Fault monitoring method and system applied to smart grid

Abstract

The application provides a fault monitoring method and monitoring system applied to a smart power grid, through acquiring a multi-source operation time sequence data set of the smart power grid; performing time window division and feature extraction processing on the multi-source operation time sequence data set to generate a multi-scale time sequence feature sequence; inputting the multi-scale time sequence feature sequence into a fault evolution analysis model to perform state transition learning processing, generating a power grid state transition feature set; inputting the power grid state transition feature set into a fault type identification network to perform feature classification and position probability calculation processing, obtaining a fault type label and a fault position probability distribution; and determining a fault positioning result of the smart power grid according to the fault type label and the fault position probability distribution. Through the application, the accuracy of fault positioning can be effectively improved, and the accurate identification and positioning of the smart power grid fault can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fault monitoring, in particular to a fault monitoring method and system applied to a smart grid. BACKGROUND

[0002] With the continuous development of smart grid technology, fault monitoring is widely used, through the collection and analysis of power grid operation data, the technology of identifying and positioning faults in the power grid is realized, at present, the smart grid fault monitoring is usually through setting fixed threshold or simple feature comparison to judge the fault type and position, but it is difficult to fully capture the dynamic characteristics and evolution trend when the fault occurs, which easily leads to the limitation of the accuracy of fault identification and the precision of positioning, in the complex and changeable power grid operation environment, how to realize the accurate monitoring of faults has become a challenge to guarantee the reliable operation of the power grid. SUMMARY

[0003] The present application provides a fault monitoring method and system applied to a smart grid.

[0004] In a first aspect, the present application provides a fault monitoring method applied to a smart grid, comprising: acquiring a multi-source operation time series data set of the smart grid; performing time window division and feature extraction processing on the multi-source operation time series data set to generate a multi-scale time series feature sequence; inputting the multi-scale time series feature sequence into a fault evolution analysis model for state transition learning processing to generate a power grid state transition feature set; inputting the power grid state transition feature set into a fault type identification network for feature classification and position probability calculation processing to obtain a fault type label and a fault position probability distribution; determining a fault positioning result of the smart grid according to the fault type label and the fault position probability distribution.

[0005] In a second aspect, the present application provides a monitoring system, comprising: a memory, the memory storing a computer program; a processor for loading the computer program to realize the fault monitoring method applied to the smart grid as described above.

[0006] The application provides a fault monitoring method applied to a smart power grid, acquires a multi-source operation time sequence data set of the smart power grid, can comprehensively collect operation data of different monitoring objects in the power grid in continuous time, performs time window division and feature extraction processing on the multi-source operation time sequence data set to generate a multi-scale time sequence feature sequence, can capture feature performance of the fault under different time granularities through division of different time windows and feature extraction, can reflect instantaneous fluctuation features when the fault occurs and can also reflect trend change features in the fault development process, and avoids limitations of feature extraction under a single time scale, inputs the multi-scale time sequence feature sequence into a fault evolution analysis model to perform state transition learning processing to generate a power grid state transition feature set, can dynamically depict an evolution process of the fault from occurrence to development through learning of transition rules between power grid operation states, breaks through limitations of traditional static fault feature matching, and is more suitable for dynamic characteristics of the power grid fault, inputs the power grid state transition feature set into a fault type recognition network to perform feature classification and position probability calculation processing to obtain a fault type label and a fault position probability distribution, can simultaneously output a qualitative fault type and quantitative position probability information, and provides multi-dimensional judgment basis for fault positioning, determines a fault positioning result of the smart power grid according to the fault type label and the fault position probability distribution, and through collaborative decision of the type and the position probability, can effectively improve accuracy of fault positioning, and makes accurate identification and positioning of the smart power grid fault. BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 is a flow chart of a fault monitoring method applied to a smart power grid provided by an embodiment of the application.

[0008] Figure 2 is a composition schematic diagram of a monitoring system provided by an embodiment of the application. DETAILED DESCRIPTION

[0009] Please refer to Figure 1 , Figure 1 is a flow chart of a fault monitoring method applied to a smart power grid provided by an embodiment of the application. The method can be executed by a monitoring system and includes the following steps:

[0010] Step S100: acquiring a multi-source operation time sequence data set of the smart power grid.

[0011] The multi-source running time series data set is a data set arranged in time sequence from multiple different data sources of the smart grid. These data sources can cover various aspects of the smart grid, including but not limited to operation state monitoring data of power equipment, power transmission data of the power grid, voltage and current data, etc. For example, the operation state monitoring data of power equipment can be the oil temperature, winding temperature, vibration condition, etc. of the transformer; the power transmission data of the power grid can be the active power and reactive power of each transmission line; and the voltage and current data are the voltage amplitude, phase and current size, etc. of each node. These data are arranged in time sequence to form a data set with time series characteristics, which can reflect the operation state of the smart grid at different time points. The multi-source running time series data set can be obtained by deploying sensors at key positions of the smart grid. For the operation state monitoring data of power equipment, temperature sensors can be installed on the transformer to collect oil temperature and winding temperature data, and vibration sensors can be installed to obtain vibration condition data; for the power transmission data of the power grid, power sensors can be installed on the transmission line to measure active power and reactive power; and for the voltage and current data, voltage transformers and current transformers can be installed at each node to collect the corresponding data. The collected data is transmitted to the data storage center through a data transmission network, such as a wired network or a wireless network, for storage and management, and finally forms a multi-source running time series data set.

[0012] Step S200: performing time window division and feature extraction processing on the multi-source running time series data set to generate a multi-scale time series feature sequence.

[0013] Time window division is the process of dividing the continuous multi-source running time series data set into multiple non-overlapping or overlapping data segments according to a certain time interval. The size and sliding step of the time window can be set according to the specific application requirements and data characteristics. Feature extraction processing is to extract key features that can reflect the operation state of the smart grid from each divided data segment. The multi-scale time series feature sequence is a series of feature sequences obtained by time window division and feature extraction of different scales, which can reflect the operation state change of the smart grid from different time scales.

[0014] As an implementation, step S200 can specifically include steps S210-S280:

[0015] Step S210: obtaining a preset time window parameter set, performing window division on the multi-source running time series data set based on the short time window parameters in the time window parameter set to obtain a plurality of short time window data segments.

[0016] The preset time window parameter set is a set of time window parameters including short time window parameters, medium time window parameters, and long time window parameters. The short time window parameters are used to define the size and sliding step of the short time window, and can capture short-term changes in the smart grid operation state. The short time window data segment is a series of data segments obtained by dividing the multi-source operation time series data set by the short time window parameters, and each data segment contains multi-source operation time series data in a short time. The preset time window parameter set can be obtained by reading a configuration file, and the configuration file stores the values of various time window parameters.

[0017] As an implementation, step S210 can specifically include steps S211-S216:

[0018] Step S211: Extract short time window parameters from the time window parameter set, the short time window parameters including window length parameters and window sliding step parameters.

[0019] The time window parameter set is a set of time window parameters, and the short time window parameters are used to define the size and sliding rules of the short time window. The window length parameter determines the number of sampling data included in each short time window, reflecting the data range that can be observed in a short time window. The window sliding step parameter specifies the number of sampling data intervals between adjacent short time windows, and by adjusting this parameter, the overlap or interval size of window division can be controlled. The short time window parameters can be extracted from the time window parameter set by traversing the time window parameter set and obtaining the corresponding window length parameter and window sliding step parameter according to the preset parameter identifier. For example, in a time window parameter set stored in the form of key-value pairs, the key identifier of the window length parameter of the short time window is "short_window_length", and the key identifier of the window sliding step parameter is "short_window_step". By looking up the values corresponding to these two keys, the short time window parameters can be extracted.

[0020] Step S212: Determine the number of sampling data included in each short time window based on the window length parameter, and determine the number of sampling data intervals between adjacent short time windows based on the window sliding step parameter.

[0021] The window length parameter determines the number of sampling data covered by each short time window. Through this parameter, it can be determined how many sampling data of the smart grid operation state can be obtained in a short time window. The window sliding step parameter is also a numerical value, which is used to determine the sampling data interval between adjacent two short time windows. Based on the two parameters, the multi-source runtime sequence data set can be accurately windowed. In specific implementation, the window length parameter and the window sliding step parameter can be used as variables for calculation and processing in the program to determine the start and end positions of each short time window.

[0022] Step S213: Taking the starting sampling data of the multi-source runtime sequence data set as the starting point of the first short time window, the sampling data is intercepted according to the window length parameter to obtain the first short time window data segment.

[0023] The multi-source runtime sequence data set is a series of sampling data arranged in time sequence, and the starting sampling data is the first data point in the entire data set. Taking this starting sampling data as the starting point of the first short time window, according to the number of sampling data determined by the window length parameter, the corresponding number of sampling data is intercepted from the starting point to obtain the first short time window data segment. In actual operation, the starting point and the intercepted data can be located by data index, and the intercepted data is stored in a new data structure as the first short time window data segment.

[0024] Step S214: Taking the starting point of the first short time window as the reference, moving a corresponding number of sampling data backward according to the window sliding step parameter to determine the starting point of the second short time window, and intercepting the sampling data according to the window length parameter to obtain the second short time window data segment.

[0025] After obtaining the first short time window data segment, the starting point of the second short time window is determined by moving a corresponding number of sampling data backward from the starting point of the first short time window according to the window sliding step parameter. Then, the sampling data starting from the starting point is intercepted according to the window length parameter to obtain the second short time window data segment. Exemplarily, the starting point index of the second short time window can be obtained by adding the value of the window sliding step parameter to the index value of the starting point of the first short time window, and then the data interception operation is performed.

[0026] Step S215: Repeat the steps of determining the starting point of the short time window and intercepting the sampling data until the windowing of all sampling data of the multi-source runtime sequence data set is completed, and a plurality of short time window data segments are obtained.

[0027] By continuously repeating the operation of determining the start point of the short time window and intercepting the sampling data in step S214, the multi-source runtime sequence data set is continuously windowed until all sampling data is included in the corresponding short time window data segment. Each time the operation is repeated, the start point of the current short time window is used as a reference to determine the start point of the next short time window according to the window sliding step parameter, and then a corresponding amount of sampling data is intercepted. This cycle continues until the entire multi-source runtime sequence data set is divided, and a series of short time window data segments are finally obtained.

[0028] Step S216: Add a window sequence number identifier to each short time window data segment, which represents the order of the short time window in the time sequence.

[0029] The window sequence number identifier is a unique identifier for distinguishing different short time window data segments, and each short time window is numbered according to the order of the time sequence. By adding the window sequence number identifier, it is convenient to sort, search and analyze the short time window data segments subsequently. For example, the window sequence number identifier of the first short time window data segment is 1, the second is 2, and so on. Exemplarily, an additional attribute field can be added to each short time window data segment to store the window sequence number identifier. After all short time window data segments are divided, each data segment is assigned a corresponding window sequence number identifier in the order of division. In this way, in subsequent processing, the position of each short time window data segment in the time sequence can be determined according to the window sequence number identifier.

[0030] Step S220: Extract the amplitude change feature of each short time window data segment, calculate the difference between the maximum value and the minimum value of the sampling data in each short time window data segment, and obtain the short time window amplitude change feature.

[0031] The amplitude variation feature extraction is a process of extracting a feature capable of reflecting the amplitude variation of the data from each short-time window data segment. The difference between the maximum value and the minimum value of the sampling data can intuitively reflect the fluctuation degree of the data in a short-time window. The short-time window amplitude variation feature is a sequence composed of a series of feature values obtained after the amplitude variation feature extraction is performed on all short-time window data segments. Exemplarily, each short-time window data segment can be traversed first, and for the currently traversed short-time window data segment, all sampling data contained therein are extracted. Then, the sampling data with the maximum value is filtered out from the extracted sampling data, and is determined as the window maximum value. The sampling data with the minimum value is filtered out from the extracted sampling data, and is determined as the window minimum value. The difference between the window maximum value and the window minimum value is calculated to obtain the amplitude variation of the current short-time window data segment. Next, the amplitude variation is stored in association with the window serial number identifier of the current short-time window data segment, so as to be sorted according to the window serial number identifier subsequently. Finally, after the traversal of all short-time window data segments is completed, the amplitude variations of all short-time window data segments are sorted according to the order of the window serial number identifiers, and the sorted amplitude variation sequence is determined as the short-time window amplitude variation feature.

[0032] As an implementation, step S220 can specifically include steps S221-S224:

[0033] Step S221: Each short-time window data segment is traversed, and for the currently traversed short-time window data segment, all sampling data contained therein are extracted.

[0034] The short-time window data segment is a series of data segments obtained after window division, and each data segment contains a certain number of sampling data. Traversing each short-time window data segment means sequentially processing each data segment. For the currently traversed short-time window data segment, all sampling data contained therein need to be extracted so as to be filtered for the maximum value and the minimum value subsequently. Exemplarily, a loop structure can be used to traverse all short-time window data segments. For each short-time window data segment, the sampling data therein can be extracted through the access method of the data storage structure thereof. For example, if the short-time window data segment is stored in an array, the index of the array can be used to traverse and extract each element, i.e., the sampling data.

[0035] Step S222: The sampling data with the maximum value is filtered out from the extracted sampling data, and is determined as the window maximum value. The sampling data with the minimum value is filtered out from the extracted sampling data, and is determined as the window minimum value. The difference between the window maximum value and the window minimum value is calculated to obtain the amplitude variation of the current short-time window data segment.

[0036] It is a common data processing operation to filter out the maximum and minimum values from the extracted sampling data. The maximum and minimum values can be recorded by traversing the sampling data, comparing the size of each data, and recording the maximum and minimum values. The window maximum value is the maximum value among all the sampling data in a short-time window data segment, and the window minimum value is the minimum value. The difference between the window maximum value and the window minimum value can be calculated to obtain the amplitude variation of the current short-time window data segment. Exemplarily, a variable can be used to record the current maximum and minimum values, and the values of the two variables are updated constantly when traversing the sampling data, and finally the difference between them is calculated.

[0037] Step S223: Store the amplitude variation in association with the window sequence identifier of the current short-time window data segment.

[0038] The associated storage is to establish a corresponding relationship between the amplitude variation and the corresponding window sequence identifier, so as to sort and process the amplitude variation according to the window sequence identifier subsequently. Data structures such as dictionaries or lists can be used to realize the associated storage. For example, the window sequence identifier is stored as the key of the dictionary, and the corresponding amplitude variation is stored as the value. In this way, in subsequent operations, the corresponding amplitude variation can be quickly found through the window sequence identifier. Exemplarily, after calculating the amplitude variation of each short-time window data segment, it is stored in the dictionary as a key-value pair with the window sequence identifier of the data segment.

[0039] Step S224: After completing the traversal of all short-time window data segments, sort the amplitude variations of all short-time window data segments according to the order of the window sequence identifiers, and determine the sorted amplitude variation sequence as the short-time window amplitude variation feature.

[0040] After completing the extraction of the amplitude variation features of all short-time window data segments, the obtained amplitude variations need to be sorted to ensure that they are arranged in the order of time sequence. The window sequence identifier is the key basis for determining the time sequence. By sorting the amplitude variations according to the order of the window sequence identifiers, an ordered amplitude variation sequence can be obtained. This sequence is the short-time window amplitude variation feature. Exemplarily, the data structure of the associated storage of the amplitude variation and the window sequence identifier can be sorted according to the window sequence identifier, and then the sorted amplitude variation is extracted to form a new sequence, which is the short-time window amplitude variation feature. For example, the amplitude variation and the window sequence identifier are stored using a Python dictionary, and by sorting the keys (window sequence identifiers) of the dictionary, and then extracting the corresponding values (amplitude variations) according to the sorted key order, the short-time window amplitude variation feature sequence can be obtained.

[0041] Step S230: window dividing the multi-source runtime sequence data set based on the medium time window parameter in the time window parameter set, to obtain a plurality of medium time window data segments.

[0042] The medium time window parameter is a parameter in the time window parameter set for defining the medium time window size and sliding rule. The medium time window data segment is a data segment obtained by dividing the multi-source runtime sequence data set based on the medium time window parameter, which can reflect the operation state change of the smart grid at a medium time scale. The process of window dividing the multi-source runtime sequence data set based on the medium time window parameter is similar to the process of dividing based on the short time window parameter. First, the medium time window parameter is extracted from the time window parameter set, which also contains the window length parameter and the window sliding step parameter. Then, taking the starting sampling data of the multi-source runtime sequence data set as the starting point of the first medium time window, the sampling data is truncated according to the window length parameter to obtain the first medium time window data segment. Next, taking the starting point of the first medium time window as the reference, the corresponding number of sampling data is moved backward according to the window sliding step parameter to determine the starting point of the second medium time window, and the sampling data is truncated according to the window length parameter to obtain the second medium time window data segment. Repeat the steps of determining the starting point of the medium time window and truncating the sampling data until the window division of all sampling data of the multi-source runtime sequence data set is completed, to obtain a plurality of medium time window data segments. Finally, add a window serial number identifier to each medium time window data segment to represent its order in the time sequence.

[0043] Step S240: trend change feature extraction is performed on each medium time window data segment, and the change slope of the sampling data in each medium time window data segment is calculated by a linear fitting method to obtain a medium time window trend change feature.

[0044] Trend change feature extraction is a process of extracting features from each medium time window data segment that can reflect the trend change of the data. Linear fitting method is a data analysis method, which finds a straight line to approximate the trend of data change, and the slope of the straight line can reflect the change rate of the data. The medium time window trend change feature is a sequence composed of a series of change slopes obtained by trend change feature extraction on all medium time window data segments.

[0045] As an implementation mode, step S240 can specifically include the following steps S241-S246:

[0046] Step S241: traverse each medium time window data segment, and for the medium time window data segment currently traversed, extract all sampling data and corresponding sampling time points contained therein.

[0047] The medium time window data segment is obtained by dividing the multi-source runtime sequence data set by the medium time window parameter. Traversing each medium time window data segment means processing each data segment in turn. For the currently traversed medium time window data segment, all the sampling data contained therein and the corresponding sampling time points need to be extracted. The sampling data is a specific numerical value reflecting the operation state of the smart grid, and the sampling time point records the collection time of each sampling data. Exemplarily, the sampling data and the corresponding sampling time point can be extracted by traversing the data storage structure of the medium time window data segment. For example, if the medium time window data segment is stored in a list, each element in the list is a tuple containing sampling data and sampling time points, then the information can be extracted by traversing the list.

[0048] Step S242: converting each sampling time point into a time series index value, the time series index value representing the relative order of the sampling time point in the medium time window.

[0049] The sampling time point is stored in the actual time format, such as date and time. In order to facilitate linear fitting analysis, each sampling time point needs to be converted into a time series index value. The time series index value is an integer, reflecting the relative position of the sampling time point in the medium time window. The conversion can be achieved by sorting the sampling time points and then assigning a continuous integer index value to each sampling time point. Exemplarily, the sampling time points can be sorted first, and then a counter can be used to assign an index value to each sampling time point.

[0050] Step S243: based on the time series index value and the corresponding sampling data, a two-dimensional data point set is constructed, each data point in the two-dimensional data point set containing a time series index value and a corresponding sampling data value.

[0051] The two-dimensional data point set is a set composed of multiple data points, each data point being a two-tuple containing a time series index value and a corresponding sampling data value. By constructing the two-dimensional data point set, the time information and the sampling data information can be combined together, facilitating subsequent linear fitting processing. Exemplarily, the time series index value and the sampling data can be combined into a two-tuple by traversing them, and then added to a list to form the two-dimensional data point set.

[0052] Step S244: performing linear fitting processing on the two-dimensional data point set using the least squares method to obtain a fitting straight line equation, the slope of the fitting straight line equation being a change slope, extracting the slope value from the fitting straight line equation to determine the trend change slope of the current medium time window data segment.

[0053] The least square method is a linear fitting method, which aims to find a straight line such that the sum of the squares of the perpendicular distances of all data points to the straight line is minimized. By performing least square linear fitting on a set of two-dimensional data points, a fitting straight line equation can be obtained. The general form of the fitting straight line equation is y = kx + b, where k is the slope and b is the intercept. The slope k reflects the rate of change of the data, i.e., the change slope. By extracting the slope value from the fitting straight line equation, the trend change slope for the current medium time window data segment can be determined. Exemplarily, a least square function in a mathematical library can be used to perform the fitting calculation, and then the slope value is extracted from the fitting result.

[0054] Step S245: Store the trend change slope in association with the window sequence identifier of the current medium time window data segment.

[0055] The association storage is to establish a corresponding relationship between the trend change slope and the corresponding window sequence identifier, so that the trend change slope can be sorted and processed according to the window sequence identifier in the subsequent operation. Data structures such as dictionaries or lists can be used to implement the association storage. For example, the window sequence identifier is stored as the key of the dictionary, and the corresponding trend change slope is stored as the value. In this way, in the subsequent operation, the corresponding trend change slope can be quickly found through the window sequence identifier. Exemplarily, after calculating the trend change slope of each medium time window data segment, it is stored in the dictionary as a key-value pair with the window sequence identifier of the data segment.

[0056] Step S246: After completing the traversal of all medium time window data segments, sort the trend change slopes of all medium time window data segments according to the order of the window sequence identifiers, and determine the sorted trend change slope sequence as the medium time window trend change feature.

[0057] After completing the trend change feature extraction of all medium time window data segments, the obtained trend change slopes are sorted to ensure that they are arranged in the order of time sequence. The window sequence identifier is a key basis for determining the time sequence. By sorting the trend change slopes according to the order of the window sequence identifiers, an ordered trend change slope sequence can be obtained. This sequence is the medium time window trend change feature. Exemplarily, the data structure of the associated storage of the trend change slope and the window sequence identifier can be sorted according to the window sequence identifier, and then the sorted trend change slope is extracted to form a new sequence, which is the medium time window trend change feature. For example, the trend change slope and the window sequence identifier are stored in a Python dictionary, the keys (window sequence identifiers) of the dictionary are sorted, and then the values (trend change slopes) corresponding to the sorted keys are extracted to obtain the medium time window trend change feature sequence.

[0058] Step S250: window partitioning the multi-source runtime time series data set based on a long time window parameter in the time window parameter set to obtain a plurality of long time window data segments.

[0059] The long time window parameter is a parameter in the time window parameter set for defining the size and sliding rule of the long time window. The long time window data segment is a data segment obtained by partitioning the multi-source runtime time series data set based on the long time window parameter, and can reflect the change of the operation state of the smart grid in a long time scale. The process of window partitioning the multi-source runtime time series data set based on the long time window parameter is similar to the process of partitioning based on the short time window parameter and the medium time window parameter. First, the long time window parameter is extracted from the time window parameter set, and the long time window parameter includes the window length parameter and the window sliding step parameter. Then, the starting sampling data of the multi-source runtime time series data set is taken as the starting point of the first long time window, the sampling data is truncated according to the window length parameter to obtain the first long time window data segment. Next, the starting point of the first long time window is taken as the reference, and the corresponding number of sampling data is moved backward according to the window sliding step parameter to determine the starting point of the second long time window, and the sampling data is truncated according to the window length parameter to obtain the second long time window data segment. The steps of determining the starting point of the long time window and truncating the sampling data are repeatedly executed until the window partitioning of all sampling data of the multi-source runtime time series data set is completed, and a plurality of long time window data segments are obtained. Finally, a window serial number identifier is added to each long time window data segment to represent its order in the time series.

[0060] Step S260: extracting fluctuation correlation features from each long time window data segment to calculate the standard deviation of the change rate between adjacent sampling data in each long time window data segment to obtain long time window fluctuation correlation features.

[0061] The fluctuation correlation feature extraction is a process of extracting features from each long time window data segment that can reflect the data fluctuation correlation. The standard deviation of the change rate between adjacent sampling data is an index for measuring the degree of data fluctuation and stability, and can reflect the fluctuation correlation characteristics of the data in a long time window. The long time window fluctuation correlation features are a sequence of a series of standard deviations of change rates obtained by extracting fluctuation correlation features from all long time window data segments.

[0062] As an implementation manner, step S260 can specifically include steps S261-S265:

[0063] Step S261: Traverse each long time window data segment, and for the current traversed long time window data segment, extract all the sampling data contained therein, and arrange the sampling data in sequence according to the sampling time to form a sampling data sequence.

[0064] The long time window data segment is a data segment obtained by dividing the multi-source runtime sequence data set according to the long time window parameter. Traversing each long time window data segment means processing each data segment in turn. For the current traversed long time window data segment, all the sampling data contained therein need to be extracted and arranged in sequence according to the sampling time to form a sampling data sequence. The sampling data is a specific numerical value reflecting the operation state of the smart grid. Arranging the sampling data in sequence according to the sampling time can ensure the time sequence of the data. Exemplarily, the sampling data can be extracted by traversing the data storage structure of the long time window data segment and stored in a list, and the sampling data sequence can be formed according to the storage order of the data, i.e., the sampling time sequence.

[0065] Step S262: Calculate the change rate between adjacent two sampling data in the sampling data sequence. For the i-th sampling data and the i+1-th sampling data in the sampling data sequence, the change rate is calculated by subtracting the i-th sampling data from the i+1-th sampling data and dividing the result by the i-th sampling data. All the change rates between adjacent sampling data are combined to form a change rate sequence.

[0066] The change rate between adjacent sampling data is calculated to measure the degree of change of the data between adjacent time points. For the i-th sampling data and the i+1-th sampling data in the sampling data sequence, the change rate between the two adjacent sampling data can be obtained by subtracting the i-th sampling data from the i+1-th sampling data and dividing the result by the i-th sampling data. After calculating the change rates between all adjacent sampling data, a change rate sequence is formed. Exemplarily, a loop structure can be used to traverse the sampling data sequence, calculate the change rate between adjacent sampling data, and add it to a new list to form the change rate sequence.

[0067] Step S263: Calculate the standard deviation of the change rate sequence, and the standard deviation represents the dispersion degree of the data in the change rate sequence. The standard deviation is determined as the fluctuation correlation characteristic value of the current long time window data segment.

[0068] The standard deviation is a statistical index used to measure the dispersion degree of the data. For the change rate sequence, calculating the standard deviation can obtain a value reflecting the fluctuation of the change rate. The larger the value, the more intense the fluctuation of the change rate; the smaller the value, the more stable the change rate. The calculated standard deviation is determined as the fluctuation correlation characteristic value of the current long time window data segment. The standard deviation of the change rate sequence can be calculated using the standard deviation calculation formula in statistics.

[0069] Step S264: Store the volatility correlation feature value in association with the window sequence identifier of the current long-time window data segment.

[0070] The association storage is to establish a corresponding relationship between the volatility correlation feature value and the corresponding window sequence identifier, so as to sort and process the volatility correlation feature value according to the window sequence identifier subsequently. A data structure such as a dictionary or a list can be used to realize the association storage. For example, the window sequence identifier is stored as the key of the dictionary, and the corresponding volatility correlation feature value is stored as the value. In this way, the corresponding volatility correlation feature value can be quickly found through the window sequence identifier. Exemplarily, after calculating the volatility correlation feature value of each long-time window data segment, it is stored in the dictionary as a key-value pair with the window sequence identifier of the data segment.

[0071] Step S265: After completing the traversal of all long-time window data segments, sort the volatility correlation feature values of all long-time window data segments according to the order of the window sequence identifiers, and determine the sorted volatility correlation feature value sequence as the long-time window volatility correlation feature.

[0072] After completing the extraction of the volatility correlation features of all long-time window data segments, the obtained volatility correlation feature values need to be sorted to ensure that they are arranged in the order of time sequence. The window sequence identifier is the key basis for determining the time sequence. By sorting the volatility correlation feature values according to the order of the window sequence identifiers, an ordered volatility correlation feature value sequence can be obtained. This sequence is the long-time window volatility correlation feature. Exemplarily, the data structure of the association storage of the volatility correlation feature values and the window sequence identifiers can be sorted according to the window sequence identifiers, and then the sorted volatility correlation feature values are extracted to form a new sequence, which is the long-time window volatility correlation feature. For example, the volatility correlation feature values and the window sequence identifiers are stored using a Python dictionary, the keys (window sequence identifiers) of the dictionary are sorted, and then the values (volatility correlation feature values) corresponding to the sorted keys are extracted to obtain the long-time window volatility correlation feature sequence.

[0073] Step S270: Concatenate the short-time window amplitude change feature, the medium-time window trend change feature, and the long-time window volatility correlation feature in the order of time sequence to form an initial time sequence feature sequence.

[0074] The features of different time scales are concatenated in time sequence to integrate them into a complete feature sequence for further analysis and processing. The short-time window amplitude variation feature, the medium-time window trend variation feature and the long-time window fluctuation correlation feature respectively reflect the operation state changes of the smart grid from different time scales, and their concatenation can more comprehensively describe the operation state of the smart grid. The initial time sequence feature sequence is a new feature sequence obtained by concatenating the three feature sequences in time sequence.

[0075] As an implementation, step 270 can specifically include steps S271-S275:

[0076] Step S271: Obtain the window sequence number identification sequence corresponding to the short-time window amplitude variation feature, the window sequence number identification sequence corresponding to the medium-time window trend variation feature and the window sequence number identification sequence corresponding to the long-time window fluctuation correlation feature.

[0077] The window sequence number identification sequence is a sequence composed of a series of window sequence number identifications corresponding to each feature sequence, which records the position of each feature value corresponding time window in the time sequence. By obtaining the window sequence number identification sequences corresponding to the three feature sequences, the corresponding relationship of each feature value in time can be determined, providing a basis for subsequent time alignment and concatenation operations. The window sequence number identification sequence can be obtained by accessing the data structure storing the association between feature values and window sequence number identifications. For example, in the previous steps, if a dictionary is used to store the short-time window amplitude variation feature values and the corresponding window sequence number identifications, the window sequence number identification sequence can be obtained by extracting the keys of the dictionary.

[0078] Step S272: Determine the time alignment reference point based on the window sequence number identification sequences, and the time alignment reference point is the earliest common time point appearing in the three window sequence number identification sequences.

[0079] The time alignment reference point is a starting point for determining the alignment of the three feature sequences in time. By finding the earliest common time point in the three window sequence identification sequences, it can be ensured that the three feature sequences can be correctly aligned in time in the subsequent concatenation process. The specific method for determining the time alignment reference point can be to traverse the three window sequence identification sequences, find their intersection, and then select the earliest appearing time point from the intersection as the time alignment reference point. For example, the window sequence identification sequence corresponding to the short time window amplitude change feature is [1, 2, 3, 4], the window sequence identification sequence corresponding to the medium time window trend change feature is [2, 3, 4, 5], and the window sequence identification sequence corresponding to the long time window fluctuation correlation feature is [3, 4, 5, 6], their intersection is [3, 4], and the earliest common time point is 3, so the time alignment reference point is 3.

[0080] Step S273: Taking the time alignment reference point as the starting point, arrange the short time window amplitude change feature, the medium time window trend change feature, and the long time window fluctuation correlation feature in time sequence.

[0081] Taking the time alignment reference point as the starting point, arranging the three feature sequences in time sequence is to ensure their consistency in the time dimension. In the arrangement process, the time sequence of each feature value is determined according to the window sequence identification. For example, taking the time alignment reference point 3 as the starting point, for the short time window amplitude change feature sequence, find the feature values with window sequence identification of 3 and later, and arrange them in the order of window sequence identification; the same operation is performed for the medium time window trend change feature sequence and the long time window fluctuation correlation feature sequence.

[0082] Step S274: For feature parts with non-overlapping time ranges, use zero padding to supplement, so that the three features are consistent in the time dimension, and the arranged and supplemented short time window amplitude change feature, medium time window trend change feature, and long time window fluctuation correlation feature are merged in the feature dimension to form a three-dimensional feature matrix.

[0083] Since the characteristic sequences of different time scales can have different time ranges, in order to make them consistent in the time dimension, the characteristic parts that do not overlap in time range need to be supplemented. Zero padding is to fill the missing characteristic values with zeros. In this way, the lengths of the three characteristic sequences in the time dimension will be the same. Then, the arranged and supplemented short-time window amplitude variation characteristics, medium-time window trend variation characteristics and long-time window fluctuation correlation characteristics are combined in the feature dimension to form a three-dimensional feature matrix. The three dimensions of the three-dimensional feature matrix correspond to the time dimension, the feature type dimension and the feature value dimension respectively. For example, the short-time window amplitude variation characteristic sequence is [0, 0, 10, 12, 15] after zero padding, the medium-time window trend variation characteristic sequence is [0, 2, 3, 4, 0], and the long-time window fluctuation correlation characteristic sequence is [0, 0, 5, 6, 7]. Combining them forms a three-dimensional feature matrix, where the first dimension represents time, the second dimension represents the feature type (short-time, medium-time, long-time), and the third dimension represents the feature value.

[0084] Step S275: converting the three-dimensional feature matrix into a one-dimensional feature vector sequence to determine the initial time sequence characteristic sequence.

[0085] Converting the three-dimensional feature matrix into a one-dimensional feature vector sequence is to convert the multi-dimensional feature data into a one-dimensional form suitable for subsequent processing. Each element of the three-dimensional feature matrix can be arranged in a one-dimensional vector sequence by traversing them in order. For example, for the above three-dimensional feature matrix, the three feature values at each time point are extracted in time order to form a one-dimensional feature vector sequence. This one-dimensional feature vector sequence is the initial time sequence characteristic sequence. Exemplarily, a loop structure can be used to traverse the three-dimensional feature matrix, add elements to a new list to form a one-dimensional feature vector sequence.

[0086] Step S280: performing feature alignment processing on the initial time sequence characteristic sequence, and determining the initial time sequence characteristic sequence after the feature alignment processing as the multi-scale time sequence characteristic sequence.

[0087] Feature alignment processing is a further adjustment and optimization of the initial time series feature sequence to ensure consistency and comparability between features. The initial time series feature sequence after feature alignment processing can better reflect the changes in the operation state of the smart grid at different time scales, so it is determined as the multi-scale time series feature sequence. Feature alignment processing can include normalization processing, smoothing processing and other operations. Normalization processing can unify the range of feature values to an interval, such as [0, 1], to eliminate the dimensional differences between different features. Smoothing processing can remove noise and outliers in the feature sequence, making the feature sequence smoother and more stable. For example, the initial time series feature sequence is normalized using the min-max normalization method, and each feature value is mapped to the [0, 1] interval. Then, the normalized feature sequence is smoothed using the moving average smoothing method to remove noise. After these processes, the obtained feature sequence is the multi-scale time series feature sequence.

[0088] Step S300: input the multi-scale time series feature sequence into the fault evolution analysis model for state transition learning processing to generate a power grid state transition feature set.

[0089] The fault evolution analysis model is a model for analyzing the fault evolution process of the smart grid, which can learn the transition rules between different states of the smart grid. State transition learning processing is to analyze and learn the multi-scale time series feature sequence through the model to determine the state transition of the smart grid at different times. The power grid state transition feature set is a series of sets reflecting the power grid state transition features obtained after state transition learning processing.

[0090] As an implementation, step S300 can specifically include steps S310-S390:

[0091] Step S310: input the multi-scale time series feature sequence into the state initialization layer of the fault evolution analysis model, and construct an initial state vector through the state initialization layer to obtain an initial power grid state vector.

[0092] The state initialization layer is used to construct an initial state vector according to the input multi-scale time series feature sequence. The initial state vector is a vector that can reflect the operation state of the smart grid at the initial time, which provides a basis for subsequent state transition simulation and analysis. The specific method of constructing an initial state vector through the state initialization layer to the multi-scale time series feature sequence can be feature extraction and integration of the multi-scale time series feature sequence. For example, the feature values of the first few time scales in the multi-scale time series feature sequence can be selected and combined into a vector as the initial power grid state vector. Illustratively, the state initialization layer can use a fully connected layer or a convolutional layer to process the multi-scale time series feature sequence and convert it into an initial power grid state vector.

[0093] Step S320: input the initial power grid state vector to the transition rule learning layer of the fault evolution analysis model, learn the state transition law in the historical fault case data through the transition rule learning layer, and generate a state transition rule set.

[0094] The main task of the transition rule learning layer is to learn the transition law of the smart grid between different states from the historical fault case data and generate a corresponding state transition rule set. The historical fault case data is a data set that records the past fault events and state transition processes of the smart grid. By analyzing and learning these data, the rules of fault occurrence and evolution can be found. The state transition rule set is a set of rules that contains the transition probabilities and transition conditions between different states, which can guide the subsequent state transition simulation process.

[0095] As an implementation, in step S320, the state transition rule set is generated by learning the state transition law in the historical fault case data through the transition rule learning layer. Specifically, it can include the following steps S321-S329:

[0096] Step S321: obtain historical fault case data from a preset database, and the historical fault case data contains complete state transition process records of multiple historical fault events.

[0097] The preset database is a database that stores a large amount of historical fault case data, which can be a local database or a remote database. The historical fault case data is detailed data that records past fault events and state transition processes of the smart grid, including fault occurrence time, fault type, and state before and after the fault. When obtaining historical fault case data from the preset database, a database query statement can be used. For example, use an SQL query statement to select all historical fault case data records from the database. When obtaining data, the integrity and accuracy of the data need to be ensured, and necessary cleaning and preprocessing of the data are performed, such as removing missing values and outliers.

[0098] Step S322: perform event segmentation processing on the historical fault case data, and divide the complete state transition process record of each historical fault event into a pre-fault state sequence, a fault occurrence state sequence, and a post-fault state sequence.

[0099] The event segmentation processing is to divide the complete state transition process record of each historical fault event according to the time point of fault occurrence, and divide it into a pre-fault state sequence, a fault occurrence state sequence and a post-fault state sequence. The pre-fault state sequence records the operation state of the smart grid before the fault occurs, the fault occurrence state sequence records the state change at the fault occurrence time, and the post-fault state sequence records the recovery state of the smart grid after the fault occurs. Through the event segmentation processing, the evolution process of the fault can be more clearly analyzed. For example, according to the fault occurrence time mark in the historical fault case data, the state record of each fault event is divided into three parts. Exemplarily, the time point of fault occurrence can be determined using the timestamp information, and then the state record is segmented according to the time point.

[0100] Step S323: classifying the segmented state transition process records according to fault types to obtain a state transition sub-sequence set corresponding to different fault types.

[0101] The classification of the segmented state transition process records according to fault types is to classify fault events of the same type into a category, so as to subsequently analyze different types of faults separately. The state transition sub-sequence set corresponding to different fault types is a set of state transition sub-sequences under the same fault type. Through the classification processing, the state transition rules of different types of faults can be more targetedly learned. For example, the fault types are divided into short-circuit faults, overload faults, etc., and then the segmented state transition process records are classified according to these fault types. Exemplarily, each state transition sub-sequence can be assigned to the corresponding fault type set according to the fault type mark in the historical fault case data.

[0102] Step S324: performing time sequence alignment processing on each state transition sub-sequence set to make the state transition sub-sequences under the same fault type consistent in length in the time dimension.

[0103] The time sequence alignment processing is to make the state transition sub-sequences under the same fault type comparable in the time dimension. Since different fault events can have different durations and state transition processes, through the time sequence alignment processing, their lengths can be unified for subsequent statistics and analysis. The time sequence alignment processing can use interpolation or truncation methods. For example, for shorter state transition sub-sequences, a linear interpolation method can be used to interpolate in the time dimension to make their length consistent with that of longer sub-sequences; for longer sub-sequences, truncation processing can be performed to make their length consistent with that of shorter sub-sequences. Exemplarily, a suitable time sequence alignment method can be selected according to specific requirements.

[0104] Step S325: Calculate the transition frequency between adjacent states in each state transition sub-sequence, and count the number of transitions from the same starting state to the same ending state.

[0105] The transition frequency is the frequency of transitions between adjacent states in a state transition sub-sequence. Counting the number of transitions from the same starting state to the same ending state can provide a basis for subsequent calculation of state transition probabilities. The specific method of calculating transition frequency and counting transitions is to traverse each state transition sub-sequence and record the transitions between adjacent states. For example, for a state transition sub-sequence [S1, S2, S3, S2, S4], count the number of transitions from S1 to S2, the number of transitions from S2 to S3, etc. Exemplarily, a dictionary can be used to store the number of transitions from the same starting state to the same ending state, with the key being the combination of the starting state and the ending state, and the value being the number of transitions.

[0106] Step S326: Calculate the state transition probability based on the number of transitions, the state transition probability is the ratio of the number of transitions from the same starting state to the same ending state to the total number of possible transitions from the starting state, and smooth the calculated state transition probability to avoid abnormal fluctuations in probability values caused by insufficient sample size.

[0107] The state transition probability is an indicator that measures the likelihood of transitioning from one state to another. Calculating the state transition probability from the number of transitions can reflect the transition rules between different states in historical fault case data. Smoothing the calculated state transition probability is to avoid abnormal fluctuations in probability values caused by insufficient sample size, making the probability values more stable and reliable. Smoothing can use Laplace smoothing, specifically, the Laplace smoothing method adds a smoothing parameter, usually set to 1, to the count of each state transition when calculating the state transition probability.

[0108] Step S327: Extract the key feature change amount in each state transition process, the key feature change amount is the core feature index that triggers state transition.

[0109] Key feature change amount is the change of feature indicators that play a key role in the state transition process. In the fault evolution process of smart grid, different state transitions may be related to specific feature changes. For example, sudden drop of voltage amplitude, sharp increase of current, etc. may be the key factors that trigger state transition. Extracting key feature change amount can be achieved by analyzing the difference of feature values before and after state transition. Specifically, for each state transition process, record the feature vector before state transition and the feature vector after state transition, and then calculate the difference of corresponding feature values. In order to determine which feature changes are key, feature selection methods such as correlation analysis, chi-square test, etc. can be used. For example, by calculating the correlation coefficient between feature change amount and state transition, select the feature change with high correlation as the key feature change amount.

[0110] Step S328: store the state transition probability, key feature change amount and corresponding fault type in association, and form the initial state transition rule.

[0111] Association storage is to establish a corresponding relationship between state transition probability, key feature change amount and corresponding fault type, so as to be used and queried subsequently. Database or data structure such as dictionary can be used to realize association storage. For example, use the key of the dictionary to store the combination of fault type and state transition, and use the value to store the corresponding state transition probability and key feature change amount. Specifically, the key of the dictionary can be expressed as “fault type-starting state-terminating state”, and the value is a tuple containing state transition probability and key feature change amount. In this way, in subsequent fault analysis, the corresponding state transition probability and key feature change amount can be quickly found according to the fault type and state transition condition.

[0112] Step S329: iteratively optimize the initial state transition rule by the deep belief network of the transition rule learning layer, adjust the weight coefficient of the key feature change amount, and determine the optimized initial state transition rule as the state transition rule set.

[0113] A deep belief network (DBN) is a deep learning model composed of multiple stacked Restricted Boltzmann Machines (RBMs). It can learn complex distributions and feature representations of data. In the transition rule learning layer, a deep belief network is used to iteratively optimize the initial state transition rules. The goal is to adjust the weight coefficients of the key feature changes, so that the state transition rules more accurately reflect the actual fault evolution process of the smart grid. For example, the key feature changes and state transition probabilities in the initial state transition rules are first input into the deep belief network as input data. The training goal of the deep belief network is to minimize the error between the predicted state transition results and the actual state transition results. During the training process, the deep belief network continuously adjusts the weight parameters of the network through the backpropagation algorithm, thereby indirectly adjusting the weight coefficients of the key feature changes. For example, for a certain state transition process, the deep belief network may find that a certain key feature change has a greater impact on state transition, and thus increase the weight coefficient of that feature change. After multiple iterations of training, when the error reaches an acceptable threshold, the training is stopped, and the resulting state transition rules are the optimized state transition rule set. In practical applications, optimization algorithms such as stochastic gradient descent can be used to update the weight parameters of the deep belief network to improve training efficiency and convergence speed.

[0114] Step S330: Based on the state transition rule set, simulate the state transition of the initial power grid state vector to obtain the power grid state vector at the first time.

[0115] State transition simulation is to predict the state of the initial power grid state vector after a state transition according to the state transition rule set. The initial power grid state vector represents the operating state of the smart grid at the initial time, and through state transition simulation, the power grid state vector at the next time can be obtained. Specifically, for the initial power grid state vector, a possible transition state is randomly selected according to the state transition probability corresponding to the state in the state transition rule set.

[0116] Step S340: Compare the first time power grid state vector with the feature data at the corresponding time in the multi-scale time series feature sequence, and calculate the state deviation value.

[0117] The state deviation value is an index that measures the difference between the simulated first time power grid state vector and the feature data at the corresponding time in the actual multi-scale time series feature sequence. By calculating the state deviation value, the accuracy of the state transition simulation can be evaluated. The specific calculation method is, for example, to calculate the distance between two vectors, such as Euclidean distance, Manhattan distance, etc. The larger the state deviation value, the greater the difference between the simulated state and the actual state, and the lower the accuracy of the state transition simulation; on the contrary, the smaller the state deviation value, the closer the simulation result to the actual situation.

[0118] Step S350: Adjust the state transition rule set according to the state deviation value, update the state transition condition and transition probability.

[0119] The adjustment of the state transition rule set according to the state deviation value is to make the state transition rule more consistent with the actual fault evolution process. If the state deviation value is large, it means that the current state transition rule may not be accurate and needs to be adjusted. The specific adjustment method includes updating the state transition condition and transition probability. For example, if it is found that the simulation result of a certain state transition process deviates greatly from the actual situation, the key feature change quantity and state transition probability of the state transition can be checked to see if they are reasonable. If the key feature change quantity is not accurate, the weight coefficient of the key feature change quantity can be recalculated or adjusted; if the state transition probability is not reasonable, the transition probability can be corrected according to the actual situation. For example, increase or decrease the probability of a certain transition state. During the adjustment process, the adjustment range can be determined according to the size of the state deviation value. For example, when the state deviation value exceeds a preset threshold, a large adjustment is made; when the state deviation value is small, a small adjustment is made. By continuously adjusting the state transition rule set, the accuracy of the state transition simulation can be improved.

[0120] Step S360: Based on the adjusted state transition rule set, the next state transition simulation of the first time grid state vector is performed to obtain the second time grid state vector.

[0121] After adjusting the state transition rule set, the next state transition simulation of the first time grid state vector is performed using the adjusted rule set. The simulation process is similar to step S330, a possible transition state is randomly selected according to the adjusted state transition probability, and then the feature value of the first time grid state vector is updated combined with the key feature change quantity to obtain the second time grid state vector. By continuously repeating the process of state transition simulation and rule adjustment, the actual fault evolution process of the smart grid can be gradually approached, and the state change of the grid can be more accurately predicted.

[0122] Step S370: Repeat the process of state transition simulation, state deviation calculation and transition rule adjustment until the state transition simulation of all monitoring times is completed, and a sequence of grid state vectors at multiple times is obtained.

[0123] The process of repeating state transition simulation, state deviation calculation and transition rule adjustment is to realize the simulation of the state evolution process of the smart grid in the entire monitoring time period. Starting from the initial grid state vector, state transition simulation is performed in sequence to obtain the grid state vector at each time. After each simulation, the state deviation value is calculated, and the state transition rule set is adjusted according to the deviation value. This continuous iteration is repeated until the state transition simulation of all monitoring times is completed.

[0124] Step S380: Feature difference processing is performed on the power grid state vector sequence to calculate the difference features between adjacent time instants, and state transition difference features are obtained.

[0125] Feature difference processing is to calculate the difference between adjacent time instants of the power grid state vector to highlight the feature changes in the state transition process. By calculating the difference features, the state changes of the smart grid at different time instants can be more clearly observed.

[0126] Step S390: Feature fusion is performed on the state transition difference features and the transition probabilities in the state transition rule set to generate a power grid state transition feature set.

[0127] Feature fusion is to combine the state transition difference features and the transition probabilities in the state transition rule set to form a more comprehensive feature set. The state transition difference features reflect the feature changes in the state transition process, while the transition probabilities reflect the possibility of state transition. Fusion of the two can more accurately describe the fault evolution process of the smart grid. The specific fusion method can be splicing or weighted combination of the state transition difference features and the transition probabilities. For example, the state transition difference feature vector and the corresponding transition probability can be taken as a new feature vector. In the case of weighted combination, different weights can be assigned to the state transition difference features and the transition probabilities according to different application requirements. For example, for some fault analysis tasks that are more sensitive to transition probabilities, the weight of transition probability can be increased. The power grid state transition feature set obtained through feature fusion can provide more abundant information for subsequent fault type identification and location.

[0128] Step S400: The power grid state transition feature set is input into the fault type identification network for feature classification and location probability calculation processing to obtain the fault type label and the fault location probability distribution.

[0129] The fault type identification network is a network model for identifying the fault type of the smart grid and calculating the fault location probability distribution. The power grid state transition feature set is input into the network, and through the feature classification and location probability calculation functions of the network, the fault type label and the fault location probability distribution can be obtained. The fault type label is used to determine the fault type of the smart grid, such as short circuit fault, overload fault, etc.; the fault location probability distribution represents the probability of fault occurrence at different locations. The specific processing process includes feature standardization, deep feature extraction, type classification, and location probability calculation steps.

[0130] As an implementation, step S400 can specifically include the following steps S410-S440:

[0131] Step S410: input the power grid state transition feature set to the feature input layer of the fault type recognition network, and perform feature standardization processing on the power grid state transition feature set through the feature input layer.

[0132] The feature standardization processing is to make the feature values in the power grid state transition feature set have the same scale and range, so as to avoid the influence of the large scale difference of the feature values on the subsequent feature extraction and classification effect. The feature input layer is the first layer of the fault type recognition network, which is responsible for receiving the power grid state transition feature set and performing standardization processing. The feasible feature standardization methods include z-score standardization and minimum-maximum standardization. In the feature input layer, a suitable standardization method can be selected according to the specific situation. For example, for feature data obeying normal distribution, z-score standardization can be more appropriate; for the case where the range of feature values is known, minimum-maximum standardization can be more convenient. Through feature standardization processing, the feature data can be more easily learned and processed by the network model.

[0133] Step S420: input the standardized power grid state transition feature set to the feature extraction layer of the fault type recognition network, and perform deep feature extraction on it through the convolutional neural network of the feature extraction layer to obtain fault deep features.

[0134] The convolutional neural network (CNN) of the feature extraction layer is used to extract deep features from the standardized power grid state transition feature set. Deep features are those that can more essentially reflect the fault type and fault location information. The convolutional neural network extracts and reduces the dimension of the input feature data through convolutional layers, pooling layers, etc. Specifically, the convolutional layer extracts local features through convolution operation on the input data by using convolution kernels. Convolution kernels of different sizes can extract features of different scales. For example, small-size convolution kernels extract detailed features, and large-size convolution kernels can extract global features. The pooling layer is used to reduce the dimension of the output of the convolutional layer, reduce the number of features, and retain important feature information. Through multiple convolution and pooling operations, deep features can be gradually extracted.

[0135] As an implementation manner, step S420 can specifically include the following steps S421-S4210:

[0136] Step S421: convert the standardized power grid state transition feature set into a three-dimensional feature tensor, and the dimensions of the three-dimensional feature tensor include time dimension, feature dimension and channel dimension.

[0137] The standardized power grid state transition feature set is converted into a three-dimensional feature tensor to meet the input requirements of the convolutional neural network. The time dimension of the three-dimensional feature tensor represents different monitoring times, the feature dimension represents different feature types, and the channel dimension can represent different feature subsets or feature combinations. For example, for a power grid state transition feature set containing multiple time points, each time point has a feature vector with n feature values, which can be converted into a T x n x 1 three-dimensional feature tensor, where T is the number of monitoring times. In some cases, the channel dimension can be increased as needed. For example, if there are different types of features (such as voltage features, current features, etc.), each type of feature can be treated as a channel to form a T x n x m three-dimensional feature tensor, where m is the number of channels. By converting the feature set into a three-dimensional feature tensor, the spatial feature extraction capability of the convolutional neural network can be better utilized.

[0138] Step S422: input the three-dimensional feature tensor into the first convolutional layer of the convolutional neural network, and perform parallel convolution operations on the three-dimensional feature tensor using multiple convolution kernels of different sizes to generate multi-scale feature maps.

[0139] The first convolutional layer is the first layer of the convolutional neural network, which is used to perform preliminary feature extraction on the input three-dimensional feature tensor. By using multiple convolution kernels of different sizes for parallel convolution operations, multi-scale feature maps can be generated. Convolution kernels of different sizes can capture feature information of different scales. During parallel convolution operations, each convolution kernel independently convolves the three-dimensional feature tensor to generate a corresponding feature map.

[0140] Step S423: perform batch normalization processing on each scale of the feature map to adjust the distribution range of the values in the feature map.

[0141] Batch normalization processing is used to accelerate the training process of the convolutional neural network and improve the stability of the model. In the convolutional neural network, as the number of network layers increases, the numerical distribution of the feature map may change significantly, leading to problems such as gradient vanishing or gradient explosion. Batch normalization processing adjusts the numerical distribution of the feature map to a relatively stable range by normalizing each batch of feature maps. Specifically, batch normalization processing calculates the mean and standard deviation of each batch of feature maps, then subtracts the mean and divides by the standard deviation to obtain the normalized feature map. In addition, batch normalization processing also introduces learnable scaling factors and offset factors to adjust the scale and offset of the normalized feature map. Through batch normalization processing, the training of the network becomes more stable, the training time is reduced, and the generalization ability of the model is improved.

[0142] Step S424: input the batch-normalized multi-scale feature maps into an activation function layer to enhance the expression ability of the features through a nonlinear activation function.

[0143] The activation function layer is used to introduce a nonlinear factor to enhance the feature expression ability of the convolutional neural network. In the convolution operation process, linear convolution operation can only extract linear features, while actual fault features are often nonlinear. By using a nonlinear activation function, the network can learn more complex feature patterns. Feasible nonlinear activation functions include ReLU (Rectified Linear Unit), Sigmoid function, and Tanh function, etc. Taking the ReLU function as an example, it is defined as f(x) = max(0, x). For the batch-normalized multi-scale feature maps, each feature value is input into the ReLU function, and when the feature value is greater than 0, it remains unchanged; when the feature value is less than or equal to 0, it is set to 0. In this way, a nonlinear transformation can be introduced to enable the network to capture more rich feature information. Different activation functions have different characteristics and application scenarios, and in actual application, appropriate activation functions can be selected according to the specific network structure and task requirements.

[0144] Step S425: input the activated multi-scale feature maps into a pooling layer to perform dimension reduction processing on the feature maps through a max-pooling operation to retain key feature information.

[0145] The pooling layer is used to perform dimension reduction processing on the feature maps to reduce the number of features while retaining important feature information. The max-pooling operation selects the maximum value in a window of a fixed size on the feature map as the pooling result. Through the max-pooling operation, some redundant feature information can be removed, the computational complexity can be reduced, and important features in the feature map can be highlighted. When performing max-pooling on the activated multi-scale feature maps, the feature maps of each scale are independently subjected to the pooling operation to obtain the dimension-reduced feature maps. In this way, the key features can be retained while reducing the dimension of the features, and the efficiency of subsequent processing can be improved.

[0146] Step S426: perform feature concatenation on the pooled multi-scale feature maps to fuse feature information of different scales.

[0147] Feature concatenation on the pooled multi-scale feature maps is performed to fuse feature information of different scales so that the network can utilize both local detailed features and global features. Feature concatenation is performed on the channel dimension of the feature maps of different scales to form a higher-dimensional feature map. Through feature concatenation, feature information of different scales can be integrated together to provide more rich feature representations for subsequent deep feature extraction. In actual implementation, a tensor concatenation operation can be used to complete the feature concatenation process.

[0148] Step S427: input the spliced feature map into the second convolutional layer of the convolutional neural network, perform deep convolution operation on the feature map by using a plurality of dilated convolution kernels with different dilation rates, and capture long-range dependencies in the feature map.

[0149] The second convolutional layer uses dilated convolution kernels with different dilation rates to perform deep convolution operation on the spliced feature map, in order to capture long-range dependencies in the feature map. Dilated convolution is a convolution method that inserts holes in the convolution kernel, and the receptive field size of the convolution kernel can be controlled by adjusting the dilation rate. Dilated convolution kernels with different dilation rates can capture feature information at different distances. For example, a dilated convolution kernel with a dilation rate of 1 is the same as a normal convolution kernel and can only capture local features; a dilated convolution kernel with a dilation rate of 2 can capture more extensive feature information and help discover long-range dependencies in the feature map. In the second convolutional layer, a plurality of dilated convolution kernels with different dilation rates are used to perform convolution operation on the spliced feature map in parallel, generating a plurality of feature maps. These feature maps contain feature information at different scales and distances, and can more comprehensively describe fault features. In this way, the network's ability to capture fault features, especially those related to long-range dependencies, can be improved.

[0150] Step S428: perform second batch normalization and activation processing on the feature map after deep convolution operation.

[0151] The purpose of the second batch normalization and activation processing on the feature map after deep convolution operation is similar to the first time, which is to speed up network training and enhance the expression ability of features. After deep convolution operation, the numerical distribution of the feature map may change again, and batch normalization processing can adjust it to a stable range. Then, a nonlinear activation function is used to activate the batch-normalized feature map, introducing a nonlinear factor so that the network can learn more complex feature patterns. For example, ReLU function can still be used as the activation function. Through the second batch normalization and activation processing, the performance and stability of the network can be further improved.

[0152] Step S429: input the processed feature map into the global average pooling layer, compress the spatial dimension of the feature map, and generate a feature vector.

[0153] The global average pooling layer is used to compress the processed feature map into a feature vector. The global average pooling operation is to perform global average on each channel of the feature map, and average the feature map of each channel into a numerical value. For example, for a feature map of HxWxC (where H and W are the height and width of the feature map respectively, and C is the number of channels), the global average pooling operation will average the HxW feature values of each channel to obtain a feature vector of length C. Through the global average pooling operation, the spatial dimension of the feature map can be compressed, reducing the number of features while retaining the main feature information of each channel. In this way, the two-dimensional feature map can be converted into a one-dimensional feature vector, facilitating subsequent fully connected layers for classification and regression processing.

[0154] Step S4210: dropout regularization processing is performed on the feature vector to prevent overfitting in the feature extraction process, and the feature vector after regularization processing is determined as the fault deep feature.

[0155] Dropout regularization is a feasible method to prevent overfitting of neural networks. In the feature extraction process, if the network model is too complex, overfitting may occur, that is, the model performs well on training data but poorly on test data. Dropout regularization processing randomly discards the output of some neurons, so that the network cannot rely too much on certain neurons during training, thereby improving the generalization ability of the network. Specifically, during training, for the input feature vector, some elements are randomly set to 0 with a certain probability p. During testing, no dropout operation is needed, but the output of all neurons is scaled by 1-p. Through dropout regularization processing, the network can learn more robust feature representation and reduce the risk of overfitting. The feature vector after dropout regularization processing is the fault deep feature, which can more accurately reflect the fault features of the smart grid and provide a reliable basis for subsequent fault type classification and location probability calculation.

[0156] Step S430: input the fault deep feature into the type classification layer of the fault type recognition network, and perform classification processing on the fault deep feature through the fully connected network of the type classification layer, output the probability value corresponding to each fault type, and select the label corresponding to the fault type with the highest probability value as the fault type label.

[0157] The full connection network of the type classification layer is used for classifying the deep fault features. The full connection network is composed of multiple full connection layers, and each neuron is connected to all neurons of the previous layer. The deep fault features are input into the full connection network, and the probability values corresponding to each fault type are output through the forward propagation calculation of the network. Specifically, the last layer of the full connection network usually uses a Softmax activation function to convert the output of the network into a probability distribution. The label corresponding to the fault type with the highest probability value is selected as the fault type label. For example, if the output probability distribution is [0.1, 0.2, 0.7], corresponding to three fault types, the probability value with the highest value is 0.7, and the corresponding fault type label is the fault type. In this way, the fault type currently occurring in the smart grid can be accurately identified.

[0158] Step S440: input the deep fault features into the position prediction layer of the fault type identification network, and perform time series analysis on the deep fault features through the recurrent neural network of the position prediction layer to output the fault occurrence probability corresponding to each monitoring point. The fault occurrence probability corresponding to all monitoring points forms a probability distribution vector as the fault location probability distribution.

[0159] The position prediction layer uses a recurrent neural network (RNN) to perform time series analysis on the deep fault features, because the fault features of the smart grid usually have time series characteristics. The recurrent neural network can process sequence data, and through the recurrent structure of the network, the time dependence in the data can be captured. The deep fault features are input as input sequences into the recurrent neural network, and the network will update the state and calculate the output according to the historical information of the sequence and the current input. In the position prediction layer, the output layer of the recurrent neural network can use a Sigmoid activation function to convert the output into a probability value. For each monitoring point, the recurrent neural network outputs a corresponding fault occurrence probability. All fault occurrence probabilities corresponding to the monitoring points form a probability distribution vector, which is the fault location probability distribution.

[0160] Step S500: determine the fault location result of the smart grid according to the fault type label and the fault location probability distribution.

[0161] The fault location result of the smart grid is determined according to the fault type label and the fault location probability distribution in order to accurately find the location of the fault and the corresponding processing priority. The fault type label clearly indicates the type of the fault, and different fault types may require different processing methods and priorities. The fault location probability distribution provides probability information of each monitoring point occurring a fault. By considering these two factors comprehensively, the fault location can be more accurately located, and the corresponding processing measures can be determined.

[0162] As an implementation, step S500 can specifically include steps S510-S540:

[0163] Step S510: Extract the fault occurrence probability corresponding to each monitoring point from the fault location probability distribution.

[0164] The fault location probability distribution is a vector containing the fault occurrence probability of each monitoring point. The fault occurrence probability corresponding to each monitoring point is extracted from the vector for subsequent sorting and screening of the fault location. In actual implementation, the probability value of each monitoring point can be extracted from the probability distribution vector using an index operation.

[0165] Step S520: Sort all monitoring points according to fault occurrence probability from high to low, and select a preset number of monitoring points at the top of the ranking as candidate fault locations.

[0166] The sorting of all monitoring points according to fault occurrence probability from high to low is to find the monitoring points most likely to have faults. The preset number is a value determined according to actual demand and experience, and usually the top few monitoring points are selected as candidate fault locations. For example, if the preset number is 3, the 3 monitoring points with the highest fault occurrence probability are selected as candidate fault locations. In this way, the range of fault location can be narrowed down and the efficiency of fault location can be improved. In the sorting process, sorting algorithms such as quicksort, merge sort, etc. can be used to sort the fault occurrence probability of the monitoring points, while recording the index or identification of each monitoring point.

[0167] Step S530: Find the processing priority corresponding to the fault type label from the preset fault type and processing priority correspondence table.

[0168] The preset fault type and processing priority correspondence table is a pre-defined table that records the processing priority corresponding to different fault types. The processing priority indicates the urgency of the processing required for different types of faults. For example, for a short circuit fault, immediate processing may be required, and its processing priority is high; for some minor overload faults, the processing priority may be relatively low. According to the fault type label, the corresponding processing priority is found in this correspondence table. For example, if the fault type label is "short circuit fault", the processing priority corresponding to "short circuit fault" is found to be "high" in the correspondence table. In this way, the urgency of processing different fault types can be determined, providing guidance for subsequent fault processing.

[0169] Step S540: Associate and integrate the candidate fault location, the corresponding fault occurrence probability and the processing priority to obtain the fault location result of the smart grid.

[0170] The candidate fault location, the corresponding fault occurrence probability and the processing priority are associated and integrated to unify the management and presentation of the related information of fault location. The specific integration process includes the steps of obtaining the topology structure information of the smart grid, calculating the electrical distance, weighting and fusing the fault occurrence probability, feature matching, adjusting the sorting position, field association and generating the fault identification code.

[0171] As an implementation, step S540 can specifically include steps S541-S547:

[0172] Step S541: Obtain the topology structure information of the smart grid, which contains the connection relationship and distance information between each monitoring point.

[0173] The topology structure information of the smart grid describes the physical connection relationship and distance information between each monitoring point in the power grid. The topology structure information can be obtained by querying the geographic information system (GIS) or the topology database of the power grid. For example, the GIS stores the geographic position information of each device and line in the power grid and their connection relationship, and the electrical distance and connection information between each monitoring point can be obtained through the GIS. The topology database specifically stores the topology structure data of the power grid, including the connection matrix and distance matrix between nodes, etc. By querying these data sources, complete topology structure information can be obtained.

[0174] Step S542: Calculate the electrical distance between each candidate fault location and the adjacent monitoring point based on the topology structure information.

[0175] The electrical distance between each candidate fault location and the adjacent monitoring point is calculated based on the topology structure information to further evaluate the influence range and propagation path of the fault. The electrical distance is the equivalent distance between two monitoring points in the power system, which not only considers the physical distance, but also considers the impedance, voltage level and other factors of the line. The electrical distance can be calculated using the method in circuit theory, such as calculating the equivalent impedance between nodes according to the resistance, reactance and admittance of the line, and then calculating the electrical distance according to the impedance value. For example, for two adjacent monitoring points A and B, the line parameters connecting them are obtained by querying the topology structure information, and then the electrical distance is calculated using the formula (where R is the line resistance and X is the line reactance). For each candidate fault location, the electrical distance between it and the adjacent monitoring point is calculated and recorded.

[0176] Step S543: Weight and fuse the electrical distance and the fault occurrence probability corresponding to the candidate fault location to adjust the value of the fault occurrence probability.

[0177] The electrical distance and the fault occurrence probability corresponding to the candidate fault position are weighted and fused in order to comprehensively consider the two factors of electrical distance and fault occurrence probability, and more accurately evaluate the possibility of the candidate fault position. The method of weighted fusion can select different weight distribution methods according to actual conditions. For example, a certain weight can be allocated to the electrical distance and the fault occurrence probability, and then they are combined by weighting. In this way, the numerical value of the fault occurrence probability can be more reasonable, and the influence of the electrical distance on the fault propagation is considered.

[0178] Step S544: The typical fault features corresponding to the fault type label are retrieved from the preset fault handling knowledge base, and the adjusted fault occurrence probability, the electrical distance and the typical fault features are matched for feature matching to verify the rationality of the candidate fault position.

[0179] The preset fault handling knowledge base is a knowledge base that stores typical fault features corresponding to different fault types. The typical fault features corresponding to the fault type label are retrieved from the knowledge base, and then the adjusted fault occurrence probability, the electrical distance and the typical fault features are matched for feature matching in order to verify the rationality of the candidate fault position. For example, for a certain fault type, the typical fault features can include voltage anomaly, current mutation, etc. The adjusted fault occurrence probability, the electrical distance of the candidate fault position are compared and matched with these typical fault features. If the related information of the candidate fault position is consistent with the typical fault features, it is considered that the candidate fault position is reasonable; otherwise, it is considered that the candidate fault position can be unreasonable. In the feature matching process, similarity calculation methods such as cosine similarity, Euclidean distance, etc. can be used to measure the similarity between the information of the candidate fault position and the typical fault features.

[0180] Step S545: For the candidate fault position with a matching degree lower than a preset threshold, the sorting position of the candidate fault position in the candidate list is reduced.

[0181] If the matching degree of the candidate fault position is lower than the preset threshold, it means that the similarity between the candidate fault position and the typical fault features is low, and it can not be the real fault position. Therefore, the sorting position of the candidate fault position in the candidate list is reduced. The preset threshold is a value determined according to actual experience and experimental results, which is used to judge whether the matching degree is high enough. In this way, the candidate points that are more likely to be the fault position can be further screened, and the accuracy of fault positioning is improved.

[0182] Step S546: The candidate fault positions are rearranged according to the adjusted sorting positions, and the rearranged candidate fault positions, the corresponding adjusted fault occurrence probability, the electrical distance and the handling priority are associated in fields.

[0183] The candidate fault locations are rearranged according to the adjusted ranking positions so as to make the candidate fault location list more in line with the actual fault possibility order. The rearranged candidate fault locations, the corresponding adjusted fault occurrence probabilities, the electrical distances and the processing priorities are field-associated so as to integrate these related information together to form a complete fault location record. For example, a database table or a data structure can be used to store these information, and each record contains the identification of the candidate fault location, the adjusted fault occurrence probability, the electrical distance and the processing priority and the like.

[0184] Step S547: generating a unique fault identification code for each associated candidate fault location, integrating all the associated candidate fault location information in the order of the fault identification codes to form a structured fault location data set, and determining the structured fault location data set as the result of the association and integration.

[0185] Generating a unique fault identification code for each associated candidate fault location is to facilitate the management and tracking of the fault location result. The fault identification code can be a string or a number, which is unique. For example, the fault identification code can be generated by combining the time stamp and the identification of the candidate fault location. Integrating all the associated candidate fault location information in the order of the fault identification codes to form a structured fault location data set. The structured fault location data set can be a list or an array, and each element is a record containing the candidate fault location information. In this way, the fault location result can be stored and presented in a standardized manner, which is convenient for subsequent fault handling and analysis. Finally, this structured fault location data set is the result of the association and integration, which accurately reflects the fault location situation of the smart grid.

[0186] It can be understood that the various algorithms involved in the above introduction of the embodiments of the present application, such as the Euclidean distance algorithm, the cosine distance algorithm and the like, can be known from the related contents in the prior art. In order to save space, the above-mentioned algorithms are not expanded in the embodiments of the present application. In addition, those skilled in the art can supplement the details according to the common knowledge in the art when implementing the scheme of the present application. For example, according to the common knowledge in the art, the dimensional conflict before feature fusion can be eliminated by normalization, the dimension difference can be eliminated by interpolation, the threshold can be reasonably set by combining historical data, experience or business scenario requirements, the model can be trained based on a general model training method, the number of layers in the model structure can be set based on actual needs, the activation function can be selected, and the like. The present application will not make redundant introduction to the too detailed implementation process.

[0187] Please refer to Figure 2 , Figure 2A structural diagram of a monitoring system is provided for an embodiment of the present application. The monitoring system can be a server or a computer system arranged in a background of a smart grid, and at least includes a processor 101, a communication interface 102 and a memory 103. The processor 101, the communication interface 102 and the memory 103 can be connected through a bus or other manners. The processor 101 (or a central processing unit (CPU)) is a calculation core and a control core of the monitoring system, which can analyze various instructions in the monitoring system and process various data of the monitoring system. The communication interface 102 can optionally include a standard wired interface, a wireless interface (such as WI-FI, a mobile communication interface, etc.), and can be used for transmitting and receiving data under the control of the processor 101; the communication interface 102 can also be used for transmitting and interacting data in the monitoring system. The memory 103 is a memory device in the monitoring system, which is used for storing programs and data. It can be understood that the memory 103 can include a built-in memory of the monitoring system, and of course can also include an extended memory supported by the monitoring system. The memory 103 provides a storage space, which stores an operating system of the monitoring system, and the present application is not limited thereto.

[0188] In an embodiment, the processor 101 executes the computer program in the memory 103 to implement the fault monitoring method applied to the smart grid provided in the above embodiments of the present application.

Claims

1. A fault monitoring method applied to a smart grid, characterized by, The method comprises the following steps: obtaining a multi-source operation time series data set of a smart grid; performing time window division and feature extraction processing on the multi-source operation time series data set to generate a multi-scale time series feature sequence; Specifically, the method comprises the following steps: obtaining a preset time window parameter set, performing window division on the multi-source operation time series data set based on short time window parameters in the time window parameter set to obtain a plurality of short time window data segments, performing amplitude change feature extraction on each short time window data segment, calculating the difference between the maximum value and the minimum value of the sampling data in each short time window data segment to obtain a short time window amplitude change feature, performing window division on the multi-source operation time series data set based on medium time window parameters in the time window parameter set to obtain a plurality of medium time window data segments, performing trend change feature extraction on each medium time window data segment, calculating the change slope of the sampling data in each medium time window data segment by a linear fitting method to obtain a medium time window trend change feature, performing window division on the multi-source operation time series data set based on long time window parameters in the time window parameter set to obtain a plurality of long time window data segments, performing fluctuation correlation feature extraction on each long time window data segment, calculating the standard deviation of the change rate between adjacent sampling data in each long time window data segment to obtain a long time window fluctuation correlation feature, concatenating the short time window amplitude change feature, the medium time window trend change feature and the long time window fluctuation correlation feature in time sequence to form an initial time series feature sequence, and performing feature alignment processing on the initial time series feature sequence, and determining the initial time series feature sequence after the feature alignment processing as the multi-scale time series feature sequence; inputting the multi-scale time series feature sequence into a fault evolution analysis model to perform state transition learning processing, and generating a power grid state transition feature set; inputting the power grid state transition feature set into a fault type recognition network to perform feature classification and location probability calculation processing, and obtaining a fault type label and a fault location probability distribution; determining a fault positioning result of the smart grid according to the fault type label and the fault location probability distribution.

2. The method of claim 1, wherein, The method comprises the following steps: extracting short time window parameters from the time window parameter set, wherein the short time window parameters include a window length parameter and a window sliding step parameter; determining the number of sampling data included in each short time window based on the window length parameter, and determining the interval number of sampling data between adjacent two short time windows based on the window sliding step parameter; taking the starting sampling data of the multi-source operation time series data set as the starting point of the first short time window, and obtaining the first short time window data segment by truncating the sampling data according to the window length parameter; Determining a starting point of a second short time window based on a starting point of a first short time window, moving a corresponding number of sampling data backward according to the window sliding step parameter, and obtaining a second short time window data segment by truncating the sampling data according to the window length parameter; Repeating the steps of determining the starting point of the short time window and truncating the sampling data until the window division of all sampling data of the multi-source runtime sequence data set is completed, and obtaining a plurality of short time window data segments; Adding a window serial number identifier to each short time window data segment, which represents the order of the short time window in the time sequence; The amplitude variation feature extraction of each short time window data segment, the difference between the maximum value and the minimum value of the sampling data in each short time window data segment is calculated, and the short time window amplitude variation feature is obtained, including: Traverse each short time window data segment, and extract all sampling data contained in the currently traversed short time window data segment; From the extracted sampling data, the sampling data with the maximum value is selected as the window maximum value, and the sampling data with the minimum value is selected as the window minimum value, and the difference between the window maximum value and the window minimum value is calculated to obtain the amplitude variation of the current short time window data segment; The amplitude variation is associated with the window serial number identifier of the current short time window data segment and stored; After completing the traversal of all short time window data segments, the amplitude variations of all short time window data segments are sorted according to the order of the window serial number identifier, and the sorted amplitude variation sequence is determined as the short time window amplitude variation feature.

3. The method of claim 1, wherein, The trend variation feature extraction of each medium time window data segment, the change slope of the sampling data in each medium time window data segment is calculated by linear fitting method, and the medium time window trend variation feature is obtained, including: Traverse each medium time window data segment, and extract all sampling data and corresponding sampling time points contained in the currently traversed medium time window data segment; Convert each sampling time point to a time sequence index value, which represents the relative order of the sampling time point in the medium time window; Based on the time sequence index value and the corresponding sampling data, a two-dimensional data point set is constructed, each data point in the two-dimensional data point set contains a time sequence index value and a corresponding sampling data value; The least square method is used to linearly fit the two-dimensional data point set to obtain a fitting straight line equation, the slope of the fitting straight line equation is the change slope, and the slope value is extracted from the fitting straight line equation to determine the trend change slope of the current medium time window data segment; The trend change slope is associated with the window serial number identifier of the current medium time window data segment and stored. After the traversal of all the medium time window data segments is completed, the trend change slopes of all the medium time window data segments are sorted in the order of the window sequence numbers, and a sorted trend change slope sequence is determined as the medium time window trend change feature.

4. The method of claim 1, wherein, The fluctuation correlation feature of each long time window data segment is extracted, the standard deviation of the change rate between adjacent sampling data in each long time window data segment is calculated, and a long time window fluctuation correlation feature is obtained, including: Each long time window data segment is traversed, and all sampling data contained in the currently traversed long time window data segment are extracted and arranged in a sampling data sequence according to the sampling time sequence; The change rate between two adjacent sampling data in the sampling data sequence is calculated, and all the change rates between adjacent sampling data are combined to form a change rate sequence; The standard deviation of the change rate sequence is calculated, and the standard deviation represents the dispersion degree of the data in the change rate sequence. The standard deviation is determined as the fluctuation correlation feature value of the current long time window data segment; The fluctuation correlation feature value is associated with the window sequence number identification of the current long time window data segment and is stored; After the traversal of all the long time window data segments is completed, the fluctuation correlation feature values of all the long time window data segments are sorted in the order of the window sequence numbers, and a sorted fluctuation correlation feature value sequence is determined as the long time window fluctuation correlation feature; The short time window amplitude change feature, the medium time window trend change feature and the long time window fluctuation correlation feature are concatenated in time sequence to form an initial time sequence feature sequence, including: The window sequence number identification sequence corresponding to the short time window amplitude change feature, the window sequence number identification sequence corresponding to the medium time window trend change feature and the window sequence number identification sequence corresponding to the long time window fluctuation correlation feature are obtained; A time alignment reference point is determined based on the window sequence number identification sequences. The time alignment reference point is the earliest common time point in the three window sequence number identification sequences; The short time window amplitude change feature, the medium time window trend change feature and the long time window fluctuation correlation feature are arranged in time sequence starting from the time alignment reference point; For feature parts with non-overlapping time ranges, zero padding is used for supplementation to make the lengths of the three features consistent in the time dimension. The arranged and supplemented short time window amplitude change feature, medium time window trend change feature and long time window fluctuation correlation feature are combined in feature dimension to form a three-dimensional feature matrix; The three-dimensional feature matrix is converted into a one-dimensional feature vector sequence, which is determined as the initial time sequence feature sequence.

5. The method of claim 1, wherein, The multi-scale time sequence feature sequence is input into a fault evolution analysis model for state transition learning processing to generate a power grid state transition feature set, including: inputting the multi-scale time sequence feature sequence into a state initialization layer of the fault evolution analysis model, constructing an initial state vector of the multi-scale time sequence feature sequence through the state initialization layer, and obtaining an initial power grid state vector; inputting the initial power grid state vector into a transition rule learning layer of the fault evolution analysis model, learning a state transition rule in historical fault case data through the transition rule learning layer, and generating a state transition rule set; based on the state transition rule set, simulating state transition of the initial power grid state vector to obtain a power grid state vector at a first time; comparing the power grid state vector at the first time with feature data at a corresponding time in the multi-scale time sequence feature sequence, and calculating a state deviation value; adjusting the state transition rule set according to the state deviation value, updating a state transition condition and a transition probability; based on the adjusted state transition rule set, simulating next state transition of the power grid state vector at the first time to obtain a power grid state vector at a second time; repeating the processes of state transition simulation, state deviation calculation and transition rule adjustment until state transition simulation at all monitoring times is completed, and obtaining a power grid state vector sequence at multiple times; performing feature difference processing on the power grid state vector sequence, calculating difference features between adjacent power grid state vectors, and obtaining state transition difference features; performing feature fusion on the state transition difference features and the transition probability in the state transition rule set to generate a power grid state transition feature set.

6. The method of claim 1, wherein, The inputting the power grid state transition feature set into a fault type identification network for feature classification and position probability calculation processing to obtain a fault type label and a fault position probability distribution includes: inputting the power grid state transition feature set into a feature input layer of the fault type identification network, and performing feature standardization processing on the power grid state transition feature set through the feature input layer; inputting the standardized power grid state transition feature set into a feature extraction layer of the fault type identification network, and performing deep feature extraction on the power grid state transition feature set through a convolutional neural network of the feature extraction layer to obtain fault deep features; inputting the fault deep features into a type classification layer of the fault type identification network, performing classification processing on the fault deep features through a fully connected network of the type classification layer, outputting probability values corresponding to each fault type, and selecting a label corresponding to a fault type with the highest probability value as the fault type label; inputting the fault deep features into a position prediction layer of the fault type identification network, performing time series analysis on the fault deep features through a recurrent neural network of the position prediction layer, outputting a fault occurrence probability corresponding to each monitoring point, and grouping fault occurrence probabilities corresponding to all monitoring points into a probability distribution vector as the fault position probability distribution.

7. The method of claim 1, wherein, The determining a fault positioning result of the smart power grid according to the fault type label and the fault position probability distribution includes: extracting a fault occurrence probability corresponding to each monitoring point from the fault position probability distribution; Sort all monitoring points according to fault occurrence probability from high to low, select the top pre-set number of monitoring points as candidate fault locations; From the pre-set fault type and processing priority correspondence table, find the processing priority corresponding to the fault type label; Integrate the candidate fault locations, corresponding fault occurrence probabilities and processing priorities to obtain the fault location result of the smart grid.

8. The method of claim 5, wherein, The state transition rule set is generated by learning the state transition law in the historical fault case data through the transition rule learning layer, including: Obtain historical fault case data from a pre-set database, the historical fault case data containing complete state transition process records of multiple historical fault events; Perform event segmentation processing on the historical fault case data, and divide the complete state transition process record of each historical fault event into a pre-fault state sequence, a fault occurrence state sequence and a post-fault state sequence; Classify the segmented state transition process records according to fault types to obtain a state transition sub-sequence set corresponding to each fault type; Perform time sequence alignment processing on each state transition sub-sequence set to make the state transition sub-sequences under the same fault type have consistent lengths in the time dimension; Calculate the transition frequency between adjacent states in each state transition sub-sequence, and count the number of transitions from the same starting state to the same ending state; Calculate the state transition probability based on the number of transitions, which is the ratio of the number of transitions from the same starting state to the same ending state to the total number of possible transitions of the starting state, and perform smoothing processing on the calculated state transition probability; Extract the key feature change amount in each state transition process, which is the core feature index triggering state transition; Store the state transition probability, key feature change amount and corresponding fault type in association to form an initial state transition rule; Iteratively optimize the initial state transition rule through the deep belief network of the transition rule learning layer, adjust the weight coefficient of the key feature change amount, and determine the optimized initial state transition rule as the state transition rule set.

9. A monitoring system, characterized by It includes: A memory in which a computer program is stored; A processor for loading the computer program to implement the fault monitoring method for smart grid according to any one of claims 1-8.

Images