A source network load multi-source data anomaly detection and repair method and device

Abstract

The disclosure provides a source network load multi-source data anomaly detection and repair method and device, relating to the technical field of power system data processing. The method comprises: based on the synchronized multi-source data of the power system, determining the graph structure data at each time; reconstructing the graph structure data at each time through a graph time series reconstruction model to obtain the reconstruction result at each time; wherein the joint loss function containing the mask reconstruction term and the physical consistency term is used for parameter updating in the training stage of the graph time series reconstruction model; based on the mask reconstruction residual and the physical consistency residual of the reconstruction result, determining the comprehensive score at each time; wherein the physical consistency residual is calculated by a physical consistency operator model; the time whose comprehensive score is greater than the abnormal judgment threshold is marked as the abnormal time; and the graph structure data of the abnormal time is repaired based on the reconstruction result of the abnormal time. The disclosure can accurately detect abnormal data and effectively repair it, thereby reducing the subsequent operation error of the system.

Description

Technical Field

[0001] This disclosure relates to the field of power system data processing technology, specifically to a method and apparatus for anomaly detection and repair of multi-source data from power sources, grids, loads, and storage. Background Technology

[0002] With the rapid development of new power systems, massive amounts of multi-source, heterogeneous time-series data are continuously generated from the power source side (wind power, photovoltaic, conventional units), the grid side (bus / branch / transformer / switch measurements and topology status), the load side (user electricity consumption and operating condition characteristics), the energy storage side (battery / pumped storage SOC and charging / discharging power), and the meteorological side (temperature, wind speed, irradiance, etc.). This data is crucial input for load forecasting, power forecasting, demand response, and coordinated control and scheduling optimization of power generation, grid, load, and storage systems; its quality directly impacts grid security and system safety. Due to the high heterogeneity and unstable quality of this data, data processing is necessary to reduce system operating errors.

[0003] However, existing data processing methods, such as statistical methods, machine learning or deep learning methods, and power mechanism verification methods, are fragmented and prone to false alarms or false alarms, leading to increased system operation errors and seriously restricting the safe and efficient operation of new power systems. Summary of the Invention

[0004] This disclosure addresses the problems existing in the prior art by providing a method and apparatus for anomaly detection and repair of multi-source data from the power grid, load, and storage systems. It can solve the problem that the existing methods are fragmented and prone to false alarms or missed alarms, which leads to increased system operation errors, thereby ensuring the safe and efficient operation of the new power system.

[0005] To achieve the above objectives, the technical solution adopted in this disclosure is as follows: The first aspect of this disclosure provides a method for anomaly detection and repair of multi-source data from power grid, load, and storage systems, comprising: determining graph structure data at each time point based on multi-source data of a power system aligned to a preset time grid; wherein the graph structure data includes a graph node feature matrix and a graph mask matrix; for each time point, reconstructing the graph structure data at that time point using a pre-constructed graph time-series reconstruction model to obtain a reconstruction result for that time point; wherein the graph time-series reconstruction model uses a joint loss function including a mask reconstruction term and a physical consistency term for parameter updates during the training phase; determining a comprehensive score for each time point based on the mask reconstruction residual and the physical consistency residual of the reconstruction result; wherein the physical consistency residual is calculated by a pre-constructed physical consistency operator model; marking times with a comprehensive score greater than an anomaly judgment threshold as anomaly times to complete anomaly detection; performing physical consistency correction on the reconstruction results of the aomaly times to obtain a target repair value that satisfies preset operating boundary constraints, and repairing the graph structure data of the aomaly times based on the target repair value to complete the repair.

[0006] In one possible implementation, the graph temporal reconstruction model includes a graph temporal autoencoder; the graph temporal autoencoder includes a spatial encoder, a temporal encoder, and a decoder; by using the pre-built graph temporal reconstruction model, the graph structure data at a given time is reconstructed to obtain the reconstruction result at that time, including: based on the graph structure data of all times within the historical time window of that time, and the topology graph and parameter versions, the spatial encoder performs spatial feature aggregation on the graph structure data to obtain a spatial representation; wherein, the historical time window is a continuous time interval formed by truncating a preset time length from the given time as the ending time; based on the spatial representation, the temporal sequence of the spatial representation is modeled by the temporal encoder to obtain temporal latent variables; based on the temporal latent variables, the decoder determines the reconstruction result at that time.

[0007] In one possible implementation, the comprehensive score at a given time is determined based on the masked reconstruction residual and the physical consistency residual of the reconstruction result, including: determining the masked reconstruction residual between the reconstruction result and the graph structure data; determining the physical consistency residual between the reconstruction result and the power system physical constraints; and determining the comprehensive score at a given time based on the masked reconstruction residual and the physical consistency residual; wherein the comprehensive score is the sum of the product of the preset weighting coefficient and the physical consistency residual and the masked reconstruction residual.

[0008] In one possible implementation, physical consistency correction is performed on the reconstruction results at abnormal moments to obtain a target repair value that satisfies preset operating boundary constraints. This includes: under preset operating boundary constraints, using the reconstruction results at abnormal moments as candidate repair values, and performing physical consistency correction on the reconstruction results at abnormal moments through a predefined constraint optimization formula to obtain the target repair value; wherein, the preset operating boundary constraints include voltage boundary constraints, branch power boundary constraints, and energy storage state of charge boundary constraints.

[0009] In one possible implementation, the physical consistency residual is the aggregated value of one or more types of residuals, including power balance residual, power flow consistency residual, energy storage energy conservation residual, energy storage state of charge dynamic residual, and equipment boundary residual.

[0010] In one possible implementation, the quantile of the physical consistency residuals corresponding to all normally operating samples during the training phase is determined as the anomaly detection threshold; wherein the quantile is selected from any one of the 0.95 to 0.99 quantiles.

[0011] In one possible implementation, the graph structure data for each time moment is determined based on multi-source data of the power system aligned to a preset time grid. This includes: collecting multi-source data of the power system and aligning the multi-source data to the same preset time grid to obtain synchronization data for each time moment; for each time moment, determining the multi-dimensional observation vector and missing mask vector based on the synchronization data for that time moment; wherein the dimension of the missing mask vector is the same as the dimension of the multi-dimensional observation vector; and determining the graph node feature matrix and graph mask matrix for that time moment based on the multi-dimensional observation vector and missing mask vector, the topology graph, and the parameter version, thereby obtaining the graph structure data for each time moment.

[0012] In one possible implementation, the time interval of the preset time grid includes any one of 1 min, 5 min, and 15 min.

[0013] In one possible implementation, the multi-source data includes operational data from at least two different sources; the operational data includes power source operational data, grid-side operational data, load-side operational data, energy storage-side operational data, and meteorological-side operational data.

[0014] A second aspect of this disclosure provides an anomaly detection and repair device for multi-source data from power grid, load, and storage systems, comprising: a graph data determination unit, used to determine graph structure data at each time point based on multi-source data of a power system aligned to a preset time grid; wherein the graph structure data includes a graph node feature matrix and a graph mask matrix; a reconstruction unit, used to reconstruct the graph structure data at each time point using a pre-built graph time-series reconstruction model to obtain the reconstruction result at that time point; wherein the graph time-series reconstruction model uses a joint loss function including a mask reconstruction term and a physical consistency term for parameter updates during the training phase; a score determination unit, used to determine a comprehensive score at each time point based on the mask reconstruction residual and the physical consistency residual of the reconstruction result; an anomaly detection unit, used to mark times with a comprehensive score greater than an anomaly threshold as an anomaly time point; and a repair unit, used to perform physical consistency correction on the reconstruction result at the anomaly time point to obtain a target repair value that satisfies preset operating boundary constraints, and repair the graph structure data at the anomaly time point based on the target repair value.

[0015] This disclosure also provides an electronic device, comprising: a memory for storing at least one instruction; and a processor for calling the instruction stored in the memory to execute the anomaly detection and repair method for multi-source data of source-network-load-storage in the first aspect and any embodiment of the first aspect.

[0016] This disclosure also provides a computer-readable storage medium storing at least one executable instruction, which is loaded and executed by a processor to implement the anomaly detection and repair method for multi-source data from source, network, load, and storage in the first aspect and any embodiment of the first aspect.

[0017] This disclosure also provides a computer program product, which includes: computer program code, which, when executed by a computer, causes the computer to perform the anomaly detection and repair method for multi-source data of source-network-load-storage in the first aspect and any embodiment of the first aspect.

[0018] Compared with the prior art, this disclosure has the following beneficial effects: By structuring multi-source data into graph-structured data that integrates power grid topology correlations and measurement validity identifiers, a graph time-series reconstruction model containing mask reconstruction terms and physical consistency terms is constructed and trained to achieve robust reconstruction of multi-dimensional time-series data. During the inference phase, the mask reconstruction residuals and physical consistency residuals are used for fusion scoring, thereby improving the accuracy of anomaly detection and reducing false alarms and false negatives. During the repair phase, the reconstruction results are used as candidates, and the target repair value is solved through constraint optimization by combining operational boundary constraints and physical consistency constraints. This ensures that the repaired data meets physical rules such as power balance, energy conservation, and equipment operating boundaries, improving the usability and reliability of the repaired data. This reduces the impact of data quality issues on prediction and control applications, and is conducive to the safe, stable, and economical operation of the new power system. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for anomaly detection and repair of multi-source data from source, grid, load, and storage provided in Embodiment 1 of this disclosure; Figure 2 This is a schematic diagram of the overall processing flow of an anomaly detection and repair method for multi-source data from source, grid, load, and storage provided in Embodiment 1 of this disclosure; Figure 3 This is a flowchart of a method for determining graph structure data at various times, provided in Embodiment 2 of this disclosure; Figure 4 This is a flowchart of a method for determining a reconstruction result provided in Embodiment 3 of this disclosure; Figure 5 This is a schematic diagram of the input, output, and structure of a graph timing reconstruction model provided in Embodiment 3 of this disclosure; Figure 6 This is a flowchart of a method for determining a comprehensive score provided in Embodiment 4 of this disclosure; Figure 7 This is a structural block diagram of an anomaly detection and repair device for multi-source data from source, grid, load, and storage provided in Embodiment 5 of this disclosure. Detailed Implementation

[0020] The present disclosure will now be further described with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present disclosure and should not be construed as limiting the scope of protection of the present disclosure. It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application.

[0021] The acquisition, transmission, storage, use, and processing of data in this disclosed technical solution comply with relevant national laws and regulations. In the embodiments of this disclosure, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this disclosure, and do not imply that the applicant has already used or necessarily used such solutions.

[0022] All terms used in this disclosure have the same meaning as understood by one of ordinary skill in the art to which this disclosure pertains, unless otherwise specifically defined. It should also be understood that terms defined in general dictionaries should be interpreted as having meanings consistent with their meanings in the context of the relevant art, and not as idealized or highly formalized, unless expressly defined herein.

[0023] Due to the strong heterogeneity of multi-source data in the power system under the source-grid-load-storage scenario, such as significant differences in time granularity, dimensionality, accuracy and sampling mechanism between different sources; and unstable quality, such as sensor failure, communication delay, sampling error, point-to-meter mismatch, unsynchronized topology changes, parameter drift, etc., data may be missing, noisy, abrupt, drift, out of bounds and inconsistent with the physical data; therefore, data processing is required to reduce system operation errors.

[0024] However, existing data processing methods are fragmented and prone to false alarms or missed alarms. Specifically, these include: statistical methods, such as Z-score (Standard Score), IQR (Interquartile Range), and control charts, which are poorly adaptable to non-Gaussian distributions, multidimensional coupling, and operating condition switching, and are prone to false alarms; machine learning or deep learning methods, such as isolated forests, OCSVM (One-Class Support Vector Machine), and autoencoders, which usually lack constraints from the laws of electrical physics and are prone to misjudgment when faced with topological coupling and parameter uncertainty; and power mechanism verification, such as power flow balance, energy conservation, and equipment boundary verification, which are sensitive to noise, measurement sparsity, and parameter deviations, and are difficult to cover complex dynamic environments.

[0025] Furthermore, the above methods typically only output anomaly markers. If the processed data values ​​are directly used for system prediction and scheduling optimization, they will still generate significant system operation errors.

[0026] Example 1 provides a method for anomaly detection and repair of multi-source data from source, grid, load, and storage, such as... Figure 1 As shown, the anomaly detection and repair method may include the following steps S11 to S15.

[0027] Step S11: Based on the multi-source data of the power system that has been aligned to the preset time grid, determine the graph structure data at each time point.

[0028] It should be noted that, in this embodiment of the disclosure, multi-source data refers to various operation-related data originating from different links or monitoring devices in the power system. Multi-source data includes operation data from at least two different sources. In one possible implementation, multi-source data includes power source side operation data, grid side operation data, load side operation data, energy storage side operation data, and meteorological side operation data.

[0029] A preset time grid refers to a standardized time interval set in advance to achieve time synchronization of multi-source data. Its function is to normalize multi-source data with different acquisition frequencies and times onto the same time dimension, ensuring data consistency in the time dimension. In one possible implementation, the time interval of the preset time grid includes any one of 1 minute, 5 minutes, and 15 minutes, and the specific selection can be determined according to the operating characteristics of the power system and data acquisition requirements.

[0030] Graph structured data refers to structured data on the operating status of a power system at a specific moment. Specifically, graph structured data includes a graph node feature matrix and a graph mask matrix. The graph mask matrix is ​​used to identify the validity of each element in the graph node feature matrix, in order to distinguish between valid and invalid data and prevent interference from invalid data.

[0031] Furthermore, it should be noted that in this embodiment, graph structure data can be constructed based on multi-source data, topology diagrams, and parameter versions of the power system. This allows the topology information and operational data of the power system to be integrated into a unified structured form, adapting to the input requirements of subsequent graph time-series reconstruction models. The topology diagrams and parameter versions are pre-acquired and maintained; the topology diagrams depict the connection relationships between various devices in the power system, while the parameter versions map the device parameter configurations under different operating conditions.

[0032] Step S12: For each time step, determine the reconstruction result of that time step using the pre-built graph time series reconstruction model.

[0033] It should be noted that the graph time-series reconstruction model is used to reconstruct the time-series data of the graph structure of the power system. In this embodiment of the disclosure, a segment of continuous time-series data from a historical time window can be selected as input to provide time-series context information for the graph time-series reconstruction model, helping the model to capture the time-series evolution patterns of the data and improve the time-series rationality of the reconstruction results.

[0034] Specifically, the graph time-series reconstruction model can output a value most likely to occur under normal operating conditions based on historical patterns (time dependence), topological adjacency relationships (spatial coupling), and the normal manifold learned during training within a time window. Therefore, even if some dimensions of the graph structure data are missing or abnormal at the current moment, the graph time-series reconstruction model can still infer reasonable estimates using the remaining dimensions, adjacent nodes, and historical sequences, making the output reconstruction result statistically closer to the normal state. Thus, the reconstruction result in this embodiment naturally has the effect of "replacing abnormal or missing values ​​with reasonable estimates," which is equivalent to performing preliminary repair on the original data. However, this reconstruction result may not strictly meet all the physical constraints of the power system, so further optimization is still required.

[0035] Furthermore, it should be noted that the graph time-series reconstruction model can use a joint loss function that includes a mask reconstruction term and a physical consistency term for parameter updates during the training phase. The training phase refers to the period before the graph time-series reconstruction model is officially deployed, during which it learns parameters using a large amount of sample data. The joint loss function is a loss function composed of multiple loss terms used to guide the training process of the graph time-series reconstruction model. In one possible implementation, the joint loss function includes a mask reconstruction term and a physical consistency term. The mask reconstruction term measures the deviation between the model's reconstruction result and the valid data (identified by the graph mask matrix) in the original graph structure data, while the physical consistency term measures whether the reconstruction result conforms to the physical constraints of the power system. By using a joint loss function that includes both types of loss terms for parameter updates, the dual objectives of "data fitting accuracy" and "physical rationality" can be optimized, ensuring that the reconstruction patterns learned by the model both fit the actual data distribution and conform to the operating mechanism of the power system. In another possible implementation, the joint loss function also includes a time consistency term. The time consistency term is a loss term used to measure the degree of matching between the time-series variation pattern of the reconstruction results at adjacent time points and the actual operating sequence characteristics of the power system. Adding a time consistency term to the joint loss function can further constrain the time-series smoothness and continuity of the reconstruction results on the basis of "data fitting accuracy" and "physical rationality", thus constructing a triple optimization objective of "data fitting", "physical constraint" and "time constraint".

[0036] Step S13: Determine the overall score at time based on the mask reconstruction residual and physical consistency residual of the reconstruction result.

[0037] It should be noted that in this embodiment of the disclosure, the mask reconstruction residual refers to the deviation value between the reconstruction result and the valid data in the graph structure data (identified by the graph mask matrix). The larger the deviation, the greater the difference between the reconstruction result and the original data, and the more likely there is an anomaly.

[0038] Physical consistency residuals refer to the aggregated result of various residual values ​​arising from the reconfiguration result violating the physical constraints of the power system. These physical constraints refer to the physical laws and technical limitations that the power system must follow during operation, including power balance constraints, power flow constraints, energy storage conservation constraints, energy storage state-of-charge dynamic constraints, and equipment boundary constraints. In one possible implementation, the physical consistency residual is the aggregated value of one or more of the following residual types: power balance residual, power flow consistency residual, energy storage conservation residual, energy storage state-of-charge dynamic residual, and equipment boundary residual. For example, in a specific implementation, the physical consistency residual is the aggregated value of the power balance residual, energy storage conservation residual, and equipment boundary residual.

[0039] In one possible implementation, the physical consistency residual is determined through a pre-built physical consistency operator model. It should be noted that the physical consistency operator model includes a physical consistency operator, which may further include a power balance residual / power flow consistency residual calculation module, an energy storage energy conservation residual / energy storage state of charge dynamic residual calculation module, a device boundary residual calculation module, and a residual aggregation module. These modules are used to calculate the power balance residual / power flow consistency residual, the energy storage energy conservation residual / energy storage state of charge dynamic residual, and the device boundary residual, respectively, and to aggregate the various residuals into a physical consistency residual.

[0040] In this embodiment of the disclosure, the comprehensive score is determined based on the mask reconstruction residual and the physical consistency residual. It takes into account two dimensions, namely "data fit" and "physical rationality", to comprehensively evaluate whether there is anomaly in the data at the current moment, avoid misjudgment caused by single-dimensional evaluation, and improve the accuracy of anomaly detection.

[0041] Step S14: Mark the moment when the comprehensive score is greater than the anomaly judgment threshold as an abnormal moment.

[0042] In one possible implementation, the quantile of the physical consistency residuals corresponding to all normally operating samples during the training phase is determined as the anomaly detection threshold; the quantile is selected from any one of the 0.95 to 0.99 quantiles.

[0043] Step S15: Perform physical consistency correction on the reconstruction results at the abnormal moment to obtain the target repair value that satisfies the preset running boundary constraints, and repair the graph structure data at the abnormal moment based on the target repair value.

[0044] In one possible implementation, the reconstruction results at abnormal moments are physically consistent to obtain a target repair value that satisfies preset operating boundary constraints. Specifically, this may include: under preset operating boundary constraints, using the reconstruction results at abnormal moments as candidate repair values, and through a predefined constraint optimization formula, physically consistent to correct the reconstruction results at abnormal moments to obtain the target repair value; wherein, the preset operating boundary constraints include voltage boundary constraints, branch power boundary constraints, and energy storage state of charge boundary constraints.

[0045] The following combination Figure 2 Taking graph structure data, including graph node feature matrix and graph mask matrix, as an example, the processing flow of the anomaly detection and repair method provided in Embodiment 1 of this disclosure is described in its entirety.

[0046] like Figure 2 As shown, the overall processing flow of the anomaly detection and repair method includes: obtaining the graph node feature matrix and graph mask matrix, specifically, obtaining the graph node feature matrix and graph mask matrix at the current time and historical time windows; reconstructing the graph node feature matrix, specifically, inputting the graph node feature matrix and graph mask matrix at the current time and historical time windows, along with the current power system topology (reflecting node connection relationships) and the current parameter set (reflecting equipment operating parameter configurations), into the graph time-series reconstruction model, which reconstructs the graph node feature matrix at the current time to obtain the reconstructed graph node feature matrix; and calculating... The mask reconstruction residual and physical consistency residual are calculated as follows: Based on the graph mask matrix, valid measurement points are selected, and the numerical error aggregation value between the original graph node feature matrix and the reconstructed graph node feature matrix at valid points is calculated to obtain the mask reconstruction residual. Simultaneously, through the physical consistency operator model, various residuals that deviate from the power system's physical constraints are calculated and weighted to obtain the physical consistency residual. A comprehensive score is calculated: based on the mask reconstruction residual and the physical consistency residual, a weighted sum is performed using preset weighting coefficients to obtain the comprehensive score at the current moment. This score is used to comprehensively evaluate the "data fitting accuracy" and "physical rationality" of the reconstruction result. The comprehensive score is compared with an anomaly judgment threshold. If the comprehensive score is greater than the anomaly judgment threshold, anomaly marking and repair are performed. Specifically, the current moment is marked as an anomaly moment, and based on the reconstruction result at the anomaly moment, a target repair value that satisfies the operating boundary constraints is solved using a constraint optimization algorithm. The original graph structure data at the anomaly moment is replaced with the target repair value to complete the anomaly repair.

[0047] The anomaly detection and repair method provided in Embodiment 1 of this disclosure converts multi-source data into graph-structured data that integrates topological associations and measurement validity. It then combines a graph temporal reconstruction model that balances data fitting accuracy and physical operating rules for temporal modeling and data reconstruction. Next, it calculates a comprehensive score using a two-dimensional residual index to achieve accurate anomaly determination. Finally, it completes anomaly data repair by solving for the target repair value based on constraint optimization. Not only is physical consistency residual calculated for anomaly determination during the inference phase, but physical consistency is also used as a loss term / constraint during the training phase, making the model output closer to the physical feasible region, thereby reducing false positives and false negatives and improving repair usability.

[0048] Example 2, based on Example 1, provides a method for determining graph structure data at various times, such as... Figure 3 As shown, in one possible implementation, step S11 may specifically include steps S101 to S103.

[0049] Step S101: Collect multi-source data from the power system and align the multi-source data to the same preset time grid to obtain synchronization data at each time point.

[0050] The multi-source data includes power supply side operation data, grid side operation data, load side operation data, energy storage side operation data, and meteorological side operation data. The preset time grid interval can be 5 minutes.

[0051] Step S102: For each time step, based on the synchronization data at that time step, determine the multidimensional observation vector and the missing mask vector at that time step.

[0052] The dimension of the missing mask vector is the same as the dimension of the multidimensional observation vector.

[0053] In one possible implementation, the method further includes initial screening of the synchronization data at each time point: for each time point, out-of-bounds values ​​of the synchronization data at that time point are removed, wherein out-of-bounds values ​​include at least one of negative voltage values, SOC values ​​greater than 1, and power values ​​exceeding the system's rated upper limit; and the initial screening of the synchronization data at each time point is normalized.

[0054] In one possible implementation, for each time-series of synchronized data, a multidimensional observation vector and a missing mask vector for that time are determined. Specifically, this may include: extracting at least one observed variable for each time-series of synchronized data; organizing the data into a multidimensional observation vector based on the at least one observed variable for that time, and defining a corresponding missing mask vector. The elements of the missing mask vector are either 1 or 0, where a value of 1 indicates that the corresponding data point is valid, and a value of 0 indicates that the corresponding data point is missing.

[0055] Step S103: Based on the multidimensional observation vector and missing mask vector at each time point, and the topology graph and parameter version, determine the graph node feature matrix and graph mask matrix at each time point to obtain the graph structure data at each time point.

[0056] The method for determining graph structure data at various times provided in Embodiment 2 of this disclosure transforms dispersed and heterogeneous multi-source power system data into graph structure data that combines topological correlation and measurement validity, providing a high-quality and standardized input foundation for subsequent graph time-series reconstruction models.

[0057] Example 3, based on any of the foregoing examples, provides a method for determining the reconstruction result, such as... Figure 4 As shown, in one possible implementation, the graph temporal reconstruction model may include a graph temporal autoencoder, which includes a spatial encoder, a temporal encoder, and a decoder. The above step S12 may specifically include the following steps S201 to S203.

[0058] Step S201: Based on the graph structure data of all times within the historical time window, the topology graph and parameter version are used to aggregate spatial features of the graph structure data through a spatial encoder to obtain a spatial representation.

[0059] The graph structure data includes a graph node feature matrix and a graph mask feature matrix. The historical time window is a continuous time interval formed by truncating a preset time length from the current time as the end time.

[0060] Step S202: Based on the spatial representation, the temporal sequence of the spatial representation is modeled by a time encoder to obtain the temporal latent variables.

[0061] Step S203: Based on the temporal latent variables, determine the reconstruction result at each moment using a decoder.

[0062] The reconstruction results include the reconstructed graph node feature matrix.

[0063] For ease of understanding, the following is combined with Figure 5 The processing flow of the graph time series reconstruction model is illustrated with an example.

[0064] like Figure 5As shown, the processing flow includes: inputting the graph node feature matrix and graph mask matrix corresponding to the historical time window at the current moment, the current topology graph, and the current parameter set; using the spatial encoder of the graph time-series autoencoder, combined with the input topology graph and parameter set, performing spatial dimension feature aggregation on the graph node feature matrix and graph mask matrix within the historical time window, capturing the connection relationships and spatial correlation features between nodes in the power system, and obtaining a spatial representation; inputting the spatial representation output by the spatial encoder into the time encoder, performing temporal dimension modeling on the spatial feature sequence within the historical time window, learning the temporal evolution law of power system operation data, and finally obtaining the temporal latent variable corresponding to the current moment (a high-dimensional representation that integrates spatial and temporal features); inputting the temporal latent variable into the decoder of the graph time-series autoencoder, and mapping the high-dimensional temporal latent variable back to the graph node feature space through decoding operations, and finally outputting the reconstructed graph node feature matrix at the current moment.

[0065] The method for determining the reconstruction result provided in Embodiment 3 of this disclosure reconstructs the graph node feature matrix at the current moment by combining graph time-series autoencoder with graph structure data of historical time windows. By utilizing the time correlation and dynamic continuity of source-grid-load-storage data (such as load inertia, equipment ramp-up, and SOC continuity), reasonable values ​​can still be recovered in the case of missing / noise / abrupt changes, while reducing misjudgment.

[0066] Example 4, based on any of the foregoing examples, provides a method for determining a comprehensive score, such as... Figure 6 As shown, in one possible implementation, step S13 may specifically include steps S301 to S303.

[0067] Step S301: Determine the mask reconstruction residual between the reconstruction result and the graph structure data.

[0068] In one possible implementation, the mask reconstruction residual includes the aggregated numerical error between the original graph node feature matrix and the reconstructed graph node feature matrix in the effective measurement dimension. By calculating the numerical deviation between the reconstruction result and the original data only for the effective measurement data, invalid data from missing sites is avoided from interfering with the accuracy of the residual calculation, thereby ensuring the reliability of the subsequent comprehensive score.

[0069] Step S302: Determine the physical consistency residual between the reconfiguration result and the physical constraints of the power system.

[0070] In one possible implementation, a physical consistency operator model is used to calculate the physical consistency residuals, which include the aggregated value of the consistency deviations between the reconstructed graph node feature matrix and power balance, energy conservation, and boundary constraints. By measuring whether the reconstruction results conform to the actual operating logic of the power system, situations where the reconstruction results violate physical rules are avoided.

[0071] Step S303: Determine the comprehensive score at any given time based on the mask reconstruction residual and the physical consistency residual.

[0072] In one possible implementation, the overall score is the sum of the product of a preset weighting coefficient and the physical consistency residual, and the mask reconstruction residual. The preset weighting coefficient can be set to a value between 0.1 and 0.9, and the specific value can be adjusted according to the operating scenario of the power system. Since the mask reconstruction residual and the physical consistency residual evaluate the quality of the reconstruction result from the data level and the physical level, respectively, by adjusting the preset weighting coefficient, the overall score can be made to better fit the evaluation needs of different power system operating scenarios, thereby improving the accuracy of subsequent anomaly detection.

[0073] The method for determining the comprehensive score provided in Embodiment 4 of this disclosure combines the calculation of numerical error in the effective measurement dimensions with the evaluation of consistency deviation in multi-dimensional physical constraints. This not only avoids the interference of missing data on residual calculation and ensures the accuracy of the data-level evaluation, but also filters out reconstruction results that violate the operating logic of the power system through physical consistency constraints, ensuring the comprehensiveness of the evaluation dimensions.

[0074] For ease of understanding, the following example illustrates the overall process of anomaly detection and repair for multi-source data (source, grid, load, and storage) using the methods described in the above embodiments. This overall process specifically includes steps S1 to S7.

[0075] S1. Data Acquisition and Preprocessing.

[0076] Multi-source data acquisition: Collect operational data from the power supply side, grid side, load side, energy storage side, and meteorological side.

[0077] Time synchronization and resampling: unify the data from each source to a preset time grid, such as 1min / 5min / 15min; aggregate high-frequency data and interpolate or hold forward low-frequency data.

[0078] Unified timing input definition: Synchronized data at time... Organized as multidimensional observation vectors or matrix expansion vectors: ,in, Represents the real number field. d The multidimensional observation vector representing time t The dimension / feature count refers to the total number of features obtained by splicing together or flattening the matrix of measurements such as source / grid / load / storage / meteorology according to a predetermined caliber under a unified time grid. A missing mask is also defined. ,in, Indicates the first Dimensional measurement is valid. 0 indicates missing.

[0079] Coarse screening and normalization: Remove obviously out-of-range values, such as negative voltage, SOC>1, power exceeding the rated upper limit, etc.

[0080] Standardize by feature dimension: ,in Let be the mean of the k-th feature dimension during the training period. The standard deviation of the k-th feature dimension during the training period is, i.e. , These are statistics for the training period and are fixed during the inference period.

[0081] S2. Topology graph construction and feature mapping.

[0082] Topology diagram definition: The electrical diagram is constructed based on the campus / distribution network topology. ,in This represents the set of busbars, equipment, and measuring points. It represents the connection relationships of lines, transformers, and switches; and defines the set of power grid and equipment parameters Θ, such as line impedance, turns ratio, capacity limit, energy storage capacity and efficiency, etc.

[0083] Graph node feature matrix: (Time intervals) The measurement is mapped to the graph node feature matrix. ,in, This indicates the number of nodes in the topology graph. f This indicates the number of features corresponding to each node, with each row corresponding to a node. Features such as voltage amplitude / phase angle or equivalent state quantities, active / reactive power, branch power flow, current, energy storage SOC / power, and weather conditions.

[0084] Graph mask matrix: Define the missing mask matrix This is used to identify whether the feature dimension of each node is valid.

[0085] S3. Construct a physically consistent embedded graph-time autoencoder.

[0086] Model input: Time window of As input.

[0087] Model structure: Constructing a graph temporal autoencoder, including: Spatial encoder: Graph convolutional / message passing networks for topology graphs Spatial representation is obtained by aggregating the upper adjacency relationships. ; Time encoder: Gated Recurrent Unit (GRU) model / Long Short-Term Memory (LSTM) network model / Transformer model pair Sequence modeling yields latent variables ; Decoder: by Output the feature matrix of the reconstructed graph nodes .

[0088] Calculate the mask reconstruction residual: Calculate the reconstruction error only for valid measurements, and define... ,in This is an element-wise multiplication.

[0089] S4. Construct a physical consistency operator and embed it into the training target.

[0090] Physical consistency operator definition: Constructing power mechanism consistency operators It is used to measure the consistency deviation between the reconstruction results and the power mechanism (such as power balance, energy conservation, boundary constraints, etc.).

[0091] A consistency operator contains at least one or more of the following types of residuals: Power balance / power flow consistency residuals: For nodes definition and ; in, These represent the active and reactive power injected into the node (from measurements). These represent the theoretical injected active power and power calculated from the power flow model, respectively. These represent the residuals of active and reactive power imbalance, respectively. It can be calculated by AC power flow operator or linearized power flow operator; Energy storage energy conservation and SOC dynamic residuals (e.g., for energy storage nodes) SOC dynamic residual): ; Equipment boundary residuals (taking voltage boundary as an example): ; Physically Consistent Residual Aggregation: Definition ; Joint training objective (i.e., mechanistic embedded learning): ; in These are time-series consistency terms, including rate of change constraints, smoothing terms, and SOC continuity terms; For mask reconstruction items; This is a physical consistency item.

[0092] It should be noted that the training process minimizes the weighted sum. In joint training, mechanistic constraints are incorporated into the training objective, enabling the model to learn physically feasible normal manifolds, rather than simple posterior validation.

[0093] S5. Anomaly detection and judgment during the reasoning phase.

[0094] Inference output: Calculated from the input window. and .

[0095] Comprehensive Abnormal Score: Definition ;in This is a weighting coefficient, which can be set as a constant or determined by factors such as missing rate, communication latency, and data quality score.

[0096] Threshold determination: The threshold is determined using the quantiles of normal samples during the training period. Where q is the quantile ratio parameter used to determine the anomaly detection threshold. q=0.95 corresponds to the 95th quantile, which means that the aggregated value of the physical consistency residuals of 95% of the normal samples in the training set is less than or equal to the value corresponding to this quantile.

[0097] when Judgment Time An anomaly exists.

[0098] S6, Physical Consistency Repair (Candidate Reconstruction + Constraint Correction / Optimization).

[0099] Candidate repair values: based on model reconstruction results As a candidate repair value (which has been trained under mechanistic constraints).

[0100] Constraint Correction / Optimization Repair: Solving for the final repair value .

[0101] Satisfy operating boundary constraints: Voltage boundary constraints Branch power boundary constraints Energy storage charge state boundary constraints .

[0102] in This is a weight matrix (abnormal dimensions have small weights, and trustworthy dimensions have large weights). The coefficient is the consistency enhancement coefficient; it can be solved online using projection gradient, augmented Lagrange, or ADMM (Alternating Direction Method of Multipliers).

[0103] Output confidence level: based on and The mapping generates repair confidence and records repair logs (time, measurement points, type, model version, topology / parameter version, etc.).

[0104] S7. Output results and system integration.

[0105] Mark the anomalies and repair the data. Confidence level and traceability information are written into a unified data base; APIs are provided for external systems such as load forecasting, source-grid-load-storage optimization scheduling, and power trading analysis. The interface output includes at least: raw data, repaired data, anomaly tags, confidence level, and traceability information.

[0106] Compared with the prior art, this disclosure has at least the following effects: (1) Mechanism constraint endogenization: Power balance, energy conservation, equipment boundary, etc. are embedded into the deep model training target in the form of consistency operators, so that the model learns the physically feasible normal manifold and reduces false alarms and false negatives.

[0107] (2) Dual residual interpretable detection: Simultaneously output reconstruction residual and physical consistency residual, which can distinguish between statistical anomalies and physical inconsistency anomalies, and improve the efficiency of operation and maintenance location.

[0108] (3) Physically feasible repair: Model reconstruction is used as candidate repair values, and the final repair value is generated through physical consistency correction / constraint optimization to ensure that the boundary and mechanism constraints such as voltage, power flow, and SOC are met, and it can be directly used for prediction and scheduling optimization.

[0109] (4) Adaptation to missing and heterogeneous structures: Mask reconstruction and topology graph modeling are adopted to make full use of temporal correlation and topology correlation, making it more robust to missing, noise and heterogeneous structures.

[0110] (5) Can be engineered and implemented: outputs anomaly markers, repair data and confidence level, and supports writing to a unified data base and provides APIs for downstream systems to call.

[0111] Example 5 provides an anomaly detection and repair device for multi-source data from source, grid, load, and storage, such as... Figure 7 As shown, the anomaly detection and repair device 100 may include a graph data determination unit 110, a reconstruction unit 120, a score determination unit 130, an anomaly detection unit 140, and a repair unit 150.

[0112] The graph data determination unit 110 is used to determine the graph structure data at each time point based on multi-source data of the power system that has been aligned to a preset time grid.

[0113] The reconstruction unit 120 is used to reconstruct the graph structure data of each time step using a pre-built graph temporal reconstruction model to obtain the reconstruction result of that time step. The graph temporal reconstruction model uses a joint loss function containing a mask reconstruction term and a physical consistency term for parameter updates during the training phase.

[0114] The score determination unit 130 is used to determine the comprehensive score at a given time based on the mask reconstruction residual and the physical consistency residual of the reconstruction result; wherein, the physical consistency residual is calculated by a pre-built physical consistency operator model.

[0115] Anomaly detection unit 140 is used to mark moments when the overall score is greater than the anomaly threshold as an anomaly moment.

[0116] Repair unit 150 is used to perform physical consistency correction on the reconstruction results at abnormal times, obtain target repair values ​​that meet preset running boundary constraints, and repair the graph structure data at abnormal times based on the target repair values.

[0117] For specific details and benefits of the anomaly detection and repair device for multi-source data of source, grid, load and storage provided in the embodiments of this disclosure, please refer to the above description of the anomaly detection and repair method for multi-source data of source, grid, load and storage, which will not be repeated here.

[0118] This disclosure also provides an electronic device, comprising: a memory for storing at least one instruction; and a processor for calling the instruction stored in the memory to execute the anomaly detection and repair method for multi-source data of source-network-load storage in any of the above embodiments.

[0119] This disclosure also provides a computer-readable storage medium storing at least one executable instruction, which is loaded and executed by a processor to implement the anomaly detection and repair method for multi-source data from source, network, load, and storage in any of the above embodiments.

[0120] This disclosure also provides a computer program product, which includes computer program code. When the computer program code is run by a computer, the computer executes the anomaly detection and repair method for multi-source data from source, network, load, and storage in any of the above embodiments.

[0121] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0122] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0123] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0124] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0125] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0126] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0127] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0128] It should be noted that the terms "first," "second," and similar terms used in this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different parts. Terms such as "including" or "contains" mean that the element preceding the word covers the element listed after the word, and do not exclude the possibility of covering other elements as well.

[0129] Although operations are described in a specific order in the accompanying drawings in this disclosure, it should not be construed as requiring these operations to be performed in the specific order or serial order shown, or requiring all of the shown operations to obtain the desired result. In certain environments, multitasking and parallel processing may be advantageous.

[0130] Finally, it should be noted that the above content is only used to illustrate the technical solution of this disclosure, and is not intended to limit the scope of protection of this disclosure. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of this disclosure do not depart from the substance and scope of the technical solution of this disclosure.

Claims

1. A method for anomaly detection and repair of multi-source data from source, grid, load, and storage, characterized in that, include: Based on multi-source data of the power system aligned to a preset time grid, the graph structure data at each time point is determined; wherein, the graph structure data includes a graph node feature matrix and a graph mask matrix; For each time point, the graph structure data at that time point is reconstructed using a pre-built graph time-series reconstruction model to obtain the reconstruction result at that time point. During the training phase, the graph time-series reconstruction model uses a joint loss function that includes a mask reconstruction term and a physical consistency term for parameter updates. The mask reconstruction term is a loss term used to measure the deviation between the reconstruction result and the valid data in the original graph structure data, and the physical consistency term is a loss term used to measure whether the reconstruction result conforms to the physical constraints of the power system. Based on the mask reconstruction residual and physical consistency residual of the reconstruction result, the comprehensive score at the specified moment is determined; wherein, the physical consistency residual is calculated by a pre-constructed physical consistency operator model; the physical consistency operator model includes a physical consistency operator, which includes a power balance residual / power flow consistency residual calculation module, an energy storage energy conservation residual / energy storage state of charge dynamic residual calculation module, an equipment boundary residual calculation module, and a residual aggregation module, used to calculate the power balance residual / power flow consistency residual, the energy storage energy conservation residual / energy storage state of charge dynamic residual, and the equipment boundary residual, respectively, and to aggregate the various residuals into a physical consistency residual; The moment when the comprehensive score is greater than the anomaly determination threshold is marked as an anomaly moment, and the anomaly detection is completed; Physical consistency correction is performed on the reconstruction results at the abnormal moment to obtain a target repair value that satisfies the preset running boundary constraints. Based on the target repair value, the graph structure data at the abnormal moment is repaired to complete the repair.

2. The method for anomaly detection and repair of multi-source data from source, grid, load, and storage as described in claim 1, characterized in that, The graph temporal reconstruction model includes a graph temporal autoencoder; the graph temporal autoencoder includes a spatial encoder, a temporal encoder, and a decoder; the process of reconstructing the graph structure data at the specified time using the pre-built graph temporal reconstruction model to obtain the reconstruction result at the specified time includes: Based on the graph structure data of all times within the historical time window of the stated time, and the topology and parameter versions, the spatial encoder performs spatial feature aggregation on the graph structure data to obtain a spatial representation; wherein, the historical time window is a continuous time interval formed by truncating a preset time length forward from the stated time as the end time. Based on the spatial representation, the temporal sequence of the spatial representation is modeled using the time encoder to obtain the temporal latent variables; Based on the temporal latent variables, the reconstruction result at the specified moment is determined by the decoder.

3. The method for anomaly detection and repair of multi-source data from source, grid, load, and storage as described in claim 1, characterized in that, The determination of the comprehensive score at the specified time step based on the mask reconstruction residual and physical consistency residual of the reconstruction result includes: Determine the mask reconstruction residual between the reconstruction result and the graph structure data; Determine the physical consistency residual between the reconstructed result and the physical constraints of the power system; Based on the mask reconstruction residual and the physical consistency residual, the comprehensive score at the specified time is determined; wherein the comprehensive score is the sum of the product of the preset weighting coefficient and the physical consistency residual and the mask reconstruction residual.

4. The method for anomaly detection and repair of multi-source data from source, grid, load, and storage as described in claim 1, characterized in that, The physical consistency correction of the reconstruction results at the abnormal moment to obtain the target repair value that satisfies the preset operating boundary constraints includes: Under preset operational boundary constraints, the reconstruction result at the abnormal moment is used as a candidate repair value. Through a predefined constraint optimization formula, the reconstruction result at the abnormal moment is physically consistent to obtain the target repair value. The preset operating boundary constraints include voltage boundary constraints, branch power boundary constraints, and energy storage state of charge boundary constraints.

5. The method for anomaly detection and repair of multi-source data from source, grid, load, and storage as described in any one of claims 1-4, characterized in that, The quantile of the physical consistency residuals corresponding to all normally operating samples during the training phase is determined as the anomaly detection threshold; wherein the quantile is selected from any one of the 0.95 to 0.99 quantiles.

6. The method for anomaly detection and repair of multi-source data from source, grid, load, and storage as described in claim 1, characterized in that, The determination of graph structure data at each time point based on multi-source power system data aligned to a preset time grid includes: Collect multi-source data from the power system and align the multi-source data to the same preset time grid to obtain synchronized data at each time point; For each said time moment, based on the synchronization data at said time moment, determine the multidimensional observation vector and the missing mask vector at said time moment; wherein, the dimension of the missing mask vector is the same as the dimension of the multidimensional observation vector; Based on the multidimensional observation vector and missing mask vector at the given time, and the topology graph and parameter version, the graph node feature matrix and graph mask matrix at the given time are determined, and the graph structure data at each time is obtained.

7. The method for anomaly detection and repair of multi-source data from source, grid, load, and storage as described in claim 1, characterized in that, The time interval of the preset time grid includes any one of 1 min, 5 min, and 15 min.

8. The method for anomaly detection and repair of multi-source data from source, grid, load, and storage as described in claim 1, characterized in that, The multi-source data includes operational data from at least two different sources; the operational data includes power source operational data, grid-side operational data, load-side operational data, energy storage-side operational data, and meteorological-side operational data.

9. A device for anomaly detection and repair of multi-source data from source, grid, load, and storage, characterized in that, include: The graph data determination unit is used to determine the graph structure data at each time point based on multi-source data of the power system that has been aligned to a preset time grid; wherein, the graph structure data includes a graph node feature matrix and a graph mask matrix; A reconstruction unit is used to reconstruct the graph structure data at each time point using a pre-built graph time-series reconstruction model to obtain the reconstruction result at that time point. The graph time-series reconstruction model uses a joint loss function containing a mask reconstruction term and a physical consistency term for parameter updates during the training phase. The mask reconstruction term is a loss term used to measure the deviation between the reconstruction result and the valid data in the original graph structure data, and the physical consistency term is a loss term used to measure whether the reconstruction result conforms to the physical constraints of the power system. A score determination unit is used to determine the comprehensive score at the specified moment based on the mask reconstruction residual and physical consistency residual of the reconstruction result; wherein, the physical consistency residual is calculated by a pre-constructed physical consistency operator model; the physical consistency operator model includes a physical consistency operator, which includes a power balance residual / power flow consistency residual calculation module, an energy storage energy conservation residual / energy storage state of charge dynamic residual calculation module, an equipment boundary residual calculation module, and a residual aggregation module, used to calculate the power balance residual / power flow consistency residual, the energy storage energy conservation residual / energy storage state of charge dynamic residual, and the equipment boundary residual respectively, and to aggregate the various residuals into a physical consistency residual; An anomaly detection unit is used to mark the moment when the comprehensive score is greater than the anomaly threshold as an anomaly moment; The repair unit is used to perform physical consistency correction on the reconstruction results at the abnormal moment, obtain a target repair value that satisfies the preset running boundary constraints, and repair the graph structure data at the abnormal moment based on the target repair value.

Images