Power distribution network user topology intelligent identification method based on dynamic time sequence feature clustering

Abstract

The invention provides a power distribution network user topology intelligent identification method based on dynamic time sequence feature clustering. The method comprises the following steps: 1, acquiring historical voltage data of a low-voltage power distribution area user electric meter to construct a time sequence data set; 2, performing feature extraction on the voltage time sequence by using a deep learning neural network; 3, optimizing the feature extraction network by using the double-layer triple loss based on the Soft-DTW distance; and 4, the Soft-DTW distance between the user voltage feature sequences is calculated, the users are clustered by using a minimum and maximum Soft-DTW distance algorithm, and clustering labels are output to obtain user topology information. The method is suitable for topology identification of the low-voltage power distribution network users, topology identification can be realized through data of the low-voltage power distribution network users, and the management level of the low-voltage power distribution network is further improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of low-voltage distribution network user topology identification, and specifically relates to a distribution network user topology intelligent identification method based on dynamic time series feature clustering. Background Art

[0002] As the final link in the power system's transmission system, the low-voltage distribution network steps down medium-voltage electricity to low voltage and transmits it to end users. Currently, the majority of power grid measurement, control, and protection is focused on medium-voltage and high-voltage lines, leaving most low-voltage lines and low-voltage users largely unmonitored. With the expansion of urban development and the widespread connection of distributed photovoltaic (PV) and charging stations to low-voltage distribution areas, grid operators face difficulties in obtaining reliable and complete topological information about the low-voltage distribution network.

[0003] Topological information for low-voltage distribution networks is crucial for grid operators. Accurate topological information facilitates the precise management of low-voltage distribution substations. One method for obtaining this information is through on-site verification and registration, but this method is time-consuming and labor-intensive, making it difficult to implement when substation topology frequently changes.

[0004] Currently, common topology identification methods include installing dedicated hardware for topology identification, performing ordered power outages within stations to obtain information about users experiencing power outages, and performing topology identification based on user electricity usage data. Installing dedicated hardware for topology identification carries high hardware and labor costs. Furthermore, due to the frequent changes in the topology of low-voltage distribution substations, obtaining topology information through ordered power outages can disrupt user lives and easily cause the information to become outdated. Using a topology identification algorithm based on user electricity usage data not only fully utilizes existing resources but also reduces system construction and maintenance costs, providing a more cost-effective solution for improving user topology information perception within low-voltage distribution substations.

[0005] Compared with the existing technology, the technical differences are as follows:

[0006] Technical comparison with patent CN108564485B "Low-voltage area user phase identification method based on voltage curve similarity analysis"

[0007] Patent CN108564485B is applicable to phase identification of users in low-voltage substations. It requires the substation to be known, with a focus on distinguishing between phases A, B, and C.

[0008] This patent can simultaneously realize the user's substation affiliation identification and phase identification. It is suitable for low-voltage distribution networks with unknown topology structures and is more adaptable to complex substations (such as multiple distribution transformers and low three-phase voltage imbalance).

[0009] There are essential differences between the two in their application scenarios.

[0010] Patent CN108564485B discloses a low-voltage substation user phase identification method based on voltage curve similarity analysis. It directly measures the similarity of the user's original voltage time series through the DTW algorithm, and uses the k-distance neighborhood to calculate the local outlier factor to screen the correct substation users. Finally, the DTW distance between the user and the distribution transformer three-phase voltage is calculated to determine the phase. No clustering algorithm is used.

[0011] This patent uses deep learning networks (1D-CNN and Bi-GRU) for training to automatically extract the features of user voltage time series, and uses a two-level triplet loss to improve the structural sensitivity of the deep learning network. It combines the differentiable Soft-DTW algorithm to measure the similarity of user feature sequences to achieve end-to-end training. Finally, it uses the minimum-maximum Soft-DTW distance clustering algorithm to perform cluster analysis on users to determine the user's station area and phase, which is more robust.

[0012] There are essential differences between the two in terms of their technical implementation solutions.

[0013] Technical comparison with patent CN112952828A "A method for analyzing user areas in low-voltage distribution networks"

[0014] Patent CN112952828A is applicable to the identification of the user's substation area in a low-voltage substation area, but cannot identify the user's phase.

[0015] This patent can simultaneously realize substation affiliation identification and phase identification, and is suitable for the overall topological relationship identification of low-voltage substation user substations, with a wider range of applicability.

[0016] There are essential differences between the two in their application scenarios.

[0017] Patent CN112952828A uses matrix completion, DTW algorithm to process asynchronous data and fuzzy C-means clustering to achieve substation identification, focusing on solving the problems of missing data and time series asynchrony. However, it relies on manual data preprocessing and lacks flexibility when facing multi-substation and multi-user scenarios.

[0018] This patent automatically extracts the features of voltage time series through a deep learning model, uses the differentiability of Soft-DTW to train and optimize the model, and adopts minimum and maximum Soft-DTW distance clustering for cluster analysis. It supports the identification of substations and phases and is suitable for substations with a large number of users and a large amount of data.

[0019] There are essential differences between the two in terms of their technical implementation solutions.

[0020] Technical comparison with patent CN116454856A "A method for identifying household transformer topology in low-voltage distribution networks based on SC-DTW algorithm"

[0021] Patent CN116454856A is applicable to phase identification of low-voltage distribution network users. It requires the known substation affiliation and is suitable for substations with obvious differences in voltage curve morphology, and has high requirements for timing alignment.

[0022] This patent can simultaneously realize the user's substation affiliation identification and phase identification, and has high robustness, and is suitable for scenarios with high similarity in voltage curve morphology.

[0023] There are essential differences between the two in their application scenarios.

[0024] Patent CN116454856A uses a shape context cost matrix instead of the Euclidean distance matrix used in DTW to improve the accuracy of morphological similarity measurement and uses the k-means clustering algorithm to cluster to obtain the household transformer topology of the low-voltage distribution station area. However, it relies on the manually designed pole logarithmic coordinate histogram feature and is sensitive to differences in voltage curve morphology.

[0025] This patent uses a deep learning model to dynamically model timing characteristics, utilizes the gradient optimization capability of Soft-DTW to improve model robustness, strengthens topological hierarchical constraints through dual-level loss, and uses the minimum-maximum Soft-DTW distance clustering algorithm to simultaneously achieve end-to-end identification of substations and phases. It can achieve more accurate user substation attribution and phase identification in scenarios with noise and low voltage curve differences.

[0026] There are essential differences between the two in terms of their technical implementation solutions. Summary of the Invention

[0027] To address the problem of frequent changes in user topology in low-voltage distribution network areas, this paper proposes an intelligent identification method for distribution network user topology based on dynamic time series feature clustering. It extracts the features of user voltage time series through a deep learning algorithm and clusters users using the minimum-maximum Soft-DTW distance clustering algorithm, thereby achieving accurate identification of user topology in low-voltage distribution network areas.

[0028] To achieve the above object, the technical solution adopted by the present invention is:

[0029] The intelligent identification method of distribution network user topology based on dynamic time series feature clustering has the following specific steps:

[0030] Step 1: Read the measured values ​​of user voltage data in the low-voltage distribution network and establish a historical data set;

[0031] Step 2: Establish a deep learning model for extracting user voltage time series features. Use the data in the training set to train the deep learning model to obtain the feature sequence F = f(X) of the voltage time series.

[0032] Step 3: Optimize the feature extraction network using a two-level triplet loss based on Soft-DTW distance;

[0033] Step 4: Calculate the Soft-DTW distance between user voltage feature sequences and use the minimum-maximum Soft-DTW distance algorithm to cluster users and output cluster labels to obtain user topology information.

[0034] As a further improvement of the present invention, the establishment of the neural network of the deep learning model in step 2 includes the following process:

[0035] (2-1) The first layer of one-dimensional convolution: kernel = 3, stride = 1, channel = 32;

[0036] (2-2) Pooling layer 1: Max-pooling, pool size = 2;

[0037] (2-3) Second layer of one-dimensional convolution: kernel = 3, stride = 1, channel = 64;

[0038] (2-4) Pooling layer 2: Max-pooling, pool size = 2;

[0039] (2-5) Bi-GRU structure: hidden size = 128, bidirectional;

[0040] (2-6) Fully connected layer: maps the 256-dimensional output to 64 dimensions;

[0041] (2-7) Delayed pooling: Average pooling with a stride of 2 is used on the sequence during training, and the original resolution is retained during testing.

[0042] As a further improvement of the present invention, the two-level triplet loss based on Soft-DTW distance in step 3 includes the following process:

[0043] (3-1) The district-level triplet is selected from the district T k , the voltage time series X of phase Z a As the anchor point, select the voltage time series that belongs to the same substation as the anchor point As a positive sample, select the voltage time series that does not belong to the same substation as the anchor point As a negative sample, the district-level triplet is recorded as The area-level triplet loss is defined as:

[0044]

[0045] Where, δ sis a marginal hyperparameter used to constrain the distance difference between the positive and negative sample pairs of the station-level samples.

[0046] (3-2) Phase-level triplet selection: Voltage time series that belongs to the same substation and the same phase as the anchor point As a positive sample, select the voltage time series that belongs to the same substation but not the same phase as the anchor point As a negative sample, the phase-level triplet is recorded as Define the phase-level triplet loss as:

[0047]

[0048] Where, δ p is a marginal hyperparameter used to constrain the distance difference between the positive and negative sample pairs of the phase-level samples;

[0049] (3-3) The two-level triplet loss is defined as:

[0050]

[0051] Where λ s and λ p is a weighting factor that can adjust the weights of the station area identification task and the phase identification task;

[0052] (3-4) Soft-DTW distance The calculation process is as follows:

[0053] For two feature sequences U=[u1, u2, ..., u n ],V=[v1,v2,...,v m ], under the premise of allowing the time axis to stretch or compress, find an optimal matching path W = [w1, w2, ..., w K ] makes:

[0054]

[0055] Where w k =(i, j) is the element u in the feature sequence U i and element v in V j The coordinate points of the elements, d(w k )=d(u i , v j )=(u i —v j ) 2 For element u i and v j The distance between them, DTW(U,V) is the DTW distance between feature sequences U and V;

[0056] To solve It is necessary to construct an n×m distance cost matrix D, which is calculated as follows:

[0057]

[0058] Where i = 1, 2, ..., n, j = 1, 2, ..., m, and the initial condition is D(0, 0) = 0. The last element D(n, m) obtained by accumulating the distance cost matrix is ​​the final DTW distance, that is:

[0059] DTW(U,V)=D(n,m)

[0060] The generalized min operator with smoothing parameter γ is:

[0061]

[0062] Using the generalized min operator, Soft-DTW distance SoftDTW γ (U, V) can be expressed as:

[0063]

[0064] As a further improvement of the present invention, clustering users based on the minimum-maximum Soft-DTW distance algorithm in step 4 includes the following process:

[0065] (4-1) According to the number of distribution transformers deployed in the low-voltage distribution network, the number of gathering points is set to N;

[0066] (4-2) Calculate the voltage characteristic sequence F of all users i =f(X i ) between the Soft-DTW distance, where X i is the voltage time series of the i-th user;

[0067] (4-3) Select the two users with the largest Soft-DTW distance among all users as the initial gathering points;

[0068] (4-4) Select the user with the largest Soft-DTW distance from the existing clustering points as the new clustering point;

[0069] (4-5) Repeat step (4-4) until the number of cluster points reaches N;

[0070] (4-6) Assign each user to the nearest cluster based on the Soft-DTW distance, and output the cluster label to which each user belongs;

[0071] (4-7) Calculate the characteristic sequence F of the cluster points of each user clusteri =f(X i ) and the characteristic sequence of the average value of the three-phase voltage on the low-voltage side of each distribution transformer Soft-DTW distance between where R j is the time series of the average three-phase voltage on the low-voltage side of the j-th distribution transformer. Each cluster is classified into the substation with the smallest Soft-DTW distance to it, and the substation to which the users in each cluster belong is determined to complete the substation identification;

[0072] (4-8) Calculate the feature sequence F of each user separately i =f(X i ) and the characteristic sequence of the voltage time series of each phase on the low-voltage side of the distribution transformer where it is located Soft-DTW distance, where P k is the voltage time series of the kth phase;

[0073] (4-9) The voltage time series of each phase on the low-voltage side of the distribution transformer is used as a cluster point, and each user is classified into the cluster with the closest Soft-DTW distance. The cluster label of each user is output to complete phase identification.

[0074] Beneficial effects:

[0075] The present invention discloses a method for intelligent identification of distribution network user topology based on dynamic time series feature clustering. This method uses the voltage data collected by the user meters in the low-voltage distribution station area as the input of the deep learning network, and uses the measured voltage time series as the feature extraction training benchmark of the deep learning network. The extracted feature sequence can be used as the benchmark for calculating the Soft-DTW distance in clustering. By constructing a two-level triplet loss, the feature sequence output by the neural network feature extraction module is made more structurally sensitive in the feature space, reducing the inter-class differences of users in the same station area and the same phase, and strengthening the inter-class differences of users in different stations and different phases. This can improve the robustness of topology recognition within a certain range and reduce the probability of misidentification in the case of random errors or low three-phase voltage imbalance in the station area. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a flowchart of intelligent identification of user topology of low-voltage distribution network provided by the present invention;

[0077] Figure 2 This is the structural diagram of the deep learning neural network proposed in this invention. DETAILED DESCRIPTION

[0078] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0079] The present invention discloses a method for intelligently identifying user topology of a distribution network based on dynamic time series feature clustering. The flow chart of the method of the present invention is as follows: Figure 1 As shown, the deep learning neural network structure diagram is as follows Figure 2 As shown, the specific steps include:

[0080] Step 1: Historical voltage data collection. Collect voltage data from user meters in the low-voltage distribution area and generate a voltage time series. Label and classify this data for subsequent analysis and identification.

[0081] Step 2: Extract voltage time series features. Establish a deep learning model for extracting user voltage time series features, use the data in the training set to train the deep learning model, and obtain the feature sequence F=f(X) of the voltage time series. The deep learning model structure established in the present invention is as follows: Figure 2 As shown, the specific steps include:

[0082] (2-1) The first layer of one-dimensional convolution: kernel = 3, stride = 1, channel = 32;

[0083] (2-2) Pooling layer 1: Max-pooling, pool size = 2;

[0084] (2-3) Second layer of one-dimensional convolution: kernel = 3, stride = 1, channel = 64;

[0085] (2-4) Pooling layer 2: Max-pooling, pool size = 2;

[0086] (2-5) Bi-GRU structure: hidden size = 128, bidirectional;

[0087] (2-6) Fully connected layer: maps the 256-dimensional output to 64 dimensions;

[0088] (2-7) Delayed pooling: Average pooling with a stride of 2 is used on the sequence during training, and the original resolution is retained during testing.

[0089] Step 3: Construct a two-level triplet loss. Construct a two-level triplet loss based on Soft-DTW distance to optimize the feature extraction network. The specific steps include:

[0090] (3-1) The district-level triplet is selected from the district T k , the voltage time series X of phase Z a As the anchor point, select the voltage time series that belongs to the same substation as the anchor point As a positive sample, select the voltage time series that does not belong to the same substation as the anchor point As a negative sample, the district-level triplet is recorded as The area-level triplet loss is defined as:

[0091]

[0092] Where, δ s is a marginal hyperparameter used to constrain the distance difference between the positive and negative sample pairs of the station-level samples.

[0093] (3-2) Phase-level triplet selection: Voltage time series that belongs to the same substation and the same phase as the anchor point As a positive sample, select the voltage time series that belongs to the same substation but not the same phase as the anchor point As a negative sample, the phase-level triplet is recorded as Define the phase-level triplet loss as:

[0094]

[0095] Where, δ p is a margin hyperparameter used to constrain the distance difference between the positive and negative sample pairs of the discrimination level samples.

[0096] (3-3) The two-level triplet loss is defined as:

[0097]

[0098] Where λ s and λ p It is a weighting factor that can adjust the weights of the area identification task and the phase identification task.

[0099] In the training of the established deep learning model, the marginal hyperparameter δ s The value is 2.0, and the marginal hyperparameter δ p The value is 1.0, and the weighting factor λ s and λ p The value is 0.5.

[0100] (3-4) Soft-DTW distance The calculation process is as follows:

[0101] For two feature sequences U=[u1, u2, ..., u n ],V=[v1,v2,...,v m ], under the premise of allowing the time axis to stretch or compress, find an optimal matching path W = [w1, w2, ..., w K ] makes:

[0102]

[0103] Where w k =(i, j) is the element u in the feature sequence U i and element v in V j The coordinate points of the elements, d(w k )=d(u i , v j )=(u i -v j ) 2 For element u i and v j The distance between them, DTW(U,V) is the DTW distance between feature sequences U and V;

[0104] To solve It is necessary to construct an n×m distance cost matrix D, which is calculated as follows:

[0105]

[0106] Where i = 1, 2, ..., n, j = 1, 2, ..., m, and the initial condition is D(0, 0) = 0. The last element D(n, m) obtained by accumulating the distance cost matrix is ​​the final DTW distance, that is:

[0107] DTW(U,V)=D(n,m)

[0108] The generalized min operator with smoothing parameter γ is:

[0109]

[0110] Using the generalized min operator, Soft-DTW distance SoftDTW γ (U, V) can be expressed as:

[0111]

[0112] In the present invention, the value of the smoothing parameter γ is 1.

[0113] Step 4: User cluster analysis. Calculate the Soft-DTW distance between user voltage feature sequences and use the minimum-maximum Soft-DTW distance algorithm to cluster users and output cluster labels to obtain user topology information. The specific steps include:

[0114] (4-1) According to the number of distribution transformers deployed in the low-voltage distribution network, the number of gathering points is set to N;

[0115] (4-2) Calculate the voltage characteristic sequence F of all users i =f(X i) between the Soft-DTW distance, where X i is the voltage time series of the i-th user;

[0116] (4-3) Select the two users with the largest Soft-DTW distance among all users as the initial gathering points;

[0117] (4-4) Select the user with the largest Soft-DTW distance from the existing clustering points as the new clustering point;

[0118] (4-5) Repeat step (4-4) until the number of cluster points reaches N;

[0119] (4-6) Assign each user to the nearest cluster based on the Soft-DTW distance, and output the cluster label to which each user belongs;

[0120] (4-7) Calculate the characteristic sequence F of the cluster points of each user cluster i =f(X i ) and the characteristic sequence of the average value of the three-phase voltage on the low-voltage side of each distribution transformer Soft-DTW distance between where R j is the time series of the average three-phase voltage on the low-voltage side of the j-th distribution transformer. Each cluster is classified into the substation with the smallest Soft-DTW distance to it, and the substation to which the users in each cluster belong is determined to complete the substation identification;

[0121] (4-8) Calculate the feature sequence F of each user separately i =f(X i ) and the characteristic sequence of the voltage time series of each phase on the low-voltage side of the distribution transformer where it is located Soft-DTW distance, where P k is the voltage time series of the kth phase;

[0122] (4-9) The voltage time series of each phase on the low-voltage side of the distribution transformer is used as a cluster point, and each user is classified into the cluster with the closest Soft-DTW distance. The cluster label of each user is output to complete phase identification.

[0123] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent variation based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for intelligent identification of distribution network user topology based on dynamic time series feature clustering, characterized by: The specific steps are as follows: Step 1: Read the measured values ​​of user voltage data in the low-voltage distribution network and establish a historical data set; Step 2: Establish a deep learning model for extracting user voltage time series features. Use the data in the training set to train the deep learning model to obtain the feature sequence F = f(X) of the voltage time series. Step 3: Optimize the feature extraction network using a two-level triplet loss based on Soft-DTW distance; Step 4: Calculate the Soft-DTW distance between user voltage feature sequences and use the minimum-maximum Soft-DTW distance algorithm to cluster users and output cluster labels to obtain user topology information.

2. The method for intelligently identifying distribution network user topology based on dynamic time series feature clustering according to claim 1 is characterized in that: The establishment of the neural network of the deep learning model in step 2 includes the following processes: (2-1) The first layer of one-dimensional convolution: kernel = 3, stride = 1, channel = 32; (2-2) Pooling layer 1: Max-pooling, pool size = 2; (2-3) Second layer of one-dimensional convolution: kernel = 3, stride = 1, channel = 64; (2-4) Pooling layer 2: Max-pooling, pool size = 2; (2-5) Bi-GRU structure: hidden size = 128, bidirectional; (2-6) Fully connected layer: maps the 256-dimensional output to 64 dimensions; (2-7) Delayed pooling: Average pooling with a stride of 2 is used on the sequence during training, and the original resolution is retained during testing.

3. The method for intelligently identifying distribution network user topology based on dynamic time series feature clustering according to claim 1 is characterized in that: The two-level triplet loss based on Soft-DTW distance in step 3 includes the following process: (3-1) The district-level triplet is selected from the district T k , the voltage time series X of phase Z a As the anchor point, select the voltage time series that belongs to the same substation as the anchor point As a positive sample, select the voltage time series that does not belong to the same substation as the anchor point As a negative sample, the district-level triplet is recorded as The area-level triplet loss is defined as: Where, δ s is a marginal hyperparameter used to constrain the distance difference between the positive and negative sample pairs of the substation level samples. (3-2) The phase level triplet selects the voltage time series that belongs to the same substation and the same phase as the anchor point. As a positive sample, select the voltage time series that belongs to the same substation but not the same phase as the anchor point As a negative sample, the phase-level triplet is recorded as Define the phase-level triplet loss as: Where, δ p is a marginal hyperparameter used to constrain the distance difference between the positive and negative sample pairs of the phase-level samples; (3-3) The two-level triplet loss is defined as: Where λ s and λ p is a weighting factor that can adjust the weights of the station area identification task and the phase identification task; (3-4) Soft-DTW distance The calculation process is as follows: For two feature sequences U=[u1, u2, ..., u n ],V=[v1,v2,...,v m ], under the premise of allowing the time axis to stretch or compress, find an optimal matching path W = [w1,w2,...,w K ] makes: Where w k =(i,j) is the element u in the feature sequence U i and element v in V j The coordinate points of the elements, d(w k )=d(u i , v j )=(u i -v j ) 2 For element u i and v j The distance between them, DTW(U, V) is the DTW distance between feature sequences U and V; To solve It is necessary to construct an n×m distance cost matrix D, which is calculated as follows: Where i = 1, 2, ..., n, j = 1, 2, ..., m, and the initial condition is D(0, 0) = 0. The last element D(n, m) obtained by accumulating the distance cost matrix is ​​the final DTW distance, that is: DTW(U,V)=D(n,m) The generalized min operator with smoothing parameter γ is: Using the generalized min operator, Soft-DTW distance SoftDTW γ (U,V) can be expressed as:

4. The method for intelligently identifying distribution network user topology based on dynamic time series feature clustering according to claim 1 is characterized in that: In step 4, clustering users based on the minimum-maximum Soft-DTW distance algorithm includes the following process: (4-1) According to the number of distribution transformers deployed in the low-voltage distribution network, the number of gathering points is set to N; (4-2) Calculate the voltage characteristic sequence F of all users i =f(X i ) between the Soft-DTW distance, where X i is the voltage time series of the i-th user; (4-3) Select the two users with the largest Soft-DTW distance among all users as the initial gathering points; (4-4) Select the user with the largest Soft-DTW distance from the existing clustering points as the new clustering point; (4-5) Repeat step (4-4) until the number of cluster points reaches N; (4-6) Assign each user to the nearest cluster based on the Soft-DTW distance, and output the cluster label to which each user belongs; (4-7) Calculate the characteristic sequence F of the cluster points of each user cluster i =f(X i ) and the characteristic sequence of the average value of the three-phase voltage on the low-voltage side of each distribution transformer Soft-DTW distance between where R j is the time series of the average three-phase voltage on the low-voltage side of the j-th distribution transformer. Each cluster is classified into the substation with the smallest Soft-DTW distance to it, and the substation to which the users in each cluster belong is determined to complete the substation identification; (4-8) Calculate the feature sequence F of each user separately i =f(X i ) and the characteristic sequence of the voltage time series of each phase on the low-voltage side of the distribution transformer where it is located Soft-DTW distance, where P k is the voltage time series of the kth phase; (4-9) The voltage time series of each phase on the low-voltage side of the distribution transformer is used as a cluster point, and each user is classified into the cluster with the closest Soft-DTW distance. The cluster label of each user is output to complete phase identification.

Images