Power distribution network multi-source data fusion method and system based on clustering technology

Abstract

A power distribution network multi-source data fusion method based on clustering technology, comprising the following steps: obtaining a plurality of data source sequences; dividing each sequence into a plurality of continuous sequence segments; clustering the sequence segments to generate clustering clusters; selecting a reference sequence segment in each clustering cluster and obtaining the time sequence offset of each other sequence segment in the clustering cluster after time sequence translation, when the sequence segment has the highest similarity with the reference sequence segment; traversing all clustering clusters to obtain the time sequence offset of each sequence segment; obtaining the global time sequence offset of each data source based on all time sequence offsets of each data source, and adjusting the time stamp of the sequence of the corresponding data source based on the global time sequence offset; based on the sequence of the plurality of data sources after adjusting the time stamp, iterating the above clustering adjustment step until the sequence corresponding to each data source after the i-th adjustment is obtained, and outputting as a target sequence; and obtaining at least one fusion data set of the same event based on the target sequence of the plurality of data sources.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and system for fusing multi-source data of power distribution networks based on clustering technology. Background Technology

[0002] During the current transition phase of rapid digitalization in power distribution networks, a large number of old and new devices coexist: on one end are high-frequency sensing terminals equipped with µPMU (micro-Phasor Measurement Unit), smart ring network cabinets, IoT sensors, etc., while on the other end are outdated RTU (Remote Terminal Unit), smart meters, and SCADA (Supervisory Control and Data Acquisition) measurement systems lacking GPS / PTP (Precision Time Protocol, such as IEEE1588) timing capabilities. Because underground substations, tunnels, and long rural feeders often cannot reliably receive satellite signals, or communication links (4G, PLC, narrowband IoT) exhibit significant jitter, traditional hardware-based time synchronization methods (GPS, IRIG-B (Inter-Range Instrumentation Group time code B), NTP (Network Time Protocol)) are costly, slow to implement, and have limited coverage, resulting in multi-source measurement timestamp errors ranging from milliseconds to minutes. Summary of the Invention

[0003] This invention provides a method for fusing multi-source data in a power distribution network based on clustering technology, comprising the following steps:

[0004] S01. Obtain sequences from multiple data sources, wherein at least two of the multiple data sources record at least one identical event.

[0005] S02. Divide each sequence into several consecutive sequence segments, with any two sequence segments corresponding to the same sampling duration;

[0006] S03. Clustering sequence fragments to generate at least one cluster, wherein any cluster comprises at least two sequence fragments;

[0007] S04. Select a reference sequence fragment in each cluster, and obtain the temporal offset of other sequence fragments in each cluster when they have the highest similarity to the reference sequence fragment after temporal translation.

[0008] S05. For each sequence segment of each data source, traverse all clusters and obtain the temporal offset of each sequence segment.

[0009] S06. Based on all time series offsets corresponding to each data source, obtain the global time series offset for each data source, and adjust the timestamp of the corresponding data source sequence based on the global time series offset.

[0010] S07. Based on the sequences of multiple data sources after adjusting the timestamps, repeat steps S02 to S05 until the sequence corresponding to each data source after the i-th adjustment is obtained, and output as the target sequence.

[0011] S08. Based on the target sequence from multiple data sources, obtain a fused dataset of at least one identical event, wherein the timestamps of any two sets of data in the fused dataset are consistent.

[0012] In some examples provided by this invention, the partitioning of any sequence includes the following steps:

[0013] Set a time window and a moving step size, wherein the time window and the moving step size are the same sampling duration;

[0014] Using the time window, several continuous sequence segments are extracted from this sequence based on the moving step size.

[0015] In some examples provided by this invention, the clustering sequence fragment generates at least one cluster, comprising the following steps:

[0016] Extract the feature vector for each sequence segment, the feature vector including time-domain features and / or frequency-domain features;

[0017] All feature vectors are mapped to the same feature space, and clusters representing different events are formed in the feature space, where all sequence fragments in any cluster express the same event.

[0018] In some examples provided by this invention, the selection of a reference sequence fragment in any cluster includes the following steps:

[0019] Obtain the similarity between any two sequence segments in each cluster;

[0020] Based on any given sequence segment, summarize its similarity to other sequence segments;

[0021] The sequence segment with the highest sum of similarity values ​​with other sequence segments is selected as the reference sequence segment in this cluster.

[0022] In some examples provided by this invention, the temporal offset of another sequence segment in a cluster that has the highest similarity to a reference sequence segment after temporal translation satisfies the following acquisition steps:

[0023] Based on the interval accuracy, within the sampling duration range, multiple candidate timing drift values ​​are set one by one from small to large.

[0024] For the other sequence segment, obtain the sequence segment after temporal shift based on each candidate temporal drift amount;

[0025] The similarity between each adjusted sequence segment and the reference sequence segment is obtained, and the candidate time drift that makes the adjusted sequence segment have the highest similarity with the reference sequence segment is selected as the time offset.

[0026] In some examples provided in this invention, the similarity between any two sequence segments is obtained through the following steps:

[0027] The amplitude of the two sequence segments is normalized.

[0028] Based on the sequence segments after amplitude normalization, the peak value of the cross-correlation function between the two is obtained;

[0029] Based on the sequence segments after amplitude normalization, the corresponding main frequency components are extracted respectively, and the cosine similarity of the main frequency components of the two is obtained.

[0030] The similarity between two corresponding sequence segments is obtained by weighted summing of the peak value of the cross-correlation function and the cosine similarity.

[0031] In some examples provided in this invention, the global timing offset of any data source is the average timing offset of all timing offsets corresponding to this data source.

[0032] In some examples provided by the present invention, the iteration number i in S07 is a fixed iteration number.

[0033] In some examples provided by the present invention, in step S07, the difference between the global timing offset of each data source after the i-th adjustment and the global timing offset of the corresponding data source after the (i-1)-th adjustment is less than or equal to the preset offset.

[0034] In some embodiments provided by the present invention, based on the above-described example of the distribution network multi-source data fusion method based on clustering technology, a distribution network multi-source data fusion system based on clustering technology is also provided, which includes an input device, a processor, a memory, and an output device;

[0035] The input device, the processor, the memory, and the output device are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to call the program instructions to execute the above-mentioned multi-source data fusion method for power distribution networks based on clustering technology.

[0036] The multi-source data fusion method and system for power distribution networks based on clustering technology provided by this invention includes the following gains:

[0037] This invention can automatically identify the same event cluster by dividing time-series data from multiple sources in a power distribution network into equal-length sequence segments, extracting time-domain / frequency-domain features, and clustering them in a unified feature space. Simultaneously, within each cluster, the most representative reference segment is selected, and the time offset of each segment relative to the reference is calculated using cross-correlation and cosine similarity of the dominant frequency. This allows for the calculation of the average global time-series offset of each data source and correction of the original timestamp. Furthermore, based on multiple iterations until the offset converges, time-series alignment of multi-source data is achieved. Attached Figure Description

[0038] Figure 1 A flowchart of a multi-source data fusion method for power distribution networks based on clustering technology, as provided in this invention; and

[0039] Figure 2 This is a schematic diagram of a multi-source data fusion system for power distribution networks based on clustering technology, provided as an example of the present invention. Detailed Implementation

[0040] In the following description, specific details such as particular systems, structures, and technologies are presented for illustrative purposes rather than limiting, in order to provide a thorough understanding of the examples in this application.

[0041] Those skilled in the art will understand that this application can also be implemented in other examples without these specific details.

[0042] In the description of this application, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary details.

[0043] Additionally, it should be noted that the terms "first" and "second" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0044] During the current transition phase of rapid digitalization in power distribution networks, the coexistence of numerous old and new devices can lead to situations where there is no unified time synchronization or GPS synchronization issues. To eliminate time discrepancies between multi-source measurement data, this invention provides a multi-source data fusion method for power distribution networks based on clustering technology.

[0045] See one example. Figure 1 , Figure 1 The flowchart illustrates the multi-source data fusion method for power distribution networks based on clustering technology, as an example of the present invention.

[0046] like Figure 1 As shown, the multi-source data fusion method for power distribution networks based on clustering technology provided by this invention includes the following steps:

[0047] S01. Obtain sequences from multiple data sources, wherein at least two of the multiple data sources record at least one identical event.

[0048] Furthermore, any of the data sources mentioned in this example are independent devices, equipment, or systems that generate and store power distribution network transportation data, such as SCADA master stations, PMU master stations, DFRs (fault recorders), smart meter concentrators, IoT data sensors, etc.

[0049] It should be noted that, in this example, any data source corresponding to the first sequence refers to the original sequence segment output by this data source within a given time window, specifically including the original timestamp and the original measurement value; further, it should be noted that the original measurement values ​​of multiple data sources can be the measurement values ​​of the same physical quantity or the measurement values ​​of different physical quantities.

[0050] In one example, step S01 obtains the first sequence of data source S1, data source S2, and data source SN. The first sequence of each data source records at least one physical event that occurred in the same distribution network, and all of them are the same physical quantity, such as voltage value.

[0051] In one or more other examples, there exists a first sequence of different physical quantities that record the same physical event, and there also exists a sequence that records other physical quantities, such as the first sequence of data source S1 recording voltage, the sequence of data source S2 recording current, and the third sequence of data source S3 recording voltage.

[0052] Furthermore, it should be noted that the event proposed in any example of the present invention refers to a physical process or human operation that occurs in the distribution network within a finite time window and can cause a sudden change or anomaly in at least one measurable physical quantity, and such a sudden change or anomaly is simultaneously observed in the original sequences of at least two independent data sources.

[0053] S02. Divide each sequence into several consecutive sequence segments, with any two sequence segments corresponding to the same sampling duration.

[0054] In this example, the partitioning of any sequence includes the following steps:

[0055] S021. Set the time window and the moving step size, wherein the time window and the moving step size are the same sampling duration.

[0056] Specifically, based on the minimum common sampling period of each data source (e.g., 1s, 5s, or 10s), the "time window W" and "movement step size S" are determined, and W=S, both equal to the minimum sampling period. This ensures that each segment covers the original sequence of the same length, avoiding additional overlap due to a step size smaller than the window, and also avoiding gaps due to a step size larger than the window, thus preventing time discontinuities or data redundancy during subsequent alignment.

[0057] S022. Using the time window, extract several continuous sequence segments from this sequence based on the moving step size.

[0058] Starting from the beginning of the sequence t0, the time window W is shifted sequentially in steps S, extracting consecutive segments P1=[t0,t0+W), P2=[t0+S,t0+S+W), ..., until the end of the sequence. For regions at the end of the sequence that are less than a complete window, a "zero-padding" or "back-forward" strategy can be used to uniformly pad them to length W, ensuring that all segments have equal length.

[0059] Furthermore, to improve segmentation quality and event coverage, the window length and step size can be dynamically adjusted based on the sampling frequency and event mutation characteristics of the data source: during periods of high noise or dense fault mutations, W can be amplified and S increased to enhance signal energy within the segment; under stable operating conditions, W can be reduced to obtain higher temporal resolution. After segmentation, each segment is labeled with its original sequence number, start timestamp, and source ID, providing an accurate index and alignment basis for downstream feature extraction and cross-source clustering.

[0060] S03. Clustering sequence fragments to generate at least one cluster, wherein any cluster comprises at least two sequence fragments.

[0061] In this example, the clustering sequence fragment generates at least one cluster, including the following steps:

[0062] S031. Extract the feature vector of each sequence segment respectively, wherein the feature vector includes time domain features and / or frequency domain features.

[0063] Furthermore, multidimensional features are extracted for each equal-length sequence segment. These features include both the segment's statistics in the time domain (such as mean, variance, peak value, and distribution of abrupt change points) and its frequency domain features (such as dominant frequency, harmonic energy ratio, and spectral entropy). In some examples, wavelet packet or empirical mode decomposition can also be combined to obtain a multi-scale description in the time and frequency domains.

[0064] S032. Map all feature vectors to the same feature space, and in the feature space, cluster different events into clusters, where all sequence fragments in any cluster express the same event.

[0065] Furthermore, the standardized feature vectors of all sequence fragments are projected onto the same feature space, and several clusters are automatically identified in this space using algorithms such as density clustering, spectral clustering, or Gaussian mixture model. All fragments in each cluster represent the same physical event, while noise or isolated samples are labeled separately for subsequent processing.

[0066] In some examples, the entire clustering process adaptively optimizes the clustering parameters using metrics such as the silhouette coefficient or the Davies-Bouldin index, and can perform preliminary clustering in both the time and frequency domain feature subspaces. Finally, the subclusters are merged into highly reliable event clusters using consistent clustering, thus laying a solid foundation for subsequent alignment and correction.

[0067] S04. Select a reference sequence fragment in each cluster, and obtain the temporal offset of other sequence fragments in each cluster when they have the highest similarity to the reference sequence fragment after temporal translation.

[0068] In this example, a most representative "reference fragment" is automatically determined within each cluster for subsequent time-series offset calculations. The reference sequence fragment for each cluster is obtained through the following steps:

[0069] S041. Obtain the similarity between any two sequence segments in the cluster.

[0070] S042. Based on any sequence segment, summarize its similarity to other sequence segments.

[0071] Understandably, for each sequence fragment in a cluster, its similarity to all other sequence fragments in the same cluster is summed one by one to generate a cluster centrality score for that sequence fragment.

[0072] S043. Select the sequence segment with the highest summation value of similarity with other sequence segments as the reference sequence segment in this cluster.

[0073] Furthermore, the segment with the highest score is selected as the reference sequence segment for this cluster, which not only ensures the best coverage of the common patterns of each segment within the cluster, but also enhances robustness against noise and isolated anomalies.

[0074] The unsupervised, adaptive reference segment selection mechanism provided in this example enables each cluster to automatically lock onto the timing reference that best represents the event pattern without relying on external priors, thus providing support for subsequent dynamic time alignment and global correction.

[0075] In this example, the temporal offset of another sequence segment in a cluster that has the highest similarity to the reference sequence segment after temporal translation is obtained by the following steps:

[0076] S044. Based on the interval accuracy, within the sampling duration range, multiple candidate timing drift values ​​are set one by one from small to large.

[0077] Furthermore, within the time range corresponding to the aforementioned time window, a series of candidate time-series drift values ​​are generated sequentially from smallest to largest, using the minimum sampling interval among multiple data sources as the precision.

[0078] S045. For the other sequence segment, obtain the sequence segment after it has been shifted in time based on each candidate time drift amount.

[0079] Specifically, for each candidate drift, the entire sequence segment to be corrected is shifted along the time axis by the candidate time drift, and the matching degree between the shifted sequence and the reference sequence is calculated based on the same amplitude normalization strategy.

[0080] S046. Obtain the similarity between each adjusted sequence segment and the reference sequence segment, and select the candidate time drift amount that makes the adjusted sequence segment have the highest similarity with the reference sequence segment as the time offset.

[0081] Specifically, by traversing all candidate drift values ​​and comparing their corresponding similarities, the time drift value that enables the adjusted sequence to achieve the highest similarity with the reference sequence is selected as the time offset of the segment. This not only helps to align the same event from different sources, but also takes into account the deep consistency of the spectral structure, thus providing a high-precision foundation for subsequent global clock correction and multi-source data fusion.

[0082] Furthermore, the similarity between any two sequence segments, such as the similarity between a reference sequence segment and another sequence segment, is obtained through the following steps:

[0083] S0411. Normalize the amplitude of the two sequence segments.

[0084] Specifically, normalizing the amplitude of any two equal-length sequence segments within a cluster can eliminate the effects of differences in source dimensions and amplitudes.

[0085] S0412. Based on the sequence segments after amplitude normalization, obtain the peak value of the cross-correlation function between the two.

[0086] Specifically, by calculating the cross-correlation function of the normalized signals in the time domain and extracting the peak value, the similarity index of the two sequences at the current alignment position can be obtained.

[0087] S0413. Based on the sequence segments after amplitude normalization, extract the corresponding main frequency components and obtain the cosine similarity of the main frequency components of the two.

[0088] Specifically, by performing a Fast Fourier Transform on each sequence segment, extracting its main frequency component, and calculating the cosine similarity between the two main frequency component vectors, the similarity of the signals in terms of spectral morphology can be reflected.

[0089] S0414. Weight the peak value of the cross-correlation function and the cosine similarity to obtain the similarity between the two corresponding sequence segments.

[0090] Specifically, according to a pre-set weighting ratio, the cross-correlation peak value and the frequency domain cosine similarity are weighted to synthesize a unified similarity score. It should be noted that the pre-set weighting ratio is determined in advance during the offline stage through historical alignment accuracy verification or expert experience. For example, the time domain similarity weight is set to 0.7 and the frequency domain similarity weight is set to 0.3 to balance the accuracy of time domain alignment with the consistency of frequency domain patterns.

[0091] S05. For each sequence segment of each data source, traverse all clusters and obtain the temporal offset of each sequence segment.

[0092] Specifically, each sequence fragment of each data source is traversed sequentially, and the records that appear in all clusters are extracted and summarized one by one.

[0093] Furthermore, for a sequence fragment, all clusters generated in the aforementioned steps are scanned to find a list of clusters containing the fragment; then, for each cluster, the temporal offset calculated for the sequence fragment in step S04 is read, and the confidence index of the corresponding cluster (such as the mean intra-cluster similarity or silhouette coefficient) is obtained simultaneously, which is used as the weight of the offset confidence; subsequently, all candidate offsets are weighted and averaged or weighted and median is calculated to obtain the final temporal offset value of the fragment.

[0094] It should be noted that for sequence segments that are not identified in any cluster (noise points), nearest neighbor cluster offset interpolation or default values ​​(such as zero offset) are used for compensation.

[0095] In this example, the above method not only ensures that each segment obtains the most matching offset that integrates multi-cluster information, but also utilizes dynamic weighting based on cluster quality to further improve the robustness and accuracy of offset estimation, laying a reliable foundation for subsequent timestamp correction based on global offset.

[0096] S06. Based on all time series offsets corresponding to each data source, obtain the global time series offset for each data source, and adjust the timestamp of the sequence of the corresponding data source based on the global time series offset.

[0097] In this example, the global time offset of any data source is the average time offset of all time offsets corresponding to this data source.

[0098] Specifically, firstly, the temporal offsets of all sequence segments under the same data source are aggregated. After outlier removal from these offsets, the arithmetic mean is calculated as the global temporal offset of the data source. Then, this global offset is rounded to an integer multiple consistent with the minimum sampling interval, and all original timestamps are uniformly shifted by this amount, thereby completing the overall correction of the sequence of the data source.

[0099] S07. Based on the sequences of multiple data sources after adjusting the timestamps, repeat steps S02 to S05 until the sequence corresponding to each data source after the i-th adjustment is obtained, and output as the target sequence.

[0100] In this example, in step S07, the difference between the global timing offset of each data source after the i-th adjustment and the global timing offset of the corresponding data source after the (i-1)-th adjustment is less than or equal to the preset offset.

[0101] Furthermore, in each completion of the global offset ΔT （n） After calculation, it is compared with the offset ΔT from the previous round. （n-1） Perform a source-by-source comparison. When all data sources satisfy |ΔT （n） –ΔT （n-1） When |≤ε (ε is the pre-set minimum detectable offset threshold, such as 0.01s), the timing correction can be considered to have converged. At this point, the iteration stops, and the corrected multi-source sequences are output as the final target sequence. The choice of threshold ε should take into account both the maximum latency tolerance of the target application and the convergence speed of the algorithm.

[0102] In some other examples, the iteration number i in S07 is a fixed number of iterations.

[0103] Furthermore, in these real-time or computationally resource-constrained scenarios, the maximum number of iterations, i, can be preset. Even if the difference has not fully converged, the iteration is forcibly terminated after the i-th global offset correction to ensure that the overall process duration is controllable. At this time, the output target sequence may have a slight clock drift, but it can meet the accuracy requirements of most engineering applications.

[0104] In some other embodiments, if the offset oscillates repeatedly or exceeds the preset maximum offset range during the iteration process, an abnormal warning is triggered, indicating that there may be an abnormal source clock or clustering grouping error, requiring manual intervention or separate diagnosis.

[0105] It should be noted that once the iteration stops, the final corrected time series of each data source becomes the target sequence, which can be directly used for subsequent cross-source data fusion, state estimation, or fault diagnosis. At the same time, the clock stability of each data source can be evaluated through the iteration log, providing a reference for subsequent operation and maintenance.

[0106] S08. Based on the target sequence from multiple data sources, obtain a fused dataset of at least one identical event, wherein the timestamps of any two sets of data in the fused dataset are consistent.

[0107] Furthermore, step S08, based on the multi-source target sequence obtained after the previous iterative correction, locks at least one event interval in which all sources respond together through unified event identification or time indexing, and constructs a common time grid within this interval.

[0108] Furthermore, at each moment of the aforementioned public event grid, observations from each data source are extracted synchronously. For missing data caused by incomplete sampling or data gaps, strategies such as linear interpolation or preserving previous values ​​are used to fill in the gaps. At the same time, obvious abnormal peaks and valleys are removed or marked.

[0109] Finally, the observation vectors from all sources at each moment are concatenated one-to-one according to the timestamp to form a complete fusion dataset. In this dataset, any two data points correspond to the same timestamp, which can support subsequent state estimation and fault diagnosis, and also provide highly reliable, multi-dimensional input for accurate load forecasting and optimized scheduling.

[0110] This invention dynamically identifies data sequences of the same event from different data sources through clustering, and finds the optimal time series within the data sequence cluster under the same event to align the time offset, thereby enabling clock mutual calibration among data sources under the same event. This results in the formation of a time-calibrated multi-source unified dataset at the event granularity level, achieving optimal data fusion at the event level.

[0111] In yet another example, see Figure 2 , Figure 2 This is a schematic diagram of a multi-source data fusion system for power distribution networks based on clustering technology.

[0112] like Figure 2 As shown, based on the above-described example, the present invention provides a distribution network multi-source data fusion method based on clustering technology, and also provides a distribution network multi-source data fusion system based on clustering technology, which includes an input device, a processor, a memory, and an output device.

[0113] Furthermore, the input device, the processor, the memory, and the output device are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to call the program instructions to execute the above-described distribution network multi-source data fusion method based on clustering technology.

[0114] In the examples above, the descriptions of each example have their own emphasis. For parts that are not described or recorded in detail in a certain example, please refer to the relevant descriptions in other examples.

[0115] It should be noted that the above examples can be freely combined as needed. The above are merely preferred embodiments of the present invention; it should be observed that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for fusing multi-source data from a power distribution network based on clustering technology, characterized in that, Includes the following steps: S01. Obtain sequences from multiple data sources, wherein at least two of the multiple data sources record at least one identical event. S02. Divide each sequence into several consecutive sequence segments, with any two sequence segments corresponding to the same sampling duration; S03. Clustering sequence fragments to generate at least one cluster, wherein any cluster comprises at least two sequence fragments; S04. Select a reference sequence fragment in each cluster, and obtain the temporal offset of other sequence fragments in each cluster when they have the highest similarity to the reference sequence fragment after temporal translation. S05. For each sequence segment of each data source, traverse all clusters and obtain the time offset of each sequence segment. S06. Based on all time series offsets corresponding to each data source, obtain the global time series offset for each data source, and adjust the timestamp of the corresponding data source sequence based on the global time series offset. S07. Based on the sequences of multiple data sources after adjusting the timestamps, repeat steps S02 to S05 until the sequence corresponding to each data source after the i-th adjustment is obtained, and output as the target sequence. S08. Based on target sequences from multiple data sources, obtain a fused dataset with at least one identical event, wherein the timestamps of any two sets of data in the fused dataset are consistent. The selection of a reference sequence fragment in any cluster in step S04 includes the following steps: obtaining the similarity between any two sequence fragments in the cluster; summarizing the similarity between any sequence fragment and other sequence fragments; and selecting the sequence fragment with the highest sum of similarity values ​​with other sequence fragments as the reference sequence fragment in this cluster. In step S04, the temporal offset of another sequence segment in any cluster that has the highest similarity to the reference sequence segment after temporal translation satisfies the following acquisition steps: based on the interval accuracy, multiple candidate temporal drift amounts are set one by one from small to large within the sampling duration; for the other sequence segment, the sequence segment after temporal translation based on each candidate temporal drift amount is obtained; the similarity between each adjusted sequence segment and the reference sequence segment is obtained, and the candidate temporal drift amount that makes the adjusted sequence segment have the highest similarity to the reference sequence segment is selected as the temporal offset. Further, in step S04, the similarity between any two sequence segments is obtained through the following steps: normalizing the amplitude of the two sequence segments; obtaining the peak value of the cross-correlation function of the two segments based on the amplitude-normalized sequence segments; extracting the corresponding main frequency components based on the amplitude-normalized sequence segments and obtaining the cosine similarity of the main frequency components of the two segments; and weighted summing the peak value of the cross-correlation function and the cosine similarity to obtain the similarity between the two corresponding sequence segments.

2. The method for multi-source data fusion of power distribution networks based on clustering technology according to claim 1, characterized in that, The partitioning of any sequence includes the following steps: Set a time window and a moving step size, wherein the time window and the moving step size are the same sampling duration; Using the time window, several continuous sequence segments are extracted from this sequence based on the moving step size.

3. The method for multi-source data fusion of power distribution networks based on clustering technology according to claim 1, characterized in that, The clustering sequence fragment generates at least one cluster, including the following steps: Extract the feature vector for each sequence segment, the feature vector including time-domain features and / or frequency-domain features; All feature vectors are mapped to the same feature space, and clusters representing different events are formed in the feature space, where all sequence fragments in any cluster express the same event.

4. The method for multi-source data fusion of distribution networks based on clustering technology according to claim 1, characterized in that, The global time series offset of any data source is the average time series offset of all time series offsets corresponding to this data source.

5. The method for multi-source data fusion of distribution networks based on clustering technology according to claim 1, characterized in that, The iteration number i in S07 is a fixed number of iterations.

6. The method for multi-source data fusion of power distribution networks based on clustering technology according to claim 1, characterized in that, In step S07, the difference between the global timing offset of each data source after the i-th adjustment and the global timing offset of the corresponding data source after the (i-1)-th adjustment is less than or equal to the preset offset.

7. A multi-source data fusion system for power distribution networks based on clustering technology, characterized in that, The method includes an input device, a processor, a memory, and an output device, wherein the input device, the processor, the memory, and the output device are interconnected, the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the multi-source data fusion method for distribution networks based on clustering technology as described in any one of claims 1 to 6.

Images