Voltage sag feature-fused power metering box voltage failure classification and diagnosis method

Abstract

The present application relates to the technical field of power quality intelligent monitoring, and particularly relates to a power metering box voltage loss fault classification and diagnosis method fusing voltage sag characteristics, and specifically as follows: three-phase voltage and three-phase current are collected at the power supply monitoring end to form a dynamic electrical quantity sequence, static sag description quantities are extracted and fault category labels are configured to build a sample data set; then, based on the stage evolution law of voltage sag events, the mechanism of three-phase system symmetrical components and other physical priors, physical re-characterization without training parameters is performed on the samples to obtain a dynamic time series representation matrix and a global event semantic vector; then, a stage-aware double-flow interaction graph coding model is constructed to predict the fault category and complete model training; finally, the trained model is deployed on various monitoring terminals or platforms to realize online identification of voltage loss faults. The present application can improve the accuracy and robustness of power metering box voltage loss fault diagnosis by fusing physical mechanism and deep learning technology.

Description

Technical Field

[0001] This invention relates to the field of intelligent power quality monitoring technology, and in particular to a method for classifying and diagnosing power metering box undervoltage faults that integrates voltage sag characteristics. Background Technology

[0002] With the deepening of smart grid construction, electricity metering boxes, as key nodes connecting the user side and the distribution network, directly affect the accuracy of electricity billing, electricity safety, and the quality of power supply services. However, in actual operating environments, metering boxes have long faced multiple hidden dangers, such as loose wiring terminals, aging and broken wires, deterioration of voltage transformer performance, and abnormal sampling links. The resulting voltage loss faults have become an important issue affecting the stable operation of low-voltage distribution networks.

[0003] Existing technologies objectively suffer from the following shortcomings: Existing methods often simply concatenate dynamic sequences with static event-level descriptors, resulting in redundant static descriptors. This redundant input not only increases the model's burden but may also obscure the true details of dynamic changes, making it difficult for the model to focus on key processes. Traditional segmentation methods are mostly based on fixed thresholds for single voltage amplitudes, making it difficult to adapt to different sag depths and strong noise scenarios. They are prone to misjudging short-term fluctuations as event boundaries and cannot effectively utilize known event-level descriptors to improve segmentation stability. Existing data-driven models often directly use the original three-phase voltage and current as input, forcing the model to learn complex physical laws such as phase-to-phase coupling and symmetry disruption, which is particularly difficult when samples are limited and lacks the ability to distinguish similar faults. Conventional time-series models (such as RNNs / CNNs) focus on local temporal dependencies, making it difficult to capture the overall patterns between non-adjacent points within the same fault stage. Furthermore, they typically lack an effective mechanism to use global event-level information to guide the model to focus on the most discriminative key segments in the dynamic process.

[0004] Therefore, this invention proposes a method for classifying and diagnosing voltage sag faults in power metering boxes that integrates voltage sag characteristics to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a method for classifying and diagnosing voltage sag faults in power metering boxes by integrating voltage sag characteristics. By combining physical mechanisms with deep learning technology, this invention can improve the accuracy and robustness of voltage sag fault diagnosis in power metering boxes.

[0006] The technical solution of this invention is a method for classifying and diagnosing voltage sag faults in power metering boxes, which integrates voltage sag characteristics, and includes the following steps: S1. Construct voltage sag event samples. Collect three-phase voltage and three-phase current at the power supply monitoring end to form a dynamic electrical quantity sequence. Extract static sag descriptive quantities for each voltage sag time sample and configure real fault category labels. A dataset is composed of multiple samples. S2. Based on the stage evolution law of voltage sag events, the mechanism of symmetrical components of three-phase systems, and the consistency relationship between dynamic measurement process and static sag descriptor, physical re-representation without training parameters is performed on the samples in the dataset to obtain dynamic time series representation matrix and global event semantic vector. S3. Construct a phase-aware dual-stream interactive graph coding model, use dynamic temporal representation matrix to model the pressure loss fault process, use global event semantic vector to model event-level physical priors, and use the event semantic-guided key time step aggregation mechanism to filter the most discriminative temporal evidence to predict the pressure loss fault category. The model is trained under multi-class supervision, and the optimal parameters of the model are determined to obtain the trained model. S4. Deploy the trained model to power quality monitoring terminals, distribution-side edge computing devices, or centralized fault diagnosis platforms to identify undervoltage faults.

[0007] S1 is as follows: The voltage sag event sample includes dynamic electrical quantity sequences, static sag descriptive quantities, and actual fault category labels; The dynamic electrical quantity sequence is a sequence of three-phase voltage RMS values ​​collected and arranged in chronological order within the observation window. , , and three-phase current RMS value sequence , , Let T represent the total number of time steps in a single observation window, and t represent the index of T. , , These represent the effective voltage sequences of phases A, B, and C at the t-th time step within the observation window, respectively. , , These represent the effective current sequences of phases A, B, and C at the t-th time step of the observation window, respectively. Static sag descriptors include the depth of voltage sag. Duration of temporary descent Phase jump angle of the descent initiation point The static sag descriptor is directly output by the power quality monitoring device, or obtained by offline calculation based on the three-phase voltage waveform, RMS curve and fundamental phase angle change; The actual fault category labels include normal state, single-phase voltage loss, two-phase voltage loss, three-phase voltage loss, voltage loss due to internal wire breakage in the metering box, and abnormal voltage loss of the current transformer.

[0008] S2 is as follows: S2.1. The initial judgment threshold is adaptively determined by using the voltage sag depth and the phase jump angle of the sag initiation point in the voltage sag event sample, and a hysteresis interval is formed by using a recovery judgment threshold higher than the initial judgment threshold. The entire observation window is then divided into the pre-fault period, the fault duration period, and the fault recovery period. S2.2. Based on the stage statistics and rate of change features extracted during the pre-fault, fault duration and fault recovery periods, an event stage feature vector is constructed. S2.3 Map the three-phase voltage and three-phase current to the positive sequence, negative sequence and zero sequence component spaces to construct the voltage imbalance sequence and the current imbalance sequence, and then construct the physical enhancement time sequence feature vector. Stack the physical enhancement time sequence feature vectors of all time steps in time order to obtain the physical enhancement time sequence matrix. S2.4 Calculate the depth consistency coefficient, duration consistency coefficient, phase consistency coefficient and high impedance anomaly indicator coefficient respectively, and construct the physical consistency verification vector; S2.5 Construct an event stage indicator matrix based on the current stage, and concatenate it with the physical enhancement time series matrix to obtain a dynamic time series representation matrix; concatenate the event stage feature vector, static temporary sag descriptor, and physical consistency check vector to obtain a global event semantic vector, and perform unified processing on its execution scale.

[0009] S2.1 is as follows: Obtain the three-phase voltage RMS value sequence from the current voltage sag event sample. If the three-phase voltage RMS values ​​have not yet been converted to per-unit values, normalize them first using the rated voltage as a reference. Then, read the voltage RMS values ​​of the three phases at each time step and select the minimum value to obtain the minimum phase voltage sequence. Determine the sag initiation judgment threshold based on the voltage sag depth and the phase jump angle at the sag initiation point in the static sag descriptor, and determine the sag recovery judgment threshold based on the voltage sag depth. Then, scan the minimum phase voltage sequence along the time sequence to determine the sag event initiation time step. Continue scanning the minimum phase voltage sequence after the sag event initiation time step to determine the sag event end time step. Finally, based on the sag event initiation time step and the sag event end time step, divide the entire observation window into the pre-fault period, the fault duration period, and the fault recovery period.

[0010] S2.2 is as follows: During the early stage of the fault, the average level and standard deviation of the effective values ​​of phase A, phase B, and phase C voltages are statistically analyzed to obtain the average voltage characteristics and voltage fluctuation characteristics during the early stage of the fault. Then, a local time neighborhood is selected around the start time step of the transient event, and linear fitting is performed on the three-phase voltage effective value sequences to obtain the voltage drop rate characteristics. During the fault duration, the dispersion of the three-phase current effective values ​​is statistically analyzed to obtain the current dispersion characteristics during the fault duration. Then, a local time neighborhood is selected around the end time step of the transient event, and linear fitting is performed on the three-phase voltage effective value sequences to obtain the voltage recovery rate characteristics. Finally, the average voltage characteristics, voltage fluctuation characteristics, voltage drop rate characteristics, current dispersion characteristics, and voltage recovery rate characteristics during the fault duration are concatenated in a fixed order to obtain the event stage feature vector.

[0011] S2.3 is as follows: At each time step, the effective values ​​of the three-phase voltage and the effective values ​​of the three-phase current are read to form the original three-phase voltage amplitude vector and the original three-phase current amplitude vector. The fundamental phasors of the three-phase voltage and the three-phase current corresponding to each time step are obtained. Then, Fortescue symmetrical component transformation is performed on the three-phase voltage fundamental phasors to obtain the zero-sequence voltage phasor, the positive-sequence voltage phasor, and the negative-sequence voltage phasor. Then, the amplitudes of the zero-sequence voltage phasor, the positive-sequence voltage phasor, and the negative-sequence voltage phasor are respectively taken to obtain the zero-sequence voltage component, the positive-sequence voltage component, and the negative-sequence voltage component. Then, in the same way as the voltage, symmetrical component transformation is performed on the three-phase current fundamental phasors to obtain the zero-sequence current component, the positive-sequence current component, and the negative-sequence current component. Based on the positive-sequence voltage component and the negative-sequence voltage component, the voltage imbalance is calculated to quantify the proportion of the reverse-sequence voltage disturbance relative to the normal positive-sequence component at the current time step. At the same time, based on the positive-sequence current component and the negative-sequence current component, the current imbalance is calculated to quantify the proportion of the reverse-sequence current disturbance relative to the normal positive-sequence component at the current time step. Finally, the original three-phase voltage amplitude vector, the original three-phase current amplitude vector, the voltage sequence component, the current sequence component, the voltage imbalance, and the current imbalance are combined in a fixed order to obtain the physical enhancement time series feature vector for each time step. The physical enhancement time series feature vectors corresponding to all time steps are stacked in time order to obtain the physical enhancement time series matrix.

[0012] S2.4 is as follows: During the fault duration, the minimum phase voltage sequence is read to determine the lowest voltage level obtained by dynamic measurement. The theoretical lowest voltage level is calculated based on the voltage sag depth in the static sag descriptor and compared with the low voltage level to obtain the depth consistency coefficient. Based on the start time step, end time step and sampling time interval of the transient landing event, the duration of the transient landing is calculated by dynamic measurement. The duration of the transient landing is then compared with the duration of the transient landing in the static transient landing descriptor to obtain the duration consistency coefficient. Around the start time step of the transient landing event, select the phase statistics window before the event and the phase statistics window after the event respectively, calculate the actual phase change, and then compare the actual phase change with the phase jump angle of the transient landing start point in the static transient landing descriptor to obtain the phase consistency coefficient. Based on the zero-sequence components of current and voltage, a zero-sequence normalized admittance sequence is constructed. The zero-sequence normalized admittance sequence is averaged during the early stage of the fault and the fault duration, respectively, to obtain the mean zero-sequence normalized admittance during the early stage of the fault and the mean zero-sequence normalized admittance during the fault duration. The absolute difference between the two is then calculated to obtain the high-resistance anomaly indication coefficient. Finally, the depth consistency coefficient, duration consistency coefficient, phase consistency coefficient, and high impedance anomaly indicator coefficient are combined in a fixed order to obtain the physical consistency verification vector.

[0013] S2.5 is as follows: Based on the pre-fault phase, fault duration, and fault recovery phase, an event phase indicator vector is constructed for each time step. The event phase indicator vectors corresponding to all time steps are stacked in chronological order to obtain the event phase indicator matrix. Then, the physical augmentation time series matrix and the event phase indicator matrix are concatenated according to the feature dimension to obtain the dynamic time series representation matrix. At the same time, the event phase feature vector, static sag descriptor, and physical consistency check vector are concatenated in a fixed order to obtain the global event semantic vector. Before entering the subsequent classification model, the continuous numerical features in the global event semantic vector are subjected to scale unification processing.

[0014] S3 is as follows: The phase-aware dual-stream interaction graph coding model includes a dual-stream input module, a temporal stream processing module, an event stream processing module, a dual-stream interaction and fusion module, and a classification output module. The dual-stream input module receives a dynamic temporal representation matrix and a global event semantic vector. It projects the dynamic temporal representation matrix onto a unified hidden dimension through an input linear mapping layer to obtain the initial node features of each time node. For the global event semantic vector, it determines the event stage category to which each time node belongs. Then, it establishes local temporal edges based on temporal proximity and stage association edges based on event stage consistency. It merges the local temporal edges and stage association edges to construct the connection relationship of the stage-aware dynamic graph. The time step is used as a node by the dynamic graph encoder in the time-series stream processing module to construct a graph structure that integrates time proximity edges and stage association edges. Then, the nodes are encoded by a multi-layer graph attention network to output a dynamic graph node encoding matrix that integrates local and stage context information. The semantic compressor in the event stream processing module compresses the global event semantic vector into a fixed-dimensional summary vector, which serves as the query vector. The dual-stream interaction and fusion module first uses the summary vector output by the event stream as a query, performs an attention mechanism on all nodes of the time-series output, and weights and aggregates the key dynamic process vectors. Then, it concatenates the aggregated key dynamic process vectors with the global event semantic summary vector to obtain the fused vector. The classification output module inputs the fused vector into a fully connected classification layer and outputs the predicted probability of each category.

[0015] The specific steps for model training in S3 are as follows: A multi-class supervised training method is used to optimize the stage-aware two-stream interactive graph encoding model end-to-end. The training samples are sequentially input into the stage-aware two-stream interactive graph encoding model. After stage-aware dynamic graph construction, graph attention message passing, global event semantic compression, key time step aggregation guided by event semantics, and two-stream feature fusion, the corresponding fault category prediction probability vector is obtained. Multi-class cross-entropy loss is calculated based on the predicted probability vector and the real fault category label, and the trainable parameters in the input mapping layer, graph attention encoding layer, global event semantic compression layer, event semantic guided aggregation module and classification layer are updated through backpropagation. The Adam optimizer is used for model training. During the training process, the learning rate and the number of training epochs are adjusted according to the classification accuracy on the validation set, the macro average F1 value, or the decrease in loss. The model parameters corresponding to the highest macro average F1 value are selected and saved to obtain the trained model.

[0016] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: This invention discloses a method for classifying and diagnosing voltage sag faults in power metering boxes. It decouples and reconstructs the original mixed electrical quantity sequences and static descriptive quantities into a clearly structured dynamic temporal representation matrix and a global event semantic vector based on the event stage evolution law and symmetric component mechanism. This achieves effective separation of the temporal process and event-level semantics. The method uses static sag descriptive quantities (depth, phase jump) to dynamically calculate hysteresis thresholds, replacing fixed thresholds for event stage division. A physical consistency verification vector is constructed to quantify the deviation between dynamic measurements and static descriptions, thereby stably identifying boundaries and discovering atypical fault clues. A dynamic graph integrating temporal proximity and stage consistency relationships is constructed, and a graph attention network is used to learn the temporal correlations within different stages. A dual-stream architecture is designed, using global semantics to guide the dynamic process for attention aggregation at key time steps, achieving a deep integration of physical mechanisms and data-driven approaches. The method clarifies the boundary between deterministic re-representation based on electrical mechanisms and trainable classification models. Physical knowledge encoding enhances the physical separability of faults, and end-to-end training is used for joint optimization, balancing physical interpretability and model learning capabilities. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0019] Figure 2 Figure 1 shows a schematic diagram of the effective values ​​of three-phase voltage and current. Figure 2 shows the effective values ​​of three-phase voltage and Figure 3 shows the effective values ​​of three-phase current.

[0020] Figure 3 This is a schematic diagram of the zero-order normalized admittance sequence.

[0021] Figure 4 This is a schematic diagram of segmented detection of transient events in the minimum phase voltage sequence.

[0022] Figure 5 Figure 1 shows a schematic diagram of the analysis of the amplitude of the symmetrical component and the unbalance sequence. Figure 2 shows the amplitude of the symmetrical voltage component, Figure 3 shows the voltage unbalance sequence, and Figure 4 shows the current unbalance sequence. Detailed Implementation

[0023] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0024] like Figure 1As shown, a method for classifying and diagnosing voltage sag faults in energy metering boxes includes the following steps: S1. Construct voltage sag event samples. Collect three-phase voltage and three-phase current at the power supply monitoring end to form a dynamic electrical quantity sequence. Extract static sag descriptive quantities for each voltage sag time sample and configure real fault category labels. A dataset is composed of multiple samples. S2. Based on the stage evolution law of voltage sag events, the mechanism of symmetrical components of three-phase systems, and the consistency relationship between dynamic measurement process and static sag descriptor, physical re-representation without training parameters is performed on the samples in the dataset to obtain dynamic time series representation matrix and global event semantic vector. S3. Construct a phase-aware dual-stream interactive graph coding model, use dynamic temporal representation matrix to model the pressure loss fault process, use global event semantic vector to model event-level physical priors, and use the event semantic-guided key time step aggregation mechanism to filter the most discriminative temporal evidence to predict the pressure loss fault category. The model is trained under multi-class supervision, and the optimal parameters of the model are determined to obtain the trained model. S4. Deploy the trained model to power quality monitoring terminals, distribution-side edge computing devices, or centralized fault diagnosis platforms to identify undervoltage faults.

[0025] In a specific implementation, the construction and category labeling of S1 voltage sag event samples are as follows: First, voltage sag event samples are constructed for undervoltage fault identification. Each voltage sag event sample simultaneously includes a dynamic electrical quantity sequence, a static sag descriptor, and a true fault category label. The dynamic electrical quantity sequence describes the voltage and current changes before the event, during the event duration, and during the event recovery period. The static sag descriptor describes the overall sag depth, sag duration, and sag initiation phase jump angle. The true fault category label supervises the subsequent classification model in learning the discrimination rules between different undervoltage faults.

[0026] In practical implementation, the dynamic electrical quantity sequence refers to the time-series arrangement of the effective values ​​of three-phase voltage and three-phase current within a specific observation window, specifically including: the effective value sequence of three-phase voltage ( , , ) and the three-phase current RMS value sequence ( , , These six sequences together constitute a dynamic electrical quantity sequence, used to characterize the complete dynamic process of an event's evolution over time. Specifically, by collecting continuous monitoring data of three-phase voltage and three-phase current at the power supply monitoring terminal, the effective value sequences of the three-phase voltage are denoted as follows: , and The effective value sequences of the three-phase currents are denoted as follows: , and .in, Indicates the time step index. ; This represents the total number of time steps in a single observation window, with examples of possible values. .

[0027] in, The sequence of effective values ​​of phase A voltage refers to the sequence of values ​​within the observation window. Each time step, the effective value of phase A voltage (e.g., 220V or per unit value). The sequence of effective values ​​of phase B voltage refers to the sequence of values ​​within the observation window. The effective value of phase B voltage at each time step; The sequence of effective values ​​of phase C voltage refers to the sequence of values ​​within the observation window. The effective value of phase C voltage at each time step; The sequence of effective values ​​of phase A current refers to the sequence of values ​​within the observation window. Each time step, the effective value of phase A current (e.g., 10A); The sequence of effective values ​​of phase B current refers to the sequence of values ​​within the observation window. The effective value of phase B current at each time step; The sequence of effective values ​​of phase C current refers to the sequence of values ​​within the observation window. The effective value of the C-phase current at each time step.

[0028] In one implementation, the three-phase voltage RMS sequence and the three-phase current RMS sequence can be output by a power quality monitoring device, a smart metering terminal, or an edge acquisition module. If the original acquisition result is an instantaneous sampling waveform, the RMS value can be calculated first within a fixed sliding window, and then the corresponding RMS value sequence can be formed in chronological order. If different devices have different rated voltage levels, it is preferable to use the rated voltage as a reference and convert the three-phase voltage RMS sequence into per-unit values ​​so that the voltage RMS value under normal operating conditions is close to the rated voltage. This is to enable unified modeling across different samples.

[0029] In the specific implementation, a static sag descriptor is extracted for each voltage sag event sample. The static sag descriptor includes the voltage sag depth. Duration of temporary descent Phase jump angle of the descent initiation point .in, This refers to the percentage drop in voltage relative to the rated voltage during an event, expressed in per-unit values, with the preferred range being... ; This refers to the duration of the temporary landing event, measured in milliseconds. It refers to the amplitude of the phase disturbance at the moment of temporary descent triggering, and the unit is degrees.

[0030] In one implementation, the static sag descriptor can be directly output by the power quality monitoring device; in another, it can be obtained through offline calculation based on the three-phase voltage waveform, RMS curve, and fundamental phase angle change. It should be noted that the static sag descriptor is an event-level description and should not be repeatedly filled in at each time step before being directly concatenated with the dynamic electrical quantity sequence. Otherwise, the event-level characteristics may be repeatedly amplified, reducing the model's ability to reflect the actual dynamic changes.

[0031] In the specific implementation, a true fault category label is assigned to each voltage sag event sample. The number of undervoltage fault categories is denoted as... , This refers to the total number of fault categories in the classification task, which can be taken in one implementation method. The six categories are: normal state, single-phase voltage loss, two-phase voltage loss, three-phase voltage loss, voltage loss due to internal disconnection in the metering box, and abnormal voltage loss of the current transformer. Specifically, normal state indicates that the three-phase voltage maintains a normal power supply level without any obvious voltage dips; single-phase voltage loss indicates a significant decrease or disappearance of voltage in one phase; two-phase voltage loss indicates a simultaneous and significant voltage loss in two phases; three-phase voltage loss indicates a synchronous drop in all three phases; voltage loss due to internal disconnection in the metering box indicates a localized voltage loss caused by loose terminals, broken wires, or internal contact faults; and abnormal voltage loss of the current transformer indicates an abnormal voltage loss due to a faulty voltage transformer, an abnormal sampling link, or a distorted measurement circuit.

[0032] The dynamic electrical quantity sequence, static sag descriptor, and actual fault category label are combined to form a voltage sag event sample. The sample of voltage sag events is denoted as The corresponding actual fault category label is denoted as The training sample set is denoted as... , It refers to a data set consisting of multiple tagged voltage sag event samples.

[0033] In practice, the training sample set should cover different load levels, different sag depths, different sag durations, and different phase transition intensities. For example, sag depth can cover... to The duration of the temporary descent can cover to Phase jump angle can cover to By expanding the range of sample operating conditions, the risk of the model relying solely on a single fixed operating condition feature for classification can be reduced, thereby improving the adaptability of subsequent classification models to actual operating environments.

[0034] In a specific implementation, the decoupling of the physical characteristics of the S2 voltage sag event and the re-characterization of heterogeneous data are as follows: The voltage sag event samples contain both dynamic electrical quantity sequences and static sag descriptors. The dynamic electrical quantity sequences describe the voltage and current evolution before the fault occurs, during the fault duration, and during the fault recovery period; the static sag descriptors describe event-level information such as sag depth, sag duration, and phase jump angle at the sag initiation point. The two differ significantly in data format, temporal attributes, and physical meaning.

[0035] Based on the stage evolution law of voltage sag events, the symmetrical component mechanism of three-phase systems, and the consistency relationship between dynamic measurement processes and static sag descriptors, this invention performs a physical re-representation of the original samples without training parameters to obtain a dynamic time-series representation matrix. With global event semantic vector This provides a clearly structured and physically meaningful input for the subsequent underpressure fault classification model. The specific steps are as follows: S2.1 Event Phase Segmentation Based on Temporary Sag Descriptor Constraints Voltage sag events typically progress through a pre-fault phase, a fault duration phase, and a fault recovery phase. If segmentation is based solely on instantaneous changes in the effective value of the three-phase voltage, short-term noise fluctuations, local spikes, or minor load disturbances can easily be misjudged as event boundaries. This invention adaptively determines the initial judgment threshold using the voltage sag depth and the phase jump angle at the sag initiation point, and uses a recovery judgment threshold higher than the initial judgment threshold to form a hysteresis interval, making event boundary identification more stable. The specific steps are as follows: 1) Obtain the sequence of three-phase voltage RMS values ​​from the current voltage sag event sample. , and If the three-phase voltage RMS value has not yet been converted to per-unit value, it should first be normalized based on the rated voltage to make the voltage RMS value under normal power supply conditions close to the rated voltage. .

[0036] In one embodiment, for example, suppose one monitoring point has a rated voltage of 220V, and its effective voltage value is approximately 220V during normal operation; another monitoring point has a rated voltage of 110V, and its effective voltage value is approximately 110V during normal operation. If nominal values ​​are used directly, the sag threshold for the former might need to be set to 154V (dropping to 0.7), and for the latter to be set to 77V, which is inconsistent. However, after per-unit scaling, the normal voltages for both are close to 1.0; setting the sag threshold to 0.7 applies to all monitoring points, achieving unified modeling. For example, if the voltage of a 220V system drops to 154V, the per-unit value is 0.7; if the voltage of a 110V system drops to 77V, the per-unit value is also 0.7.

[0037] In practical implementation, per-unit processing can eliminate the absolute amplitude differences caused by different rated voltage levels, allowing subsequent sag thresholds to be set directly based on a uniform numerical range. For example, for monitoring points with different rated voltages, as long as their normal supply voltage is normalized to a value close to... Then the same set of segmentation logic can be reused.

[0038] 2) Read the effective voltage values ​​of the three phases at each time step, and select the minimum value to obtain the minimum phase voltage sequence. , Characterizing the first The phase voltage level that drops most severely in the three-phase system at the time step is used to trigger sag boundary identification.

[0039] In one implementation, a three-phase comparison operation can be directly performed to obtain the result. Using the minimum phase voltage sequence for event boundary location can prioritize the response to the earliest obvious drop phenomenon in asymmetrical faults such as single-phase loss of voltage and two-phase loss of voltage, and avoid the true sag initiation being masked by the other phases maintaining higher voltage.

[0040] In one embodiment, for example, suppose that at some point, the voltage of phase A drops to 0.5 (per unit), but phases B and C remain above 0.9; if we look at each phase individually, phases B and C do not trigger any thresholds and may be ignored. However, through calculation... If the minimum value of 0.5 is obtained, the temporary sag boundary identification logic is directly triggered, thereby capturing the starting point of this single-phase loss of pressure and avoiding the fault being masked by the normality of other phases.

[0041] 3) Based on voltage sag depth Phase jump angle of the descent initiation point Determine the threshold for initiating temporary descent , It refers to the boundary threshold used to determine whether the voltage has entered a sag state.

[0042] In one implementation, the initial boundary correction coefficient is first constructed based on the phase jump angle of the temporary descent initiation point. Then increase the voltage sag depth The correction coefficient is mapped to the initial judgment threshold, and is expressed as follows: ; ; in, This indicates the strength of the phase perturbation's correction to the sag initiation boundary; when When it is large, Increase accordingly, so that It is more suitable for capturing sudden sag events accompanied by obvious phase disturbances; based on this, event-level phase information can be introduced into the boundary identification process in advance, reducing the probability of misjudging complex sag boundaries when segmenting based solely on voltage amplitude.

[0043] 4) Based on voltage sag depth Determine the threshold for temporary descent recovery , It refers to the boundary threshold used to determine whether the voltage has left the sag state.

[0044] In practical implementation, the recovery determination threshold can be set to... ;in, This represents the recovery boundary relaxation coefficient, used to control the degree to which the voltage recovers before the fault duration is considered to have ended; to form a stable hysteresis interval, it is preferable to set... ;For example, Desirable .when hour, This indicates that the fault recovery period can only begin after the voltage has returned to near normal levels.

[0045] It should be noted that the settings and Two different thresholds can construct hysteresis judgment logic between "entering the sag state" and "exiting the sag state", thereby reducing stage boundary jitter caused by small fluctuations near the low-voltage platform.

[0046] 5) Scan the minimum phase voltage sequence in time sequence Determine the start time of the temporary landing event. Specifically, when continuous The time step is lower than the initial judgment threshold. When the first time step that satisfies the continuity condition is selected, the time step that initiates the temporary descent event is determined. .

[0047] in, This indicates the continuous acknowledgment length of the initial boundary, used to suppress false triggering caused by isolated noise points or transient spikes. An example value is shown below. or For example, when At that time, only continuous Each time step satisfies the condition below Only when certain conditions are met will the temporary landing event be confirmed as truly underway.

[0048] Furthermore, starting from the start time of the temporary landing event... Then continue scanning the minimum phase voltage sequence Determine the end time of the temporary landing event. Specifically, when continuous The number of time steps is higher than the recovery judgment threshold. When the first time step that satisfies the continuity condition is determined as the end time step of the temporary descent event, the event is terminated. .

[0049] in, Indicates the continuous confirmation length of the end boundary, with the following example values: or This continuous confirmation strategy can prevent short-term voltage rises during the sag period from being mistaken for fault recovery completion.

[0050] 6) Based on the start time of the temporary landing event and the end time of the temporary landing event The entire observation window is divided into the pre-fault period, the fault duration period, and the fault recovery period.

[0051] Specifically, the set of time steps in the early stage of the fault is denoted as... This is used to describe the normal operating baseline before the fault occurs; the set of time steps for the fault duration is denoted as... This is used to describe the voltage anomaly maintenance phase; the set of time steps for the fault recovery period is denoted as... It is used to describe the process by which a system returns to a stable state from an abnormal state.

[0052] In one embodiment, for example, if , ,but Approximately , Approximately ;like ,but .when First time multiple consecutive time steps below When the start time step of the temporary landing event is determined, and multiple subsequent time steps are higher than the start time step, the start time step is determined. At that time, determine the end time of the temporary landing event.

[0053] It should be noted that this step does not simply segment based on a single fixed threshold, but uses static sag descriptors to constrain the event boundary identification process and reduces boundary jitter through hysteresis determination, thereby improving the stability of event stage segmentation when there are large sag depth changes, significant phase perturbations, or strong waveform noise.

[0054] In one embodiment, such as Figure 2 As shown, the analysis of the three-phase voltage and current effective value sequence consists of two sub-figures (a) and (b), which respectively show the change process of the three-phase voltage effective value sequence (per unit value) and the three-phase current effective value sequence (ampere) with time step within the observation window. Figure 2 The starting time of the fault sag is marked with a red dashed line, and the ending time of the fault sag is marked with a green dashed line, which intuitively shows the dynamic changes of electrical quantities before the fault occurs, during its duration, and during the recovery period. Subgraph (a) has the horizontal axis as the time step (no unit, index) and the vertical axis as the voltage (pu, per unit value); Subgraph (b) has the horizontal axis as the time step and the vertical axis as the current (A, ampere).

[0055] S2.2 Statistical Feature Extraction and Process Description Construction for Event Stages The differences between various undervoltage faults are not only reflected in the depth of the sag, but also in the rate of undervoltage occurrence, the current dispersion during the fault duration, and the fault recovery speed. Using only static sag descriptors is insufficient to adequately distinguish between similar events such as single-phase undervoltage, two-phase undervoltage, undervoltage due to internal wiring disconnection in the metering box, and abnormal undervoltage of the instrument transformer. This invention extracts stage statistics and rate of change features based on the pre-fault, fault duration, and fault recovery periods to construct an event stage feature vector. The specific steps are as follows: 1) In the early stage of the fault Within this period, the average levels of the effective voltage values ​​of phase A, phase B, and phase C were statistically analyzed to obtain the average voltage characteristics during the pre-fault period. , It consists of the average voltage values ​​corresponding to the three phases and is used to characterize the normal power supply reference before a voltage dip event occurs.

[0056] In practical implementation, the arithmetic mean of the effective voltage value sequence of each phase during the early stage of the fault can be calculated. This feature can distinguish between two situations: "there was already a slight undervoltage before the event occurred" and "a sudden loss of voltage occurred after the event was triggered", thus providing more realistic baseline information before the fault for subsequent models.

[0057] In one embodiment, as an example Two scenarios can be distinguished: Scenario A (Sudden Fault): Early stage of the fault, average three-phase voltage The values ​​were [0.99, 1.00, 0.98], close to normal levels, after which the voltage suddenly dropped; Scenario B (fault after slow voltage drop): In the early stages of the fault, the system itself has undervoltage. The voltage is [0.91, 0.92, 0.90], and then it drops further into a sag.

[0058] The model passes The characteristics can reveal the baseline state before the fault, avoiding misjudging the early low voltage of scenario B as part of a sag event.

[0059] 2) In the early stage of the fault Within this period, the standard deviations of the effective values ​​of phase A, phase B, and phase C voltages were statistically analyzed to obtain the voltage fluctuation characteristics in the early stage of the fault. , It consists of the voltage fluctuation levels corresponding to the three phases and is used to characterize the stability of the power supply state before the event occurs.

[0060] In practical implementation, the early stage of the fault can be addressed. The standard deviation of the sample voltage RMS value sequence for each phase is calculated. Taking phase A as an example, its standard deviation is... The calculation method is expressed as follows: ; in, Indicates the early stage of the fault Total number of time steps included This represents the average voltage of phase A during the early stage of the fault. The calculation methods for phases B and C are the same as for phase A. Arranging the standard deviations of the three phases in order yields the voltage fluctuation characteristics during the early stage of the fault. .

[0061] In one embodiment, for example, if the voltage of phase A fluctuates slightly within the range of [0.98, 1.02] in the early stage of a fault, the calculated standard deviation may be 0.012; if the voltage fluctuates irregularly between [0.95, 1.05] due to load switching, the calculated standard deviation may be 0.035. The latter indicates that the power supply stability is poor before the event occurs, and the voltage surge near its boundary is more likely to be caused by load disturbances rather than fault triggering.

[0062] It should be noted that if the voltage fluctuations are strong in the early stages of a fault, then the local voltage changes near the event boundary cannot be directly equated with a sudden fault occurrence; Incorporating event phase feature vectors helps the model distinguish between "original operational fluctuations" and "actual pressure loss triggers".

[0063] 3) Steps regarding the start time of the temporary landing event By selecting a local time neighborhood and performing linear fitting on the effective value sequences of the three-phase voltages respectively, the voltage drop rate characteristics are obtained. , It is used to describe how quickly the voltage changes from a normal state to a low voltage state.

[0064] In one implementation, the initial fitting window can be set to... front and back Each time step, total Each time step. Least squares linear fitting is performed on the effective voltage value sequences of phases A, B, and C respectively to obtain the fitting slopes for the three phases; then the average of the absolute values ​​of the three slopes is taken as the... .when A larger value usually indicates a more sudden decompression process; when When the value is smaller, it is more likely to correspond to non-sudden scenarios such as contact anomalies, slow detachment, or measurement link drift.

[0065] In one embodiment, for example, suppose The initial fitting window is from time step 75 to time step 85. Least squares linear fitting is then performed on the effective voltage values ​​of phases A, B, and C within this window, resulting in three fitted lines with slopes. The average of the absolute values ​​of these three slopes yields the voltage drop rate characteristic. ,For example: Sudden voltage loss (e.g., short circuit): Phase A voltage drops rapidly from 0.98 to 0.45. A linear fit is performed on these 11 points, yielding a slope of approximately -0.07 pu / step, whose absolute value... , Larger; Slow voltage drop (e.g., high-resistance contact): Phase A voltage slowly drops from 0.98 to 0.85, with a fitting slope of approximately -0.01 pu / step. , The magnitude is relatively small. The model can use this to distinguish the suddenness of pressure loss.

[0066] 4) During the fault duration Within this period, the dispersion of the three-phase current effective values ​​is statistically analyzed to obtain the current dispersion characteristics during the fault duration. , Used to describe the overall level of three-phase current difference during the fault duration phase.

[0067] In the specific implementation, for each time step within the fault duration, read... , and Determine the maximum and minimum phase currents in the current time step and calculate their ratio; then average all ratios over the fault duration to obtain the result. If the minimum phase current is close to Then, add the minimum stability constant to the denominator. , The stability constant can be calculated using the current ratio, which can be taken as... .

[0068] In one embodiment, for example, if the maximum phase current at a certain time step during the fault duration is 12.5A and the minimum phase current is 4.0A, then the current dispersion ratio at that time step is 12.5 / 4.0 = 3.125. If the current ratio at each time step is around this level throughout the entire fault duration, then... The ratio will be significantly greater than 1; if the three-phase currents are relatively balanced and the ratio is close to 1, then... It will approach 1.

[0069] It should be noted that the current discrete characteristics during the fault duration are... The larger the value, the more unbalanced the three-phase current. This feature is helpful in identifying atypical undervoltage scenarios such as internal open circuits, high-resistance contacts, and transformer abnormalities, because these scenarios are often accompanied by asymmetrical enhancement of the current response.

[0070] 5) Regarding the end time of the temporary landing incident By selecting a local time neighborhood and performing linear fitting on the effective value sequences of the three-phase voltages respectively, the voltage recovery rate characteristics are obtained. , Used to describe the speed at which voltage recovers from an abnormal state to a normal state.

[0071] In one implementation, the recovery fitting window can be set to... front and back Each time step; least squares linear fitting is performed on the effective value sequence of the three-phase voltages, the absolute value of the recovery slope of each of the three phases is extracted, and then the average value is calculated to obtain... ;like A larger value indicates a rapid recovery; if A smaller value may reflect a gradual change in contact impedance or a slow recovery of the measurement link.

[0072] In one embodiment, for example, suppose If the fitting window is from time step 195 to 205, then least squares linear fitting is performed on the effective value sequence of the three-phase voltages within this window, and the average of the three absolute slopes is taken to obtain the voltage recovery rate characteristic. ,For example: The fault was cleared, and the voltage quickly recovered: the phase A voltage rapidly rose from 0.5 to 0.98, and the absolute value of the fitted slope was very large. Larger; Poor contact, slow voltage rise: Phase A voltage slowly rises from 0.5 to 0.75, with a relatively small absolute value of the fitted slope. Smaller size helps identify atypical recovery processes.

[0073] 6) Average the characteristics of the pre-fault voltage in a fixed order. (A, B, and C phases, 3 dimensions in total) Characteristics of voltage fluctuations in the early stage of a fault (Three phases A, B, and C, totaling 3 dimensions) Voltage sag rate characteristics (1-dimensional scalar), discrete characteristics of fault duration current (1-dimensional scalar) and voltage recovery rate characteristics By concatenating (1-dimensional scalars), we obtain the event stage feature vector. , The dimension is It is used to comprehensively characterize the operating baseline in the early stage of a fault, the sag triggering process, the degree of current imbalance during the fault duration, and the fault recovery process.

[0074] In one embodiment, for example, if the sampling time interval is The temporary landing event started at step number 1. The time step can be in the first... Time step to the 1st Within each time step, the effective value variation curves of the three-phase voltage are fitted, and the average of the three slopes is calculated. If the maximum phase current at a certain time step during the fault duration is The minimum phase current is Then the current discrepancy ratio at that time step is If a high current dispersion ratio is maintained across multiple time steps, then It will increase significantly.

[0075] It should be noted that the event phase feature vector Instead of performing rough statistics on data across all time periods, it preserves the process differences between the early stage of a fault, the duration of a fault, and the recovery period. This allows subsequent models to use event evolution logic for discrimination, rather than relying solely on a single static amplitude feature.

[0076] S2.3, Mapping of Symmetrical Components of Three-Phase Electrical Quantities and Construction of Unbalance Sequence Different undervoltage faults exhibit different symmetrical failure modes in a three-phase system. Single-phase undervoltage typically causes a significant increase in the negative-sequence component; two-phase undervoltage is often accompanied by changes in both the negative-sequence and zero-sequence components; and simultaneous three-phase undervoltage is more likely to manifest as an overall decrease in the positive-sequence component with a weaker change in imbalance. If only the original RMS values ​​of the three-phase voltage and current are used, the model needs to learn the interphase coupling relationship on its own based on the training data, which increases the training difficulty and is not conducive to stable identification under small sample conditions.

[0077] This invention maps three-phase voltage and three-phase current to positive-sequence, negative-sequence, and zero-sequence component spaces, and further constructs voltage imbalance sequences and current imbalance sequences. The specific steps are as follows: 1) At each time step Read the RMS values ​​of the three-phase voltage and the RMS values ​​of the three-phase current to form the original three-phase voltage amplitude vector. and the original three-phase current amplitude vector ;in, Used to retain the The original voltage amplitude information of the time step, Used to retain the The original current amplitude information at the time step.

[0078] In practical implementation, Depend on , and Composed in sequence, Depend on , and The components are arranged sequentially; the original three-phase amplitude information is preserved, which enables subsequent models to utilize the characteristics of symmetrical components without losing the absolute drop pattern in the real monitoring values.

[0079] 2) Obtain the first The three-phase voltage fundamental phasors and three-phase current fundamental phasors corresponding to the time step; wherein, the three-phase voltage fundamental phasors are denoted as... , and The fundamental phasors of the three-phase currents are denoted as follows: , and .

[0080] In one implementation, if the acquisition device can directly output the three-phase fundamental amplitude and phase angle, then complex phasors can be constructed based on the amplitude and phase angle. In another implementation, if the acquisition device outputs instantaneous voltage and instantaneous current waveforms, then a sliding window of one fundamental period or half a fundamental period can be selected around the current time step, and the fundamental complex coefficients can be extracted through discrete Fourier transform to obtain the corresponding fundamental phasors.

[0081] In one embodiment, as an example, assuming a sampling rate of 128 points / cycle (i.e., a 50Hz system, sampling 128 points within 20ms per cycle), the current sampling rate is the [number missing]. One time step. To obtain the fundamental phasor at this moment, waveform data for a complete fundamental period (128 points) including this time step can be taken and subjected to a Discrete Fourier Transform. Assume the complex coefficients of the Discrete Fourier Transform of phase A fundamental are calculated as follows: Then the fundamental phasor of phase A The amplitude is Phase is ,plural Phasors used for subsequent symmetric component transformations.

[0082] It should be noted that symmetric component mapping must be based on the three-phase fundamental phasor containing phase information, and cannot be calculated directly based solely on the effective amplitude of the three phases; otherwise, the physical relationship between the positive sequence, negative sequence, and zero sequence cannot be correctly represented.

[0083] 3) Perform Fortescue symmetrical component transformation (i.e., symmetrical component method) on the three-phase voltage fundamental phasors to obtain the zero-sequence voltage phasors. Voltage positive sequence phasor and negative sequence voltage phasor In practical implementation, the following transformation relationship can be adopted: ; in, Indicates the three-phase rotation factor; It represents the imaginary unit.

[0084] Furthermore, by taking the amplitude values ​​of the zero-sequence voltage phasor, positive-sequence voltage phasor, and negative-sequence voltage phasor respectively, the zero-sequence voltage component is obtained. Positive sequence voltage component and negative sequence voltage components .

[0085] in, Used to describe the main voltage components that maintain normal phase sequence relationships; Used to describe the reverse-order perturbation components caused by three-phase asymmetry; It is used to describe the superimposed components of the same phase and has an auxiliary characterization effect on grounding-related anomalies and special open circuit conditions.

[0086] In one embodiment, as an example, if the calculation yields... The complex value of is Its amplitude V; similarly, it can be calculated that and amplitude , .

[0087] 4) Perform symmetrical component transformation on the fundamental phasors of the three-phase currents in the same manner as for the voltage to obtain the zero-sequence components of the currents. Positive sequence component of current and negative sequence component of current .

[0088] in, Used to describe the normal symmetrical component in three-phase current; Used to describe the reverse-sequence disturbance component generated by current imbalance; Used to describe the in-phase composite components of three-phase currents, it has an auxiliary role in identifying internal open circuits, high-resistance contacts, and grounding-related anomalies.

[0089] 5) Based on positive sequence voltage components and negative sequence voltage components Calculate voltage imbalance This is used to quantify the proportion of the reverse-sequence voltage disturbance at the current time step relative to the normal forward-sequence component, and is calculated as follows: ; in, This represents the stability constant used in calculating voltage unbalance, and is used to prevent the denominator from becoming too small when the positive sequence voltage component is too small. Example of a value: ; The larger the value, the higher the value. The more asymmetrical the three-phase voltages are in time step, the more unbalanced they become.

[0090] 6) Based on the positive sequence component of current and negative sequence component of current Calculate the current imbalance This is used to quantify the proportion of the reverse-sequence current disturbance relative to the normal positive-sequence component at the current time step, and is calculated as follows: ; in, The stability constant used to calculate the current unbalance can be taken as... ; The larger the value, the more pronounced the three-phase current asymmetry under fault conditions.

[0091] 7) Convert the original three-phase voltage amplitude vector Original three-phase current amplitude vector Voltage sequence component, current sequence component, voltage imbalance and current imbalance Combining them in a fixed order yields the [number]. Physically augmented temporal feature vectors of time steps , It refers to the first The temporal feature vector after physical re-representation of the time step has a dimension of This vector not only preserves the original three-phase measurement amplitude, but also adds positive sequence, negative sequence, zero sequence and unbalance characteristics, enabling subsequent models to directly utilize input features with clear electrical meaning.

[0092] 8) Convert the physical augmentation timing feature vectors corresponding to all time steps. Stack them in chronological order to obtain the physical enhancement time series matrix. , The dimension is It is used to describe the dynamic physical evolution of underpressure faults within the observation window.

[0093] In one embodiment, for example, after a single-phase loss of voltage event occurs, the negative sequence voltage component... Typically, the voltage imbalance increases rapidly. Significantly increased; after a three-phase synchronous undervoltage event, the positive sequence voltage component... Overall decrease, but negative voltage sequence component The voltage imbalance remains at a low level, therefore... Typically small; this difference can directly enhance the physical separability between different types of pressure loss at the input layer.

[0094] It should be noted that this step encodes the three-phase electrical mechanism of undervoltage fault into the input features through symmetric component mapping and unbalance construction. This can reduce the burden of the model relying entirely on data to fit the interphase coupling relationship, and is especially suitable for application scenarios with a limited number of samples or where the boundaries between fault categories are close.

[0095] S2.4 Consistency Verification between Dynamic Measurement Process and Static Suspension Descriptor The dynamic process of a real voltage sag event should be reasonably consistent with the static sag description. If the sag depth, duration, or phase jump description significantly deviates from the actual measurement process, it may indicate that the sample is not a typical external power supply side voltage loss, but is related to atypical operating conditions such as internal wiring breaks, high-resistance contacts, or transformer malfunctions. This invention constructs a physical consistency verification vector. This is used to describe the deviation between dynamic electrical quantity processes and static sag description quantities. The specific steps are as follows: 1) During the fault duration Internal reading of minimum phase voltage sequence Determine the lowest voltage level obtained from dynamic measurements. , Characterizes the most severe actual voltage drop level during the fault duration.

[0096] In practical implementation, the duration of the fault can be directly monitored. The minimum value is the low voltage level. This represents the most severe voltage drop level monitored during the event, and this feature is used to provide a dynamic measurement basis for comparison with the static sag depth.

[0097] 2) Based on voltage sag depth Calculate the theoretical minimum voltage level and with By comparing the results, we can obtain the deep consistency coefficient. , This characterizes the degree of deviation between the dynamically measured minimum voltage and the derived results of the static sag depth.

[0098] In practical implementation, the theoretical minimum voltage based on static descriptors can be calculated. ) and the actual lowest voltage based on dynamic measurements ( The absolute difference between the two values ​​is used to calculate the deep consistency coefficient through standardization. ,Right now ; The smaller the value, the more consistent the static descent depth is with the dynamic measurement process; The larger the value, the more significant the deviation between the event description and the actual voltage drop process.

[0099] 3) Based on the start time of the temporary landing event The temporary landing event ended at a certain time. and sampling time interval Calculate the duration of the descent obtained from dynamic measurement. The calculation method is expressed as ;in, Characterizes the sampling time interval between adjacent time steps, with examples of values ​​as follows: ; Characterizes the dynamic duration obtained based on the stage segmentation results. For example, if , Sampling interval ,but .

[0100] 4) The duration of the descent obtained by dynamic measurement Compared to the duration of sag in static sag descriptors By comparing the results, the duration consistency coefficient is obtained. , Used to describe the relative deviation between the results of dynamic event segmentation and the description of static duration.

[0101] In one implementation, a relative error method can be used for calculation. and in When the value is small, add the minimum stability constant. ,Right now ,in, It is a very small value (e.g., 0.001). ), to prevent division by zero, This represents the function that takes the maximum value. The smaller the value, the more consistent the duration relationship between the dynamic process and the static descriptive quantity. For example, static duration. Dynamic measurement of duration ,but The value is very small, indicating good consistency.

[0102] 5) Steps regarding the start time of the temporary landing event Select the phase statistics window before the event and the phase statistics window after the event respectively, and calculate the actual phase change. , The actual phase abrupt change is characterized by measurements based on dynamic voltage phasors.

[0103] In one implementation, the voltage positive-sequence phasor is started first. Extract the positive sequence phase angle; then respectively in The phase angle mean is calculated within the short time window before and after the phase angle mean; finally, the circumferential angle difference is calculated on the two phase angle mean values ​​to obtain the result. The calculation of the difference in circumferential angles is used to avoid... and Mistakenly judged as phase difference Instead, it should be correctly understood as the difference. .

[0104] In one embodiment, as an example, take The average phase angle of the first 5 time steps is , The average phase angle of the last 5 time steps is If the absolute difference is calculated directly, it is... This does not conform to physical reality; by using the method of calculating the difference of circumferential angles, the two angles can first be normalized to... The interval is then used to calculate the absolute difference; the normalized angles are respectively... and The absolute difference is , That is .

[0105] 6) The actual phase change Phase jump angle of the sag in the static sag descriptor By comparison, the phase consistency coefficient is obtained. , Used to describe the degree of matching between dynamic measurement phase perturbations and event-level phase descriptions.

[0106] In practical implementation, the phase jump is specifically calculated by dynamically measuring the phase transition. Compared with static descriptor The absolute difference is then normalized to obtain the phase coherence coefficient. , represented as For example, static descriptors Dynamic measurement obtained ;but , completely consistent; if ,but There are discrepancies.

[0107] It should be noted that, The smaller the value, the more consistent the dynamic phase change is with the static sag description; The larger the value, the more likely there is a deviation in the event description, or the sample is closer to atypical scenarios such as transformer anomalies or measurement link anomalies.

[0108] 7) Based on the zero-sequence component of current and voltage zero sequence component Constructing a zero-order normalized admittance sequence , Used to describe the The response strength of the zero-sequence current at the time step relative to the zero-sequence voltage is calculated as follows: ; in, The zero-sequence normalized admittance is used to calculate the stability constant, which can be taken as... If the zero-sequence voltage is low and the zero-sequence current relationship is significantly abnormal during the event, then It will produce significant changes, which helps to characterize special pressure loss conditions such as internal wire breaks and high-resistance contacts.

[0109] 8) In the early stages of the fault and duration of failure Within, for the zero-order normalized admittance sequence The mean of the zero-sequence normalized admittance in the early stage of the fault is obtained by averaging. and the mean of zero-order normalized admittance during fault duration Then calculate the absolute difference between the two to obtain the high-resistivity anomaly indication coefficient. ,Right now .

[0110] in, Used to describe the degree of change in the zero-sequence admittance relationship after an event occurs relative to the pre-fault stage; if A significant increase indicates a sudden change in the zero-sequence electrical behavior during the fault duration, which has auxiliary identification value for internal wire breaks, high contact impedance, or abnormal measurement links in the metering box.

[0111] In one embodiment, for example, if the fault is in its early stages It remains around 0.001, with its mean... The value is 0.001; during the fault duration, a high-resistance ground fault occurred. It suddenly increased to 0.05, its mean If the value is 0.05, then the high resistance anomaly indication coefficient is... This indicates that the zero-sequence admittance relationship changed significantly before and after the fault.

[0112] 9) Combine depth consistency coefficients in a fixed order Duration Consistency Coefficient Phase consistency coefficient and high resistance anomaly indicator coefficient The physical consistency verification vector is obtained. , This refers to the comprehensive consistency characteristics between the dynamic measurement process and the static sag descriptor, with dimensions of 1. .

[0113] In one embodiment, for example, if the static sag duration The dynamic measurement duration calculated based on the stage segmentation results Then the duration consistency coefficient A smaller value indicates a good consistency between the dynamic process and the static descriptive quantity. If an event simultaneously has a larger value... and larger If the result is negative, it indicates that the event may not be a typical external power supply loss, but is more likely related to an internal disconnection, high-resistance contact, or abnormal measurement link.

[0114] In one embodiment, such as Figure 3 As shown, the analysis of the zero-order normalized admittance sequence This demonstrates the variation of the zero-sequence normalized admittance sequence, calculated from the ratio of the zero-sequence current component to the zero-sequence voltage component, throughout the entire observation window. A significant abrupt change occurs in this sequence during the fault duration, indicating a substantial alteration in the zero-sequence electrical behavior before and after the fault. The zero-sequence normalized admittance's ability to indicate atypical undervoltage scenarios such as high-resistance contacts and internal open circuits serves as the high-resistance anomaly indication coefficient in the physical consistency verification vector. Provides the data foundation. The horizontal axis represents the time step, and the vertical axis represents the admittance (S, Siemens).

[0115] S2.5, Heterogeneous Structured Characterization Output In the original voltage sag event samples, dynamic electrical quantity sequences and static sag descriptions are stored together, which can easily lead to interference between timing information and event-level semantics. This invention separates and organizes these two types of information, outputting a dynamic timing representation matrix. and global event semantic vector The specific steps are as follows: 1) Based on the early stage of the fault Fault duration and fault recovery period Construct an event stage indicator vector for each time step. If the... If the time step is in the early stage of the fault, then the stage indication vector is taken as follows: If it falls within the fault duration period, then take... If it is within the fault recovery period, then take .

[0116] Stack the event stage indicator vectors corresponding to all time steps in chronological order to obtain the event stage indicator matrix. , Dimensions This is used to identify the event stage to which each time step belongs.

[0117] In one embodiment, for example, suppose an observation window has 5 time steps, where steps 1 and 2 belong to the pre-fault phase, steps 3 and 4 belong to the fault duration phase, and step 5 belongs to the fault recovery phase, then the event phase indication matrix... for: ; Each row represents a time step, and the three columns correspond to "pre-fault stage", "fault duration", and "fault recovery stage" respectively.

[0118] 2) Physical enhancement timing matrix Event Phase Indicator Matrix Concatenating the features along their dimensions yields the dynamic temporal representation matrix. , Dimensions It is used to describe the changes in the original three-phase electrical quantities, the changes in the symmetrical components, the changes in the unbalance, and the event stage to which the time step belongs within the entire observation window.

[0119] 3) Divide the event stage feature vector Voltage sag depth Duration of temporary descent Phase jump angle at the descent start point and physical consistency verification vector By concatenating them in a fixed order, a global event semantic vector is obtained. , Dimensions It is used to comprehensively characterize event phase statistics, static transient descriptive quantities, and the consistency relationship between dynamic measurement processes and static descriptive quantities.

[0120] It should be noted that a fixed order refers to the sequence of event phase feature vectors. (9-dimensional), voltage sag depth (1-dimensional), duration of temporary descent (1D), Phase jump angle at the descent initiation point (1-dimensional) and physical consistency verification vector (4-dimensional) Concatenating them in this order yields a vector of 9+1+1+1+4=16 dimensions, i.e. .

[0121] 4) Before entering the subsequent classification model, process the global event semantic vector. The continuous numerical features in the training set undergo scaling uniformity processing. Specifically, the mean and standard deviation of each feature dimension can be calculated based on the statistics of the training set, and then the corresponding dimensions are standardized to bring features with different dimensions, such as sag duration, phase jump angle, and admittance bias, into a similar numerical range.

[0122] It should be noted that standardized statistics should only be calculated on the training set, and the statistics saved during the training phase should be used consistently during the validation and online inference phases. The normalization range should not be recalculated for individual test events, otherwise the relative scale relationships between events will be altered. It should also be noted that this step separates dynamic process information and event-level semantic information into... and Two independent data streams avoid the repeated filling of static slack descriptors on the time axis, reduce redundant input, and provide a structured foundation for subsequent dual-stream interactive modeling.

[0123] In one embodiment, such as Figure 4 As shown, the minimum phase voltage sequence is illustrated. The time series curve is superimposed with the initial decision threshold adaptively calculated based on the static sag descriptor. and recovery judgment threshold . Figure 4 The detected descent start time step was selected. and end time step The fault duration is marked with a gray semi-transparent area, demonstrating the ability to segment stable event stages using hysteresis thresholds and static descriptor constraints, effectively avoiding boundary jitter caused by noise fluctuations. The horizontal axis represents the time step, and the vertical axis represents the voltage (pu, per unit value).

[0124] In one embodiment, such as Figure 5As shown, the analysis of the symmetric component amplitude and imbalance sequence includes three subplots: the first subplot (a) shows the evolution of the voltage positive-sequence, negative-sequence, and zero-sequence component amplitudes over time; the second subplot (b) shows the voltage imbalance. Sequence; the third subplot (c) shows the current imbalance. The sequence is shown. All subplots are labeled with the start and end times of the temporary sag. The significant changes in the negative sequence component under single-phase undervoltage faults demonstrate that symmetric component mapping can encode three-phase asymmetry as input features, enhancing the physical separability of different fault types. The horizontal axis of the three subplots represents the time step, the vertical axis of the first subplot (a) represents the amplitude, and the vertical axis of the remaining subplots (b) and (c) represents the unbalance degree (dimensionless).

[0125] In a specific implementation, the training of the undervoltage fault classification model based on stage-aware dual-stream interactive graph encoding in S3 is as follows: Dynamic time series representation matrix Describes the complete evolution of a fault from before it occurs to after it is recovered; global event semantic vector. This describes the severity of the event, stage statistical characteristics, and dynamic-static consistency. These correspond to "time-series process information" and "event-level global semantic information," respectively. This invention constructs a stage-aware two-stream interactive graph coding model, using a dynamic temporal representation matrix... Modeling the underpressure fault process using global event semantic vectors We model event-level physical priors and use an event semantic-guided key time step aggregation mechanism to filter the most discriminative temporal evidence to predict the type of pressure loss fault. The specific steps are as follows: S3.1, Construction of Stage-Aware Dynamic Graph and Modeling of Local Temporal Relationships The underpressure fault process exhibits both local continuity and stage-based overall characteristics. Adjacent time steps reflect short-term evolution trends, while time steps spaced far apart within the same fault stage reflect the overall pattern within that stage. This invention maps time steps to graph nodes and constructs a stage-aware dynamic graph. The specific steps are as follows: 1) Dynamic time series representation matrix The first in The row feature is regarded as the first The original input features at the nth time point are obtained and projected onto a unified hidden dimension through an input linear mapping layer to obtain the nth time point. Initial node characteristics of each time point , It refers to the first The initial feature representation of the time step in the graphical model; the dimension of the hidden features is denoted as . Example of a value: .

[0126] In its implementation, the linear mapping layer is a trainable module used to convert the original... The dynamic temporal representation is uniformly transformed into latent space features that facilitate graph encoding. Specifically, the linear mapping layer is implemented by a single-layer fully connected network, without activation functions, and contains one dimension. The weight matrix and a dimension of The bias vector is calculated as follows: [The process involves] setting the first... The original feature vectors (17 dimensions) at each time step are multiplied by the weight matrix, and then the bias vector is added to obtain the dimension... Initial node features .

[0127] 2) Based on the event phase indicator matrix Determine the event stage category to which each time point belongs. ;in, One of the following can be selected: the early stage of the fault, the duration of the fault, or the recovery stage, for subsequent establishment of stage relationships.

[0128] In practical implementation, the value can be directly read from the event phase indicator vector. The location is used to determine the stage category corresponding to the current time node.

[0129] In one embodiment, following the example above, The third row is [0,1,0], and the position with a value of 1 is the second column; according to the mapping relationship (first column - pre-fault stage, second column - fault duration, third column - fault recovery stage), the third time step belongs to the stage category. It refers to the "fault duration".

[0130] 3) Establish local temporal edges based on temporal proximity. Specifically, for any two time nodes, if the difference in their time step indices is not greater than the radius of their local temporal neighborhood... Then a local time edge is established between the two; This indicates the scope of local short-term dependency modeling, with examples of possible values. .

[0131] In one embodiment, as an example, suppose For the 40th time step, local time edges will be established with the 37th, 38th, 39th, 41st, 42nd, and 43rd time steps (index difference ≤ 3). For example, the 40th and 42nd time steps will be connected, but the 40th and 44th time steps (index difference = 4) will not be connected.

[0132] It should be noted that local time edges can express the continuous changing trend of voltage and current between adjacent sampling points, making it easier for the model to capture sag trigger edges, recovery edges, and short-time current abrupt changes.

[0133] 4) Establish stage-related edges based on the consistency relationship between event stages. Specifically, for any two time points, if they belong to the same stage category, i.e. Then, a phase-related edge is established between the two, where, It refers to the first The event phase category to which each time point belongs.

[0134] It should be noted that stage-related edges enable information exchange between time points that are far apart within the same event stage. For example, although the first and second phases of the fault duration are far apart in time, both are on a low-voltage anomaly platform. Establishing stage-related edges can help the model learn the overall stability pattern of the fault duration stage, rather than just focusing on local and nearby changes.

[0135] 5) Merge local time edges with stage-related edges to construct the connection relationships of the stage-aware dynamic graph, and form an adjacency matrix from these connections. , Used to describe whether graph message passing is allowed between any two time points.

[0136] In practical implementation, the adjacency matrix It is a dimension If the square matrix, the first The and the first If there is an edge between the nth time nodes (whether it is a local time edge or a stage-related edge), then the nth time node in the matrix... Line 1 Column elements Otherwise, it is 0. For example, if the 40th and 42nd time steps are connected due to their proximity in time, then... If the 40th and 90th time steps are connected because they both belong to the fault duration, then .

[0137] In one embodiment, if the first Time step and the The index difference between time steps is ,and Then the two establish a local time edge; if the first Time step and the All time steps belong to the fault duration period, and even if the two are far apart, a stage association edge is established; based on this, the resulting stage-aware dynamic graph can simultaneously express local short-term changes and the overall pattern within the stage.

[0138] It should be noted that the graph structure constructed in this step is neither a fixed fully connected graph nor a simple adjacency time series graph, but rather a graph that integrates both "temporal proximity" and "stage consistency". Based on this, the model's understanding of the stage structure of the pressure loss event can be enhanced, which is especially helpful in distinguishing between different physical processes such as the sustained period of stable low pressure and the recovery period of rising fluctuations.

[0139] S3.2 Graph Attention Message Passing and Stage Context Encoding After establishing the phase-aware dynamic graph, it is necessary to perform neighborhood aggregation on the features of each time node. Since different neighboring nodes have different importance to the current node, fault trigger boundaries, low-voltage extreme value intervals, and recovery boundaries are usually more valuable for classification than stable intervals. This invention adopts a graph attention message passing mechanism to assign adaptive weights to neighboring nodes. The specific steps are as follows: 1) Input the node features from the previous layer into the graph attention encoding layer, perform a linear feature transformation on each time node to obtain message mapping features for subsequent relevance calculations. The dimension of the message mapping features is the same as that of the hidden features. (e.g., 64-dimensional); the number of graph attention encoding layers is denoted as... Example of a value: .

[0140] In its implementation, the graph attention encoding layer is implemented using a single-layer graph attention network (GAT). The input to each layer is a set of node features, and the output is new node features after neighborhood aggregation and feature transformation, while maintaining the same dimensionality. The first layer of graph attention encoding receives initial node features. The second layer graph attention encoding receives node features from the output of the first layer. .

[0141] In practical implementation, for the first-layer graph attention encoding, the node features of the previous layer refer to the initial node features. For the second layer, the node features of the previous layer refer to the node features output by the graph attention encoding of the first layer. That is, the first-layer graph attention encoding receives the initial node features. The second layer graph attention encoding receives node features from the output of the first layer; the linear feature transformation is used to project the hidden representations of different nodes onto a unified comparison space, providing a basis for calculating neighbor weights.

[0142] 2) Calculate the correlation score between the current node and its neighboring nodes. The correlation score comprehensively reflects the degree of matching between the features of the current node and the features of the neighboring nodes. It is a scalar used to measure the relative importance of a specific neighboring node to the current central node. It is calculated by analyzing the features of the two nodes.

[0143] In one implementation, a shared attention scoring function from graph attention networks can be used. This function concatenates the features of the current node with those of its neighbors to calculate a score, and then obtains a relevance score through non-linear activation. The shared attention scoring function is a core computational mechanism in graph attention networks, used to quantify the feature relevance between the center node and its neighbors. A typical implementation involves first concatenating the features of the current node with those of its neighbors... Features with neighboring nodes Features Each is transformed by a learnable linear transformation matrix. The two transformed feature vectors are mapped to the same space, then concatenated, and finally combined with a learnable attention weight vector. We perform an inner product and finally obtain the non-normalized correlation score using the LeakyReLU activation function. The entire process can be described as follows: First, calculate the concatenation vector. Then calculate .in, Indicates the central node The feature vectors, corresponding to the dynamic graph node encoding matrix The Middle The node features at each time step, with dimensions of [dimensionality]. ; Representing neighboring nodes The feature vectors, corresponding to the dynamic graph node encoding matrix The Middle The node features at each time step, with the same dimension. ; The linear transformation matrix representing the node features. This represents a weight vector used to map the concatenated high-dimensional features to a scalar score. yes transpose, This represents a vector concatenation operation. This represents the activation function of the linear unit with leakage correction.

[0144] In one embodiment, for example, suppose the hidden dimension Central node Features Given a 64-dimensional vector, neighbor nodes Features It is also a 64-dimensional vector; the weight matrix The dimension is This can map the two into a new 64-dimensional feature; attention weight vector The dimension is This is used to perform an inner product with the concatenated 128-dimensional vector, ultimately yielding a scalar fraction. This score represents the score of neighboring nodes in the unnormalized case. For the central node The importance of.

[0145] 3) Softmax normalization is performed on the relevance scores between the current node's neighbor set and its own nodes to obtain the neighbor attention weights. These neighbor attention weights describe the degree of dependence of the current node on different neighbor nodes when aggregating information, and the sum of all normalized weights corresponding to the same node is... For example, nodes located near the initial boundary of a transient event may be assigned higher weights to nodes experiencing rapid drops; nodes located in the middle of the fault duration may be assigned higher weights to nodes with stable low-voltage platforms during the same phase.

[0146] It should be noted that the neighbor attention weight is applied to the current node. A set of scalar values, where each neighbor (including itself) corresponds to a weight, and the total dimension is equal to the node's scalar value. The size of the neighbor set; for a given node, the sum of these scalars is 1.

[0147] 4) Based on the neighbor attention weights, the message mapping features of neighbor nodes are weighted and summed, and then passed through a non-linear activation function to obtain the updated node features. After multi-layer graph attention encoding, the dynamic graph node encoding matrix is ​​obtained. , The encoding result representing all time points after passing through the stage-aware graph structure and attention message passing, with dimensions of ; Each row in the code corresponds to a node encoding for a time step, which integrates both local temporal neighborhood changes and contextual information from distant time nodes in the same stage.

[0148] In the actual implementation, the node is assigned attention weights based on the calculated neighbor attention weights. All neighboring nodes (including node) The message mapping features of the node itself are weighted and summed to obtain an aggregated feature vector; then, this aggregated feature vector is input into a non-linear activation function (such as ELU or ReLU) to obtain the updated node features. ,in This represents the layer index of the current graph attention encoding layer; after multiple layers of graph attention encoding, the dynamic graph node encoding matrix is ​​obtained. Each row corresponds to a node encoding for a time step, which integrates contextual information from its local time neighborhood and distant time nodes in the same stage.

[0149] In one embodiment, as an example, suppose the current coding layer The central node is at time step 80. Its neighboring nodes include itself, time steps 78, 79, 81, and 82 (local time edges), and time steps 120 and 150 (stage-related edges, assumed to belong to the same fault duration). Assuming the attention weights calculated for these neighboring nodes are 0.1, 0.05, 0.15, 0.4, 0.2, 0.05, and 0.05 respectively, when updating node features, the message mapping feature of each neighboring node is multiplied by its corresponding weight. Then, all these weighted feature vectors are summed to obtain an aggregate vector. This aggregate vector is then subjected to ELU non-linear activation to finally obtain the value at time step 80. Layer output features It is evident that even though step 120 is relatively far away, its information can still be aggregated and participate in the updating of the central node's features because it belongs to the same fault duration period.

[0150] It should be noted that graph attention message passing is a trainable relation selection mechanism that can adaptively highlight key neighbor nodes within a given graph connection range. Compared with simple average aggregation, this method is more suitable for the data characteristics of "few key boundaries and many stable intervals" in underpressure faults.

[0151] S3.3 Global Event Semantic Compression Global event semantic vector It simultaneously includes stage statistics, static transient descriptive quantities, and consistency verification quantities. These features come from different sources and have different scales. If directly involved in subsequent fusion, they might lead to certain features with large numerical ranges dominating model learning. This invention, through multi-layer fully connected mapping and layer normalization, compresses the global event semantic vector into a fixed-dimensional global event semantic summary. The specific steps are as follows: 1) Transfer the global event semantic vector The first fully connected layer is input to obtain intermediate semantic features. The first fully connected layer is used to extend the original event-level semantic features to a higher-dimensional representation space to enhance the ability to express combinations of different physical quantities.

[0152] In one implementation, when Dimensions Hidden feature dimensions At that time, the first layer fully connected mapping can dimensional input mapping is Intermediate semantic features. Specifically, the first fully connected layer consists of a weight matrix and a bias vector, without an activation function, and its weight matrix has a dimension of [dimensionality missing]. ,in , .

[0153] 2) Perform nonlinear activation on the intermediate semantic features to obtain a nonlinear event semantic representation with the same dimensionality as the intermediate semantic features, which is 128 dimensions.

[0154] In practical implementation, either the GeLU activation function or the ReLU activation function can be used; nonlinear activation is used to enhance the combined modeling capability between event phase features, static sag descriptors, and consistency deviations.

[0155] 3) Input the semantic representation of the nonlinear event into the second fully connected layer and compress it to the hidden feature dimension. The compressed global semantic features are obtained. This compression process is used to ensure that the global event semantics and the dynamic graph node encoding are in the same latent space dimension, which facilitates subsequent interaction.

[0156] In one implementation, the second fully connected layer consists of a weight matrix and a bias vector, without an activation function. The dimension of its weight matrix is... ,in , It is used to compress 128-dimensional nonlinear event semantic representations into 64-dimensional global semantic features.

[0157] 4) Perform layer normalization on the compressed global semantic features to obtain a global event semantic summary. , This refers to the compressed expression of the overall physical semantics of an event, with a dimension of It is used as a query vector in the subsequent key time step filtering process.

[0158] In one implementation, when At that time, " can be adopted The mapping structure of ""; if the global event semantic vector simultaneously exhibits a large depth consistency coefficient. and a higher high-resistance anomaly indication coefficient ,but During the training process, semantic directions related to atypical pressure loss anomalies will gradually be formed, providing guidance for subsequent key time step screening.

[0159] S3.4, Key Time Step Aggregation Guided by Event Semantics Dynamic graph node encoding matrix It includes latent features across all time steps, but the contribution of different time steps to undervoltage fault classification varies. Fault trigger boundaries, low-voltage extreme ranges, current surge ranges, and fault recovery boundaries typically have stronger discriminative power. This invention utilizes global event semantic summarization. For querying, the encoding matrix of dynamic graph nodes. The key time steps of the execution event semantic guidance are aggregated to obtain the event guidance time sequence aggregation vector. The specific steps are as follows: 1) Using global event semantic summaries As a query vector, and for encoding the nodes of the dynamic graph. Each time point in the encoding is used to calculate a semantic matching score, which measures the degree of correlation between the dynamic features of that time step and the overall physical semantics of the current event.

[0160] In one implementation, the semantic matching score can be calculated using a scaled dot product method, that is... After performing an inner product with the encoding of each time node, the model is then scaled according to the hidden dimension. This method is simple to compute, stable to train, and suitable for retrieving time step features using event-level semantics.

[0161] In the specific implementation, the query vector (dimension is) ) and dynamic graph node encoding matrix (dimension is) Perform matrix multiplication to obtain a matrix of length . A vector, where each element corresponds to an initial matching score at a time step; then, each element of this vector is divided by... Scaling is performed to obtain the semantic matching score at each time step.

[0162] In one embodiment, for example, if Then the scaling factor Assuming If the inner product of the encoding of a node at a certain time step is 6.4, then the semantic matching score of that time step is... .

[0163] 2) Perform Softmax normalization on the semantic matching scores for all time steps to obtain the event semantic attention weights. , It refers to the first The importance of each time step to the current fault classification task is determined by the sum of the weights of all time steps. When the node encoding at a certain time step matches the current global event semantic summary better, its A higher value indicates that the time step contains more valuable classification evidence.

[0164] Furthermore, based on event semantic attention weights By weighting and summing the codes of all time nodes, an event-guided time-series aggregation vector is obtained. , This refers to the key dynamic process expression after global semantic filtering, with dimensions of .

[0165] In one embodiment, as an example, assume the total number of time steps in the observation window The calculated semantic matching score vector is After Softmax normalization, the event semantic attention weights are obtained. Approximately This result indicates that the fourth time step has the highest weight (0.49), and the model considers it to be the most important for the current classification task. Furthermore, the dynamic graph node encoding matrix... (dimension) The node encodings at time steps 1 to 5 in the process are multiplied by their corresponding attention weights. to Then, these five weighted 64-dimensional vectors are added element by element to obtain a final 64-dimensional event-guided timing aggregation vector. This vector focuses on expressing the dynamic features of the key time steps with the highest weights (such as step 4).

[0166] It should be noted that this step realizes "event-level semantic-guided dynamic process retrieval". Compared with directly performing average pooling on all time steps, this invention can actively filter more discriminative local dynamic evidence by combining the global physical background, thereby improving the ability to distinguish between similar underpressure faults.

[0167] S3.5, Dual-stream Feature Fusion and Underpressure Fault Category Prediction Event-guided timing aggregation vector Used to express key dynamic processes, global event semantic summary Used to express the overall physical meaning of the event, both support the classification of pressure loss faults from the two levels of "local dynamic evidence" and "global event evidence," respectively. This invention uses event-guided time-series aggregation vectors. Global event semantic summary The fusion process outputs the fault category prediction results, and the specific steps are as follows: 1) Generate event-guided time-series aggregation vectors Global event semantic summary Concatenate the features along their respective dimensions to obtain the fused classification feature vector. , Dimensions It is used to simultaneously express key time process features and event-level global semantic features.

[0168] Furthermore, the classification feature vectors will be fused. Input the classification fully connected layer to obtain the fault category score vector. , Each element in the table corresponds to an unnormalized score for a type of underpressure fault.

[0169] In one implementation, the classification fully connected layer is implemented by a single fully connected network with an input dimension of . The output dimension is the total number of underpressure fault categories. (For example, 6), this layer contains a dimension of The weight matrix and a dimension of The bias vector does not contain a non-linear activation function.

[0170] 2) Score vector for fault categories Performing Softmax normalization yields the fault category prediction probability vector. , Each component in the equation represents the predicted probability that the current voltage sag event sample belongs to the corresponding undervoltage fault category. The sum of the predicted probabilities for all categories is... .

[0171] In practical implementation, the model uses the category with the highest predicted probability as the predicted result for the current sample's underpressure fault. For example, the number of underpressure fault categories is set to... The following conditions are listed in sequence: normal state, single-phase voltage loss, two-phase voltage loss, three-phase voltage loss, voltage loss due to internal wire breakage in the metering box, and abnormal voltage loss of the current transformer; if If the location corresponding to the internal disconnection and pressure loss in the metering box is the most probable, then the current sample is determined to be due to an internal disconnection and pressure loss in the metering box.

[0172] S3.6 Multi-class supervised training and model parameter determination To enable the model to jointly learn dynamic temporal evolution, global event semantics, and the relationship between the two, this invention employs a multi-class supervised training method to perform end-to-end optimization of the stage-aware two-stream interaction graph encoding model. The specific steps are as follows: 1) Construct the training input. Specifically, each training sample includes a dynamic temporal representation matrix. Global event semantic vector And actual fault category labels; actual fault category labels can be represented using one-hot encoding, if the total number of categories is Then the length of each sample is The label vector, with only the true category position taking the value. The values ​​for the remaining positions are... .

[0173] 2) Input the training samples into the stage-aware dual-stream interactive graph encoding model in sequence. After stage-aware dynamic graph construction, graph attention message passing, global event semantic compression, key time step aggregation guided by event semantics, and dual-stream feature fusion, the corresponding fault category prediction probability vector is obtained.

[0174] 3) Calculate the multi-class cross-entropy loss based on the predicted probability vector and the real fault category label, and update the trainable parameters in the input mapping layer, graph attention encoding layer, global event semantic compression layer, event semantic guidance aggregation module and classification layer through backpropagation.

[0175] It should be noted that the dynamic time series representation matrix in S2 and global event semantic vector The construction process does not depend on training parameters and belongs to deterministic re-representation based on electrophysical knowledge; the stage-aware two-stream interactive graph coding model in S3 belongs to trainable classification modules.

[0176] In one implementation, the Adam optimizer is used for model training, and the initial learning rate can be set to... The training batch size can be selected as follows: During training, the learning rate and number of training epochs can be adjusted based on the validation set classification accuracy, macro-average F1 score, or loss degradation. For example, when the validation set loss is continuous... If the learning rate does not decrease, reduce it to its original value. times.

[0177] 4) After each training cycle, the model performance is evaluated using the validation set, and the model parameters corresponding to the highest macro-average F1 value are preferably saved. The macro-average F1 value can reflect the recognition effect of different fault categories at the same time, avoiding the model from performing well only in categories with a large number of samples, while ignoring the underpressure fault category with a small number of samples.

[0178] Through the above training process, the stage-aware dynamic graph coding module can learn the temporal correlation between different stages of the fault, the global event semantic compression module can extract the physical information of temporary events, and the event semantic guided aggregation module can adaptively select key time steps according to global semantics, ultimately forming a stage-aware dual-stream interactive graph coding classification model for undervoltage fault identification.

[0179] It should be noted that the underpressure fault classification model adopts a stage-aware two-stream interactive graph coding model, the overall structure of which is as follows: a) Dual-stream input: Receives dynamic temporal representation matrix (temporal stream) and global event semantic vector (event stream).

[0180] b) Timing stream processing (dynamic graph encoder): b-1) Graph construction: Using time steps as nodes, construct a graph structure that integrates "time proximity edges" and "stage association edges"; b-2) Graph Coding: Nodes are encoded through a multi-layer graph attention network, and the output is a dynamic graph node encoding matrix that integrates local and stage context information.

[0181] c) Event Stream Processing (Semantic Compressor): Compresses the global event semantic vector into a fixed-dimensional summary vector through a multi-layer fully connected network, which serves as the query vector.

[0182] d) Dual-stream interaction and fusion: d-1) Event-guided aggregation: Using the summary vector output by the event stream as the query, an attention mechanism is executed on all nodes of the time-series stream output to generate a weighted aggregation of key dynamic process vectors; d-2 feature fusion: concatenates the aggregated key dynamic process vector with the global event semantic summary vector.

[0183] e) Classification output: Input the fused vector into a fully connected classification layer and output the predicted probability of each category.

[0184] In a specific implementation, the S4 application process for identifying underpressure faults in online monitoring scenarios is as follows: After model training is complete, the trained undervoltage fault classification model is deployed to power quality monitoring terminals, distribution-side edge computing devices, or centralized fault diagnosis platforms. During online operation, the system continuously receives three-phase voltage, three-phase current, and event-level sag descriptions, and outputs undervoltage fault identification results according to the same feature re-representation and classification inference processes as during the training phase. The specific steps are as follows: 1) Extract the observation window corresponding to the current voltage sag event from the online monitoring data stream to obtain the voltage sag event sample to be diagnosed. Sample of voltage sag events to be diagnosed It includes dynamic electrical quantity sequences and static sag description quantities. The dynamic electrical quantity sequences include the three-phase voltage RMS value sequence and the three-phase current RMS value sequence; the static sag description quantities include voltage sag depth, sag duration, and sag initiation point phase jump angle.

[0185] 2) Sample of voltage sag events to be diagnosed The physical re-representation process is executed in a manner completely consistent with the training phase. Specifically, event phase segmentation is performed first to obtain the pre-fault phase, fault duration phase, and fault recovery phase; then, event phase feature vectors are extracted. Then construct the physical enhancement timing matrix. Physical consistency verification vector Finally, output the dynamic time-series representation matrix of the target to be diagnosed. and the semantic vector of the global event to be diagnosed .

[0186] It should be noted that the standardized statistics saved during the training phase should be used during the online inference phase. Scaling should be performed, and the mean or standard deviation should not be re-estimated based on a single sample to be diagnosed, otherwise the feature scaling relationship formed during the training phase will be destroyed, affecting the stability of diagnosis.

[0187] 3) The dynamic time-series representation matrix to be diagnosed and the semantic vector of the global event to be diagnosed Input the trained undervoltage fault classification model to obtain the predicted probability vector of the fault category to be diagnosed. , Used to describe the probability distribution of the current voltage sag event belonging to each undervoltage fault category.

[0188] Furthermore, the probability vector is predicted based on the category of the fault to be diagnosed. Output the fault identification result for the current event. If the maximum predicted probability is not less than the confidence threshold... If the maximum predicted probability is less than the confidence threshold, then the fault category corresponding to the maximum predicted probability will be output; if the maximum predicted probability is less than the confidence threshold, then the fault category corresponding to the maximum predicted probability will be output. If the event is not found, a low confidence level message will be output, and the event will be marked as an event to be reviewed. Confidence threshold This represents the minimum probability threshold for a classification result to be accepted. An example value is... to .

[0189] In one embodiment, for example, if ,and Then the maximum prediction probability is If the fault is due to a broken wire and loss of pressure inside the metering box, the system will output this fault category. Then the maximum prediction probability is If the value is below the confidence threshold, the system outputs a low confidence warning.

[0190] 4) Perform time consistency verification on the output results of multiple consecutive temporary drops or multiple consecutive diagnostic windows. The number of consecutive verification windows is denoted as... , This refers to the number of windows that need to continuously exhibit the same fault category. An example value is... to When the same undervoltage fault category occurs consecutively All windows contain the highest probability category, and the corresponding highest predicted probability is not less than the confidence threshold. If the fault category changes frequently in adjacent windows, an alarm will be output; if the fault category changes frequently in adjacent windows, the alarm will not be triggered for the time being, and the subsequent monitoring results will continue to be observed.

[0191] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for classifying and diagnosing voltage sag faults in energy metering boxes, characterized by: Includes the following steps: S1. Construct voltage sag event samples. Collect three-phase voltage and three-phase current at the power supply monitoring end to form a dynamic electrical quantity sequence. Extract static sag descriptive quantities for each voltage sag time sample and configure real fault category labels. A dataset is composed of multiple samples. S2. Based on the stage evolution law of voltage sag events, the mechanism of symmetrical components of three-phase systems, and the consistency relationship between dynamic measurement process and static sag descriptor, physical re-representation without training parameters is performed on the samples in the dataset to obtain dynamic time series representation matrix and global event semantic vector. S3. Construct a phase-aware dual-stream interactive graph coding model, use dynamic temporal representation matrix to model the pressure loss fault process, use global event semantic vector to model event-level physical priors, and use the event semantic-guided key time step aggregation mechanism to filter the most discriminative temporal evidence to predict the pressure loss fault category. The model is trained under multi-class supervision, and the optimal parameters of the model are determined to obtain the trained model. S4. Deploy the trained model to power quality monitoring terminals, distribution-side edge computing devices, or centralized fault diagnosis platforms to identify undervoltage faults.

2. The method for classifying and diagnosing voltage sag faults in power metering boxes according to claim 1, characterized in that, S1 is as follows: The voltage sag event sample includes dynamic electrical quantity sequences, static sag descriptive quantities, and actual fault category labels; The dynamic electrical quantity sequence is a sequence of three-phase voltage RMS values ​​collected and arranged in chronological order within the observation window. , , and three-phase current RMS value sequence , , Let T represent the total number of time steps in a single observation window, and t represent the index of T. , , These represent the effective voltage sequences of phases A, B, and C at the t-th time step within the observation window, respectively. , , These represent the effective current sequences of phases A, B, and C at the t-th time step of the observation window, respectively. Static sag descriptors include the depth of voltage sag. Duration of temporary descent Phase jump angle of the descent initiation point The static sag descriptor is directly output by the power quality monitoring device, or obtained by offline calculation based on the three-phase voltage waveform, RMS curve and fundamental phase angle change; The actual fault category labels include normal state, single-phase voltage loss, two-phase voltage loss, three-phase voltage loss, voltage loss due to internal wire breakage in the metering box, and abnormal voltage loss of the current transformer.

3. The method for classifying and diagnosing power metering box undervoltage faults based on integrated voltage sag characteristics as described in claim 1, characterized in that, S2 Specifically as follows: S2.

1. Adaptively determine the initial judgment threshold by using the voltage sag depth and the phase jump angle of the sag initiation point in the voltage sag event sample, and use the recovery judgment threshold higher than the initial judgment threshold to form the hysteresis interval, thereby dividing the entire observation window into the pre-fault period, the fault duration period and the fault recovery period. S2.

2. Based on the stage statistics and rate of change features extracted during the pre-fault, fault duration, and fault recovery periods, an event stage feature vector is constructed. S2.3 Map the three-phase voltage and three-phase current to the positive sequence, negative sequence and zero sequence component spaces to construct the voltage imbalance sequence and the current imbalance sequence, and then construct the physical enhancement time sequence feature vector. Stack the physical enhancement time sequence feature vectors of all time steps in time order to obtain the physical enhancement time sequence matrix. S2.4 Calculate the depth consistency coefficient, duration consistency coefficient, phase consistency coefficient and high impedance anomaly indicator coefficient respectively, and construct the physical consistency verification vector; S2.5 Construct an event stage indicator matrix based on the current stage, and concatenate it with the physical enhancement time series matrix to obtain a dynamic time series representation matrix; concatenate the event stage feature vector, static temporary sag descriptor, and physical consistency check vector to obtain a global event semantic vector, and perform unified processing on its execution scale.

4. The method for classifying and diagnosing voltage sag faults in energy metering boxes according to claim 2, characterized in that, S2.1 is as follows: Obtain the three-phase voltage RMS value sequence from the current voltage sag event sample. If the three-phase voltage RMS values ​​have not yet been converted to per-unit values, normalize them first using the rated voltage as a reference. Then, read the voltage RMS values ​​of the three phases at each time step and select the minimum value to obtain the minimum phase voltage sequence. Determine the sag initiation judgment threshold based on the voltage sag depth and the phase jump angle at the sag initiation point in the static sag descriptor, and determine the sag recovery judgment threshold based on the voltage sag depth. Then, scan the minimum phase voltage sequence along the time sequence to determine the sag event initiation time step. Continue scanning the minimum phase voltage sequence after the sag event initiation time step to determine the sag event end time step. Finally, based on the sag event initiation time step and the sag event end time step, divide the entire observation window into the pre-fault period, the fault duration period, and the fault recovery period.

5. The method for classifying and diagnosing voltage sag faults in energy metering boxes according to claim 2, characterized in that, S2.2 is as follows: During the early stage of the fault, the average level and standard deviation of the effective values ​​of phase A voltage, phase B voltage, and phase C voltage are statistically analyzed to obtain the average voltage characteristics and voltage fluctuation characteristics during the early stage of the fault. Next, a local time neighborhood is selected around the start time step of the transient event, and linear fitting is performed on the three-phase voltage RMS value sequence to obtain the voltage drop rate feature. Then, during the fault duration, the dispersion between the three-phase current RMS values ​​is statistically analyzed to obtain the fault duration current dispersion feature. Then, a local time neighborhood is selected around the end time step of the transient event, and linear fitting is performed on the three-phase voltage RMS value sequence to obtain the voltage recovery rate feature. Finally, the average voltage feature in the early stage of the fault, the voltage fluctuation feature in the early stage of the fault, the voltage drop rate feature, the fault duration current dispersion feature, and the voltage recovery rate feature are concatenated in a fixed order to obtain the event stage feature vector.

6. The method for classifying and diagnosing voltage sag faults in energy metering boxes according to claim 2, characterized in that, S2.3 is as follows: At each time step, the effective values ​​of the three-phase voltage and the effective values ​​of the three-phase current are read to form the original three-phase voltage amplitude vector and the original three-phase current amplitude vector. The fundamental phasors of the three-phase voltage and the three-phase current corresponding to each time step are obtained. Then, Fortescue symmetrical component transformation is performed on the three-phase voltage fundamental phasors to obtain the zero-sequence voltage phasor, the positive-sequence voltage phasor, and the negative-sequence voltage phasor. Then, the amplitudes of the zero-sequence voltage phasor, the positive-sequence voltage phasor, and the negative-sequence voltage phasor are respectively taken to obtain the zero-sequence voltage component, the positive-sequence voltage component, and the negative-sequence voltage component. Then, in the same way as the voltage, symmetrical component transformation is performed on the three-phase current fundamental phasors to obtain the zero-sequence current component, the positive-sequence current component, and the negative-sequence current component. Based on the positive-sequence voltage component and the negative-sequence voltage component, the voltage imbalance is calculated to quantify the proportion of the reverse-sequence voltage disturbance relative to the normal positive-sequence component at the current time step. At the same time, based on the positive-sequence current component and the negative-sequence current component, the current imbalance is calculated to quantify the proportion of the reverse-sequence current disturbance relative to the normal positive-sequence component at the current time step. Finally, the original three-phase voltage amplitude vector, the original three-phase current amplitude vector, the voltage sequence component, the current sequence component, the voltage imbalance, and the current imbalance are combined in a fixed order to obtain the physical enhancement time series feature vector for each time step. The physical enhancement time series feature vectors corresponding to all time steps are stacked in time order to obtain the physical enhancement time series matrix.

7. The method for classifying and diagnosing voltage sag faults in energy metering boxes according to claim 2, characterized in that, S2.4 is as follows: During the fault duration, the minimum phase voltage sequence is read to determine the lowest voltage level obtained by dynamic measurement. The theoretical lowest voltage level is calculated based on the voltage sag depth in the static sag descriptor and compared with the low voltage level to obtain the depth consistency coefficient. Based on the start time step, end time step and sampling time interval of the transient landing event, the duration of the transient landing is calculated by dynamic measurement. The duration of the transient landing is then compared with the duration of the transient landing in the static transient landing descriptor to obtain the duration consistency coefficient. Around the start time step of the transient landing event, select the phase statistics window before the event and the phase statistics window after the event respectively, calculate the actual phase change, and then compare the actual phase change with the phase jump angle of the transient landing start point in the static transient landing descriptor to obtain the phase consistency coefficient. Based on the zero-sequence components of current and voltage, a zero-sequence normalized admittance sequence is constructed. The zero-sequence normalized admittance sequence is averaged during the early stage of the fault and the fault duration, respectively, to obtain the mean zero-sequence normalized admittance during the early stage of the fault and the mean zero-sequence normalized admittance during the fault duration. The absolute difference between the two is then calculated to obtain the high-resistance anomaly indication coefficient. Finally, the depth consistency coefficient, duration consistency coefficient, phase consistency coefficient, and high impedance anomaly indicator coefficient are combined in a fixed order to obtain the physical consistency verification vector.

8. The method for classifying and diagnosing voltage sag faults in energy metering boxes according to claim 2, characterized in that, S2.5 is as follows: Based on the pre-fault phase, fault duration, and fault recovery phase, an event phase indicator vector is constructed for each time step. The event phase indicator vectors corresponding to all time steps are stacked in chronological order to obtain the event phase indicator matrix. Then, the physical augmentation time series matrix and the event phase indicator matrix are concatenated according to the feature dimension to obtain the dynamic time series representation matrix. At the same time, the event phase feature vector, static sag descriptor, and physical consistency check vector are concatenated in a fixed order to obtain the global event semantic vector. Before entering the subsequent classification model, the continuous numerical features in the global event semantic vector are subjected to scale unification processing.

9. The method for classifying and diagnosing voltage sag faults in power metering boxes according to claim 1, characterized in that, S3 is as follows: The phase-aware dual-stream interaction graph coding model includes a dual-stream input module, a temporal stream processing module, an event stream processing module, a dual-stream interaction and fusion module, and a classification output module. The dual-stream input module receives a dynamic temporal representation matrix and a global event semantic vector. It projects the dynamic temporal representation matrix onto a unified hidden dimension through an input linear mapping layer to obtain the initial node features of each time node. For the global event semantic vector, it determines the event stage category to which each time node belongs. Then, it establishes local temporal edges based on temporal proximity and stage association edges based on event stage consistency. It merges the local temporal edges and stage association edges to construct the connection relationship of the stage-aware dynamic graph. The time step is used as a node by the dynamic graph encoder in the time-series stream processing module to construct a graph structure that integrates time proximity edges and stage association edges. Then, the nodes are encoded by a multi-layer graph attention network to output a dynamic graph node encoding matrix that integrates local and stage context information. The semantic compressor in the event stream processing module compresses the global event semantic vector into a fixed-dimensional summary vector, which serves as the query vector. The dual-stream interaction and fusion module first uses the summary vector output by the event stream as a query, performs an attention mechanism on all nodes of the time-series output, and weights and aggregates the key dynamic process vectors. Then, it concatenates the aggregated key dynamic process vectors with the global event semantic summary vector to obtain the fused vector. The classification output module inputs the fused vector into a fully connected classification layer and outputs the predicted probability of each category.

10. The method for classifying and diagnosing voltage sag faults in energy metering boxes according to claim 9, characterized in that, The specific steps for model training in S3 are as follows: A multi-class supervised training method is used to optimize the stage-aware two-stream interactive graph encoding model end-to-end. The training samples are sequentially input into the stage-aware two-stream interactive graph encoding model. After stage-aware dynamic graph construction, graph attention message passing, global event semantic compression, key time step aggregation guided by event semantics, and two-stream feature fusion, the corresponding fault category prediction probability vector is obtained. Multi-class cross-entropy loss is calculated based on the predicted probability vector and the real fault category label, and the trainable parameters in the input mapping layer, graph attention encoding layer, global event semantic compression layer, event semantic guided aggregation module and classification layer are updated through backpropagation. The Adam optimizer is used for model training. During the training process, the learning rate and the number of training epochs are adjusted according to the classification accuracy on the validation set, the macro average F1 value, or the decrease in loss. The model parameters corresponding to the highest macro average F1 value are selected and saved to obtain the trained model.

Images