CN-121980503-A - Anomaly detection and repair method and device for source network charge storage multi-source data

CN121980503ACN 121980503 ACN121980503 ACN 121980503ACN-121980503-A

Abstract

The disclosure provides an anomaly detection and repair method and device for source network charge storage multi-source data, and relates to the technical field of power system data processing. The method comprises the steps of determining graph structure data of all moments based on multisource data after synchronization of a power system, reconstructing the graph structure data of all moments through a graph time sequence reconstruction model to obtain a reconstruction result of all the moments, wherein the graph time sequence reconstruction model adopts a joint loss function comprising a mask reconstruction item and a physical consistency item to update parameters in a training stage, determining comprehensive scores of all the moments based on mask reconstruction residual errors and physical consistency residual errors of the reconstruction result, wherein the physical consistency residual errors are calculated by a physical consistency operator model, marking the moments with the comprehensive scores larger than an abnormality judgment threshold as abnormal moments, and repairing the graph structure data of the abnormal moments based on the reconstruction result of the abnormal moments. The method and the device can accurately detect abnormal data and effectively repair the abnormal data, so that subsequent operation errors of the system are reduced.

Inventors

LI TAOTAO
SHEN QIANFENG
CHANG YIXIN
WANG YUFAN
CHEN JIAXU
ZHANG MENG
MAO RUIYAN
LI DONGZE
LUO KE

Assignees

中能智新科技产业发展有限公司
特变电工新疆新能源股份有限公司

Dates

Publication Date: 20260505
Application Date: 20260120

Claims (10)

1. The method for detecting and repairing the abnormality of the source network charge storage multi-source data is characterized by comprising the following steps: Determining graph structure data of each moment based on multi-source data of the power system aligned to a preset time grid, wherein the graph structure data comprises a graph node characteristic matrix and a graph mask matrix; reconstructing graph structure data of the moment through a pre-constructed graph time sequence reconstruction model to obtain a reconstruction result of the moment, wherein the graph time sequence reconstruction model adopts a joint loss function comprising a mask reconstruction item and a physical consistency item to update parameters in a training stage; Determining the comprehensive score of the moment based on a mask reconstruction residual error and a physical consistency residual error of the reconstruction result, wherein the physical consistency residual error is calculated by a pre-constructed physical consistency operator model; marking the time when the comprehensive score is greater than an abnormality judgment threshold as an abnormality time to finish abnormality detection; And carrying out physical consistency correction on the reconstruction result at the abnormal moment to obtain a target repair value meeting the preset operation boundary constraint, and repairing the graph structure data at the abnormal moment based on the target repair value to finish the repair.
2. The method for anomaly detection and repair of source network load storage multisource data according to claim 1, wherein the graph time sequence reconstruction model comprises a graph time sequence self-encoder, the graph time sequence self-encoder comprises a space encoder, a time encoder and a decoder, the graph structure data at the moment is reconstructed through the pre-constructed graph time sequence reconstruction model, and a reconstruction result at the moment is obtained, and the method comprises the following steps: Based on graph structure data, topological graph and parameter version of all moments in a historical time window of the moment, spatial feature aggregation is carried out on the graph structure data through the spatial encoder to obtain spatial representation, wherein the historical time window is a continuous time interval formed by taking the moment as a termination moment and intercepting a preset time length forwards; Modeling a time sequence of the spatial representation through the time encoder based on the spatial representation to obtain a time sequence hidden variable; and determining a reconstruction result of the moment by the decoder based on the time sequence hidden variable.
3. The anomaly detection and repair method for source network load stored multi-source data according to claim 1, wherein the determining the composite score of the moment based on the mask reconstruction residual and the physical consistency residual of the reconstruction result comprises: Determining a mask reconstruction residual error of the reconstruction result and the graph structure data; determining a physical consistency residual error of the reconstruction result and the physical constraint of the power system; And determining the comprehensive score of the moment based on the mask reconstruction residual and the physical consistency residual, wherein the comprehensive score is the sum of the product of a preset weight coefficient and the physical consistency residual and the mask reconstruction residual.
4. The method for detecting and repairing anomalies of source network load storage multi-source data according to claim 1, wherein the step of performing physical consistency correction on the reconstruction result at the anomaly time to obtain a target repair value satisfying a preset operation boundary constraint comprises the steps of: Under the constraint of a preset operation boundary, taking a reconstruction result of the abnormal moment as a candidate restoration value, and carrying out physical consistency correction on the reconstruction result of the abnormal moment through a predefined constraint optimization to obtain the target restoration value; The preset operation boundary constraint comprises a voltage boundary constraint, a branch power boundary constraint and an energy storage charge state boundary constraint.
5. The method for anomaly detection and repair of source network charge storage multisource data according to any one of claims 1 to 4, wherein the physical consistency residual is an aggregate value of one or more types of residual among power balance residual, power flow consistency residual, stored energy conservation residual, stored state of charge dynamic residual, and equipment boundary residual.
6. The method for detecting and repairing anomalies of source network load multi-source data according to any one of claims 1 to 4, wherein the number of digits of the physical consistency residual errors corresponding to all normal operation samples in a training stage is determined as the anomaly judgment threshold, and the number of digits is selected from any one of 0.95 to 0.99 digits.
7. The anomaly detection and repair method for source network charge storage multisource data according to claim 1, wherein determining graph structure data at each time based on multisource data of a power system aligned to a preset time grid includes: the multi-source data of the power system are collected, and are aligned to the same preset time grid to obtain synchronous data of all moments; For each moment, based on the synchronous data of the moment, determining a multi-dimensional observation vector and a missing mask vector of the moment, wherein the dimension of the missing mask vector is the same as the dimension of the multi-dimensional observation vector; and determining a graph node characteristic matrix and a graph mask matrix at the moment based on the multidimensional observation vector and the missing mask vector at the moment, the topological graph and the parameter version, and obtaining graph structure data at each moment.
8. The method for detecting and repairing anomalies of source network charge storage multi-source data according to claim 1, wherein the time interval of the preset time grid comprises any one of 1min, 5min and 15 min.
9. The method for detecting and repairing anomalies of source network load storage multi-source data according to claim 1, wherein the multi-source data comprises at least two different sources of operation data, and the operation data comprises power source side operation data, power grid side operation data, load side operation data, energy storage side operation data and weather side operation data.
10. An anomaly detection and repair device for source network charge storage multi-source data is characterized by comprising: a graph data determining unit for determining graph structure data of each moment based on multi-source data of the power system aligned to a preset time grid, wherein the graph structure data comprises a graph node characteristic matrix and a graph mask matrix; A reconstruction unit, configured to reconstruct, for each time, graph structure data of the time through a pre-constructed graph time sequence reconstruction model, to obtain a reconstruction result of the time, where the graph time sequence reconstruction model performs parameter update in a training stage by adopting a joint loss function including a mask reconstruction term and a physical consistency term; The score determining unit is used for determining the comprehensive score of the moment based on a mask reconstruction residual error and a physical consistency residual error of the reconstruction result, wherein the physical consistency residual error is calculated by a pre-constructed physical consistency operator model; An anomaly detection unit for marking the time when the integrated score is greater than an anomaly threshold as an anomaly time; And the repairing unit is used for carrying out physical consistency correction on the reconstruction result at the abnormal moment to obtain a target repairing value meeting the preset operation boundary constraint, and repairing the graph structure data at the abnormal moment based on the target repairing value.

Description

Anomaly detection and repair method and device for source network charge storage multi-source data Technical Field The disclosure relates to the technical field of power system data processing, in particular to an anomaly detection and repair method and device for source network charge storage multi-source data. Background With the rapid development of the novel power system, a power source side (wind power, photovoltaic, conventional units), a power grid side (bus/branch/transformer/switch measurement and topology state), a load side (user electricity utilization and working condition characteristics), an energy storage side (battery/pumping and accumulating SOC and charging and discharging power) and a meteorological side (temperature, wind speed, irradiance and the like) continuously generate massive, multi-source and heterogeneous time sequence data. The data are important inputs of load prediction, power prediction, demand response, source network load storage cooperative control and scheduling optimization, and the quality of the data directly influences the safety of a power grid and the safe operation of a system. Because the data has strong isomerism and unstable quality, data processing is needed to reduce system operation errors. However, the existing data processing methods, such as a statistical method, a machine learning or deep learning method, an electric power mechanism checking method and the like, are split and are easy to report by mistake or miss, so that the operation error of the system is increased, and the safe and efficient operation of the novel electric power system is seriously restricted. Disclosure of Invention The invention provides an anomaly detection and repair method and device for source network load storage multi-source data, aiming at the problems in the prior art, and can solve the problem that in the prior art, the system operation error is increased due to method fracture and easy false alarm or false omission, thereby ensuring the safe and efficient operation of a novel power system. In order to achieve the above purpose, the technical scheme adopted in the present disclosure is as follows: The first aspect of the disclosure provides an anomaly detection and repair method for source network load storage multi-source data, which comprises the steps of determining graph structure data at each moment based on multi-source data of a power system which is aligned to a preset time grid, wherein the graph structure data comprise a graph node characteristic matrix and a graph mask matrix, reconstructing the graph structure data at each moment through a pre-constructed graph time sequence reconstruction model to obtain a reconstruction result at each moment, wherein the graph time sequence reconstruction model carries out parameter updating by adopting a joint loss function comprising a mask reconstruction item and a physical consistency item in a training stage, determining a comprehensive score at each moment based on a mask reconstruction residual and a physical consistency residual of the reconstruction result, wherein the physical consistency residual is obtained by calculating a pre-constructed physical consistency operator model, marking the moment with the comprehensive score larger than an anomaly determination threshold as the anomaly moment, completing anomaly detection, carrying out physical consistency correction on the reconstruction result at each anomaly moment to obtain a target repair value meeting a preset operation boundary constraint, and carrying out repair on the graph structure data at the anomaly moment based on the target repair value, and completing repair. In one possible implementation mode, the graph time sequence reconstruction model comprises a graph time sequence self-encoder, the graph time sequence self-encoder comprises a space encoder, a time encoder and a decoder, graph structure data of time is reconstructed through the pre-constructed graph time sequence reconstruction model to obtain a time reconstruction result, the graph structure data of all time in a time history time window based on time, a topological graph and a parameter version are subjected to space feature aggregation through the space encoder to obtain space representation, the history time window is a continuous time interval formed by taking the time as a termination time and intercepting a preset time length forwards, the time sequence of the space representation is modeled through the time encoder based on the space representation to obtain a time sequence hidden variable, and the time reconstruction result is determined through the decoder based on the time sequence hidden variable. In one possible implementation, the method comprises the steps of determining a mask reconstruction residual of a reconstruction result and a physical consistency residual of graph structure data, determining a physical consistency residual of the reconstruction result and physical constr