Multi-device collaborative line loss anomaly collaborative detection method and system

Abstract

The present invention discloses a collaborative line loss anomaly detection method and system for multiple devices, belonging to the field of power technology. It includes collecting original time series data in multiple dimensions and performing preprocessing, generating a dynamic spatio-temporal graph based on the time series data to reflect the collaborative working characteristics of the devices, extracting the collaborative feature vectors of the devices and performing fault detection, while performing self-optimization and adjusting the system strategy, and finally outputting the results to the devices. This solution overcomes the problem that the existing methods cannot adapt to the dynamic changes of power equipment in real time, and also enhances the adaptability of the system to the complexity of the power network through the adaptive optimization of intelligent algorithms.

Description

Technical Field

[0001] The present invention belongs to the technical field of electric power, and particularly relates to a method and system for collaborative detection of abnormal line losses with multi-device collaboration. Background Art

[0002] In the power system, line losses (i.e., the loss of energy during power transmission and distribution) have always been an important issue that power companies are concerned about. With the continuous expansion and intelligent development of the power grid, the detection of abnormal line losses has become increasingly complex. Most traditional line loss detection methods rely on manual inspection, equipment status monitoring, and rule detection based on experience. These methods have many limitations. For example, they can only monitor the abnormal status of a single device and cannot obtain the collaborative working conditions between multiple devices in real time; they only rely on static rules or threshold judgments, resulting in a slow response to abnormal signals in complex systems; and it is difficult to cope with the increasingly complex interrelationships and real-time status changes between devices in the power system.

[0003] At present, some existing anomaly detection techniques based on data analysis, such as machine learning models (such as support vector machines, decision trees, etc.), have improved this problem to a certain extent and can identify abnormal patterns through historical data. However, when facing the complexity and real-time requirements of multi-device collaboration in the power system, these methods often cannot achieve accurate and effective detection. There are many types of devices in the power system, which are widely distributed, and there are strong correlations between devices. Simply relying on traditional device monitoring methods and static machine learning models cannot comprehensively capture the collaborative effects between devices, resulting in low accuracy of anomaly detection. In addition, the time-varying nature of device status and complex environmental factors make it difficult for existing methods to adapt to the dynamic changes of the power system. Therefore, the adaptability and robustness of the existing technology have significant deficiencies.

[0004] Therefore, we propose a method and system for collaborative detection of abnormal line losses with multi-device collaboration to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to solve the limitations existing in the prior art in the collaborative detection of abnormal line losses with multi-device collaboration, and to propose a method and system for collaborative detection of abnormal line losses with multi-device collaboration.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions:

[0007] A method for collaborative detection of abnormal line losses with multi-device collaboration includes:

[0008] S1: Collect original time series data in multiple dimensions, perform time synchronization processing on all original time series data, eliminate the outliers, and then perform per-feature normalization processing to obtain time series feature data;

[0009] S2: Based on the time series feature data obtained in step S1, establish a dynamic cooperation degree matrix to obtain the dynamic cooperation degree between devices. Based on the adjacency relationship and dynamic cooperation degree of the devices, establish a dynamic adjacency matrix, and generate a dynamic spatio-temporal graph to reflect the collaborative perception characteristics of the devices;

[0010] S3: Extract the collaborative perception features of the devices to obtain the collaborative feature vectors of the devices, and send the collaborative feature vectors of the devices into the device anomaly detection branch and the line loss collaborative anomaly detection branch respectively, and output the device anomaly detection result A and the line loss collaborative anomaly detection result S respectively;

[0011] S4: Generate a policy adjustment factor F based on the device anomaly detection result A, the line loss collaborative anomaly detection result S, the power grid load status, and the environmental risk index, and optimize the structure of the dynamic spatio-temporal graph in step 2 based on the policy adjustment factor F to obtain an optimized dynamic spatio-temporal graph. The system synchronously generates a "policy optimization instruction" to dynamically adjust the anomaly detection thresholds of each device in the next round of detection to achieve sensitivity self-adaptation;

[0012] S5: Based on the optimized dynamic spatio-temporal graph, the device anomaly detection result A, and the line loss collaborative anomaly detection result S, execute a multi-device collaborative response mechanism, and for the detected collaborative anomalies, combine the collaborative relationships to implement partitioned, hierarchical, and dynamic response actions.

[0013] Preferably, in step S1, the multiple dimensions include the current, voltage, active power, temperature, and vibration amplitude of the device.

[0014] Preferably, in the time synchronization process of step S1, for low-sampling devices, the collaborative trend difference compensation algorithm is used to jointly compensate for missing data through the local trends of adjacent devices with collaborative relationships.

[0015] Preferably, the formula for calculating the dynamic cooperation degree in step S2 is as follows:

[0016] ;

[0017] Among them, represents the characteristic cosine similarity between the and devices at time , is the time window, is the device and the device in the past cycle of the collaborative anomaly index of abnormal events, reflecting the collaborative occurrence frequency of the abnormal behaviors of the two devices, is the compensation factor. This item significantly enhances the system's modeling ability for the "synergistic anomaly risk area" and solves the problem that pure "similarity" modeling cannot capture hidden synergistic anomalies.

[0018] Preferably, the formula for generating the dynamic spatio-temporal graph in step S2 is as follows:

[0019] ;

[0020] where is the set of device nodes, is the dynamic adjacency matrix;

[0021] The dynamic adjacency matrix is constructed by fusing the adjacency relationship and dynamic cooperation degree of the devices, and the calculation formula is:

[0022] ;

[0023] where is the device and the device the edge weight in the physical topology, is the balance factor, is the device and the device the difference term of the temporal feature distribution of the feature matrices of the devices within the most recent time steps, is the regularization factor. This calculation formula introduces a difference penalty term. When the feature changes between devices are drastic but the abnormal cooperation degree is high, the system still retains the basic dependence on the physical topology, reducing the risk of false associations, and is particularly suitable for graph relationship control when power equipment is abnormal due to environmental influences.

[0024] Preferably, after entering the device anomaly detection branch in step S3, for each collaborative perception vector, the anomaly probability of a single device is output through a single-device detection head, and when it exceeds 50%, the device is determined to be abnormal.

[0025] Preferably, after entering the line loss collaborative anomaly detection branch in step S3, aggregate the collaborative feature vectors of all devices, obtain the global collaborative feature through collaborative weighting, input the global collaborative feature into the global detection head, output the line loss collaborative anomaly probability, and when it exceeds 50%, determine that the whole network has a collaborative anomaly.

[0026] Preferably, in step S4, based on the policy adjustment factor F, the structure of the dynamic spatio-temporal graph in step 2 is optimized to obtain an optimized dynamic spatio-temporal graph, specifically:

[0027] ;

[0028] where is the policy adjustment factor for device i, is the policy adjustment factor for device j, is the original edge weight of the dynamic spatio-temporal graph, is the adaptive adjustment coefficient. Comprehensively reflects the device The current "abnormal risk perception intensity". By dynamically adjusting Achieve "strengthening of collaborative relationships" in high-risk areas, enhance the attention to abnormal-prone areas in subsequent detections, and through the risk grading mechanism, dynamically adjust the detection sensitivity to achieve adaptive sensitivity management, and improve the detection accuracy and robustness in the dynamic power grid scenario.

[0029] And a line loss anomaly collaborative detection system for multi-device collaboration, including:

[0030] The data acquisition module is configured to collect multi-device multi-source data and preprocess the collected original data, including performing time synchronization processing on the original data, removing outliers, and performing per-feature normalization processing to obtain time-series feature data, and output the time-series feature data to the dynamic graph module and the detection module;

[0031] The dynamic graph module is configured to receive the time-series feature data from the data acquisition module, construct a dynamic collaboration degree matrix based on this data, calculate the dynamic collaboration degree, and fuse the adjacency relationship of the devices and the dynamic collaboration degree to establish a dynamic adjacency matrix, generate a dynamic spatio-temporal graph and output it to the detection module;

[0032] The detection module is configured to receive the dynamic spatio-temporal graph and the time-series feature data of the devices, perform feature propagation through the dynamic spatio-temporal graph, extract the collaborative feature vectors of each device, and send the collaborative feature vectors to the device anomaly detection branch and the line loss collaborative anomaly detection branch respectively, and output the device anomaly detection result A and the line loss collaborative anomaly detection result S respectively, and output them to the adaptive module and the collaboration module;

[0033] The adaptive module is configured to generate a policy adjustment factor based on the current device anomaly detection result A and the line loss collaborative anomaly detection result S, combined with the real-time power grid load status and the environmental risk index and use the policy adjustment factor to optimize the graph structure of the edge weights of the dynamic spatio-temporal graph, and synchronously detect the range where the policy adjustment factor is located to dynamically adjust the anomaly detection thresholds of each device in the next round of detection, and output them to the collaboration module;

[0034] Collaboration module: It is configured to execute a multi-device collaborative response mechanism based on the optimized dynamic spatiotemporal graph and the device's anomaly detection result A and line loss collaborative anomaly detection result S. It implements partitioned, hierarchical, and dynamic response actions for the detected collaborative anomalies in combination with the collaborative relationship.

[0035] In summary, the technical effects and advantages of the present invention are as follows: This solution can comprehensively monitor the operating status of each device in the power system by introducing a dynamic detection mechanism for multi-device collaboration, combined with real-time data analysis and adaptive learning algorithms, and can capture line loss anomalies caused by the collaborative work of devices in real time, and can autonomously optimize the anomaly detection strategy according to the actual status of the equipment and environmental changes, and improve real-time performance and accuracy. Through deep data fusion, the system can not only analyze the individual status of each device, but also comprehensively evaluate the synergy between devices, thereby identifying line loss anomalies caused by device interactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the method flow in the present invention;

[0037] Figure 2 It is a schematic diagram of the system structure in the present invention. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0039] like Figure 1 As shown, the multi-device collaborative line loss anomaly collaborative detection method includes:

[0040] S1: Collect original time series data of multiple dimensions, perform time synchronization processing on all original time series data, remove outliers, and then perform feature-by-feature normalization processing to obtain time series feature data;

[0041] S2: Based on the time series feature data obtained in step S1, a dynamic coordination degree matrix is ​​established to obtain the dynamic coordination degree between devices. Based on the adjacency relationship and dynamic coordination degree of the devices, a dynamic adjacency matrix is ​​established to generate a dynamic time-space graph to reflect the collaborative working characteristics of the devices.

[0042] S3: extract the collaborative perception features of the device to obtain the collaborative feature vector of the device, and send the collaborative feature vector of the device to the device anomaly detection branch and the line loss collaborative anomaly detection branch respectively, and output the device anomaly detection result A and the line loss collaborative anomaly detection result S respectively;

[0043] S4: Generate a policy adjustment factor F based on the abnormal detection result A of the device, the collaborative abnormal detection result S of line loss, the power grid load status, and the environmental risk index, and optimize the structure of the dynamic spatio-temporal graph in step 2 based on the policy adjustment factor F to obtain an optimized dynamic spatio-temporal graph. The system synchronously generates a "policy optimization instruction" to dynamically adjust the abnormal detection thresholds of each device in the next round of detection, realizing sensitivity self-adaptation;

[0044] S5: Based on the optimized dynamic spatio-temporal graph, the abnormal detection result A of the device, and the collaborative abnormal detection result S of line loss, execute a multi-device collaborative response mechanism, and for the detected collaborative abnormalities, combine the collaborative relationships to implement partitioned, hierarchical, and dynamic response actions.

[0045] Example 1

[0046] In step S1, the system first performs time synchronization processing, using the highest sampling frequency in the whole network (for example, 1 second) as the benchmark, and uniformly resamples all device data to a 1-second granularity. For low-sampling devices, a collaborative trend interpolation algorithm is used. By considering the local trends of neighbor devices (the set is denoted as ) that have physical or electrical collaborative relationships, the missing data is jointly interpolated. For example, when the data of device 3 is missing at seconds, the system refers to the historical mean of device 3 within and the means of neighbor devices 1 and 2 in the same time period 、 、 , and is complemented through the following collaborative weighting formula:

[0047] ;

[0048] Among them, is the interpolation value of device 3 at , is the harmonic coefficient, and the value range is , for example to strengthen its own data trend, is the set of collaborative neighbors, represents device at the local mean within window.

[0049] After interpolation, the system introduces a collaborative anomaly rejection mechanism. Based on the local mean of device and its neighbor and the standard deviation , define the rejection threshold , for example . If a certain data point , it is regarded as an outlier and removed. For example, when the current value of Device 1 is at , but the current range of adjacent devices is , the system determines that this point is an outlier across the collaborative interval and directly removes it.

[0050] Finally, the system performs per-feature normalization on the cleaned data , maps all feature values to the interval, eliminates the differences in feature scales of different devices, and forms structured time-series aligned data . Each is a matrix of ( is the unified step size, is the feature dimension), which serves as the input for the next step of "spatiotemporal dependence modeling".

[0051] In step S2, the system receives the output from the previous step, that is, the time-series aligned and normalized multi-dimensional feature matrix of devices in the power system, as the input for dynamic graph modeling. To achieve the goal of "abnormal line loss detection under multi-device collaboration", a dynamic spatiotemporal collaborative graph modeling mechanism (DSTGM mechanism) is proposed, which is specifically customized for the collaborative characteristics and actual scenarios of power system devices to address the deficiencies of existing technologies in dealing with dynamic collaborative relationships. The collaborative relationships between devices in the power system are not fixed. Affected by factors such as geographical location, device operating status, and environmental load fluctuations, the collaborative intensity of devices has significant spatiotemporal dynamics, and traditional static topology graphs or simple graphs based on physical connections cannot accurately reflect this change.

[0052] Therefore, in this step, a dynamically updatable graph structure is designed, where is the set of device nodes, is the dynamic adjacency matrix, and the specific generation process is as follows:

[0053] The system first uses the output from step 1 to construct a dynamic collaboration degree matrix based on the device collaboration relationship to measure the recent spatiotemporal dependence relationship between devices. The calculation of the collaboration degree not only considers the local feature similarity between devices but also introduces a collaboration compensation factor specifically designed for the power system to reflect the abnormal consistency contribution in physical or electrical relationships.

[0054] Define the following collaboration degree calculation formula:

[0055] ;

[0056] where Indicates and Equipment at time The feature cosine similarity at each moment, is the time window, For equipment With equipment in the past The coordinated abnormality index of abnormal events (such as abnormal line loss and frequent load fluctuations) within a cycle reflects the frequency of coordinated abnormal behaviors of the two devices. is the compensation factor (e.g. ), improve the system's sensitivity to "potential collaborative anomalies".

[0057] For example, when device 1 and device 2 are in the nearest If there are 6 coordinated anomalies in the abnormal events within the hour, then , if the cosine similarity is ,but Compensated to This item significantly enhances the system's modeling capabilities for "collaborative abnormality risk areas" and solves the problem that pure "similarity" modeling cannot capture implicit collaborative abnormality risks.

[0058] Next, the system integrates static physical adjacencies Dynamic synergy with the above , construct a dynamic adjacency matrix , the formula is:

[0059] ;

[0060] in, For equipment and equipment Edge weights in physical topology, is the balance factor (e.g. ), For equipment and equipment The feature matrix is ​​recently The difference term of the time series feature distribution within time steps (Frobenius norm), is the regularization factor (e.g. ), to prevent excessive expansion of edge weights caused by abnormal similarity. This formula introduces a difference penalty term. When the characteristics between devices change drastically but the abnormal coordination is high, the system still retains the basic dependence on the physical topology, reducing the risk of false associations. It is particularly suitable for graph relationship control when power equipment is affected by the environment (such as sudden weather, system-level fluctuations) and abnormalities occur.

[0061] Illustrated by an example, when device 1 (transformer) and device 3 (distribution line) are in under the physical relationship, under the influence of the abnormal collaboration index , the collaboration degree is increased to , but due to the large fluctuations in their characteristics within time steps ( ), the final dynamic edge weight , and the edge weight is moderately suppressed to balance the collaborative risk perception and feature differences.

[0062] Through the construction of the above dynamic adjacency matrix , the system generates a dynamic spatio-temporal graph , which can accurately reflect the collaborative working characteristics, abnormal propagation risks, and potential weak or strong dependence relationships of power system devices during the current period, significantly improving the reliability of subsequent anomaly detection. Based on the high-quality data after cleaning, this step establishes a dynamic graph structure with the characteristics of the power industry, namely "collaborative anomaly perception + temporal difference penalty", specifically to solve the industry pain points of "difficult to detect weak collaborative anomalies and unable to dynamically update the graph structure" in the collaborative anomaly detection of multiple power devices, significantly improving the system's ability to detect hidden problems such as line loss anomalies.

[0063] Finally, is input into the subsequent detection model as a complete dynamic graph structure, providing graph structure support for realizing efficient and intelligent anomaly detection under multi-device collaboration.

[0064] In step S3, this step receives the dynamic graph structure and the time series feature matrix of devices output from step 2, and designs a power collaborative multi-task detection mechanism (ECMTD mechanism) with "collaborative perception ability" to realize the synchronous detection of "single-device anomalies" and "line loss anomalies under multi-device collaboration" in the power system.

[0065] The system first uses the graph structure for feature propagation to extract the collaborative perception features of each device, specifically:

[0066] ;

[0067] Among them, is the graph edge weight between device and its neighbor , is the trainable graph convolution parameter, is the activation function. This mechanism integrates the features of device and its neighbor devices to obtain , that is, the device 's collaborative eigenvector.

[0068] Subsequently, the system enters the dual-branch detection stage, and the of each device will be sent into two detection branches respectively:

[0069] 1. Device anomaly detection branch: For each , through the single-device detection head , output the anomaly prediction probability of a single device , that is:

[0070] ;

[0071] Among them, and are the parameters of the device anomaly detection branch, is the Sigmoid activation, is the th device's anomaly probability, taking values , then determines that the device is abnormal.

[0072] 2. Line loss collaborative anomaly detection branch: The system aggregates the collaborative features of all devices , and forms the global collaborative feature through the collaborative weighting mechanism:

[0073] ;

[0074] Input into the global detection head , and output the system-level line loss collaborative anomaly prediction value , that is:

[0075] ;

[0076] Among them, and are the parameters of the global detection branch, , as the prediction probability of whether there is "collaborative line loss anomaly" in the whole network during the current period, determines that there is a global collaborative anomaly.

[0077] During the training process of the system, the loss function consists of the following three parts:

[0078] 1. Single-device anomaly detection loss ;

[0079] 2. System collaborative anomaly detection loss ;

[0080] 3. Cooperative consistency regular , used to avoid "weak association false alarms" in the graph.

[0081] The complete loss expression is:

[0082] ;

[0083] This step, as the core detection module of the patent, directly realizes the goal of "line loss anomaly detection under multi-device cooperation", and the output and provide input data for the subsequent "dynamic optimization strategy" and "cooperative response mechanism", realizing intelligent cooperative anomaly management in the multi-device environment of the power system.

[0084] The technical solutions in the embodiments of the present application at least have the following technical effects or advantages: By introducing a dynamic detection mechanism of multi-device cooperation, combining real-time data analysis and adaptive learning algorithms, this solution can comprehensively monitor the operating states of various devices in the power system, and capture line loss anomalies caused by the cooperative work of devices in real time. Moreover, it can autonomously optimize the anomaly detection strategy according to the actual states of the devices and environmental changes, improving the real-time performance and accuracy. Through deep data fusion, the system can not only analyze the individual states of each device, but also comprehensively evaluate the cooperative effects between devices, thereby identifying line loss anomalies caused by device interactions.

[0085] Embodiment 2

[0086] The embodiments of the present application also provide a line loss anomaly cooperative detection system for multi-device cooperation, as Figure 2 shown, including:

[0087] A data acquisition module, configured to collect multi-source data of multiple devices and preprocess the collected original data, including performing time synchronization processing on the original data, removing outliers, and performing per-feature normalization processing to obtain time-series feature data, and output the time-series feature data to the dynamic graph module and the detection module;

[0088] A dynamic graph module, configured to receive the time-series feature data from the data acquisition module, construct a dynamic cooperation degree matrix based on this data, calculate the dynamic cooperation degree, and establish a dynamic adjacency matrix by fusing the adjacency relationship of the devices and the dynamic cooperation degree, generate a dynamic spatio-temporal graph and output it to the detection module;

[0089] The detection module is configured to receive the dynamic spatio-temporal graph and the time-series feature data of the devices, perform feature propagation through the dynamic spatio-temporal graph, extract the collaborative feature vectors of each device, and send the collaborative feature vectors to the device anomaly detection branch and the line loss collaborative anomaly detection branch respectively, output the device anomaly detection result A and the line loss collaborative anomaly detection result S respectively, and output them to the adaptive module and the collaborative module;

[0090] The adaptive module is configured to generate a policy adjustment factor based on the current anomaly detection result A and the line loss collaborative anomaly detection result S, combined with the real-time power grid load status and the environmental risk index , and use the policy adjustment factor to optimize the graph structure of the edge weights of the dynamic spatio-temporal graph, and synchronously detect the range where the policy adjustment factor is located to dynamically adjust the anomaly detection thresholds of each device in the next round of detection, and output it to the collaborative module;

[0091] The collaborative module: is configured to execute a multi-device collaborative response mechanism based on the optimized dynamic spatio-temporal graph, the device anomaly detection result A and the line loss collaborative anomaly detection result S, and implement partitioned, hierarchical, and dynamic response actions for the detected collaborative anomalies in combination with the collaborative relationships.

[0092] The adaptive module receives the (device anomaly detection result) and (system collaborative line loss anomaly detection result) output from the detection module, and designs a power dynamic collaborative optimization mechanism (EDCOM mechanism) to adaptively optimize the subsequent detection sensitivity and dynamic graph structure according to the detection results and the current state of the power grid, and improve the adaptability and robustness of the system in a changing power grid environment. The actual operating environment of the power system is highly dynamic. Affected by seasonal load fluctuations, device operating hours, sudden failures or weather, there are significant differences in the collaborative relationships and anomaly risks among devices at different stages. Therefore, the system designs an adaptive strategy optimization mechanism driven by "detection results - environmental characteristics", and simultaneously updates the "dynamic graph structure " and the "detection strategy parameters" in real time.

[0093] First of all, the system is based on the current detection results and , combined with the real-time power grid load status (such as the current load rate of the whole network) and environmental factors (such as temperature, humidity, wind speed, etc.), to generate a policy adjustment factor for dynamically adjusting the sensitivity of the device and the graph relationship. Defined as:

[0094] ;

[0095] Among them, is the policy coefficient (such as , etc.), is the environmental risk index of the device (dynamically calculated based on the device's surrounding environment). Comprehensively reflects the device current "abnormal risk perception intensity".

[0096] Subsequently, the system uses to optimize the edge weights in the dynamic graph for "graph structure optimization based on detection feedback", realizing a cross-step graph structure adaptive mechanism, specifically:

[0097] ;

[0098] Among them, is the original edge weight of the dynamic graph constructed by the dynamic graph module, is the adaptive adjustment coefficient (such as ). This mechanism dynamically increases the edge weights of "high-risk collaborative areas" according to the current detection output and environmental situation of the device, forming a "graph strengthening mechanism" for sensitive areas.

[0099] For example, when the between device 1 and device 2, and the of the two devices, , then , the system automatically strengthens the collaboration intensity, and the subsequent detection model will pay more attention to the collaborative anomalies in this area, improving the detection accuracy.

[0100] The system synchronously generates "policy optimization instructions" to dynamically adjust the anomaly detection thresholds of each device in the next round of detection , realizing sensitivity self-adaptation. The specific rules are:

[0101] If (high-risk threshold), reduce ;

[0102] If (low-risk threshold), appropriately increase ;

[0103] Maintain the current sensitivity in the middle interval.

[0104] For example, if the device 's and , then 's next-round anomaly judgment threshold will be adjusted from down to , to achieve "improved detection sensitivity of key equipment" and respond to periodic sudden risks in the power system.

[0105] In the adaptive module, through "cross-step graph structure optimization + detection strategy tuning", the detection results can be used to reversely adjust the graph model. The innovation of this mechanism is:

[0106] Combining multi-source environmental characteristics ( and ) and test results to construct a "multi-factor driven risk index ”;

[0107] Graph structure optimization: dynamic adjustment Achieve "synergistic relationship strengthening" in high-risk areas and increase the attention of subsequent testing to abnormal high-incidence areas;

[0108] Strategy tuning: Through the risk grading mechanism, the detection sensitivity is dynamically adjusted to achieve adaptive sensitivity management, thereby improving the detection accuracy and robustness in dynamic power grid scenarios.

[0109] Finally, the system outputs:

[0110] Optimized graph structure , as the input of the next round of collaborative detection model;

[0111] The optimized detection strategy for each device (such as the adjusted ).

[0112] The collaborative module receives the optimized dynamic graph structure output by the adaptive module , detection strategy parameters , and the detection results of the detection module and , execute the multi-device coordinated response mechanism, and implement partitioned, hierarchical, and dynamic response actions for the detected coordinated anomalies in combination with the current coordinated relationship to achieve accurate abnormal handling of the power system.

[0113] First, the system is based on , the collaborative subgraph division is carried out in the power grid equipment network according to the following:

[0114] ;

[0115] in , indicating that devices with a synergy strength higher than 0.6 will be included in the same sub-graph , forming a "cooperative response unit" under a physical or electrical coupling relationship.

[0116] In this embodiment, the substation 4 downstream power distribution equipment , due to tight electrical connections and high load coordination, a sub - graph is formed ;

[0117] The electrical consumption terminal at the far end and the adjacent terminal , with a coordination degree , are divided into the sub - graph .

[0118] Next, based on the detection results, the system implements a response action of zoning + hierarchy.

[0119] Response at the single - device level: If , the device directly enters the local response and performs actions such as power cut, load reduction, and tie - line isolation; if , the device maintains normal operation and only records the alarm.

[0120] If the power distribution equipment has a detection result , the system immediately triggers a power - down action to reduce the load of this node by 30%.

[0121] Co - operative response within the sub - graph: If , or , then multiple devices within the sub - graph perform a "higher - order joint response", such as branch isolation and sub - graph - level load shunting.

[0122] If in the sub - graph ( ), and are both abnormal, and , the system automatically executes 's "regional grouping isolation" strategy, cuts off the relevant branches, and at the same time shunts the load of to the neighbor with the highest coordination degree .

[0123] Linkage sequence and action allocation: Based on the edge weights inside, give priority to controlling devices with high coordination intensity to avoid the spread of abnormalities. For device pairs with edge weights , perform a "linked load adjustment", such as the load of the faulty device , which is preferentially shunted to .

[0124] If and have an edge weight , when After a triggered power outage, the system preferentially transfers its remaining load to , ensuring the stable local power supply of the system. Cross-subgraph response coordination:

[0125] If there is a high coordination relationship between subgraphs and (for example ), but the current value has not reached the abnormal threshold, then the response strategy is mitigated or delayed to avoid false triggering.

[0126] For example (downstream of the substation) and (remote power consumption terminal) have a coordination relationship ;

[0127] When a power outage response is triggered, if the abnormal conditions are not met, the system only sends a "warning signal" to and plans to reconfirm whether to perform actual actions in the next cycle to prevent chain misoperations.

[0128] System final output:

[0129] Subgraph-level response action table , each contains "linkage sequence + response type + shunt scheme"; roughly as follows:

[0130] Subgraph :

[0131] - Linkage sequence: v1 ➡ v3 ➡ v2

[0132] - Response type:

[0133] v1: Power outage + load transfer to v3

[0134] v3: Load takeover + tie line disconnection

[0135] v2: Load reduction by 20%

[0136] - Shunt scheme:

[0137] After v1 power outage, 80% of the load is transferred to v3, and 20% of the load is shunted to v2.

[0138] Meaning:

[0139] Clarify the execution order between the response actions of multiple devices within the subgraph (who goes first and who goes second to avoid chaotic responses);

[0140] Specify the specific action type for each device (power off / load reduction / load takeover, etc.);

[0141] Describe the shunt strategy and specify how the load is transferred between devices with high coordination intensity.

[0142] The specific response instructions for each device, including the local action type (power off / downregulation / shunt) and action priority; roughly as follows:

[0143] Device v1:

[0144] - Action type: Power off

[0145] - Priority: High (to be executed first)

[0146] Device v3:

[0147] - Action type: Take over the load + disconnect the tie line

[0148] - Priority: Medium

[0149] Device v2:

[0150] - Action type: Reduce the load by 20%

[0151] - Priority: Low

[0152] Note:

[0153] The local response actions of a single device, with priorities, control the execution order;

[0154] Can be directly parsed by the scheduling or automation system to execute precise device-level actions.

[0155] Cross-subgraph slow-release control strategy to reduce the risk of "coordination link misresponse";

[0156] Subgraph and :

[0157] - Coordination intensity: 0.65

[0158] - Slow-release strategy:

[0159] When performing hard isolation, delay the response action by 5 minutes;

[0160] Suspend power off and only enter the "load downregulation + status observation" mode;

[0161] If the abnormal risk does not increase within 5 minutes, cancel the response action.

[0162] Description:

[0163] For the high coordination relationship between different sub - figures, the system automatically generates control strategies of "response delay" or "observation slow - release" to avoid cross - region mis - triggering.

[0164] The technical solutions in the embodiments of the present application at least have the following technical effects or advantages: It overcomes the problem that the existing methods cannot adapt to the dynamic changes of power equipment in real time, and also enhances the adaptability of the system to the complexity of the power network through the adaptive optimization of intelligent algorithms.

[0165] The working principle is as follows: By introducing a dynamic detection mechanism for multi - device coordination, combined with real - time data analysis and adaptive learning algorithms, it can comprehensively monitor the operating states of various devices in the power system, and capture in real time the abnormal line losses caused by the collaborative work of devices. According to the actual states of the devices and environmental changes, it autonomously optimizes the abnormal detection strategy to improve real - time performance and accuracy. Through deep data fusion, the system can not only analyze the individual states of each device, but also comprehensively evaluate the collaborative effects between devices, so as to identify the abnormal line losses caused by the interaction of devices.

[0166] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solutions and inventive concepts of the present invention, makes equivalent substitutions or changes, and all should be covered by the protection scope of the present invention.

Claims

1. A multi-device collaborative line loss anomaly detection method, characterized in that: include: S1: Collect original time series data of multiple dimensions, perform time synchronization processing on all original time series data, remove outliers, and then perform feature-by-feature normalization processing to obtain time series feature data; S2: Based on the time series feature data obtained in step S1, a dynamic synergy matrix is ​​established to obtain the dynamic synergy between devices. A dynamic adjacency matrix is ​​established based on the adjacency relationship and dynamic synergy of the devices, and a dynamic spatiotemporal graph is generated to reflect the synergy perception characteristics of the devices. The dynamic synergy calculation formula is as follows: ； in, Indicates and Equipment at time The feature cosine similarity at each moment, is the time window, For equipment With equipment in the past The coordinated anomaly index of abnormal events within a cycle reflects the frequency of coordinated abnormal behaviors of two devices. is the compensation factor; The formula for generating a dynamic space-time graph is as follows: ； in, is a collection of device nodes. is the dynamic adjacency matrix; The dynamic adjacency matrix is ​​constructed by integrating the adjacency relationship and dynamic coordination degree of the devices, and the calculation formula is: ； in, For equipment and equipment Edge weights in physical topology, is the balance factor, For equipment and equipment The feature matrix is ​​recently The time series feature distribution difference term within time steps, is the regularization factor; S3: extract the collaborative perception features of the device to obtain the collaborative feature vector of the device, and send the collaborative feature vector of the device to the device anomaly detection branch and the line loss collaborative anomaly detection branch respectively, and output the device anomaly detection result A and the line loss collaborative anomaly detection result S respectively; S4: Generate a strategy adjustment factor F based on the abnormal detection result A of the device, the line loss collaborative abnormal detection result S, the power grid load status and the environmental risk index, and perform structural optimization on the dynamic space-time graph in step 2 based on the strategy adjustment factor F to obtain an optimized dynamic space-time graph. The system synchronously generates a strategy optimization instruction to dynamically adjust the abnormal detection threshold of each device in the next round of detection to achieve sensitivity adaptation; wherein, the structural optimization of the dynamic space-time graph in step 2 based on the strategy adjustment factor F to obtain an optimized dynamic space-time graph is specifically: ； in, is the policy adjustment factor of device i, is the policy adjustment factor of device j, is the original edge weight of the dynamic space-time graph, is the adaptive adjustment coefficient, is the edge weight of the optimized dynamic spatiotemporal graph; S5: Based on the optimized dynamic space-time diagram and the device anomaly detection result A and line loss collaborative anomaly detection result S, a multi-device collaborative response mechanism is executed. For the detected collaborative anomalies, combined with the collaborative relationship, partitioned, hierarchical, and dynamic response actions are implemented.

2. The multi-device collaborative line loss anomaly detection method according to claim 1, characterized in that: The multiple dimensions in step S1 include current, voltage, active power, temperature and vibration amplitude of the equipment.

3. The multi-device collaborative line loss anomaly detection method according to claim 1, characterized in that: In the time synchronization process in step S1, for low-sampling devices, a collaborative trend difference compensation algorithm is used to jointly compensate for missing data through local trends of neighboring devices in a collaborative relationship.

4. The multi-device collaborative line loss anomaly detection method according to claim 1, characterized in that: After entering the device anomaly detection branch in step S3, the anomaly probability of a single device is output through a single device detection head for each collaborative perception vector, and the device is determined to be abnormal when it exceeds 50%.

5. The multi-device collaborative line loss anomaly detection method according to claim 1, characterized in that: After entering the line loss collaborative anomaly detection branch in step S3, the collaborative feature vectors of all devices are aggregated, and the global collaborative feature is obtained by collaborative weighting. The global collaborative feature is input into the global detection head, and the line loss collaborative anomaly probability is output. When it exceeds 50%, the whole network collaborative anomaly is determined.

6. A multi-device collaborative line loss anomaly collaborative detection system, characterized in that: include: The data acquisition module is configured to collect multi-device multi-source data and pre-process the collected raw data, including time synchronization processing of the raw data, removal of outliers, and feature-by-feature normalization processing to obtain time series feature data, and output the time series feature data to the dynamic graph module and the detection module; The dynamic graph module is configured to receive the time series feature data from the data acquisition module, and construct a dynamic coordination degree matrix based on the data, calculate the dynamic coordination degree, and integrate the adjacency relationship and dynamic coordination degree of the device to establish a dynamic adjacency matrix, generate a dynamic space-time graph and output it to the detection module; wherein the dynamic coordination degree calculation formula is as follows: ； in, Indicates and Equipment at time The feature cosine similarity at each moment, is the time window, For equipment With equipment in the past The coordinated anomaly index of abnormal events within a cycle reflects the frequency of coordinated abnormal behaviors of two devices. is the compensation factor; The formula for generating a dynamic space-time graph is as follows: ； in, is a collection of device nodes. is the dynamic adjacency matrix; The dynamic adjacency matrix is ​​constructed by integrating the adjacency relationship and dynamic coordination degree of the devices, and the calculation formula is: ； in, For equipment and equipment Edge weights in physical topology, is the balance factor, For equipment and equipment The feature matrix is ​​recently The time series feature distribution difference term within time steps, is the regularization factor; The detection module is configured to receive the dynamic spatiotemporal graph and the time series feature data of the device, perform feature propagation through the dynamic spatiotemporal graph, extract the collaborative feature vector of each device, and send the collaborative feature vector to the device anomaly detection branch and the line loss collaborative anomaly detection branch respectively, output the device anomaly detection result A and the line loss collaborative anomaly detection result S respectively, and output them to the adaptive module and the collaborative module; The adaptive module is set to generate a strategy adjustment factor based on the current anomaly detection result A and the line loss collaborative anomaly detection result S, combined with the real-time grid load status and the environmental risk index , and adjust the factor Optimize the edge weights of the dynamic spatiotemporal graph and simultaneously detect the strategy adjustment factors The abnormal detection threshold of each device in the next round of detection is dynamically adjusted according to the range of the device, and output to the collaborative module; wherein, the dynamic spatiotemporal graph in step 2 is structurally optimized based on the strategy adjustment factor F to obtain the optimized dynamic spatiotemporal graph, which is specifically: ； in, is the policy adjustment factor of device i, is the policy adjustment factor of device j, is the original edge weight of the dynamic space-time graph, is the adaptive adjustment coefficient, is the edge weight of the optimized dynamic spatiotemporal graph; Collaboration module: It is configured to execute a multi-device collaborative response mechanism based on the optimized dynamic spatiotemporal graph and the device's anomaly detection result A and line loss collaborative anomaly detection result S. It implements partitioned, hierarchical, and dynamic response actions for the detected collaborative anomalies in combination with the collaborative relationship.

Images