A distribution network state estimation method in the absence of PMU measurement data

Abstract

The present invention discloses a method for estimating the state of a distribution network under the condition of missing PMU measurement data, comprising: constructing and training a measurement voltage compensation model to obtain an optimized measurement voltage compensation model; a control center receiving real-time measurement voltage data, determining whether there is missing voltage data in the real-time measurement voltage data; if so, using the optimized measurement voltage compensation model to compensate for the measurement voltage data; after compensation, performing distribution network topology identification based on the compensated measurement voltage data; if not, performing distribution network topology identification directly based on the real-time measurement voltage data; combining the complete real-time measurement voltage data and the identified distribution network topology information to perform distribution network state estimation, thereby obtaining a distribution network state estimation result. The present invention can accurately estimate the state of a distribution network even when PMU measurement data is missing.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network state estimation, and in particular to a distribution network state estimation method in the absence of PMU measurement data. Background Art

[0002] With the increasing penetration of distributed energy resources in distribution networks, accurate and real-time identification of distribution network topology is crucial. However, the loss of physical measurements due to meter failures or communication channel denial-of-service (DoS) attacks poses a significant challenge to topology identification. Current distribution networks employ radial connections, unidirectional power flows, and predictable load patterns, making control and monitoring less important in traditional distribution systems. However, the rapid growth of renewable energy sources (RES), particularly solar and wind power, is transforming distribution systems from passive load services to active market entities. This also complicates distribution management systems (DMSs), as it is difficult to track the dynamic changes in RES penetration levels over time. Therefore, distribution system state estimation (DSSE) is crucial for continuous monitoring of complex distribution systems in this context. By determining the correct topology, the network topology processor then provides structural data to the state estimator for better state estimation.

[0003] Traditional topology identification methods mostly rely on the status of switches and circuit breakers. With the rapid development of distribution system automation, the availability of measurement data such as phasor measurement units (PMUs), SCADA advanced metering infrastructure (AMI) data, and line current sensors has increased. Furthermore, denial of service (DoS) attacks are a major obstacle to accurate topology identification and state estimation solutions. DoS attacks disrupt communication channels by releasing interference noise, resulting in the loss of measurement data in the control center. Consequently, topology misjudgment and measurement data loss may occur. Power transmission systems are mostly mesh-connected networks, and topology changes rarely occur. However, in medium voltage (MV) and low voltage (LV) distribution systems, topology changes are more frequent, making it difficult to determine the correct topology using traditional methods. Therefore, the question of how to accurately identify the topology in the absence of measurement data and thus obtain accurate distribution network state estimation is urgently needed. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for estimating the state of a distribution network when PMU measurement data is missing.

[0005] To achieve the above objectives, the technical solutions provided by the present invention are:

[0006] A method for estimating distribution network state in the absence of PMU measurement data, comprising:

[0007] Constructing and training a measurement voltage compensation model to obtain an optimized measurement voltage compensation model;

[0008] The control center receives real-time measured voltage data and determines whether there is any missing voltage data in the real-time measured voltage data. If so, the optimized measured voltage compensation model is used to compensate for the measured voltage data. After the compensation is completed, the distribution network topology is identified based on the compensated measured voltage data. If not, the distribution network topology is directly identified based on the real-time measured voltage data.

[0009] The distribution network state estimation is performed by combining the complete real-time measured voltage data and the identified distribution network topology information to obtain the distribution network state estimation result.

[0010] Furthermore, in the process of training the measured voltage compensation model, historical measured voltage data is used to train the measured voltage compensation model, and the measured voltage compensation model captures the spatiotemporal scale characteristics by clarifying the nonlinear relationship between the measured voltages.

[0011] Furthermore, the measurement voltage compensation model includes a measurement compensation mechanism, which is an architecture based on an attention mechanism and includes an encoder and a decoder.

[0012] Furthermore, the encoder includes a Prob-Sparse attention mechanism to improve computational complexity, i.e.

[0013] In the Prob-Sparse attention mechanism, the probability of the mth query paying attention to the key is p(k j |q m ), its combination with the value is the output of the model; use the following formula to calculate p(k j |q m ) and uniform distribution p(k j |q m ) between:

[0014]

[0015] If the i-th query has a higher Its attention probability p is more diverse and contains the main pairs in the title field of the long-term self-attention distribution; according to this measurement method, the proposed compensation mechanism only focuses on the top u main queries for each k value:

[0016]

[0017] Where d is the input dimension, q i It is represented by the value of Q in the i-th row, k j is the value of K in row j, represents a sparse matrix containing only u queries; the main features of the feature map are constructed in subsequent layers; the entire expression of the self-attention distillation process is:

[0018]

[0019] [·] att is the attention block. For step 2 refinement, a maximum pooling layer is added after each convolution layer, and then the sampling added to half of it; as a result of this refinement, memory usage is reduced to Where λ is a set number.

[0020] Furthermore, the decoder adopts a generative inference method to reduce the speed of long sequence prediction; the output of the decoder is given to the fully connected layer, which generates the final output; in the generative inference feature, an L token Long sequence, that is, input to the decoder Where X0 is the target sequence output timestamp, the measured voltage compensation model decodes in a generative manner, and the decoder predicts the output using a single-forward approach; the measured voltage compensation model passes the following vector to the decoder:

[0021]

[0022] in, is the i-th input sequence, is the marker for the beginning of the i-th sequence, is the placeholder for the i-th target sequence.

[0023] Furthermore, the mean square error loss function is used to train the measurement voltage compensation model, and the loss is back-propagated from the output of the decoder to the entire measurement voltage compensation model architecture; the measurement voltage compensation model uses the Gaussian error linear unit activation function, the form of the GELU function:

[0024]

[0025] in,

[0026] Furthermore, when the optimized measured voltage compensation model is used to compensate the measured voltage data, it is assumed that the original voltage measurement value obtained by all nodes is U, where the pollution matrix is ​​Ω = random(U, γ), and γ is the pollution level. Then the measurement results of all nodes received by the control center are:

[0027] U c =Ω*U (8)

[0028] Among them, Ω is a matrix composed of 0 and 1, 0 means no measurement value, 1 means there is a measurement value; for the predicted value

[0029]

[0030] The predicted values ​​replace the measured values ​​of the previously missing positions and form the final compensated measurement matrix with the previous non-missing value measurements

[0031]

[0032] Furthermore, based on the complete real-time measured voltage data, the distribution network topology is identified by a multilayer perceptron, which uses a focal loss function for classification. The focal loss function is expressed as

[0033] FL=-[α(1-p) β log(p)+(1-α)(p) β log(1-p)] (11)

[0034] Among them, α is the weight parameter and β is the focusing parameter.

[0035] Compared with the existing technology, the principles and advantages of this technical solution are as follows:

[0036] 1. Verify the real-time measurement data received by the control center to determine whether there are any missing measurements. In the case of incomplete measurement data, this technical solution can quickly restore the system's missing measured power data based on past measurement samples. The missing values ​​in the original measurement data are replaced with predicted values ​​to obtain compensated measurement data. At the same time, the compensated measured voltage data is input into the classifier, and the classifier uses the complete measurement values ​​to identify the correct topology. Finally, the state of the distribution network is estimated based on the compensated measurement values ​​and the system's topology information. This technical solution can retrieve missing measured voltage data to maintain observability and correct topology identification, thereby achieving accurate estimation of the state of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the services required for use in the embodiments or the prior art descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0038] Figure 1 This is a flow chart of a method for estimating the state of a distribution network in the absence of PMU measurement data according to the present invention;

[0039] Figure 2 Improved IEEE 37-node distributed test system single-line diagram;

[0040] Figure 3 Schematic diagram showing the comparison between the predicted value of the three-phase missing position of node 2 in the improved IEEE 37-bus system and the actual data (taking phases A and B as an example). DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with specific embodiments:

[0042] like Figure 1 As shown, the method for estimating the state of a distribution network in the absence of PMU measurement data described in this embodiment includes the following steps:

[0043] S1. Build and train a measurement voltage compensation model to obtain an optimized measurement voltage compensation model;

[0044] In the process of training the measured voltage compensation model, historical measured voltage data is used to train the measured voltage compensation model, which captures the spatiotemporal scale characteristics by clarifying the nonlinear relationship between the measured voltages.

[0045] Specifically, the measurement voltage compensation model includes a measurement compensation mechanism, which is an architecture based on an attention mechanism and includes an encoder and a decoder.

[0046] The encoder includes the Prob-Sparse attention mechanism to improve the computational complexity, i.e.

[0047] In the Prob-Sparse attention mechanism, the probability of the mth query paying attention to the key is p(k j |q m ), its combination with the value is the output of the model; use the following formula to calculate p(k j |q m ) and uniform distribution p(k j |q m ) between:

[0048]

[0049] If the i-th query has a higher Its attention probability p is more diverse and contains the main pairs in the title field of the long-term self-attention distribution; according to this measurement method, the proposed compensation mechanism only focuses on the top u main queries for each k value:

[0050]

[0051] Where d is the input dimension, qi It is represented by the value of Q in the i-th row, k j is the value of K in row j, represents a sparse matrix containing only u queries; the main features of the feature map are constructed in subsequent layers; the entire expression of the self-attention distillation process is:

[0052]

[0053] [·] att is the attention block. For step 2 refinement, a maximum pooling layer is added after each convolution layer, and then the sampling added to half of it; as a result of this refinement, memory usage is reduced to Where λ is a set number.

[0054] The decoder uses a generative inference method to reduce the speed of long sequence prediction; the output of the decoder is given to the fully connected layer, which generates the final output; in the generative inference feature, an L is sampled in the input token Long sequence, that is, input to the decoder Where X0 is the target sequence output timestamp, the measured voltage compensation model decodes in a generative manner, and the decoder predicts the output using a single-forward approach; the measured voltage compensation model passes the following vector to the decoder:

[0055]

[0056] in, is the i-th input sequence, is the marker for the beginning of the i-th sequence, is the placeholder for the i-th target sequence.

[0057] The mean square error loss function is used to train the measurement voltage compensation model, and the loss is back-propagated from the output of the decoder to the entire measurement voltage compensation model architecture; the measurement voltage compensation model uses the Gaussian error linear unit activation function, the form of the GELU function:

[0058]

[0059] in,

[0060] S2. The control center receives real-time measured voltage data and determines whether there is any missing voltage data in the real-time measured voltage data. If so, the optimized measured voltage compensation model is used to compensate for the measured voltage data. After the compensation is completed, the distribution network topology is identified based on the compensated measured voltage data. If not, the distribution network topology is directly identified based on the real-time measured voltage data.

[0061] In this step, when the optimized measured voltage compensation model is used to compensate the measured voltage data, it is assumed that the original voltage measurement value obtained by all nodes is U, where the pollution matrix is ​​Ω = random (U, γ), and γ is the pollution level. Then the measurement results of all nodes received by the control center are:

[0062] U c =Ω*U (8)

[0063] Among them, Ω is a matrix composed of 0 and 1, 0 means no measurement value, 1 means there is a measurement value; for the predicted value

[0064]

[0065] The predicted values ​​replace the measured values ​​of the previously missing positions and form the final compensated measurement matrix with the previous non-missing value measurements

[0066]

[0067] In this step, based on the complete real-time measured voltage data, the distribution network topology is identified by a multilayer perceptron, which uses the focus loss function for classification. The focus loss function is expressed as

[0068] FL=-[α(1-p) β log(p)+(1-α)(p) β log(1-p)] (11)

[0069] Among them, α is the weight parameter and β is the focusing parameter.

[0070] S3. Combining the complete real-time measured voltage data and the identified distribution network topology information to perform distribution network state estimation, thereby obtaining a distribution network state estimation result.

[0071] To test the effectiveness of the method described in the present invention, the present invention uses the root mean square error (RMSE) metric, as shown in formula (12). This is because RMSE has a quadratic loss function, and results with a large deviation from the true value are more valuable than results with a small deviation.

[0072]

[0073] Where e(n) = t(n) - y(n), M is the number of samples in the test set, and t(n) and y(n) are the expected and actual outputs of the model for the nth sample, respectively.

[0074] In the case of topological classification, the accuracy metric is considered to show the effectiveness of the proposed missing value compensation based on the compensation method. The mathematical representation of the accuracy metric is shown in Equation (13).

[0075]

[0076] TP: true positive, which is judged as a positive sample and is actually a positive sample; TN: true negative, which is judged as a negative sample and is actually a negative sample; FP: false positive, which is judged as a positive sample, but is actually a negative sample; FN: false negative, which is judged as a negative sample, but is actually a positive sample.

[0077] For the traditional IEEE 37-node unbalanced distribution system, such as Figure 2 As shown in the figure, the following modifications are made: a) Distributed loads are treated as concentrated loads, ignoring transformers and voltage regulators; b) The switching lines in the IEEE 37-node system are 9-34, 14-5, 17-30, 26-32, and 32-20. These switching lines are selected to meet two conditions: they maintain radial distribution characteristics without creating loops, and they do not isolate any nodes. Based on the above assumptions, the number of topological configurations generated by the IEEE 37-node system is 25. These topologies ensure that no node loses its fixed line due to failure or maintenance, resulting in isolation. In the IEEE 37-node system, the distributed generation unit systems are randomly selected at nodes 4, 8, 18, 24, 25, 26, 29, 30, 32, and 35, respectively.

[0078] Hourly data from the National Renewable Energy Laboratory (NREL) is used for 50-day cycle load data for commercial, residential and industrial loads. These load data are normalized to the scale of the IEEE-37 node test system. Subsequently, voltage measurement data are generated by feeding the load data into the distribution power flow algorithm. The total number of voltage measurements obtained for the modified IEEE37 node system is 108. A total of 25 topologies are considered, each containing 1200 hours of voltage measurement data, forming a dimensional data set of (30000, 108). The proposed method predicts the measured data and the predicted value of the three-phase missing position of node 2 of the modified IEEE37 node system is compared with the real data. Figure 3 shown.

[0079] The embodiments described above are only preferred embodiments of the present invention and are not intended to limit the scope of implementation of the present invention. Therefore, any changes made based on the shape and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for estimating the state of a distribution network in the absence of PMU measurement data, characterized in that: include: Constructing and training a measurement voltage compensation model to obtain an optimized measurement voltage compensation model; The control center receives real-time measured voltage data and determines whether there is any missing voltage data in the real-time measured voltage data. If so, the optimized measured voltage compensation model is used to compensate for the measured voltage data. After the compensation is completed, the distribution network topology is identified based on the compensated measured voltage data. If not, the distribution network topology is directly identified based on the real-time measured voltage data. Combine the complete real-time measured voltage data and the identified distribution network topology information to perform distribution network state estimation, thereby obtaining the distribution network state estimation result; In the process of training the measured voltage compensation model, the historical measured voltage data is used to train the measured voltage compensation model. The measured voltage compensation model captures the spatiotemporal scale characteristics by clarifying the nonlinear relationship between the measured voltages. The measurement voltage compensation model includes a measurement compensation mechanism, which is an attention mechanism-based architecture including an encoder and a decoder; When using the optimized measured voltage compensation model to compensate for measured voltage data, assume that the original voltage measurement value obtained by all nodes is U, where the pollution matrix is ​​Ω = random(U, γ), and γ is the pollution level. Then the measurement results of all nodes received by the control center are: IN c =Ω*U (8) Among them, Ω is a matrix composed of 0 and 1, 0 means no measurement value, 1 means there is a measurement value; for the predicted value The predicted values ​​replace the measured values ​​of the previously missing positions and form the final compensated measurement matrix with the previous non-missing value measurements 2. The method for estimating the state of a distribution network in the absence of PMU measurement data according to claim 1, characterized in that: The encoder includes a Prob-Sparse attention mechanism to improve computational complexity, i.e. In the Prob-Sparse attention mechanism, the probability of the mth query paying attention to the key is p(k j |q m ), its combination with the value is the output of the model; use the following formula to calculate p(k j |q m ) and uniform distribution p(k j |q m ) between: If the i-th query has a higher Its attention probability p is more diverse and contains the main pairs in the title field of the long-term self-attention distribution; according to this measurement method, the proposed compensation mechanism only focuses on the top u main queries for each k value: Where d is the input dimension, q i It is represented by the value of Q in the i-th row, k j is the value of K in row j, represents a sparse matrix containing only u queries; the main features of the feature map are constructed in subsequent layers; the entire expression of the self-attention distillation process is: [·] att is the attention block. For step 2 refinement, a maximum pooling layer is added after each convolution layer, and then the sampling added to half of it; as a result of this refinement, memory usage is reduced to Where λ is a set number.

3. The method for estimating the state of a distribution network in the absence of PMU measurement data according to claim 1, characterized in that: The decoder uses a generative inference method to reduce the speed of long sequence prediction; The output of the decoder is passed to the fully connected layer, which generates the final output; In the generative reasoning feature, an L token Long sequence, that is, input to the decoder Where X0 is the target sequence output timestamp, the measured voltage compensation model decodes in a generative manner, and the decoder predicts the output using a single-forward approach; the measured voltage compensation model passes the following vector to the decoder: in, is the i-th input sequence, is the marker for the beginning of the i-th sequence, is the placeholder for the i-th target sequence.

4. The method for estimating the state of a distribution network in the absence of PMU measurement data according to claim 1, characterized in that: Based on the complete real-time measured voltage data, the distribution network topology is identified by a multilayer perceptron, which uses the focus loss function for classification. The focus loss function is expressed as FL=-[α(1-p) β log(p)+(1-α)(p) β log(1-p)] (11) Among them, α is the weight parameter and β is the focusing parameter.

Images