A low-voltage transformer area power failure anomaly detection method and device, electronic equipment and storage medium

Abstract

The application discloses a low-voltage transformer area power failure anomaly detection method and device, electronic equipment and storage medium, and particularly relates to the technical field of low-voltage power distribution network operation monitoring, the method extracts the time sequence characteristic vector corresponding to each to-be-monitored object in the low-voltage transformer area; the time sequence characteristic vector is spliced with a preset power failure judgment threshold vector to obtain the current decision state vector of each to-be-monitored object; each current decision state vector is input into a pre-trained strategy network to make the strategy network output a power failure judgment threshold adjustment vector of each to-be-monitored object, and each power failure judgment threshold vector is updated according to the power failure judgment threshold adjustment vector; and a power failure event of the low-voltage transformer area is determined according to the updated power failure judgment threshold vector and the time sequence characteristic vector of each to-be-monitored object, so that the problem that a fixed threshold rule cannot adaptively match actual dynamic operation characteristics in the existing low-voltage power distribution network power failure detection, and power failure misjudgment and omission are prone to occur is solved.

Description

Technical Field

[0001] This invention relates to the field of low-voltage distribution network operation monitoring technology, and in particular to a method, device, electronic equipment, and storage medium for detecting power outage anomalies in low-voltage distribution areas. Background Technology

[0002] Currently, proactive power outage detection in low-voltage distribution networks is a key aspect of improving power supply reliability and user service quality. The common approach is to determine outages based on multi-source data fusion and fixed threshold rules. This involves the system collecting electrical quantities such as voltage and current from the distribution transformer and branch sides in real time, comparing their instantaneous or short-term average values ​​with preset fixed thresholds, and outputting an outage alarm based on the comparison result.

[0003] However, the operation of equipment in low-voltage distribution networks is significantly time-varying and scenario-dependent. The load level of equipment fluctuates wildly and is difficult to predict depending on the type of electricity consumption. This makes it difficult for fixed threshold rules to adaptively match the actual dynamic operating characteristics, which can easily lead to parameter mismatch. When the threshold is too strict, normal operation under low load is easily misjudged as a power outage; when the threshold is too lenient, a real power outage under low load is easily missed. Summary of the Invention

[0004] This invention proposes a method, device, electronic equipment, and storage medium for detecting power outage anomalies in low-voltage distribution areas. It addresses the problem that fixed threshold rules in existing low-voltage distribution network power outage detection cannot adaptively match actual dynamic operating characteristics, easily leading to false or missed power outage detections.

[0005] One embodiment of the present invention provides a method for detecting power outage anomalies in low-voltage distribution areas, comprising: extracting the original operating data corresponding to each monitored object in the low-voltage distribution area, and extracting the multidimensional temporal features of each monitored object within a continuous time window from the original operating data; The multidimensional temporal features are encoded to obtain the temporal representation vector of each monitored object within a continuous time window; The time-series representation vector is concatenated with the preset power outage judgment threshold vector to obtain the current decision state vector of each object to be monitored. Each current decision state vector is input into a pre-trained policy network to output a power outage judgment threshold adjustment vector for each monitored object. The power outage judgment threshold vector is then updated based on this adjustment vector. The power outage judgment threshold vector includes the phase current near-zero threshold, the effective lower limit of voltage, the current drop amplitude, and the number of objects experiencing simultaneous power outages on the same line. The policy network is trained using the decision state vector of each monitored object within a historical time window as its state space, the historical power outage threshold parameter adjustment vector as its action space, correct identification of power outage events as a positive reward, and delayed reporting, false reporting, and missed reporting as negative penalties. Based on the updated power outage judgment threshold vector and the time sequence representation vector of each monitored object, the power outage event of the low-voltage distribution area is determined.

[0006] Furthermore, the raw operational data corresponding to each monitored object within the low-voltage distribution area is extracted, including: Raw operational data from multiple data sources is obtained. Using the object identification information carried in each piece of raw operational data as an index, the raw operational data is attached to the topology node corresponding to the object identification information and its parent topology node by querying a pre-stored topology relationship mapping table, thereby obtaining the raw operational data corresponding to each object to be monitored; wherein, the object to be monitored is the entity on the topology node.

[0007] Furthermore, the policy network is trained in the following manner: Obtain the decision state vector of each object to be monitored within the historical time window; An offline sample playback environment is constructed using the decision state vector of each monitored object within a historical time window as the state space and the historical power outage judgment threshold adjustment vector as the action space. In the offline sample playback environment, the decision state vector within the historical time window is input into the policy network, and the historical power outage judgment threshold adjustment vector is output. The historical power outage judgment threshold vector is updated according to the historical power outage judgment threshold adjustment vector, and historical predicted power outage events are generated based on the updated historical power outage judgment threshold vector and the preset power outage type judgment conditions. The reward value is calculated based on historical predicted power outage events, and the parameters of the policy network are updated according to the reward value in order to train the policy network.

[0008] Furthermore, based on the updated power outage judgment threshold vector and the time-series representation vector of each monitored object, the power outage events in the low-voltage distribution area are determined, including: The real-time values ​​in the time-series representation vector of each monitored object are compared with the corresponding parameters in the updated power outage judgment threshold vector to determine whether each real-time value meets the corresponding parameter conditions, and to obtain each Boolean judgment result. Based on the Boolean decision results, if the preset power outage type determination conditions are met, a corresponding candidate power outage event is generated; Based on the electrical topology and temporal overlap between the monitored objects, the candidate power outage events are merged to obtain the final power outage event.

[0009] Furthermore, after generating candidate power outage events for each monitored object, the process also includes: The original operating data and business data of each monitored object are input into a pre-trained multi-task learning model so that the multi-task learning model outputs the power outage prediction probability and power consumption type of each monitored object. When the power consumption type is the first power consumption type and the predicted power outage probability is less than a preset probability threshold, the candidate power outage event is marked as pending verification. Wherein, when the object to be monitored is the first power consumption type, the object to be monitored must meet at least one of the following conditions: the load fluctuation rate of the object to be monitored exceeds a set load fluctuation rate threshold, the start-stop frequency of the object to be monitored exceeds a set start-stop frequency threshold within a unit time, or the duration of continuous low current operation of the object to be monitored exceeds a set operation duration threshold. When the electricity consumption type identification result is not the first electricity consumption type and the power outage prediction probability is greater than or equal to the preset probability threshold, the candidate power outage event is confirmed. Otherwise, maintain normal operation and do not output candidate power outage events.

[0010] Furthermore, the process of merging the candidate power outage events based on the electrical topology and temporal overlap relationships between the monitored objects to obtain the final power outage event includes: Candidate power outage events that satisfy both topological correlation and temporal overlap are grouped into the same event cluster. Specifically, if the monitored objects corresponding to the candidate power outage events have the same line ID, the same transformer ID, or have a hierarchical electrical affiliation, the candidate power outage events are determined to satisfy a topological relationship. If the outage time periods of the candidate power outage events overlap or the difference in the start time of the outage is less than a preset time threshold, the candidate power outage events are determined to satisfy a temporal overlap relationship. Within the same event cluster, the primary event of candidate power outage events is determined according to the preset power outage type priority. The power outage type of the main event is used as the power outage type of the final power outage event, and the power outage impact range of all candidate power outage events within the same event cluster is merged to obtain the power outage impact range of the merged final power outage event.

[0011] Furthermore, the multi-task learning model includes a shared backbone network and three parallel task branches, namely a power outage detection branch, an electricity consumption type identification branch, and a contrast projection branch. The multi-task learning model is trained in the following way: The multi-task learning model takes the operational data samples and business data samples of each monitored object as input, and outputs the power outage probability prediction result and the power consumption type identification result as output for iterative training. During each training session, the running data samples and business data samples of each object to be monitored are input into the shared backbone network of the multi-task learning model, and the shared temporal representation vector of the running data samples and business data samples of each object to be monitored is extracted. The shared temporal representation vector is input into the power outage detection branch, the electricity consumption type identification branch and the comparison projection branch respectively, and the corresponding power outage probability prediction result, electricity consumption type identification result and low-dimensional comparison vector are output. Calculate the power outage detection loss based on the power outage probability prediction result and the preset power outage detection label; calculate the power consumption type classification loss based on the power consumption type identification result and the preset power consumption type label; calculate the consistency constraint loss based on the consistency between the power outage probability prediction result and the power consumption type identification result; calculate the time series comparison constraint loss based on the low-dimensional comparison vector; The total loss is calculated based on the power outage detection loss, the power consumption type classification loss, the time series comparison constraint loss, and the consistency constraint loss. If the total loss does not meet the preset convergence condition, the network parameters of the multi-task learning model are adjusted.

[0012] An embodiment of the present invention provides a low-voltage distribution area power outage anomaly detection device, comprising: The time series characterization extraction module is used to extract the original operating data of each monitored object in the low-voltage area, and extract the multi-dimensional time series characteristics of each monitored object in a continuous time window from the original operating data. The temporal encoding module is used to encode the multidimensional temporal features to obtain the temporal representation vector of each object to be monitored within a continuous time window; The decision state construction module is used to concatenate the time-series representation vector with the preset power outage judgment threshold vector to obtain the current decision state vector of each object to be monitored. The strategy optimization module is used to input each current decision state vector into a pre-trained strategy network, so that the strategy network outputs an adjustment vector for the power outage judgment threshold of each monitored object, and updates each power outage judgment threshold vector according to the adjustment vector. The strategy network is trained with the decision state vector of each monitored object in the historical time window as the state space, the historical power outage threshold parameter adjustment vector as the action space, the correct identification of power outage events as the positive reward, and the delayed reporting, false reporting and missed reporting as the negative penalty. The power outage determination module is used to determine the power outage event in the low-voltage distribution area based on the updated power outage judgment threshold vector and the time sequence representation vector of each monitored object.

[0013] An embodiment of the present invention provides an electronic device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor; wherein, when the processor executes the computer program, it implements the low-voltage distribution area power outage anomaly detection method as described in any one of the present invention.

[0014] One embodiment of the present invention provides a storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the storage medium is located to perform the low-voltage distribution area power outage anomaly detection method as described in any one of the present invention.

[0015] The embodiments of the present invention have the following beneficial effects: This invention provides a method, device, electronic device, and storage medium for detecting power outage anomalies in low-voltage distribution areas. By quantifying the power outage judgment threshold into a dynamically updatable parameter vector, a pre-trained strategy network outputs parameter adjustment amounts in real time based on the temporal representation vectors of each monitored object to update the power outage judgment threshold. This allows the updated judgment parameters to adaptively and dynamically update with changes in operating status, avoiding misjudgments and missed judgments caused by threshold settings that are too strict or too lenient. Furthermore, this invention extracts multi-dimensional temporal features of each monitored object within a continuous time window and encodes them to form a temporal representation vector reflecting the continuous time period. Power outage events are judged based on the temporal representation within the continuous time period, reducing misjudgments caused by special power usage scenarios, decreasing false alarms caused by special power usage scenarios, and improving the accuracy of power outage anomaly detection. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a low-voltage distribution area power outage anomaly detection method provided in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of a low-voltage distribution area power outage anomaly detection device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0020] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] like Figure 1 As shown, one embodiment of the present invention provides a method for detecting power outage anomalies in low-voltage distribution areas, which may include at least: S1: Extract the raw operational data corresponding to each monitored object within the low-voltage distribution area, and extract the multi-dimensional time-series characteristics of each monitored object within a continuous time window from the raw operational data. S1 specifically includes the following processes: S11: Acquire raw operating data from low-voltage monitoring terminals, metering acquisition systems, dispatching systems, customer service systems, and equipment ledgers in the low-voltage distribution area. This raw operating data includes instantaneous three-phase voltage values, instantaneous three-phase current values, voltage quality statistics, metering alarm codes, and power outage / restoration markers.

[0022] S12: Using the object identification information carried in each piece of the original operating data as an index, the original operating data is attached to the topology node corresponding to the object identification information and its superior topology node by querying the pre-stored topology relationship mapping table, so that each object to be monitored obtains the original operating data with which it has an electrical relationship.

[0023] The preset topology refers to the hierarchical power supply connection from the distribution area (transformer) through low-voltage branches and meter boxes to the user in the low-voltage distribution network, including the equipment ID, superior node ID, line ID, and distribution area ID fields of each node. The objects to be monitored are entities on the electrical topology nodes, specifically including four categories: distribution areas, low-voltage branches, meter boxes, and users.

[0024] S13: Establish a continuous time window with a preset time granularity, map the original data after mounting to the same time base, preprocess the original running data, fill missing values ​​with linear interpolation, remove and correct abnormal jump values ​​with the 3σ criterion, and generate three types of data quality labels: missing measurement mark, breakpoint mark, and zero current mark.

[0025] S14: For each monitored object, extract multi-dimensional time-series features from the preprocessed data within the current time window to form a multi-dimensional time-series feature set, specifically including: voltage drop amplitude, current drop ratio, low voltage duration, three-phase voltage imbalance, historical same-day load difference, synchronization anomaly of distribution transformers associated with the same line, alarm density of nodes on the same line, voltage quality code, current quality code, missing measurement marker, breakpoint marker, zero current marker, and reference current before the sudden drop. Among them, the voltage drop amplitude is the ratio of the lowest phase voltage within the time window to the average phase voltage of the 10 minutes before the drop; the current drop ratio is the ratio of the average phase current within the time window to the reference current before the sudden drop; the three-phase voltage imbalance is calculated according to the formula "(maximum phase voltage - minimum phase voltage) / maximum phase voltage × 100%"; the synchronization anomaly of distribution transformers associated with the same line is the proportion of the number of distribution transformers with voltage anomalies within the current time window to the total number of distribution transformers on the same 10kV line.

[0026] S2: Encode the multidimensional temporal features to obtain the temporal representation vector of the object to be monitored within the time window.

[0027] In one embodiment of the present invention, a coding structure based on one-dimensional temporal convolutional residual blocks and bidirectional gated recurrent units (Bi-GRU) can be used to encode multi-dimensional temporal features. The specific structure of the coding network is as follows: The first layer is a one-dimensional convolutional layer with a kernel size of 3, an output channel number of 32, a stride of 1, and a ReLU activation function to extract local temporal abrupt changes such as voltage drops and current interruptions. The second layer is a residual block, which contains two one-dimensional convolutional layers with the same structure. It alleviates the gradient vanishing problem through skip connections and enhances the feature transmission of long-term operation modes such as continuous low pressure and slow recovery. The third layer is a bidirectional GRU layer with a hidden layer dimension of 32, which is used to capture the sequential dependencies of time series data and output sequence features that incorporate contextual information. The fourth layer is a global average pooling layer, which compresses the variable-length sequence features into a fixed-length 64-dimensional vector, which is the temporal representation vector corresponding to the current time window.

[0028] In large-scale application scenarios with long sequences and multiple nodes, the encoding network can also be replaced by a Transformer encoding structure based on a self-attention mechanism.

[0029] S3: Concatenate the time sequence representation vector and the preset power outage judgment threshold vector to obtain the current decision state vector of the current detection object.

[0030] S31: Configure a set of dynamically updatable power outage judgment threshold vectors for each monitored object within the current time window. All elements are threshold parameters that can be adjusted through the strategy network, and the initial values ​​are set according to power industry operation specifications and on-site operation and maintenance experience. The power outage judgment threshold vectors specifically include phase current near-zero threshold, historical load current boundary, effective lower limit of voltage, upper limit of voltage anomaly, duration of voltage loss, current drop amplitude and reference current before the drop, power restoration current stabilization time, number of objects with synchronous power outage on the same line, time difference of events on the same line, metering alarm matching time, customer service work order matching time, and low load current boundary.

[0031] S32: Concatenate the time-series representation vector and the preset power outage judgment threshold vector according to the following formula to obtain the current decision state vector of the current detection object: ; in, Let this be the current decision state vector. For time series representation vectors, This is the threshold vector for determining power outages.

[0032] S4: Input each current decision state vector into the pre-trained policy network so that the policy network outputs the power outage judgment threshold adjustment vector for each monitored object. Update each power outage judgment threshold vector according to the power outage judgment threshold adjustment vector. The policy network is trained with the decision state vector of each monitored object in the historical time window as the state space, the historical power outage threshold parameter adjustment vector as the action space, the correct identification of power outage events as the positive reward, and the delayed reporting, false reporting and missed reporting as the negative penalty.

[0033] Specifically, an incremental update method is used to update each power outage judgment threshold vector, and the update formula is as follows: ; in, This is the updated power outage judgment threshold vector; This is the threshold vector for determining the current power outage; Adjust the vector for the power outage detection threshold; To update the step size coefficient, the value range is [0.01, 0.1], which is used to control the smoothness of the threshold adjustment and avoid drastic fluctuations in the threshold that could cause judgment oscillations.

[0034] In one embodiment of the present invention, the policy network is trained based on an offline policy optimization framework. When an offline deep Q-network (DQN) is used, the power outage judgment threshold adjustment vector is discretized into a set of actions such as adjustment up, adjustment down, and maintenance. Then, the decision state vector of each monitored object within the historical time window is used as the state space, the historical power outage threshold parameter adjustment vector is used as the action space, the correct identification of power outage events is used as a positive reward, and delayed reporting, false alarms, and missed reporting are used as negative penalties. The specific training process is as follows: Obtain the decision state vector of each monitored object within the historical time window, the power outage judgment result after the power outage judgment threshold is updated, and the actual power outage event label.

[0035] An offline sample playback environment is constructed by using the decision state vector of each monitored object within a historical time window as the state space and the historical power outage judgment threshold adjustment vector as the action space.

[0036] In the offline sample playback environment, the decision state vector within the historical time window is input into the policy network, and the historical power outage judgment threshold adjustment vector is output. The historical power outage judgment threshold vector is updated according to the historical power outage judgment threshold adjustment vector, and historical predicted power outage events are generated based on the updated historical power outage judgment threshold vector and the preset power outage type judgment conditions. The reward value is calculated based on historical predicted power outage events, and the parameters of the policy network are updated according to the reward value in order to train the policy network.

[0037] In one embodiment of the present invention, the expression for the reward function of the policy network is: ; in, To optimize the strategy objectives; For expectation operators; is the discount factor; r is the time step index; This is the single-step reward value obtained at time t+r (i.e., the r-th step in the future).

[0038] ; in, This is the single-step reward value. , , and These are weighting coefficients, initially set to 1.0, 1.2, 1.5, and 0.8 respectively; To correctly identify the reward, the calculation formula is as follows: ; The penalty for false alarms is calculated using the following formula: ; The penalty for underreporting is calculated using the following formula: ; The penalty for delayed reporting is calculated using the following formula: ; To predict the type of power outage; This represents the actual type of power outage; For indicator functions; The time matching degree represents the degree of matching between the predicted power outage start / restore time and the actual time, and its value is in the range of [0,1]. The smaller the time error, the closer this value is to 1. The range matching degree represents the degree of overlap between the predicted power outage impact range and the actual impact range, and its value is in the range of [0,1]. The higher the degree of overlap, the closer this value is to 1. To predict power outage conditions; This represents a real power outage. For the reporting time; This is the actual start time of the power outage; For normalized reference time; Delayed penalty for normalization.

[0039] S5: Based on the updated power outage judgment threshold vector and the time-series representation vector of each monitored object, determine the power outage event in the low-voltage distribution area, including the following process: S51: Compare the real-time values ​​in the time-series representation vector of each monitored object with the corresponding parameters in the updated power outage judgment threshold vector to determine whether each real-time value meets the corresponding parameter conditions, and obtain each Boolean judgment result. Output "True" when the real-time value meets the corresponding parameter conditions. Specifically, this includes: If the phase current is less than or equal to the near-zero phase current threshold, the current is determined to be near zero, and True is output. If the current drop percentage is less than or equal to the historical load current boundary, the current is considered to have dropped significantly, and True is output. If the phase voltage is less than or equal to the effective lower limit of voltage, the voltage is determined to be abnormally low, and True is output. If the phase voltage is greater than or equal to the upper limit of abnormal voltage, the voltage is determined to be abnormally high, and True is output. If the duration of low voltage is greater than or equal to the voltage loss duration threshold, the low voltage is determined to be effective and True is output. If the current drop percentage is less than or equal to the current drop threshold, a current drop is determined, and True is output. If the current stabilization time after power restoration is greater than or equal to the power restoration current stabilization time threshold, the power restoration state is determined to be stable, and True is output. If the number of abnormal nodes on the same line is greater than or equal to the threshold for the number of objects that are simultaneously out of power on the same line, it is determined that there is a regional anomaly and True is output. If the time difference between adjacent abnormal events is less than or equal to the time difference threshold of events on the same line, the events are considered to be time-related, and True is output. If the time difference between a metering alarm and an electrical abnormality is less than or equal to the metering alarm matching time threshold, an alarm match is determined, and True is output. If the time difference between the customer service work order and the electrical fault is less than or equal to the customer service work order matching time threshold, the work order is considered to match, and True is output. If the reference current before the sudden drop is less than or equal to the low load current boundary, it is determined that the system is in a low load operating state, and True is output.

[0040] S52: Based on the Boolean judgment results, if the preset power outage type judgment condition is met, generate the corresponding candidate power outage event. The preset power outage type judgment condition is any one of the following: Medium-voltage line power outage: Simultaneously meets the requirements of regional anomaly, abnormally low voltage, significant current drop, and metering alarm matching; Low-voltage branch power outage: Simultaneously meets the following conditions: abnormally low voltage, near-zero current, continuous low voltage, and multiple abnormal nodes in the same branch. Low-voltage main switch tripping: simultaneously satisfying current drop, current near zero, and continuous low-voltage effectiveness; Voltage anomaly: Simultaneously meeting the following conditions: voltage abnormally low or high, low voltage continuously effective, and current not significantly decreased. Phase loss: Simultaneously meets the conditions of excessive three-phase voltage imbalance and near-zero single-phase current; Single transformer outage: Simultaneously meets the conditions of near-zero current on the user side, abnormally low voltage, and no abnormalities at the upstream node.

[0041] When any of the above-mentioned criteria are met, a candidate power outage event of the corresponding type is generated, and the event start time, end time, node to which it belongs, and power outage type are recorded.

[0042] S53: Based on the electrical topology and time overlap relationships between the monitored objects, the candidate power outage events are merged to obtain the final power outage event.

[0043] In one embodiment of the present invention, S52, after generating candidate power outage events for each object to be monitored, further includes: S521: Input the original operational data and business data corresponding to each monitored object into a pre-trained multi-task learning model, so that the multi-task learning model outputs the power outage prediction probability and power consumption type corresponding to each monitored object. The business characteristic data includes at least one of the following: planned power outage marker, dispatch order time difference, number of customer service work orders within a preset time period, metered power outage alarm marker, number of households in the transformer area, transformer capacity, phase configuration, line affiliation, and load fingerprint cluster.

[0044] S522: Perform secondary verification according to the following rules: When the electricity consumption type is the first electricity consumption type and the predicted power outage probability is less than a preset probability threshold, the candidate power outage event is marked as pending verification. Specifically, when the monitored object is the first electricity consumption type, the monitored object must satisfy at least one of the following conditions: the load fluctuation rate of the monitored object exceeds a set load fluctuation rate threshold; the start-stop frequency of the monitored object within a unit time exceeds a set start-stop frequency threshold; or the duration of continuous low-current operation of the monitored object exceeds a set operation duration threshold. The load fluctuation rate is calculated by acquiring load data for each monitored object within a preset historical time window.

[0045] When the power consumption type is not the first power consumption type and the power outage prediction probability is greater than or equal to the preset probability threshold, the candidate power outage event is confirmed. In other cases, the system remains in normal operation and no candidate power outage events are output.

[0046] In one embodiment of the present invention, step S53, merging the candidate power outage events based on the electrical topology and time overlap relationships between the monitored objects to obtain the final power outage event, includes: Candidate power outage events that satisfy both topological correlation and temporal overlap are grouped into the same event cluster. Specifically, if the monitored objects corresponding to the candidate power outage events have the same line ID, the same transformer ID, or have a hierarchical electrical affiliation, the candidate power outage events are determined to satisfy a topological relationship. If the outage time periods of the candidate power outage events overlap or the difference in the start time of the outage is less than a preset time threshold, the candidate power outage events are determined to satisfy a temporal overlap relationship. Within the same event cluster, the main event of the candidate power outage events within the event cluster is determined according to the preset power outage type priority; the power outage type priority from high to low is as follows: medium voltage line power outage, low voltage branch power outage, low voltage main switch tripping, voltage abnormality, phase loss, and single distribution transformer power outage.

[0047] The power outage type of the main event is used as the power outage type of the final power outage event, and the power outage impact range of all candidate power outage events within the same event cluster is merged to obtain the power outage impact range of the merged final power outage event.

[0048] In one embodiment of the present invention, the multi-task learning model includes a shared backbone network and three parallel task branches, namely a power outage detection branch, an electricity consumption type identification branch, and a comparison projection branch. During the inference phase, only the power outage detection branch and the electricity consumption type identification branch are activated; the comparison projection branch only participates in the model training process.

[0049] The specific structure of the multi-task learning model is as follows: Shared backbone network: The input end of the shared backbone network is used to receive the runtime sequence data and service characteristics of the objects to be monitored; the raw runtime data includes electrical quantity data such as instantaneous values ​​of three-phase voltage, instantaneous values ​​of three-phase current, voltage quality code, current quality code, missing measurement mark, breakpoint mark and zero current mark; the service characteristics include planned power outage mark, dispatch order time difference, number of customer service work orders within a preset time period, metering power outage alarm mark, number of households in the transformer area, transformer capacity, phase configuration, line affiliation and load fingerprint cluster label.

[0050] The shared backbone network first normalizes the input numerical features and embeds discrete features (such as planned power outage markers and transformer substation types) into an encoding layer. The two are then concatenated and input into the temporal coding layer. The temporal coding layer can employ an encoding structure based on one-dimensional temporal convolutional residual blocks and gated recurrent units, or it can employ an encoding structure based on a self-attention mechanism. In one embodiment of this invention, the shared backbone network adopts a convolutional plus gated recurrent unit structure consistent with the temporal coding module, specifically including: a first layer is a one-dimensional convolutional layer with a kernel size of 3 and 64 output channels, used to extract local temporal abrupt change features; a second layer is a residual block containing two identical one-dimensional convolutional layers, which enhance feature transfer in long-term operating modes through skip connections; a third layer is a bidirectional gated recurrent unit layer with a hidden layer dimension of 64, used to capture the sequential dependencies of temporal data; and a fourth layer is a global average pooling layer, which compresses variable-length sequence features into a fixed-length 128-dimensional vector, i.e., a shared temporal representation.

[0051] The output of the shared backbone network is connected to the input of the power outage detection branch, the input of the power consumption type identification branch, and the input of the comparison projection branch, respectively, and the shared timing representation is simultaneously output to the three parallel task branches.

[0052] The power outage detection branch receives the shared time series representation output from the shared backbone network at its input. Its structure consists of two fully connected layers and a sigmoid or softmax activation layer. The power outage detection branch is used to output the predicted power outage probability (within 0 to 1) for each monitored object based on the input shared time series representation, and simultaneously outputs the regression prediction values ​​for the power outage start time and recovery time.

[0053] Electricity Consumption Type Identification Branch: The input of the electricity consumption type identification branch receives the shared timing representation output by the shared backbone network. Its structure consists of two fully connected layers and a Softmax activation layer. The electricity consumption type identification branch is used to output the electricity consumption type identification result of each monitored object based on the input shared timing representation. The electricity consumption type identification result includes a first electricity consumption type and non-first electricity consumption types.

[0054] The first type of electricity consumption refers to electricity consumption scenarios where load fluctuation characteristics are easily confused with power outage events. The criteria for determining this type are as follows: Load data for each monitored object is obtained within a preset historical time window; at least one of the following quantitative indicators is calculated based on the load data: load fluctuation rate, start-stop frequency per unit time, and percentage of continuous low-current operation time. If the load fluctuation rate exceeds a set load fluctuation rate threshold, the start-stop frequency per unit time exceeds a set start-stop frequency threshold, or the percentage of continuous low-current operation time exceeds a set operation time threshold, the monitored object is determined to belong to the first type of electricity consumption; otherwise, it is determined to be a non-first type of electricity consumption. The first type of electricity consumption includes, but is not limited to, low-load users, charging pile users, cold chain storage users, and intermittent production users.

[0055] The contrastive projection branch receives the shared temporal representation output from the shared backbone network at its input, and its structure consists of two fully connected layers. During the training of the multi-task learning model, the contrastive projection branch maps the input shared temporal representation into a low-dimensional contrastive vector (e.g., 32-dimensional), and imposes constraints on the low-dimensional contrastive vector using a preset contrastive learning loss function. Specifically, positive and negative sample pairs are constructed according to a preset sample pairing strategy, and the contrastive learning loss function is calculated based on the similarity between the positive and negative sample pairs.

[0056] The sample pairing strategy is as follows: samples with the same power outage type label, samples with the same electricity consumption type label, and samples of the same monitored object within adjacent time windows are constructed as positive sample pairs; samples with different power outage type labels and different electricity consumption type labels are constructed as negative sample pairs. The contrastive learning loss function is used to shorten the distance between the low-dimensional contrast vectors of the two samples in the positive sample pair and widen the distance between the low-dimensional contrast vectors of the two samples in the negative sample pair during training. It should be noted that the contrastive projection branch only participates in the calculation during the model training phase and is bypassed during the model inference phase, not participating in any output.

[0057] The multi-task learning model is trained in the following way: Historical operational data and business data samples are acquired for each monitored object. The business data samples include outage type labels, outage start time labels, outage recovery time labels, and electricity consumption type labels. Outage type labels include categories such as normal, low-voltage main switch trip, low-voltage branch switch trip, single distribution transformer outage, medium-voltage line outage, voltage anomaly, and phase loss; electricity consumption type labels include primary electricity consumption type and non-primary electricity consumption type markings.

[0058] The multi-task learning model takes the operational data samples and business data samples of each monitored object as input, outputs the power outage probability prediction result and the power consumption type identification result as output, and uses the power outage type label, power outage start time label, power outage recovery time label and power consumption type label as supervision signals for iterative training.

[0059] In each training iteration, the running data samples and business data samples of each object to be monitored are input into the shared backbone network of the multi-task learning model, and the shared temporal representation vector of the running data samples and business data samples of each object to be monitored is extracted.

[0060] The shared temporal representation vector is input into the power outage detection branch, the electricity consumption type identification branch and the comparison projection branch respectively, and the corresponding power outage probability prediction result, electricity consumption type identification result and low-dimensional comparison vector are output respectively. Calculate the power outage detection loss based on the power outage probability prediction result and the preset power outage detection label; calculate the power consumption type classification loss based on the power consumption type identification result and the preset power consumption type label; calculate the consistency constraint loss based on the consistency between the power outage probability prediction result and the power consumption type identification result; calculate the time series comparison constraint loss based on the low-dimensional comparison vector; The total loss is calculated based on the power outage detection loss, the electricity consumption type classification loss, the time series comparison constraint loss, and the consistency constraint loss. If the total loss does not meet the preset convergence condition, the network parameters of the multi-task learning model are adjusted. The expression for the total loss function is: ; in, Total loss; For time-series comparison constraint loss, Losses are categorized by electricity usage type. To assess the damage caused by the power outage, To account for the consistency constraint loss between the power outage detection branch and the power consumption type identification branch. , and The initial weights for each task are set to 0.6, 1.0, and 0.4, respectively. The specific calculation methods for each loss term are as follows: The cross-entropy method is used to calculate the electricity consumption type classification loss. The calculation formula is as follows: ; in, Losses are categorized by electricity usage type; For the true label of the sample in the k-th electricity consumption type category, in When =1, it corresponds to the actual label of the first electricity consumption type; =0 corresponds to a true label that is not the primary electricity consumption type. This refers to the predicted probability of the corresponding type output by the model.

[0061] The power outage detection loss consists of the power outage type classification loss and the power outage / restoration boundary regression loss, and the calculation formula is as follows: ; in, For power outage detection losses; k is the power outage type index; This is the true label for the power outage type, representing the true label of the sample on the k-th power outage type, with a value of 0 or 1; The power outage probability prediction result output by the power outage detection branch is the predicted probability of the k-th power outage type output by the power outage detection branch of the model, with a value range of [0,1]. This represents the actual power outage time. The predicted start time of the power outage; This represents the actual power outage restoration time. The model predicts the power outage recovery time; T is the total duration of the time window. The initial value for the regression loss weight at the start time is set to 0.5; To recover the time regression loss weight, the initial value is set to 0.5.

[0062] The time-series contrastive constraint loss is calculated using supervised contrastive learning. For sample i, samples with the same outage type or load pattern and located in adjacent time windows or similar transformer areas are used as the positive sample set. Using actual power outages and the first type of electricity consumption as negative samples, and utilizing the comparison vector... Calculate similarity: ; in, To compare the constraint loss; N is the total number of samples; i is the index of the current sample; is the set of positive samples, which includes samples whose outage type corresponds to sample i and belongs to the same outage type or the same power consumption type; at the same time, samples in adjacent time windows or similar transformer areas can also be included in the set of positive samples; P is the positive sample index, which is a positive sample in the set of positive samples; Let i be the low-dimensional contrast vector corresponding to sample i; This is the low-dimensional contrast vector of the positive sample; Let i be a low-dimensional vector of all samples other than sample i. For similarity functions; Temperature is a parameter used to adjust the smoothness of the similarity distribution. The lower the temperature value, the sharper the distinction of the similarity distribution; the higher the temperature value, the smoother the distribution.

[0063] By using temporal comparison constraint loss, the distance between positive sample pairs in the low-dimensional representation space is shortened, while the distance between negative sample pairs is widened, thereby improving the temporal representation discrimination power of the shared backbone network.

[0064] The formula for calculating the consistency constraint loss is: ; in, Loss due to consistency constraints; Electricity consumption type category label; This is the first set of electricity consumption types. The outage prediction probability is output by the outage detection branch; Power outage label; The probability of the first electricity consumption type is output by the branch for electricity consumption type identification; To allow for confidence boundaries, a pre-set tolerance threshold is used to control the permissible degree of deviation between the two branches. Penalties are only triggered when the predicted probability exceeds this threshold, thus avoiding oversensitivity to small deviations. This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise; The consistency constraint loss is used to ensure that the power outage detection branch and the power consumption type identification branch form a mutual verification constraint relationship. When semantic contradictions occur between the outputs of the two branches (such as misjudging normal first power consumption type fluctuations as power outages, or misjudging real power outages as normal operation of special loads), the contradictory outputs are suppressed through a penalty mechanism, thereby improving the accuracy and robustness of the model in actual power outage detection.

[0065] After the multi-task branch model is trained, in the practical application (inference) stage, the original operating data and auxiliary feature data of each monitored object are input into the pre-trained multi-task learning model. At this time, the comparison projection branch does not participate in the calculation. Only the shared backbone network, the power outage detection branch and the electricity consumption type identification branch work together to output the power outage prediction probability and electricity consumption type identification result of each monitored object.

[0066] Please see Figure 2 Based on the same inventive concept as the above embodiments, one embodiment of the present invention provides a low-voltage distribution area power outage anomaly detection device, comprising at least: The time-series characterization extraction module is used to extract the raw operating data corresponding to each monitored object within the low-voltage distribution area, and to extract the multi-dimensional time-series characteristics of each monitored object within a continuous time window from the raw operating data. Specifically, the time-series characterization extraction module acquires raw operating data from low-voltage monitoring terminals, metering acquisition systems, dispatching systems, customer service systems, and equipment ledgers. Using the object identification information carried in each raw operating data as an index, the module queries a pre-stored topology mapping table to attach the raw operating data to the corresponding topology node and its superior node, so that each monitored object obtains all the raw operating data with which it has an electrical relationship. A continuous time window is established with a preset time granularity, and the multi-dimensional time-series characteristics of each monitored object within the current time window are extracted from the attached raw operating data. The multi-dimensional time-series characteristics include the low-voltage duration, the current drop ratio, the degree of synchronous change of adjacent distribution transformers, and the zero-current marker.

[0067] The timing encoding module is used to encode the multidimensional timing features to obtain the timing representation vector of each monitored object within a continuous time window. The timing encoding module can employ an encoding structure based on one-dimensional timing convolutional residual blocks and gated recurrent units, or it can employ an encoding structure based on a self-attention mechanism. Its function is to capture dynamic timing patterns such as voltage and current drops, persistent missing values, recovery slopes, phase imbalances, and changes in adjacent windows, compressing the multidimensional timing features into a fixed-length timing representation vector.

[0068] Further, based on the updated power outage judgment threshold vector and the time-series representation vector of each monitored object, the power outage events in the low-voltage distribution area are determined, including: The real-time values ​​in the time-series representation vector of each monitored object are compared with the corresponding parameters in the updated power outage judgment threshold vector to determine whether each real-time value meets the corresponding parameter conditions, and to obtain each Boolean judgment result. Based on the Boolean decision results, if the preset power outage type determination conditions are met, a corresponding candidate power outage event is generated; Based on the electrical topology and temporal overlap between the monitored objects, the candidate power outage events are merged to obtain the final power outage event.

[0069] After generating candidate power outage events for each monitored object, the process also includes: The original operating data and business data of each monitored object are input into a pre-trained multi-task learning model so that the multi-task learning model outputs the power outage prediction probability and power consumption type of each monitored object. When the power consumption type is the first power consumption type and the predicted power outage probability is less than a preset probability threshold, the candidate power outage event is marked as pending verification. Wherein, when the object to be monitored is the first power consumption type, the object to be monitored must meet at least one of the following conditions: the load fluctuation rate of the object to be monitored exceeds a set load fluctuation rate threshold, the start-stop frequency of the object to be monitored exceeds a set start-stop frequency threshold within a unit time, or the duration of continuous low current operation of the object to be monitored exceeds a set operation duration threshold. When the power consumption type is not the first power consumption type and the power outage prediction probability is greater than or equal to the preset probability threshold, the candidate power outage event is confirmed. Otherwise, maintain normal operation and do not output candidate power outage events.

[0070] The decision state construction module is used to concatenate the time-series representation vector with a preset power outage judgment threshold vector to obtain the current decision state vector of each monitored object. The power outage judgment threshold vector includes at least the following parameters: phase current near-zero threshold, effective lower limit of voltage, reference current threshold before sudden drop, and number of objects experiencing synchronous power outages on the same line.

[0071] The strategy optimization module is used to input each current decision state vector into a pre-trained strategy network, so that the strategy network outputs a power outage judgment threshold adjustment vector for each monitored object, and updates each power outage judgment threshold vector according to the power outage judgment threshold adjustment vector; wherein, the strategy network is trained with the decision state vector of each monitored object within a historical time window as the state space, the historical power outage threshold parameter adjustment vector as the action space, the correct identification of power outage events as a positive reward, and the delayed reporting, false reporting, and missed reporting as negative penalties.

[0072] The policy network is pre-trained in the following manner: An offline sample playback environment is constructed using the decision state vector of each monitored object within a historical time window as the state space and the historical power outage threshold parameter adjustment vector as the action space. In the offline sample playback environment, the decision state vector within the historical time window is input into the policy network, which outputs the historical power outage judgment threshold adjustment vector. The historical power outage judgment threshold vector is updated based on this adjustment vector, and historical predicted power outage events are generated based on the updated historical power outage judgment threshold vector and preset power outage type judgment conditions. A reward value is calculated based on the historical predicted power outage events, and the policy network parameters are updated based on the reward value. Correct identification of power outage events is used as a positive reward, while delayed reporting, false alarms, and missed reporting are used as negative penalties during training.

[0073] The power outage determination module is used to determine the power outage event in the low-voltage distribution area based on the updated power outage judgment threshold vector and the time sequence representation vector of each monitored object.

[0074] Specifically, the power outage determination module compares the real-time values ​​in the time-series representation vector of each monitored object with the corresponding parameters in the updated power outage judgment threshold vector to determine whether each real-time value meets the corresponding parameter conditions, and obtains each Boolean judgment result; if the preset power outage type judgment conditions are met according to each Boolean judgment result, a corresponding candidate power outage event is generated; then, according to the electrical topology relationship and time overlap relationship between each monitored object, the candidate power outage events are merged to obtain the final power outage event.

[0075] In one embodiment of the present invention, the device further includes a multi-task verification module, used to input the original operating data and auxiliary feature data of each monitored object into a pre-trained multi-task learning model after generating candidate power outage events, so that the multi-task learning model outputs the power outage prediction probability and power consumption type corresponding to each monitored object; when the power consumption type is a first power consumption type and the power outage prediction probability is less than a preset probability threshold, the candidate power outage event is marked as pending verification; when the power consumption type is not the first power consumption type and the power outage prediction probability is greater than or equal to the preset probability threshold, the candidate power outage event is confirmed; otherwise, the normal operation state is maintained and no candidate power outage event is output. The first power consumption type refers to a power consumption scenario where load fluctuation characteristics are easily confused with power outage events, and its determination criteria are: at least one of the following: the load fluctuation rate of the monitored object exceeds a set load fluctuation rate threshold, the start-stop frequency per unit time exceeds a set start-stop frequency threshold, or the proportion of continuous low current operation time exceeds a set operation time threshold.

[0076] This invention provides a method, device, electronic equipment, and storage medium for detecting power outage anomalies in low-voltage distribution areas. By constructing a dynamically updatable parameter vector for power outage judgment thresholds and utilizing a pre-trained strategy network to output threshold adjustment amounts in real time based on the decision state vectors of each monitored object, key judgment parameters such as near-zero phase current thresholds and effective lower voltage limits can be automatically updated according to changes in operating status. This transforms the judgment criteria in fixed rules from static preset values ​​to dynamic adaptive values, enabling the same set of power outage event judgment rules to automatically match appropriate judgment boundaries under different operating scenarios, avoiding misjudgments and missed judgments caused by threshold settings that are too strict or too lenient. By judging power outage events through temporal representation over a continuous time period, misjudgments caused by special power usage scenarios are reduced, false alarms caused by special power usage scenarios are decreased, and the accuracy of power outage anomaly detection is improved. This invention, after generating candidate power outage events for each monitored object, introduces a multi-task learning model for secondary verification. It utilizes the semantic consistency between the electricity consumption type identification result and the power outage prediction probability to review or filter candidate events. The verified candidate events are then merged based on topological and temporal overlap relationships. This effectively solves the problems of splitting a single power outage into multiple duplicate alarms, difficulty in aligning power outage and restoration boundaries, and high false alarm rates under special load scenarios. This further improves the overall accuracy and robustness of power outage anomaly detection.

[0077] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0078] Those skilled in the art will clearly understand that, for convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0079] One embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the low-voltage distribution area power outage anomaly detection method of any one of the present invention.

[0080] Electronic devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. Electronic devices may include, but are not limited to, processors and memory.

[0081] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device through various interfaces and lines.

[0082] Memory can be used to store computer programs. The processor implements various functions of electronic devices by running or executing computer programs stored in memory and accessing data stored in memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc.; the data storage area can store data created based on the use of the mobile phone, etc. In addition, memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD cards), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0083] One embodiment of the present invention provides a storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the storage medium is located to execute the low-voltage distribution area power outage anomaly detection method according to any one of the present invention.

[0084] The storage medium is a computer-readable storage medium, and the computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0085] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for detecting power outage anomalies in low-voltage distribution areas, characterized in that, include: Extract the original operating data of each monitored object in the low-voltage area, and extract the multi-dimensional time series characteristics of each monitored object in a continuous time window from the original operating data; The multidimensional temporal features are encoded to obtain the temporal representation vector of each monitored object within a continuous time window; The time-series representation vector is concatenated with the preset power outage judgment threshold vector to obtain the current decision state vector of each object to be monitored. Each current decision state vector is input into a pre-trained policy network to output a power outage judgment threshold adjustment vector for each monitored object. The power outage judgment threshold vector is then updated based on this adjustment vector. The power outage judgment threshold vector includes the phase current near-zero threshold, the effective lower limit of voltage, the current drop amplitude, and the number of objects experiencing simultaneous power outages on the same line. The policy network is trained using the decision state vector of each monitored object within a historical time window as its state space, the historical power outage threshold parameter adjustment vector as its action space, correct identification of power outage events as a positive reward, and delayed reporting, false reporting, and missed reporting as negative penalties. Based on the updated power outage judgment threshold vector and the time sequence representation vector of each monitored object, the power outage event of the low-voltage distribution area is determined.

2. The low-voltage distribution area power outage anomaly detection method as described in claim 1, characterized in that, Extract the raw operational data corresponding to each monitored object within the low-voltage distribution area, including: Raw operational data from multiple data sources is obtained. Using the object identification information carried in each piece of raw operational data as an index, the raw operational data is attached to the topology node corresponding to the object identification information and its parent topology node by querying a pre-stored topology relationship mapping table, thereby obtaining the raw operational data corresponding to each object to be monitored; wherein, the object to be monitored is the entity on the topology node.

3. The method for detecting power outage anomalies in low-voltage distribution areas as described in claim 1, characterized in that, The policy network is trained in the following way: Obtain the decision state vector of each object to be monitored within the historical time window; An offline sample playback environment is constructed using the decision state vector of each monitored object within a historical time window as the state space and the historical power outage judgment threshold adjustment vector as the action space. In the offline sample playback environment, the decision state vector within the historical time window is input into the policy network, and the historical power outage judgment threshold adjustment vector is output. The historical power outage judgment threshold vector is updated according to the historical power outage judgment threshold adjustment vector, and historical predicted power outage events are generated based on the updated historical power outage judgment threshold vector and the preset power outage type judgment conditions. The reward value is calculated based on historical predicted power outage events, and the parameters of the policy network are updated according to the reward value in order to train the policy network.

4. The method for detecting power outage anomalies in low-voltage distribution areas as described in claim 1, characterized in that, Based on the updated power outage judgment threshold vector and the time-series representation vector of each monitored object, the power outage events in the low-voltage distribution area are determined, including: The real-time values ​​in the time-series representation vector of each monitored object are compared with the corresponding parameters in the updated power outage judgment threshold vector to determine whether each real-time value meets the corresponding parameter conditions, and to obtain each Boolean judgment result. Based on the Boolean decision results, if the preset power outage type determination conditions are met, a corresponding candidate power outage event is generated; Based on the electrical topology and temporal overlap between the monitored objects, the candidate power outage events are merged to obtain the final power outage event.

5. The low-voltage distribution area power outage anomaly detection method as described in claim 4, characterized in that, After generating candidate power outage events for each monitored object, the process also includes: The original operating data and business data of each monitored object are input into a pre-trained multi-task learning model so that the multi-task learning model outputs the power outage prediction probability and power consumption type of each monitored object. When the power consumption type is the first power consumption type and the predicted power outage probability is less than a preset probability threshold, the candidate power outage event is marked as pending verification. Wherein, when the object to be monitored is the first power consumption type, the object to be monitored must meet at least one of the following conditions: the load fluctuation rate of the object to be monitored exceeds a set load fluctuation rate threshold, the start-stop frequency of the object to be monitored exceeds a set start-stop frequency threshold within a unit time, or the duration of continuous low current operation of the object to be monitored exceeds a set operation duration threshold. When the power consumption type is not the first power consumption type and the power outage prediction probability is greater than or equal to the preset probability threshold, the candidate power outage event is confirmed. Otherwise, maintain normal operation and do not output candidate power outage events.

6. The low-voltage distribution area power outage anomaly detection method as described in claim 4, characterized in that, The process of merging the candidate power outage events based on the electrical topology and temporal overlap relationships between the monitored objects to obtain the final power outage event includes: Candidate power outage events that satisfy both topological correlation and temporal overlap are grouped into the same event cluster. Specifically, if the monitored objects corresponding to the candidate power outage events have the same line ID, the same transformer ID, or have a hierarchical electrical affiliation, the candidate power outage events are determined to satisfy a topological relationship. If the outage time periods of the candidate power outage events overlap or the difference in the start time of the outage is less than a preset time threshold, the candidate power outage events are determined to satisfy a temporal overlap relationship. Within the same event cluster, the primary event of candidate power outage events is determined according to the preset power outage type priority. The power outage type of the main event is used as the power outage type of the final power outage event, and the power outage impact range of all candidate power outage events within the same event cluster is merged to obtain the power outage impact range of the merged final power outage event.

7. The method for detecting power outage anomalies in low-voltage distribution areas as described in claim 5, characterized in that, The multi-task learning model includes a shared backbone network and three parallel task branches, namely a power outage detection branch, an electricity consumption type identification branch, and a contrast projection branch. The multi-task learning model is trained in the following way: The multi-task learning model takes the operational data samples and business data samples of each monitored object as input, and outputs the power outage probability prediction result and the power consumption type identification result as output for iterative training. During each training session, the running data samples and business data samples of each object to be monitored are input into the shared backbone network of the multi-task learning model, and the shared temporal representation vector of the running data samples and business data samples of each object to be monitored is extracted. The shared temporal representation vector is input into the power outage detection branch, the electricity consumption type identification branch and the comparison projection branch respectively, and the corresponding power outage probability prediction result, electricity consumption type identification result and low-dimensional comparison vector are output. Calculate the power outage detection loss based on the power outage probability prediction result and the preset power outage detection label; calculate the power consumption type classification loss based on the power consumption type identification result and the preset power consumption type label; calculate the consistency constraint loss based on the consistency between the power outage probability prediction result and the power consumption type identification result; calculate the time series comparison constraint loss based on the low-dimensional comparison vector; The total loss is calculated based on the power outage detection loss, the power consumption type classification loss, the time series comparison constraint loss, and the consistency constraint loss. If the total loss does not meet the preset convergence condition, the network parameters of the multi-task learning model are adjusted.

8. A low-voltage distribution area power outage anomaly detection device, characterized in that, include: The time series characterization extraction module is used to extract the original operating data of each monitored object in the low-voltage area, and extract the multi-dimensional time series characteristics of each monitored object in a continuous time window from the original operating data. The temporal encoding module is used to encode the multidimensional temporal features to obtain the temporal representation vector of each object to be monitored within a continuous time window; The decision state construction module is used to concatenate the time-series representation vector with the preset power outage judgment threshold vector to obtain the current decision state vector of each object to be monitored. The strategy optimization module is used to input each current decision state vector into a pre-trained strategy network, so that the strategy network outputs an adjustment vector for the power outage judgment threshold of each monitored object, and updates each power outage judgment threshold vector according to the adjustment vector. The strategy network is trained with the decision state vector of each monitored object in the historical time window as the state space, the historical power outage threshold parameter adjustment vector as the action space, the correct identification of power outage events as the positive reward, and the delayed reporting, false reporting and missed reporting as the negative penalty. The power outage determination module is used to determine the power outage event in the low-voltage distribution area based on the updated power outage judgment threshold vector and the time sequence representation vector of each monitored object.

9. An electronic device, characterized in that, include: A processor, a memory, and a computer program stored in the memory and configured to be executed by the processor; wherein, when the processor executes the computer program, it implements the low-voltage distribution area power outage anomaly detection method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the storage medium is located to perform the low-voltage distribution area power outage anomaly detection method as described in any one of claims 1-7.

Images