includes the three-phase current, voltage and frequency of electrical equipment A, B and C collected by the telemetry unit; the environmental monitoring data includes the temperature and humidity in the equipment cabin, infrared imaging temperature and noise signal strength; the structural information data includes the primary wiring diagram parsed into graphic structure data, including the connection relationship between devices.

3. The intelligent operation and maintenance method for power distribution equipment in a smart park based on big data according to claim 1 is characterized in that: In S2, a hierarchical sampling score function is introduced for the nodes in the structure graph to divide the devices into several priority levels. The dynamic granularity priority score of each node is: ； in, For nodes Total sampling score of For its structure diagram The centrality score in ; Representing a tensor Middle Dimension in time The value of is derived from the electrical or environmental sensor signal of the field device; is a unified number of time slices; A nonlinear difference function designed to enhance change detection, whose cubic term is used to emphasize sudden changes; is the change incentive factor; is the structural / dynamic equilibrium parameter.

4. The intelligent operation and maintenance method for power distribution equipment in a smart park based on big data according to claim 3 is characterized in that: After the dynamic granularity priority score is calculated, all nodes are sorted according to the dynamic granularity priority score and divided into L layers, each layer corresponding to a different sampling window length. After performing the resampling operation on each node, the length is padded to obtain a resampled tensor set.

5. The intelligent operation and maintenance method for power distribution equipment in a smart park based on big data according to claim 1 is characterized in that: In S3, the connection direction, occurrence frequency and variable correspondence of the edges of the causal structure subgraph with stability and migration ability are fixed. The structural matching area with the causal structure subgraph is found in the new campus topology structure, and the migration model is based on it.

6. The intelligent operation and maintenance method for power distribution equipment in a smart park based on big data according to claim 1 is characterized in that: In S4, a structure-guided linkage offset function is designed, and its specific calculation formula is as follows: ； in, represents a causal edge in the structure graph, is the frequency of the edge appearing in all parks, represents the standard deviation normalized offset representation of the data, It represents the standard deviation normalized offset value of variable k at time t, which is used to measure the degree of deviation of its current observation value from the historical steady-state benchmark; By setting a dynamically adjusted trigger threshold, when When the value of is greater than the trigger threshold, it is determined that the current state is a triggerable event state, and an abnormal event object is generated. The abnormal event object contains information such as the current timestamp, trigger path, abnormal level, etc., and the current data window is recorded at the same time, and the current data window is defined as a structured input.

7. The intelligent operation and maintenance method for power distribution equipment in a smart park based on big data according to claim 1 is characterized in that: In S6, the intelligent operation and maintenance response suggestions include: Low-risk suggestion: Generate a dispatch patrol work order; Medium-risk advice: Call on-duty personnel for offline inspection; High-risk advice: Immediately issue isolation and power-off instructions, and arrange for personnel evacuation.

8. The intelligent operation and maintenance system of power distribution equipment in smart parks based on big data is characterized by: Includes the following connected in sequence: Data acquisition and processing module, used to collect multi-source data of smart parks and perform pre-processing; A causal structure map construction module is used to construct a cross-campus portable causal structure map based on preprocessed data, divide the devices into several priority levels, calculate the dynamic granularity priority score of each node in the causal structure map, and perform the resampling operation on each node; The new park structure migration module is used to extract causal structure subgraphs with stability and migration capabilities, which are used as the structure for constructing the new park model; Anomaly recognition module, which is used to dynamically identify multimodal anomaly data that may occur during operation based on common causal structure graphs and online real-time data streams from multiple sensors; The anomaly tracking and positioning module is used to track anomalies at the structural path level and locate the root cause device based on the common causal structure map and structured anomaly feature input when an abnormal event occurs; The intelligent report generation module is used to generate intelligent operation and maintenance response suggestions that can be directly implemented.

Images