AI-based substation hidden danger prediction method and system

Abstract

This invention relates to the field of power system safety monitoring technology, specifically to an AI-based method and system for predicting substation hazards. The method includes: preprocessing multi-source electrical signals to form rhythmic anomaly feature signals; constructing a rhythmic disturbance time-frequency phase curve, dividing the curve into anomaly subspace and load disturbance subspace using a subspace feature splitting network, and outputting a rhythmic disturbance time-frequency phase feature expression vector; constructing a rhythmic phase migration map, and using a minimum energy drift path algorithm combined with primary equipment spatial layout mapping rules to determine the source location of substation hazards, outputting a hazard source probability vector and rhythmic disturbance spatial location results; and using a risk evolution model to calculate hazard risk values ​​and outputting hazard warning signals. This invention achieves accurate identification of rhythmic chronic hazards in substation primary equipment through rhythmic disturbance feature reconstruction and graph topology energy path analysis.

Description

Technical Field

[0001] This invention relates to the field of power system safety monitoring technology, specifically to an AI-based method and system for predicting potential hazards in substations. Background Technology

[0002] With the continuous expansion of power system scale and the improvement of substation automation level, the coupling relationship between primary and secondary equipment is becoming increasingly complex. The substation operating status is affected by multiple factors such as environment, load fluctuation, equipment aging, and electromagnetic disturbances. The occurrence of hidden risks is random and multi-source. The existing substation monitoring system mainly relies on the instantaneous alarm and static threshold judgment mechanism of protection devices, monitoring systems, and status quantity acquisition terminals. It is difficult to identify potential disturbance hazards with stable amplitude, slow change and rhythmic characteristics in a timely manner. Especially in multi-node high-frequency sampling scenarios, disturbance characteristics are often submerged in normal operating signals, causing early hidden dangers to go undetected or be misjudged as background fluctuations, thus reducing the long-term stability of the overall substation operation.

[0003] In actual operation, primary equipment may experience chronic, rhythmic disturbances under specific operating conditions or periodic external environmental influences. For example, busbar areas, circuit breaker joints, or switch contacts may experience rhythmic fluctuations in contact state due to long-term, minute thermal expansion and contraction or periodic electromagnetic stress. The amplitude of these disturbances is usually low and does not exceed the operating threshold of secondary protection and monitoring devices, thus not triggering any alarms. However, these disturbances are stably reflected in the signals acquired by secondary equipment in the form of electromagnetic coupling, manifesting as a regular increase in harmonic energy at fixed time intervals, boundary jitter in zero-sequence or negative-sequence components, rhythmic harmonic spikes and phase drift detected by the waveform recording system, and amplitude shifts in the SCADA system. Existing monitoring mechanisms typically classify such phenomena as load fluctuations or acquisition errors, failing to associate them with potential hazards in primary equipment. This results in delayed hazard identification, hindering early prediction and spatial location of hazards, and increasing potential risks in substation operation. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based method and system for predicting potential hazards in substations, in order to solve the problem mentioned in the background art that primary equipment may experience chronic hazard disturbances with rhythmic characteristics under specific operating conditions or specific periodic external environmental influences.

[0005] To achieve the above objectives, the technical solution of the present invention is: an AI-based substation hazard prediction method, comprising:

[0006] S1. Obtain multi-source electrical signals through substation monitoring devices, use time synchronization identifiers to reconstruct the multi-source electrical signals to a unified time axis, and preprocess the multi-source electrical signals to form rhythmic abnormality characteristic signals;

[0007] S2. Perform phase expansion on the rhythm abnormality feature signal to construct the rhythm disturbance time-frequency phase curve. Use the subspace feature splitting network to divide all rhythm disturbance time-frequency phase curves into abnormal disturbance subspace and load disturbance subspace. Calculate the rhythm disturbance phase consistency index and phase coherence distance. When the phase coherence distance is greater than the preset distance threshold, output the rhythm disturbance time-frequency phase feature expression vector.

[0008] S3. Construct a rhythmic phase migration map based on the rhythmic disturbance time-frequency phase feature expression vector, and use the minimum energy drift path algorithm to determine the rhythmic disturbance propagation path and source direction. Combine the spatial layout mapping rules of primary equipment to determine the source location of substation hazards, and output the hazard source probability vector and rhythmic disturbance spatial location results.

[0009] S4. Based on the rhythmic disturbance time-frequency phase feature expression vector and the hazard source probability vector, the hazard risk value is calculated using the risk evolution model, and the hazard warning signal is output.

[0010] Preferably, in step S1, the process of reconstructing and mapping the time-scale of multi-source electrical signals to a unified time axis using time synchronization identifiers includes: synchronizing and calibrating the sampling time scales of the multi-source electrical signals based on a unified time base; and aligning the time scales of all multi-source electrical signals on the same time axis by interpolating and resampling the sampling time scales of the multi-source electrical signals and compensating for time drift.

[0011] The rhythmic anomaly characteristic signal refers to the set of electrical signals used to characterize rhythmic disturbance behavior after time-scale reconstruction and unified time axis mapping are completed and preprocessed. The rhythmic anomaly characteristic signal includes the rhythmic response components of the electrical signals output by each monitoring device under the unified time axis.

[0012] Preferably, in step S2, performing phase expansion on the rhythm abnormality feature signal to construct a rhythm disturbance time-frequency phase curve specifically includes: performing a Hilbert-Huang transform on the rhythm abnormality feature signal, decomposing the rhythm abnormality feature signal into several intrinsic mode function components, extracting instantaneous phase information for each intrinsic mode function component, expanding the instantaneous phase information on a unified time axis to form a phase trajectory set, and constructing a rhythm disturbance time-frequency phase curve corresponding to each intrinsic mode function component;

[0013] The rhythmic perturbation time-frequency phase curve refers to the curve formed by the phase trajectory extracted from the multi-source rhythmic abnormality characteristic signal under a unified time axis through the Hilbert-Huang transform, which is used to characterize the dynamic change characteristics of the rhythmic perturbation signal in the time domain and phase domain.

[0014] Preferably, in S2, the subspace feature splitting network is a neural network structure constructed based on the high-dimensional embedded feature representation of the rhythmic perturbation time-frequency phase curve and the rhythmic phase coherence constraint loss function, including a feature embedding layer, a feature splitting layer and a discriminant output layer, used to divide the rhythmic perturbation time-frequency phase curve into feature spaces;

[0015] The subspace feature splitting network divides all rhythmic disturbance time-frequency phase curves into anomaly disturbance subspace and load disturbance subspace. Specifically, it includes: inputting the phase trajectory set into the feature embedding layer to extract the joint time-phase feature vector; inputting the joint time-phase feature vector into the feature splitting layer; aggregating and decoupling the joint time-phase feature vector by minimizing the rhythmic phase coherence constraint loss function to obtain the feature space partitioning result; and using the discriminant output layer to determine the convergence of the feature space partitioning result and output the subspace label.

[0016] The subspace labels include anomaly disturbance subspace and load disturbance subspace; the anomaly disturbance subspace refers to the set of phase trajectories with high rhythmic phase consistency and high coherence weight in the rhythmic disturbance time-frequency phase curve cluster; the load disturbance subspace refers to the set of phase trajectories that do not have rhythmic coherence characteristics.

[0017] Preferably, in S2, the rhythm disturbance phase consistency index is a statistic calculated based on the phase synchronization degree of the set of phase trajectories of the rhythm disturbance time-frequency phase curve under a unified time axis; the phase coherence distance is a feature space distance metric calculated based on the coherence similarity of the feature vectors between the rhythm disturbance phase consistency index and the phase trajectory, used to distinguish between abnormal disturbances and load disturbances.

[0018] The rhythmic perturbation time-frequency phase feature representation vector refers to the high-dimensional phase-time joint feature representation extracted from the embedded feature space when the phase coherence distance is greater than a preset distance threshold in the subspace feature splitting network.

[0019] Preferably, in S3, the method for constructing the rhythmic phase migration map based on the rhythmic perturbation time-frequency phase feature expression vector is as follows: taking the monitoring device acquisition point as the node, the phase migration direction and phase delay as the direction attribute of the edge, and the phase drift amount as the edge weight, the phase drift and delay relationship between adjacent acquisition points is calculated based on the rhythmic perturbation time-frequency phase feature expression vector to construct the rhythmic phase migration map;

[0020] The rhythm phase transition graph has a time-series directed graph structure, which is used to characterize the propagation path, direction and time delay of rhythm disturbances in substations.

[0021] Preferably, in S3, the minimum energy drift path algorithm is a path search and energy accumulation optimization algorithm based on rhythm phase migration graph and edge weight phase drift amount, used to find the propagation path with the minimum global energy dissipation in rhythm phase migration graph and determine the source direction of rhythm disturbance;

[0022] The process of determining the propagation path and source direction of rhythmic disturbance using the minimum energy drift path algorithm includes: calculating the phase drift cumulative energy of each node based on the directed edge set of the rhythmic phase transition graph; constructing a directed energy graph with phase drift as the weight; performing a minimum energy path search starting from the set of nodes with the smallest global in-degree in the energy graph to obtain the minimum energy path set from the source node to each affected node; determining the source direction of the rhythmic disturbance and outputting the corresponding source node and path structure based on the minimum energy path and energy cumulative gradient in the path set.

[0023] Preferably, in S3, the spatial layout mapping rule for primary equipment is a spatial mapping rule constructed based on the spatial location coordinate set, connection topology, and operational functional area division of the primary equipment in the substation. It is used to map and associate the rhythmic disturbance propagation path with the physical location and functional area of ​​the primary equipment.

[0024] The process involves determining the location of potential hazards in a substation and outputting a hazard source probability vector and the spatial location result of rhythmic disturbances. Specifically, this includes: matching the physical equipment location in the spatial layout mapping rules based on the starting node of the minimum energy propagation path in the rhythmic phase migration diagram and its associated primary equipment spatial coordinates; calculating the hazard source attribution probability for the equipment areas involved in the disturbance path to form a hazard source probability vector; and outputting the spatial location result of rhythmic disturbances based on the spatial mapping relationship between the hazard source probability vector and the primary equipment layout.

[0025] Preferably, in step S4, the risk evolution model is a time-series risk assessment model for hidden dangers constructed based on the rhythmic disturbance time-frequency phase feature expression vector, the probability vector of hidden danger sources, and the time-series disturbance response index. It is used to calculate the hidden danger risk value of rhythmic disturbance hidden dangers in the time dimension and output a hidden danger warning signal.

[0026] On the other hand, the present invention provides an AI-based substation hazard prediction system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the aforementioned AI-based substation hazard prediction method.

[0027] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects:

[0028] 1. In this invention, the joint modeling method based on the reconstruction of rhythmic disturbance time-frequency phase features and the subspace feature splitting network can accurately identify the hidden disturbance features of primary equipment in substations under rhythmic chronic disturbance conditions and effectively distinguish them from load disturbance features, breaking through the limitation of traditional monitoring mechanisms that rely on static thresholds and cannot identify low-amplitude rhythmic disturbances.

[0029] 2. In this invention, by constructing a rhythmic phase migration map, executing the minimum energy drift path algorithm, and combining it with the spatial layout mapping rules of primary equipment, the propagation path of rhythmic disturbance hazards is traced and the source direction is inverted. Then, the hazard source probability vector and spatial positioning results are output, providing a structured decision-making basis for early prediction of hazards and operation and maintenance intervention. Attached Figure Description

[0030] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0031] Example 1, as Figure 1 As shown, the specific implementation steps of the AI-based substation hazard prediction method proposed in this invention are as follows:

[0032] S1. Obtain multi-source electrical signals through substation monitoring devices, use time synchronization identifiers to reconstruct the multi-source electrical signals to a unified time axis, and preprocess the multi-source electrical signals to form rhythmic abnormality characteristic signals;

[0033] S2. Perform phase expansion on the rhythm abnormality feature signal to construct the rhythm disturbance time-frequency phase curve. Use the subspace feature splitting network to divide all rhythm disturbance time-frequency phase curves into abnormal disturbance subspace and load disturbance subspace. Calculate the rhythm disturbance phase consistency index and phase coherence distance. When the phase coherence distance is greater than the preset distance threshold, output the rhythm disturbance time-frequency phase feature expression vector.

[0034] S3. Construct a rhythmic phase migration map based on the rhythmic disturbance time-frequency phase feature expression vector, and use the minimum energy drift path algorithm to determine the rhythmic disturbance propagation path and source direction. Combine the spatial layout mapping rules of primary equipment to determine the source location of substation hazards, and output the hazard source probability vector and rhythmic disturbance spatial location results.

[0035] S4. Based on the rhythmic disturbance time-frequency phase feature expression vector and the hazard source probability vector, the hazard risk value is calculated using the risk evolution model, and the hazard warning signal is output.

[0036] In this embodiment S1, the process of reconstructing and mapping the time scale of multi-source electrical signals to a unified time axis using time synchronization identifiers includes: synchronizing and calibrating the sampling time scales of multi-source electrical signals based on a unified time base; and performing interpolation resampling and time drift compensation on the sampling time scales of multi-source electrical signals to achieve time scale alignment of all multi-source electrical signals on the same time axis.

[0037] The rhythmic anomaly characteristic signal refers to the set of electrical signals used to characterize rhythmic disturbance behavior after time-scale reconstruction and unified time axis mapping are completed and preprocessed. The rhythmic anomaly characteristic signal includes the rhythmic response components of the electrical signals output by each monitoring device under the unified time axis.

[0038] In this embodiment S1, the substation monitoring device includes a synchronous phasor measurement unit (PMU) and a measurement and control unit, a differential protection device, a fault recording system, a merging unit or a sampling value output module of an electronic instrument transformer, and a measurement and acquisition unit of the dispatching / station control system. Each device provides time-stamped sampling sequences of voltage and current samples and their derivatives, and outputs data streams for analysis through the substation communication network or local acquisition interface. Multi-source electrical signals refer to the set of time-stamped electrical quantities collected by the above-mentioned substation monitoring device during the same operating period. The electrical quantities include at least the synchronous sampling value sequences of voltage and current and characteristic quantities calculated by the device, including high-frequency harmonic components, zero-sequence components, negative-sequence components, and recording phase angles. Without changing the essence of the technical solution, multi-source electrical signals may also include derived characteristic quantities for rhythm analysis, such as frequency deviation, phase angle difference, and power factor change rate.

[0039] In this embodiment S1, the time synchronization identifier refers to the absolute time information or equivalent time base label recorded along with the multi-source electrical signals, used for cross-device time alignment and timing consistency verification. The time synchronization identifier can be generated by the station's time synchronization system and uniformly distributed. The time synchronization system can use satellite time synchronization or a clock based on the IEEE 1588 precision time protocol. The synchronization information is attached to the sampled data in the form of a combination of timestamp, sequence number, and synchronization status indication. The purpose of reconstructing the time scale of the multi-source electrical signals to a unified time axis is to eliminate the frequency offset and relative delay between the sampling clocks of different devices, so that the sampling points across devices and channels are comparable and can perform coherent analysis. The specific process is to establish a unified time base based on the time synchronization identifier, perform alignment processing on the original sampling time scale of each signal, so that they form a continuous timing data stream under the unified time axis, providing a consistent time reference for subsequent rhythm anomaly feature extraction and phase analysis.

[0040] In this embodiment S1, interpolation resampling refers to interpolating and resampling the original sampling points with non-equidistant or different sampling rates at a target sampling rate on a unified time axis, so that each channel obtains a consistent sampling point sequence on the unified time axis. The interpolation method can be linear interpolation, spline interpolation, or band-limited interpolation based on window functions. Time drift compensation refers to estimating and correcting the fixed delay and slow drift of each channel relative to a unified time base. The drift amount can be estimated by timestamp difference, cross-correlation, or synchronization message statistics, and the sampling point sequence is corrected in the form of phase or time offset. Preprocessing refers to performing a data quality improvement and feature preservation processing sequence on the multi-source electrical signals mapped to the unified time axis. This processing sequence includes denoising, band-limited filtering, DC component removal, baseline normalization, windowing and boundary processing, and frequency band selection for rhythm feature preservation.

[0041] In this embodiment S1, the rhythmic abnormality characteristic signal refers to the set of electrical signals used to characterize rhythmic disturbance behavior after the time scale reconstruction is completed and mapped to a unified time axis and preprocessed. The set consists of the sampled value sequence of multi-source electrical signals under the unified time axis and its derived rhythmic response components. The rhythmic response components include high-frequency harmonic component sequences, zero-sequence and negative-sequence component sequences, and recorded phase angle sequences.

[0042] In this embodiment S2, the rhythm abnormality feature signal is phase-expanded to construct the rhythm disturbance time-frequency phase curve. Specifically, this includes: performing Hilbert-Huang transform on the rhythm abnormality feature signal, decomposing the rhythm abnormality feature signal into several intrinsic mode function components, extracting instantaneous phase information for each intrinsic mode function component, expanding the instantaneous phase information on a unified time axis to form a phase trajectory set, and constructing the rhythm disturbance time-frequency phase curve corresponding to each intrinsic mode function component.

[0043] The rhythmic perturbation time-frequency phase curve refers to the curve formed by the phase trajectory extracted from the multi-source rhythmic abnormality characteristic signal under a unified time axis through the Hilbert-Huang transform, which is used to characterize the dynamic change characteristics of the rhythmic perturbation signal in the time domain and phase domain.

[0044] In this embodiment, the Hilbert-Huang transform is a time-frequency analysis method for nonlinear and non-stationary signals. It consists of two parts: empirical mode decomposition and Hilbert analysis. Empirical mode decomposition is used to decompose the rhythmic abnormality signal into several intrinsic mode function components and residual components, such that each intrinsic mode function component satisfies the characteristics of single narrowband and local symmetry. Hilbert analysis is used to extract instantaneous phase information and the corresponding time-frequency representation from the intrinsic mode function components. The Hilbert-Huang transform can reflect the temporal locality and phase change law of the rhythmic disturbance signal on a unified time axis.

[0045] In this embodiment, the intrinsic mode function component refers to the component with a single local oscillation scale obtained through empirical mode decomposition; the component satisfies the condition that the mean of the local upper and lower envelopes is close to zero and the number of extreme points and zero crossover points is similar; each intrinsic mode function component corresponds to a specific rhythm response scale, so that the phase changes of rhythm disturbances at different time scales can be separated; the sum of all intrinsic mode function components and residual components is reconstructed into a rhythm abnormality characteristic signal.

[0046] In this embodiment, empirical mode decomposition is used to screen rhythmic abnormality feature signals layer by layer. First, local extrema are searched in the rhythmic abnormality feature signals on a unified time axis, upper and lower envelopes are constructed, and the envelope mean is calculated. Then, candidate components are obtained by subtracting the envelope mean from the original signal, and the screening process of extrema search, envelope construction, and mean stripping is repeated for the candidate components until the stopping condition of the intrinsic mode function component is met. The stopping condition includes that the envelope mean is close to zero and the number of extrema points is close to the number of zero crossover points, or the change amplitude of adjacent cycles of the screening result is lower than a preset threshold. After the first intrinsic mode function component is completed, it is stripped from the original signal to obtain a new residual signal and enter the next round of screening until the residual signal no longer contains effective oscillatory components. In order to reduce the endpoint effect and pseudo-mode, this embodiment uses mirror extension or spline extrapolation at both ends to constrain the envelope, and sets an upper limit on the number of cycles and an energy threshold during the screening process to ensure the stability of decomposition.

[0047] In this embodiment, the instantaneous phase information includes three types of quantities: phase, phase change rate, and phase position indicator. The phase is used to characterize the instantaneous phase angle of the intrinsic mode function components on a unified time axis. The phase change rate is used to characterize the rate of phase change over time, reflecting the time-varying characteristic scale of rhythmic perturbations. The phase position indicator is used to mark the validity, continuity, and reliability of phase samples, for example, by providing an invalid identifier when the envelope degenerates or the local signal-to-noise ratio decreases. The instantaneous phase information is obtained by performing Hilbert analysis on the intrinsic mode function components. The Hilbert analysis constructs an analytical form on each intrinsic mode function component, thereby extracting the monotonic change trajectory of the phase over time and the phase change rate.

[0048] In this embodiment, the instantaneous phase information of each intrinsic mode function component is expanded. First, the instantaneous phase samples are arranged in a unified time axis order according to the time synchronization identifier, and the missing sampling points are supplemented using the aforementioned interpolation resampling strategy. Second, phase unwrapping is performed to eliminate phase jumps, and the phase trajectories are smoothly connected at moments crossing phase periods while maintaining a monotonically increasing phase. Third, abnormal phase samples are corrected or removed according to the phase position information indication, and boundary smoothing is performed within the window to suppress endpoint fluctuations. Finally, the expanded phase trajectories of all intrinsic mode function components are summarized in a unified time axis to form a phase trajectory set, which serves as the input for constructing the rhythmic perturbation time-frequency phase curve. The phase trajectory set can be indexed according to the scale order or energy proportion of the intrinsic mode function components. The phase trajectory set is generated for multi-source rhythmic abnormality characteristic signals through Hilbert-Huang... The instantaneous phase information of each eigenmode function component extracted by transformation is expanded on a unified time axis to form an ordered set of multiple phase trajectories. Each phase trajectory corresponds to the phase evolution process of an eigenmode function component in the time domain. The set of phase trajectories retains the phase response characteristics of different monitoring points and different rhythm scales, and is used to construct a cluster of rhythmic perturbation time-frequency phase curves and as input parameters for a subspace feature splitting network.

[0049] In this embodiment S2, the subspace feature splitting network is a neural network structure built based on the high-dimensional embedded feature representation of the rhythmic perturbation time-frequency phase curve and the rhythmic phase coherence constraint loss function. It includes a feature embedding layer, a feature splitting layer and a discriminant output layer, which are used to divide the rhythmic perturbation time-frequency phase curve into feature spaces.

[0050] The subspace feature splitting network divides all rhythmic disturbance time-frequency phase curves into anomaly disturbance subspace and load disturbance subspace. Specifically, it includes: inputting the phase trajectory set into the feature embedding layer to extract the joint time-phase feature vector; inputting the joint time-phase feature vector into the feature splitting layer; aggregating and decoupling the joint time-phase feature vector by minimizing the rhythmic phase coherence constraint loss function to obtain the feature space partitioning result; and using the discriminant output layer to determine the convergence of the feature space partitioning result and output the subspace label.

[0051] The subspace labels include anomaly disturbance subspace and load disturbance subspace; the anomaly disturbance subspace refers to the set of phase trajectories with high rhythmic phase consistency and high coherence weight in the rhythmic disturbance time-frequency phase curve cluster; the load disturbance subspace refers to the set of phase trajectories that do not have rhythmic coherence characteristics.

[0052] In this embodiment, the high-dimensional embedding feature representation of the rhythmic disturbance time-frequency phase curve refers to the nonlinear mapping of the rhythmic disturbance time-frequency phase curve to the phase trajectory set through a feature embedding layer, mapping the rhythmic disturbance time-frequency phase curve to an embedding feature space with high-dimensional time-phase joint feature components. This high-dimensional embedding feature representation retains the change information of the rhythmic disturbance in the time domain and the rhythmic response structure in the phase domain, and encodes the phase consistency and phase change rate for subsequent subspace splitting and aggregation operations. The rhythmic phase coherence constraint loss function is used to constrain the aggregation form of the phase trajectory set in the embedding space. Its goal is to maximize the feature similarity between trajectories with high phase consistency and high rhythmic synchronization, and minimize the feature coupling degree between non-rhythmic disturbance curves and abnormal disturbance curves, thereby achieving effective separation of abnormal disturbances and load disturbances in the feature space.

[0053] In this embodiment, the feature embedding layer is used to encode the features of the input phase trajectory set. The core structure of this layer consists of a multi-layer one-dimensional convolutional network and a temporal attention mechanism. The convolutional network is used to extract the phase change rate features and rhythmic fluctuation structure of the local phase curve, and the temporal attention mechanism is used to weight and enhance the dynamic phase changes within different time windows, thereby obtaining a stable time-phase joint feature vector representation. The feature splitting layer is a nonlinear mapping structure used to perform subspace partitioning of high-dimensional features. This layer consists of a subspace projection operator and a kernel-based feature splitting operator. The projection operator maps the time-phase joint feature vector to multiple latent features. The perturbation subspace is used by a feature splitting operator to iteratively optimize these potential subspaces using a rhythmic phase coherence constraint loss function. This gradually aggregates high-coherence features into the anomalous perturbation subspace, while decoupling low-coherence features into the load perturbation subspace. The discriminant output layer is the output layer of the subspace feature splitting network. Its structure includes a feature clustering confidence calculation unit and a subspace label mapping unit. The feature clustering confidence calculation unit is used to calculate the aggregation confidence and subspace assignment probability of each time phase joint feature vector. The subspace label mapping unit generates the final subspace partitioning label according to the aggregation convergence condition, identifying the category of each rhythmic perturbation time-frequency phase curve.

[0054] In this embodiment, the process of aggregating and decoupling the joint temporal phase feature vector by minimizing the rhythmic phase coherence constraint loss function includes: first, constructing an initial latent subspace based on the feature splitting layer, and mapping feature vectors with high coherence weights to the same subspace; second, calculating the coherence similarity and coherence penalty term between feature vectors using the rhythmic phase coherence constraint loss function; third, adjusting the distribution position of feature vectors in the latent subspace through gradient descent or adaptive iterative optimization algorithms, so that rhythmic perturbation features are automatically aggregated in the feature space, and non-rhythmic perturbation features are gradually decoupled; when the convergence condition of the loss function reaches a preset threshold, the subspace aggregation and decoupling are completed.

[0055] In this embodiment, the convergence determination of the subspace partitioning result by the discriminant output layer includes: after each iteration, calculating the feature aggregation confidence distribution and the subspace assignment probability distribution; when the change in aggregation confidence is lower than the preset convergence threshold or the assignment probability tends to stabilize, the subspace splitting process is determined to be converged; after convergence, the subspace label mapping unit generates the subspace label of each rhythmic disturbance time-frequency phase curve according to the final assignment probability, and marks it as an abnormal disturbance subspace or a load disturbance subspace.

[0056] In this embodiment S2, the rhythmic disturbance phase consistency index is a statistical quantity calculated based on the phase trajectories of the rhythmic disturbance time-frequency phase curves under a unified time axis, used to characterize the phase alignment degree and synchronization fluctuation characteristics of each phase trajectory under the action of rhythmic disturbance; the phase coherence distance is a feature space distance metric calculated based on the coherence similarity of the feature vectors between the rhythmic disturbance phase consistency index and the phase trajectory, used to distinguish between abnormal disturbances and load disturbances;

[0057] The rhythmic perturbation time-frequency phase feature representation vector refers to the high-dimensional phase-time joint feature representation extracted from the embedded feature space in the subspace feature splitting network when the phase coherence distance is greater than a preset distance threshold. It is used to characterize the rhythmic perturbation propagation path and perturbation source direction in the subsequent construction of the rhythmic phase migration map.

[0058] In this embodiment, the statistical quantity obtained from the phase synchronization degree calculation refers to the quantification result obtained by performing a synchronization measurement on the instantaneous phase information of each phase trajectory based on a unified time axis for the set of phase trajectories of the rhythmic perturbation time-frequency phase curve. Specifically, the instantaneous phase corresponding to each phase trajectory at each sampling time point is vectorized, and the phase synchronization coefficient is obtained by calculating the phase deviation between trajectories and the offset of the global phase center. The phase synchronization coefficient serves as the basic statistical quantity of the rhythmic perturbation phase consistency index in this technical solution. The feature vector between phase trajectories refers to the feature difference representation between the joint feature vectors of each time phase extracted by the feature embedding layer. This feature vector contains the difference information between trajectories in multiple component dimensions such as phase change rate, phase consistency, phase coherence weight, and trajectory scale identifier. By calculating the differences of these feature components, the relative relationship between corresponding points of any two trajectories in the high-dimensional embedding space can be characterized.

[0059] In this embodiment, the feature space distance metric refers to the high-dimensional distance quantization result calculated based on the coherent similarity between the rhythmic perturbation phase consistency index and the feature vectors between phase trajectories. The calculation process includes: first, weighting the trajectory aggregation direction in the feature space using the rhythmic perturbation phase consistency index; second, performing a coherent similarity function operation on the feature vectors between phase trajectories to map phase differences to coherent similarity; and third, normalizing the coherent similarity function result using a distance metric operator to obtain the phase coherent distance. This distance metric is used to characterize the degree of coherent aggregation of rhythmic perturbations between trajectories. A smaller coherent distance indicates load perturbation characteristics, while a larger coherent distance indicates… The system identifies anomalous perturbation characteristics. The coherent similarity function, located in the feature embedding space, calculates the similarity between the time-frequency phase curves of any two rhythmic perturbations in terms of phase synchronicity and spectral coherence. Its input consists of the joint time-phase feature vectors of the two trajectories and the rhythmic perturbation phase consistency index. The output is a scalar coherent similarity value, which is the result of the coherent similarity function. The coherent similarity function includes a phase synchronicity factor, a spectral coherence factor, and a weighted fusion operator based on the rhythmic perturbation phase consistency index. By calculating the coupling degree of the rhythmic response characteristics of the two trajectories, a coherent similarity result reflecting the degree of correlation between rhythmic perturbations is obtained and used for feature space distance measurement calculations.

[0060] In this embodiment, the preset distance threshold is determined based on the coherence distance distribution characteristics of rhythmic disturbance events in historical operating data and field-labeled samples. Specifically, it includes: firstly, extracting representative abnormal disturbances and load disturbance events from historical operating samples; secondly, calculating the rhythmic disturbance phase consistency index and phase coherence distance distribution of these samples respectively; thirdly, determining the boundary value of coherence distance based on bimodal or multimodal distribution characteristics, and determining the optimal preset distance threshold through cross-validation.

[0061] In this embodiment S3, the method for constructing a rhythmic phase migration map based on the rhythmic perturbation time-frequency phase feature expression vector is as follows: taking the monitoring device acquisition point as the node, the phase migration direction and phase delay as the direction attribute of the edge, and the phase drift amount as the edge weight, the phase drift and delay relationship between adjacent acquisition points is calculated based on the rhythmic perturbation time-frequency phase feature expression vector to construct the rhythmic phase migration map;

[0062] The rhythmic phase migration diagram is a time-series directed graph structure, used to characterize the propagation path, direction, and time delay of rhythmic disturbances in substations, providing a structured basis for the inversion of the direction of potential hazards and the location of energy drift.

[0063] In this embodiment, each monitoring device acquisition point corresponds one-to-one with a key location of the substation primary equipment. By mapping the acquisition points to nodes, the propagation path and topological distribution of rhythmic disturbances in space can be clearly represented in the graph structure. The edges in the rhythmic phase migration graph represent the rhythmic disturbance phase migration relationship between any two monitoring device acquisition points. Their direction attribute is used to characterize the temporal order of disturbance propagation, and their time delay feature is used to characterize the propagation delay of the disturbance signal between adjacent acquisition points. The direction of the edge is determined by comparing the arrival order of the phase response, and the time delay of the edge is obtained by calculating the difference in the peak time of the phase response in the rhythmic disturbance time-frequency phase feature expression vector corresponding to the two acquisition points. When the disturbance propagates from one acquisition point to another, the direction of the edge points from the former to the latter, and the time delay is recorded as one of the direction attributes in the structural information of the edge.

[0064] In this embodiment, the weight of the edge in the rhythm phase migration graph is defined as the phase drift between two acquisition points, which is used to quantify the degree of phase response change during the propagation of rhythmic disturbance. The calculation method of the phase drift includes: firstly, extracting the phase trajectory information from the rhythmic disturbance time-frequency phase feature expression vector of the two acquisition points; secondly, calculating the phase difference curve of the two trajectories under a unified time axis; and thirdly, obtaining the phase drift value based on the phase difference accumulation during the main peak time period of the disturbance response. This phase drift is recorded as the weight of the edge and is used to characterize the directional migration and amplitude change relationship of rhythmic disturbance energy. The rhythm phase migration graph is used to characterize the propagation path, direction and time delay of rhythmic disturbance between the acquisition points of each monitoring device in the substation, and describes the aggregation and diffusion characteristics of disturbance energy in space through the phase drift.

[0065] In this embodiment S3, the minimum energy drift path algorithm is a path search and energy accumulation optimization algorithm based on rhythm phase migration graph and edge weight phase drift amount. It is used to find the propagation path with the minimum global energy dissipation in the rhythm phase migration graph and determine the source direction of rhythm disturbance.

[0066] The process of determining the propagation path and source direction of rhythmic disturbance using the minimum energy drift path algorithm includes: calculating the phase drift cumulative energy of each node based on the directed edge set of the rhythmic phase transition graph; constructing a directed energy graph with phase drift as the weight; performing a minimum energy path search starting from the set of nodes with the smallest global in-degree in the energy graph to obtain the minimum energy path set from the source node to each affected node; determining the source direction of the rhythmic disturbance and outputting the corresponding source node and path structure based on the minimum energy path and energy cumulative gradient in the path set.

[0067] In this embodiment, the minimum energy drift path algorithm is a path search and energy optimization algorithm built based on the topological structure information of the rhythm phase migration graph and the phase drift of each directed edge. Its core idea is to regard the propagation of rhythm disturbance between the acquisition points of the monitoring device as a directed energy transfer chain constrained by the phase drift. By finding the path with the minimum energy dissipation in the graph structure, the location of the rhythm disturbance source direction and the unique determination of the propagation path are realized. In the case of multiple propagation paths of rhythm disturbance and superposition of energy interference, the algorithm can eliminate interference paths through energy accumulation and gradient search strategies, thereby identifying the real propagation path related to the abnormal disturbance source.

[0068] In this embodiment, the energy calculation of rhythmic perturbation is based on the edge weights of the rhythmic phase transition graph, i.e., the phase drift. Specifically, it includes: first, normalizing the phase drift of each directed edge to construct a phase drift energy weight matrix; second, mapping the energy weight matrix onto the directed graph structure, using the energy weight of each directed edge to represent the energy consumption of the rhythmic perturbation through that path; third, for any directed path between two nodes, calculating the cumulative energy weight of all edges along the path to obtain the total energy drift value of the path; this total energy drift value serves as the optimization objective function for path search, providing the minimum energy path search result. The method provides quantitative evidence; the minimum energy path search steps include the following process: identifying all nodes with the smallest in-degree in the rhythm phase transition graph as candidate source node sets; starting from each candidate source node, performing a directed path search based on the energy weight matrix for all downstream nodes, and calculating the cumulative energy drift value of each path; comparing the total energy drift values ​​of different paths, and selecting the path with the smallest energy drift value as the minimum energy path; comparing the minimum energy path sets of all candidate source nodes, and selecting the path set with the smallest energy drift and the best path convergence, providing a decision basis for determining the rhythm disturbance source.

[0069] In this embodiment S3, the spatial layout mapping rule of primary equipment is a spatial mapping rule constructed based on the spatial location coordinate set, connection topology and operation functional area division of the primary equipment in the substation. It is used to map and associate the rhythm disturbance propagation path with the physical location and functional area of ​​the primary equipment.

[0070] The process involves determining the location of potential hazards in a substation and outputting a hazard source probability vector and the spatial location result of rhythmic disturbances. Specifically, this includes: matching the physical equipment location in the spatial layout mapping rules based on the starting node of the minimum energy propagation path in the rhythmic phase migration diagram and its associated primary equipment spatial coordinates; calculating the hazard source attribution probability for the equipment areas involved in the disturbance path to form a hazard source probability vector; and outputting the spatial location result of rhythmic disturbances based on the spatial mapping relationship between the hazard source probability vector and the primary equipment layout.

[0071] In this embodiment, the primary equipment spatial layout mapping rule is constructed based on the physical spatial coordinate set of the primary equipment in the substation, the connection topology between the primary equipment, and the functional area division information. The spatial coordinate set of each primary equipment is extracted using the substation primary equipment design layout drawing and BIM model, forming a spatial reference that corresponds one-to-one with the monitoring device acquisition points. An equipment connection topology diagram is constructed based on the electrical connection and functional relationships between the primary equipment, clarifying the spatial adjacency and electrical coupling relationships of each equipment under the main electrical wiring structure. The equipment is divided into functional areas such as the main transformer area, busbar area, circuit breaker area, and switch area, forming a spatial mapping rule structure for equipment location, topology, and primary equipment function. The mapping process between the rhythmic disturbance propagation path and the primary equipment space includes: matching the nodes in the rhythmic phase migration diagram with their corresponding monitoring device acquisition points in the primary equipment spatial layout mapping rule; locating the corresponding primary equipment spatial coordinates for each node in the minimum energy propagation path based on its index number; and spatially mapping the propagation direction of the phase drift and time delay information on the path using the equipment connection topology, forming a one-to-one correspondence curve of the disturbance path in physical space.

[0072] In this embodiment, the hazard source probability vector refers to the probability distribution vector calculated based on the minimum energy propagation path starting node and the set of primary equipment in its adjacent area according to the rhythmic phase migration map. The calculation process includes: normalizing the phase drift and time delay information of the minimum energy propagation path starting node and its surrounding adjacent nodes; calculating the probability weight of each device appearing as a hazard source in the disturbance propagation path by combining the spatial layout mapping rules; and forming a normalized probability distribution by weighted fusion of the disturbance source energy gradient and the topological in-degree factor to obtain the hazard source probability vector. The output of the rhythmic disturbance spatial positioning result includes: mapping the hazard source probability vector to the coordinate system in the primary equipment spatial layout mapping rules; marking the probability weights on the corresponding physical locations of the primary equipment in a spatial annotation manner to form the inversion positioning result of the rhythmic disturbance hazard source in the substation space; when the hazard source probability exceeds the preset judgment threshold, marking the corresponding primary equipment location as the rhythmic disturbance hazard source location, and outputting the hazard source equipment number, coordinate information, and corresponding path.

[0073] In this embodiment S4, the risk evolution model is a time series assessment model for hidden danger risk based on the rhythmic disturbance time-frequency phase feature expression vector, the hidden danger source probability vector and the time series disturbance response index. It is used to calculate the hidden danger risk value of rhythmic disturbance hidden danger in the time dimension and output the hidden danger warning signal.

[0074] In this embodiment, a risk evolution model is used to calculate the risk value of a hazard and output a hazard warning signal. The specific process includes: constructing a disturbance energy accumulation function based on the temporal change of disturbance energy in the rhythmic disturbance propagation path; dynamically adjusting the disturbance risk weights for different primary equipment locations by combining the hazard source probability vector; calculating the evolution trajectory of the disturbance risk value on the time axis through a risk propagation recursive operator; and outputting a rhythmic disturbance hazard warning signal when the calculated risk value exceeds a preset threshold.

[0075] Example 2: The AI-based substation hazard prediction system proposed in this invention is applied to the AI-based substation hazard prediction method proposed in Example 1. It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the AI-based substation hazard prediction method in Example 1.

[0076] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. An AI-based method for predicting potential hazards in substations, characterized in that, Includes the following steps: S1. Obtain multi-source electrical signals through substation monitoring devices, use time synchronization identifiers to reconstruct the multi-source electrical signals to a unified time axis, and preprocess the multi-source electrical signals to form rhythmic abnormality characteristic signals; In S1, the process of reconstructing and mapping the time scale of multi-source electrical signals to a unified time axis using time synchronization identifiers includes: synchronizing and calibrating the sampling time scales of multi-source electrical signals based on a unified time base; and aligning the time scales of all multi-source electrical signals on the same time axis by interpolating and resampling the sampling time scales of multi-source electrical signals and compensating for time drift. The rhythmic anomaly characteristic signal refers to the set of electrical signals used to characterize rhythmic disturbance behavior after preprocessing following the completion of time scale reconstruction and unified time axis mapping. The rhythmic anomaly characteristic signal includes the rhythmic response components of the electrical signals output by each monitoring device under the unified time axis. S2. Perform phase expansion on the rhythm abnormality feature signal to construct the rhythm disturbance time-frequency phase curve. Use the subspace feature splitting network to divide all rhythm disturbance time-frequency phase curves into abnormal disturbance subspace and load disturbance subspace. Calculate the rhythm disturbance phase consistency index and phase coherence distance. When the phase coherence distance is greater than the preset distance threshold, output the rhythm disturbance time-frequency phase feature expression vector. In S2, the subspace feature splitting network is a neural network structure built based on the high-dimensional embedded feature representation of the rhythmic perturbation time-frequency phase curve and the rhythmic phase coherence constraint loss function. It includes a feature embedding layer, a feature splitting layer and a discriminant output layer, which are used to divide the rhythmic perturbation time-frequency phase curve into feature spaces. The subspace feature splitting network divides all rhythmic disturbance time-frequency phase curves into anomaly disturbance subspace and load disturbance subspace. Specifically, it includes: inputting the phase trajectory set into the feature embedding layer to extract the joint time-phase feature vector; inputting the joint time-phase feature vector into the feature splitting layer; aggregating and decoupling the joint time-phase feature vector by minimizing the rhythmic phase coherence constraint loss function to obtain the feature space partitioning result; and using the discriminant output layer to determine the convergence of the feature space partitioning result and output the subspace label. The subspace labels include anomaly disturbance subspace and load disturbance subspace; the anomaly disturbance subspace refers to the set of phase trajectories with high rhythmic phase consistency and high coherence weight in the rhythmic disturbance time-frequency phase curve cluster; the load disturbance subspace refers to the set of phase trajectories that do not have rhythmic coherence characteristics. S3. Construct a rhythmic phase migration map based on the rhythmic disturbance time-frequency phase feature expression vector, and use the minimum energy drift path algorithm to determine the rhythmic disturbance propagation path and source direction. Combine the spatial layout mapping rules of primary equipment to determine the source location of substation hazards, and output the hazard source probability vector and rhythmic disturbance spatial location results. In S3, the minimum energy drift path algorithm is a path search and energy accumulation optimization algorithm based on rhythm phase migration graph and edge weight phase drift amount. It is used to find the propagation path with the minimum global energy dissipation in the rhythm phase migration graph and determine the source direction of rhythm disturbance. The process of determining the propagation path and source direction of rhythmic perturbation using the minimum energy drift path algorithm includes: calculating the phase drift cumulative energy of each node based on the directed edge set of the rhythmic phase transition graph; constructing a directed energy graph with phase drift as the weight; performing a minimum energy path search in the energy graph starting from the set of nodes with the smallest global in-degree to obtain the minimum energy path set from the source node to each affected node; determining the source direction of the rhythmic perturbation and outputting the corresponding source node and path structure based on the minimum energy path and energy cumulative gradient in the path set. In S3, the spatial layout mapping rule of primary equipment is a spatial mapping rule constructed based on the spatial location coordinate set, connection topology and operation functional area division of the primary equipment of the substation. It is used to map and associate the rhythm disturbance propagation path with the physical location and functional area of ​​the primary equipment. The process involves determining the location of potential hazards in a substation and outputting a hazard source probability vector and rhythmic disturbance spatial location results. Specifically, this includes: matching the physical equipment location in the spatial layout mapping rules based on the starting node of the minimum energy propagation path in the rhythmic phase migration diagram and its associated primary equipment spatial coordinates; calculating the hazard source attribution probability for the equipment areas involved in the disturbance path to form a hazard source probability vector; and outputting the rhythmic disturbance spatial location results based on the spatial mapping relationship between the hazard source probability vector and the primary equipment layout. S4. Based on the rhythmic disturbance time-frequency phase feature expression vector and the hazard source probability vector, the hazard risk value is calculated using the risk evolution model, and the hazard warning signal is output.

2. The AI-based substation hazard prediction method according to claim 1, characterized in that: In step S2, phase expansion is performed on the rhythm abnormality feature signal to construct the rhythm disturbance time-frequency phase curve. Specifically, this includes: performing Hilbert-Huang transform on the rhythm abnormality feature signal, decomposing the rhythm abnormality feature signal into several intrinsic mode function components, extracting instantaneous phase information for each intrinsic mode function component, expanding the instantaneous phase information on a unified time axis to form a phase trajectory set, and constructing the rhythm disturbance time-frequency phase curve corresponding to each intrinsic mode function component. The rhythmic perturbation time-frequency phase curve refers to the curve formed by the phase trajectory extracted from the multi-source rhythmic abnormality characteristic signal under a unified time axis through the Hilbert-Huang transform, which is used to characterize the dynamic change characteristics of the rhythmic perturbation signal in the time domain and phase domain.

3. The AI-based substation hazard prediction method according to claim 2, characterized in that: In S2, the rhythmic disturbance phase consistency index is a statistic calculated based on the phase trajectories of the rhythmic disturbance time-frequency phase curves under a unified time axis; the phase coherence distance is a feature space distance metric calculated based on the coherence similarity of the feature vectors between the rhythmic disturbance phase consistency index and the phase trajectory, used to distinguish between abnormal disturbances and load disturbances. The rhythmic perturbation time-frequency phase feature representation vector refers to the high-dimensional phase-time joint feature representation extracted from the embedded feature space when the phase coherence distance is greater than a preset distance threshold in the subspace feature splitting network.

4. The AI-based substation hazard prediction method according to claim 3, characterized in that: In S3, the method for constructing the rhythmic phase migration map based on the rhythmic perturbation time-frequency phase feature expression vector is as follows: taking the monitoring device acquisition point as the node, the phase migration direction and phase delay as the direction attribute of the edge, and the phase drift amount as the edge weight, the phase drift and delay relationship between adjacent acquisition points is calculated based on the rhythmic perturbation time-frequency phase feature expression vector to construct the rhythmic phase migration map. The rhythm phase transition diagram is a time-series directed graph structure used to characterize the propagation path, direction, and time delay of rhythmic disturbances in substations.

5. The AI-based substation hazard prediction method according to claim 4, characterized in that: In S4, the risk evolution model is a time-series risk assessment model for hidden dangers constructed based on the rhythmic disturbance time-frequency phase feature expression vector, the probability vector of hidden danger sources, and the time-series disturbance response index. It is used to calculate the hidden danger risk value of rhythmic disturbance hidden dangers in the time dimension and output hidden danger warning signals.

6. An AI-based substation hazard prediction system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the AI-based substation hazard prediction method as described in any one of claims 1-5.

Images