An intelligent prediction and analysis method for low-voltage shunt tripping in distribution networks based on double similarity

Abstract

The present invention discloses a method for intelligent prediction and analysis of tripping of low-voltage shunts in distribution networks based on dual similarity. The method judges whether the operating status of low-voltage shunts in distribution networks is normal based on the similarity of operating modes and the similarity of time. First, a typical operating mode set is obtained based on k-means, and the similarity of operating modes is measured using Euclidean distance; then, the temporal similarity is modeled based on LSTM, and intelligent prediction and analysis of tripping is performed based on the time continuity of high-resolution continuous measurement. Through the dual similarity method, the proposed solution can effectively identify typical abnormal phenomena in the distribution network with high accuracy and recall rate. The present invention has been extensively evaluated by verifying a series of actual tripping events of 75 distribution transformers, confirming that the proposed solution is superior to traditional fault detection solutions.

Description

Technical Field

[0001] The present invention relates to state estimation of a distribution network in an electric power system, and in particular to an intelligent prediction and analysis method for low-voltage branch tripping of a distribution network based on double similarity. Background Art

[0002] To ensure reliable, safe, and economical operation of distribution networks and improve power quality for consumers, accurately sensing their operating status and tripping behavior is crucial and urgent. Low-voltage branch lines are the last stop in the power grid, directly connecting consumers and playing a crucial role. Low voltage and severe overloads in distribution substations have always been significant factors impacting power quality in distribution network substations, and the resulting impact on power stability and safety is a major concern for consumers. Furthermore, compared to other components of the power grid, substation power equipment faces challenges such as more complex environments, more diverse problems, and difficulty identifying and resolving faults. This research focuses on profiling distribution network substations, aiming to strengthen substation monitoring and enhance early warning and detection capabilities for substation faults, thereby shortening maintenance time and enabling proactive troubleshooting and repair. Therefore, improving the accuracy and speed of low-voltage branch tripping prediction is crucial and urgent for ensuring the real-time, safe, and stable operation of distribution networks. To fully tap into the information and value of massive distribution network data, employing data-driven methods to analyze distribution network metering data and improve the ability to perceive distribution network status is crucial and urgent.

[0003] Trip detection on low-voltage shunts in distribution networks is one of the most important tasks for all departments within power grid companies. Missing or incorrectly identifying abnormal events can put the safe and reliable operation of the power distribution network at risk. Accurate fault detection is crucial for emergency power supply, power system operation and maintenance, and distribution network expansion.

[0004] However, the judgment of low-voltage branch tripping in distribution networks is considered a challenging task. First, due to cost considerations, the measurement of branch circuits is usually insufficient in practice. Measuring equipment and devices are not installed on all branch circuits. Only the busbar (i.e., distribution transformer) can be measured, while the current and voltage conditions of the branches under the bus cannot be observed, resulting in inaccurate and untimely identification of abnormal conditions. Second, the complex working conditions of the distribution network pose challenges to tripping detection. The tripping of lightly loaded circuits with minimal load and current losses makes fault identification difficult. At the same time, disordered load fluctuations and mild load losses also pose obstacles to traditional logic-based fault detection. Third, the extreme imbalance between normal and abnormal samples without labels in the dataset makes it difficult to model abnormalities as extremely low-probability events.

[0005] Existing tripping judgment solutions can be divided into model-based methods and model-free methods. Model-based methods refer to methods based on the analysis of tripping mechanism methods, and model-free methods refer to data-driven models. Many model-based methods estimate the state of the power system based on given physical models and system mechanism information, such as real-time dynamic power outage detection schemes based on AC current models and statistical change detection theory. Although the performance of model-based methods depends largely on the accuracy of model parameters. However, the abnormal conditions of distribution networks are complex and diverse, so it is difficult to select appropriate settings for different situations. This dependence makes them less effective in real-world applications, and more research work focuses on model-free methods.

[0006] Furthermore, model-free methods can be primarily categorized into two groups: supervised learning and unsupervised learning. Typical supervised learning methods for fault detection include long short-term memory neural networks (LSTMs), support vector machines, decision trees, and generative adversarial networks (GANs). These studies generally identify abnormal events and classify them into different groups, which requires the type of abnormal event and a sufficient amount of labeled training data. However, defining different forms of abnormal phenomena is considered a non-trivial task and is generally addressed by converting the classification problem into a regression problem using continuous measurements, such as by applying deep learning to detect data and system anomalies using high-resolution continuous wave point or phase measurements. For unsupervised learning-based fault detection schemes, clustering is often used due to its simplicity and scalability. Similarity-based methods (SBMs) are also a commonly used fault detection algorithm. By measuring the similarity between subsequences of a time series, a state matrix can be established to characterize the operating conditions. Summary of the Invention

[0007] The purpose of this invention is to solve the problem of determining tripping of low-voltage shunts in distribution networks. A dual similarity method based on operation mode reconstruction and LSTM prediction is proposed to achieve higher fault identification accuracy and recall rate. The effectiveness of this method is verified through a large number of experiments based on historical data. It has great significance for tripping determination and situational awareness of smart distribution networks. The purpose of this invention can be achieved through the following technical solutions:

[0008] The present invention first provides a distribution network low-voltage branch tripping intelligent prediction and analysis method based on dual similarity, which includes the following steps:

[0009] 1) Obtain historical load data of the substation area, pre-process the data, and remove measurement errors;

[0010] 2) Select the corresponding features and calculate their corresponding time series changes as extended features to obtain a feature set;

[0011] 3) Divide the historical data of the feature set in step 2) into a training set and a test set, obtain an initial state matrix by clustering the training set, and obtain a final state matrix by iteratively optimizing the initial state matrix; at the same time, train an LSTM model based on the training set to obtain a substation feature prediction model that can predict the next operating condition based on the operating condition data of the previous three hours;

[0012] 4) Read the real-time data of the substation and the pre-trained substation feature prediction model, and determine whether to trip based on the error threshold.

[0013] As a preferred solution of the present invention, in step 3), the final state matrix is ​​obtained by iteratively optimizing the initial state matrix; at the same time, the LSTM model is trained based on the training set to obtain a substation feature prediction model that can predict the next moment's operating condition based on the previous three hours' operating condition data; specifically:

[0014] (a) Select historical data from 96 time points every day for the past month, with a 15-minute interval between each time point, and divide it into a training set and a test set. The test set is the feature set of tripping days when tripping is known to occur, and the training set is the feature set of normal operation.

[0015] (b) The initial state matrix is ​​input into the pattern similarity operator to calculate the reconstructed estimated value. Based on the grid operation mechanism, eight feature dimensions are analyzed: phase A current, phase A voltage, phase B current, phase B voltage, phase C current, phase C voltage, load factor, and active power. The relative error between the estimated value and the true state value of each feature dimension is calculated, and the error of each feature dimension is obtained by weighted average.

[0016] (c) Errors with an absolute value greater than 10 in the average positioning error are correspondingly eliminated from the original data and the initial state matrix and regarded as abnormal values ​​mixed into the normal working conditions;

[0017] (d) When the mean of the error is greater than 0.1 or the variance is greater than 1, and the length of the state matrix is ​​less than 80, add features to the initial state matrix. The calculation method is as follows: calculate the predicted value obtained by the original data and the state matrix through the pattern similarity operator, calculate the error between the predicted value and the true value, locate the feature with the largest error, add this feature to the state matrix, and delete it from the original data;

[0018] (e) Based on the final state matrix, calculate the mean and variance of the relative error between the estimated value and the true state value, save the mean, variance and the final state matrix, and calculate the error threshold. The calculation formula is as follows:

[0019]

[0020] Where G is the error threshold, error is the average error of the nth normal working condition, and std(·) is the variance calculation formula;

[0021] (f) Use the historical data of the most recent month as training data without excluding tripping data, perform max-min linear normalization, and use the LSTM model for training. Use the data of the first three hours to predict the data of the next 15 minutes, and use 12 data points to predict one data point in the future. Then, denormalize the predicted value, calculate the variance and mean of the error of each feature, and calculate the error threshold according to the error threshold calculation formula and save it.

[0022] As a preferred embodiment of the present invention, the step 4) is specifically as follows:

[0023] (a) Read the real-time data of the substation area; read the state matrix, variance, mean and error threshold under normal working conditions obtained and saved by pattern similarity calculation;

[0024] (b) Calculate the estimated value of the current state based on the pattern similarity function and find the relative error between the estimated value and the current value; predict the predicted value at the next moment based on the trained LSTM and find the relative error between the estimated value and the current value;

[0025] (c) Based on the two relative errors calculated in step (b), make the following two judgments:

[0026] Judgment 1: Calculate the weighted average error of the current state and determine whether it exceeds the error threshold. If not, it is determined to be normal. If it is, further determine whether there is an abnormal increase in current, that is, determine whether the current change is greater than 0.1. If there is an abnormal increase in current, filter the state and do not issue a trip alarm. If not, it is determined to be a trip.

[0027] Judgment 2: Read the substation feature prediction model obtained based on LSTM training to obtain the predicted value, calculate the relative error between the predicted value of phase A current, phase A voltage, phase B current, phase B voltage, phase C current, phase C voltage, load rate, and active power and the corresponding real-time data, and judge whether the error exceeds the error threshold. If not, it is judged to be normal; if exceeded, further judge whether there is an abnormal increase in current, that is, judge whether the current change is greater than 0.1; if there is an abnormal increase in current, filter the state and do not issue a trip alarm. If not, it is judged to be a trip;

[0028] (d) Based on the results of judgment 1 and judgment 2, if both judgments are "no tripping", no tripping has occurred; otherwise, it is determined that a tripping has occurred.

[0029] The present invention proposes a data-driven tripping judgment method based on the characteristics of the power distribution network and the technical difficulties in judging the tripping of the low-voltage shunt in the distribution network. Its innovation and technical contributions are mainly reflected in the following aspects:

[0030] (1) The present invention introduces a dual-similarity method, which includes operation mode similarity and working time similarity based on data reconstruction and prediction. The dual-similarity method can effectively improve the performance of the distribution network low-voltage shunt tripping monitoring algorithm.

[0031] (2) The present invention is a method for pattern reconstruction and prediction based on training the network and state matrix based on the normal working condition data set. It is unsupervised and does not require labeled data. Therefore, the method is not restricted by various unlabeled abnormal situations and has a low probability.

[0032] (3) This paper verifies the rationality of the proposed solution based on the real historical data of 75 distribution transformers and the real tripping event data. The experimental results confirm that it has more advantages than the traditional tripping judgment scheme in terms of accuracy and recall performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the method of the present invention;

[0034] Figure 2 It is a judgment flow chart of the algorithm of the present invention;

[0035] FIG3( a ) is a comparison diagram of the current of phase A in case 1 of the specific verification of the algorithm of the present invention;

[0036] FIG3( b ) is a comparison diagram of the current of phase B in case 1 of the specific verification of the algorithm of the present invention;

[0037] FIG3( c ) is a comparison diagram of the C-phase current in Case 1 of the specific verification of the algorithm of the present invention;

[0038] FIG3( d ) is a schematic diagram of residuals in a specific verification case 1 of the algorithm of the present invention;

[0039] FIG4( a ) is a comparison diagram of the current of phase A in case 2 of specific verification of the algorithm of the present invention;

[0040] FIG4( b ) is a comparison diagram of the current of phase B in case 2 of specific verification of the algorithm of the present invention;

[0041] FIG4( c ) is a comparison diagram of the C-phase current in Case 2 of the specific verification of the algorithm of the present invention;

[0042] FIG4( d ) is a schematic diagram of residuals in case 2 of specific verification of the algorithm of the present invention;

[0043] Figure 51 is a schematic diagram of three-phase current in a specific embodiment 1 of the present invention;

[0044] Figure 6 This is a schematic diagram of three-phase current in specific implementation example 2 of the present invention. DETAILED DESCRIPTION

[0045] The present invention will be described in detail below with reference to the accompanying drawings, and the objects and effects of the present invention will become more apparent.

[0046] Figure 1 is a flow chart of the method of the present invention, Figure 2 This is the judgment flow chart of the algorithm described in the present invention. Figure 1 As described, in the offline stage, the solution of the present invention includes the training of an LSTM model for time prediction and the construction of an optimal state matrix based on k-means. First, in order to accurately predict the next moment of comprehensive measurement, this work developed a prediction model based on LSTM, and the training of the prediction model is based on high-resolution historical operation data. LSTM is good at modeling time series and can achieve very high accuracy, so fault detection based on high-resolution continuous measurement can be obtained. Secondly, in order to reconstruct the comprehensive measurement at the current moment based on the outline of the optimal state matrix, K-means is used to initialize the optimal state matrix. The typical optimal state initialization method uses an iterative search method, which has a large amount of calculation, and the obtained state is not very discrete, so it is not enough to characterize the overall distribution.

[0047] For the online stage, the present invention mainly includes two parts: operating mode similarity and operating time similarity based on continuity. Figure 2 As described above, in the operating mode similarity module, real-time data is reconstructed into an estimated state. This data is then compared with the estimated state and fault type analysis is performed through logical judgment. In the time similarity module, the measured data in the sliding window is used as input for measurement prediction based on the LSTM model. AND logic is used to connect the operating mode similarity module and the operating time similarity module. By comparing the reconstructed and predicted data, the omission of abnormal states and the misidentification of abnormal states are reduced.

[0048] The specific steps of the intelligent prediction and analysis method for low-voltage branch tripping of distribution network based on dual similarity of the present invention are as follows:

[0049] Step 1: Obtain historical load data for the substation area, preprocess the data, and remove metering errors.

[0050] In a specific embodiment of the present invention, step 1 is specifically:

[0051] (a) Each substation selects a state matrix based on its own historical data. The number of substations is determined by the substation itself. The data source is current, voltage, load factor, and power data with a 15-minute resolution. The state matrix is ​​a sample selected from the historical data. The steps for selecting the state matrix are:

[0052] Select samples from the given historical data X and form a state matrix D. That is, a process or equipment has n related features. Suppose it is sampled at a certain time i, and the n collected features are selected as a working condition;

[0053] X(i)=[x1,x2,…,x n ] T

[0054] Combined with the changes in working conditions during the set time, m working conditions are selected and the state matrix D is formed.

[0055]

[0056] Each column vector in the state matrix represents an operating condition of the equipment.

[0057] (b) In the case of missing data, if the missing period is less than 1 hour, linear interpolation is used to fill it in; if the missing period is greater than 1 hour, the data for that day is excluded.

[0058] (c) In the case of data anomalies, the obviously abnormal moment data are directly eliminated.

[0059] Step 2: Select the corresponding features and calculate their corresponding time series changes as extended features to obtain the feature set. Step 2 is as follows:

[0060] (a) Select phase A current, phase A voltage, phase B current, phase B voltage, phase C current, phase C voltage, load factor, active power, and reactive power as features and sort them in time series;

[0061] (b) Subtract the features of the previous moment from the features of the next moment to find the corresponding changes of each feature. The features and their changes are combined into a total of 18 feature dimensions to form an n*18 matrix, where n is the number of moments.

[0062] Step 3: Divide the historical data of the feature set in step 2 into a training set and a test set. Obtain the initial state matrix by clustering the training set, and obtain the final state matrix by iteratively optimizing the initial state matrix. At the same time, train the LSTM model based on the training set to obtain a substation feature prediction model that can predict the operating conditions at the next moment based on the operating condition data of the previous three hours.

[0063] Among them, the initial state matrix is ​​obtained by clustering the training set, specifically:

[0064] Based on the k-means clustering method, m cluster centers are selected from the normal operation data of the past month to form an m*18 initial state matrix. The value of m ranges from (30, 50) and is selected according to the actual calculation efficiency and accuracy requirements. The larger the m, the longer the calculation time and the smaller the calculation error.

[0065] K-means algorithm is a commonly used clustering algorithm. Its algorithm steps are as follows:

[0066] 1. Randomly select k samples μ1, μ2, ..., μk as the initial cluster centers;

[0067] 2. For each sample x(i), calculate its distance from all cluster centers and assign it to the class c(m) corresponding to the cluster center closest to it;

[0068] c(m)=min k ||x(i)-μ m || 2

[0069] 3. For each category, take its centroid as the new cluster center.

[0070]

[0071] 4. Repeat steps 2 and 3 until the preset termination condition is reached. The termination condition can be any of the following:

[0072] 1) No (or a minimum number) of objects are reassigned to different clusters.

[0073] 2) No (or a minimum number of) cluster centers change again.

[0074] 3) The sum of squared errors is locally minimized.

[0075] The following formula can be used to determine whether the algorithm has converged.

[0076]

[0077] When E reaches the preset termination condition or the swing is very small, it means that the algorithm has stabilized.

[0078] In a specific embodiment of the present invention, the final state matrix is ​​obtained by iteratively optimizing the initial state matrix in step 3; at the same time, the LSTM model is trained based on the training set to obtain a substation feature prediction model that can predict the next moment's operating condition based on the previous three hours' operating condition data, specifically:

[0079] (a) Select historical data from 96 time points every day for the past month, with a 15-minute interval between each time point, and divide it into a training set and a test set. The test set is the feature set of tripping days when tripping is known to occur, and the training set is the feature set of normal operation.

[0080] (b) The initial state matrix is ​​input into the pattern similarity operator to calculate the reconstructed estimated value. Based on the grid operation mechanism, eight feature dimensions are analyzed: phase A current, phase A voltage, phase B current, phase B voltage, phase C current, phase C voltage, load factor, and active power. The relative error between the estimated value and the true state value of each feature dimension is calculated, and the error of each feature dimension is obtained by weighted average.

[0081] (c) Errors with an absolute value greater than 10 in the average positioning error are correspondingly eliminated from the original data and the initial state matrix and regarded as abnormal values ​​mixed into the normal working conditions;

[0082] (d) When the mean of the error is greater than 0.1 or the variance is greater than 1, and the length of the state matrix is ​​less than 80, add features to the initial state matrix. The calculation method is as follows: calculate the predicted value obtained by the original data and the state matrix through the pattern similarity operator, calculate the error between the predicted value and the true value, locate the feature with the largest error, add this feature to the state matrix, and delete it from the original data;

[0083] (e) Based on the final state matrix, calculate the mean and variance of the relative error between the estimated value and the true state value, save the mean, variance and the final state matrix, and calculate the error threshold. The calculation formula is as follows:

[0084]

[0085] Where G is the error threshold, error is the average error of the nth normal working condition, and std(·) is the variance calculation formula;

[0086] (f) Use the historical data of the most recent month as training data without excluding tripping data, perform max-min linear normalization, and use the LSTM model for training. Use the data of the first three hours to predict the data of the next 15 minutes, and use 12 data points to predict one data point in the future. Then, denormalize the predicted value, calculate the variance and mean of the error of each feature, and calculate the error threshold according to the error threshold calculation formula and save it.

[0087] It is important to note that LSTM-based measurement prediction is crucial for determining tripping in distribution network branches, as highly accurate predictions can help detect anomalies in system operation in a timely manner. Time series measurements are selected as the input vector for the prediction model, as shown in the following equation:

[0088] Xin (i)=[I A ,I B ,I C ,U A ,U B ,U C ,L rate ,P,Q] T

[0089] Among them I A ,I B and I C Represents the three-phase current of the main circuit, U A ,U B and U C Represents the three-phase voltage of the bus, L rate represents the load factor of the distribution transformer, P and Q represent active power and reactive power respectively.

[0090] Due to the different sizes of the measured values, in order to reduce the interference to the prediction model, the measurement vector should be normalized as shown in the following formula:

[0091]

[0092] Among them, represents the normalized measurement vector, and represents the maximum and minimum values ​​of the original measurement values.

[0093] In order to take advantage of high-resolution continuous measurement for fault detection, the measurement vectors of the last three hours are used to predict the measurement vector of the next sampling time slot. The prediction model is expressed as follows:

[0094] X pred =f(X′ 3h )

[0095] where X pred represents the predicted measurement value, and f(·) represents the prediction model.

[0096] All recursive neural networks in the LSTM network are a chain of repeated neural network modules, consisting of a single tanh layer that obtains the current output through the current input and the output at the previous moment; specifically, LSTM mainly contains three gates, namely the forget gate, input gate, output gate and a memory cell.

[0097] Step 4: Read the real-time data of the substation and the pre-trained substation feature prediction model, and determine whether to trip based on the error threshold.

[0098] (a) Read the real-time data of the substation area; read the state matrix, variance, mean and error threshold under normal working conditions obtained and saved by pattern similarity calculation;

[0099] (b) Calculate the estimated value of the current state based on the pattern similarity function and find the relative error between the estimated value and the current value; predict the predicted value at the next moment based on the trained LSTM and find the relative error between the estimated value and the current value;

[0100] (c) Based on the two relative errors calculated in step (b), make the following two judgments:

[0101] Judgment 1: Calculate the weighted average error of the current state and determine whether it exceeds the error threshold. If not, it is determined to be normal. If it is, further determine whether there is an abnormal increase in current, that is, determine whether the current change is greater than 0.1. If there is an abnormal increase in current, filter the state and do not issue a trip alarm. If not, it is determined to be a trip.

[0102] Judgment 2: Read the substation feature prediction model obtained based on LSTM training to obtain the predicted value, calculate the relative error between the predicted value of phase A current, phase A voltage, phase B current, phase B voltage, phase C current, phase C voltage, load rate, and active power and the corresponding real-time data, and judge whether the error exceeds the error threshold. If not, it is judged to be normal; if exceeded, further judge whether there is an abnormal increase in current, that is, judge whether the current change is greater than 0.1; if there is an abnormal increase in current, filter the state and do not issue a trip alarm. If not, it is judged to be a trip;

[0103] (d) Based on the results of judgment 1 and judgment 2, if both judgments are "no tripping", no tripping has occurred; otherwise, it is determined that a tripping has occurred.

[0104] In a specific embodiment of the present invention, the initial state matrix is ​​input into the pattern similarity operator to obtain the reconstructed estimated value, specifically:

[0105] Input condition x at a certain moment in It is composed of 1*18 feature data, where the first 9 features are phase A current, phase A voltage, phase B current, phase B voltage, phase C current, phase C voltage, load factor, active power, and reactive power. The 10th to 18th features are the changes of the first 9 features compared to the previous moment. The specific form is as follows:

[0106] x in =[x1x2,x3,x4,x5,x6,x7,x8,x9,x 1in x 2in ,x 3in ,x 4inx 5in ,x 6in ,x 7in ,x 8in ,x 9in ] T

[0107] Input working condition x in Comparing the similarity of each operating condition with the state matrix (D) will produce a similarity vector a, which contains the same number of elements as the training matrix elements stored in the state matrix;

[0108]

[0109] (b) Convert the similarity vector representing the similarity between the current operating condition and the historical normal operating condition into a weight vector (w):

[0110] w0=G -1 ·a

[0111]

[0112]

[0113] Where: is a nonlinear operator, and is chosen as the Euclidean distance between two vectors, that is:

[0114]

[0115] (c) Producing an estimate through a linear combination of samples and weights

[0116] x out =D·w

[0117] Right now:

[0118]

[0119] The present invention uses precision and recall for performance evaluation, specifically, they are defined as follows:

[0120]

[0121]

[0122] Among them, P accuracy is the accuracy, R recallis the recall rate, TP is the number of predicted tripping events that actually tripped, TN is the predicted number of predicted non-tripping events that actually did not trip, FN is the predicted number of predicted non-tripping events that actually did not trip, and FP is the number of predicted non-tripping events that actually tripped. The solution proposed in the present invention was further evaluated in terms of accuracy and recall with a typical data reconstruction method and a deep learning model for time series modeling, namely the LSTM model. The test set used 80 tripping samples, including main switch tripping, branch switch tripping, phase loss operation, etc. The actual tripping event was defined as a negative sample, and the normal operation without tripping was defined as a positive sample. There were 44 positive samples and 36 negative samples in the test set. Table 1 lists the numerical results of the performance comparison of 80 abnormal events for 75 distribution transformers. The abnormal events tested included several types: main switch tripping, branch switch tripping, phase loss, and fuse burnout. A main switch tripping is considered a low-voltage main circuit trip, while branch switch tripping, phase loss, and fuse burnout are considered low-voltage branch circuit tripping events. Table 2 shows the accuracy and recall rate of trip judgment using the proposed solution.

[0123] Table 1 Comparison of the accuracy and recall rate of the present invention and other solutions

[0124]

[0125] The judgment methods based on pattern similarity in Table 1 can be found in the following literature:

[0126] [1] S.Hu, Z.Xiao, Q.Rao and R.Liao, "An fault detection model of userbehavior based on similarity clustering," in 2018IEEE 4th Information Technology and Mechatronics Engineering Conference (ITOEC), pp.835-838, 2018.

[0127] The LSTM-based time series judgment method in Table 1 can be found in the following literature:

[0128] [2] Y.Cheng, C.Xu, D.Mashima, VLLThing, &Y.Wu, "PowerLSTM: Power DemandForecasting Using Long Short-Term Memory Neural Network," in International Conference on Advanced Data Mining&Applications, Spring, 2017.

[0129] Table 2 Precision and recall of the present invention on the test set

[0130]

[0131] As shown in Table 1, compared with the SBM and LSTM, the proposed dual-similarity-based intelligent prediction and analysis method for low-voltage shunt tripping in distribution networks improves precision by 4.28% and 7.14%, and recall by 5.23% and 11.3%, respectively. The comparison of precision and recall demonstrates the robust performance of the proposed method, supporting the high accuracy and recall achieved across diverse fault samples using a dual-similarity-based intelligent prediction and analysis method for low-voltage shunt tripping in distribution networks. As shown in Table 2, the proposed intelligent prediction and analysis method for tripping is capable of identifying major anomalies in distribution networks with high accuracy.

[0132] In a specific verification case 1 of the present invention, a branch switch tripped in the low-voltage distribution network. If there is no obvious current drop or current disappearance, such a situation is often missed when judging an abnormality because the predicted deviation is not enough to reach the discrimination threshold. In addition, in order to intuitively demonstrate the effect of the intelligent prediction and analysis method for low-voltage branch tripping in the distribution network based on dual similarity of the present invention, the intelligent prediction and analysis process of the current measurement on the day of the tripping event is shown as follows: Figure 3(a)-Figure 3(d) Specifically, the comparison between the data reconstructed by the intelligent prediction analysis method for low-voltage shunt tripping of distribution network based on double similarity of the present invention and the real data set of three-phase current is shown in FIG. Figure 3(a)-Figure 3(c) , the residuals of the three-phase currents are shown in Figure 3(d). Since the branch tripping event occurred between 12:00 and 14:00. Figure 3(a)-Figure 3(c) The three-phase currents and their respective reconstructed currents show that the branch circuit tripping event did not cause a significant drop in the total current of the distribution transformer. However, the dual-similarity-based intelligent prediction and analysis method for low-voltage shunt tripping in distribution networks presented here can identify current fluctuations, with the residuals from 12:00 to 14:00 significantly higher than those at other times. This demonstrates that the feature extraction and residuals of the present invention for tripping events can effectively determine the normal status of the distribution transformer's switch branches, reducing missed and false alarms.

[0133] In a specific verification case 2 of the present invention, a phase loss fault occurred in the low-voltage distribution network. During phase loss, no significant fluctuations were observed in the current, load factor, and power measurement profiles. However, the current in one or two phases could drop to zero. Figure 4(a)-Figure 4(d) The process of intelligent prediction and analysis of tripping of phase-loss operation fault is shown. Specifically, the comparison between the data reconstructed by the intelligent prediction and analysis method of low-voltage branch tripping of distribution network based on double similarity and the real data of three-phase current is shown in Figure 1. Figure 4(a)-Figure 4(c) , the residuals of the three-phase currents are shown in Figure 4(d). The phase loss operation event occurred between 16:00 and 18:00. Figure 4(a)-Figure 4(c) The three-phase currents and their respective reconstructed currents show that the phase loss event in phase B did not cause significant fluctuations in the currents of phases A and B, while the current of phase C experienced a sharp increase. The intelligent predictive analysis method for low-voltage shunt tripping in distribution networks based on dual similarity, described in this invention, can clearly detect abnormal events and issue alarms. This demonstrates that temporal similarity can leverage the advantages of high-resolution continuous point-to-wave analysis to accurately identify the occurrence of phase loss.

[0134] In a specific implementation example 1 of the present invention, the substation GISID is 650010998211801, the fault type is a disconnector trip, and the algorithm test results show that the maximum deviation of normal samples is 0.25, the average deviation of normal samples is 0.07, and the maximum deviation of the sample to be tested is 0.36. The abnormal time is judged to be 18:00:00 on April 19, 2022. The algorithm accurately judges it based on the large three-phase current errors, which are 0.553, 0.803, and 1.098 respectively. The three-phase current curve when the specific implementation example 1 trips is shown in the figure.

[0135] In a specific implementation case 2 of the present invention, the GISID of the substation is 650017312245901, the fault type is phase loss, the maximum deviation of the normal sample in the algorithm test result is 0.23, the average deviation of the normal sample is 0.06, the maximum deviation of the sample to be tested is 7.62, and the average deviation multiple of the abnormal sample to be tested is 2.00. It is judged that the abnormal time is 09:45:00 to 10:15:00 on April 20, 2022. The algorithm makes an accurate judgment without time error. The three-phase current curve when the specific implementation case 2 trips is as follows Figure 6 shown.

[0136] The above-described embodiments merely represent several implementation methods of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. An intelligent prediction and analysis method for low-voltage shunt tripping in distribution network based on dual similarity, characterized in that: The following steps are involved: 1) Obtain historical load data of the substation area, pre-process the data, and remove measurement errors; 2) Select the corresponding features and calculate their corresponding time series changes as extended features to obtain a feature set; 3) Divide the historical data of the feature set in step 2) into a training set and a test set, and obtain an initial state matrix by clustering the training set; obtain a final state matrix by iteratively optimizing the initial state matrix; and simultaneously, train an LSTM model based on the training set to obtain a substation feature prediction model that can predict the next operating condition based on the operating condition data of the previous three hours; 4) Reading the real-time data of the substation and the pre-trained substation feature prediction model, and judging whether to trip based on the error threshold; the step 4) is specifically as follows: (a) Read the real-time data of the substation area; read the state matrix, variance, mean and error threshold under normal working conditions obtained and saved by pattern similarity calculation; (b) Calculate the estimated value of the current state based on the pattern similarity function and find the relative error between the estimated value and the current value; predict the predicted value at the next moment based on the trained LSTM and find the relative error between the estimated value and the current value; (c) Based on the two relative errors calculated in step (b), make the following two judgments: Judgment 1: Calculate the weighted average error of the current state and determine whether the average error of the current state exceeds the error threshold. If not, it is judged to be normal; If it exceeds, it will further determine whether there is an abnormal increase in current, that is, determine whether the current change is greater than 0.1; if there is an abnormal increase in current, the state will be filtered and no trip alarm will be issued; if not, it will be determined as a trip; Judgment 2: Read the trained substation feature prediction model to obtain the predicted values. Calculate the relative error between the predicted values ​​of phase A current, phase A voltage, phase B current, phase B voltage, phase C current, phase C voltage, load factor, and active power and the corresponding real-time data. Determine whether the error exceeds the error threshold. If not, it is determined to be normal. If it exceeds, it will further determine whether there is an abnormal increase in current, that is, determine whether the current change is greater than 0.1; if there is an abnormal increase in current, the state will be filtered and no trip alarm will be issued; if not, it will be determined as a trip; (d) Based on the results of judgment 1 and judgment 2, if both judgments are "no tripping", no tripping has occurred; otherwise, it is determined that a tripping has occurred.

2. The intelligent prediction and analysis method for low-voltage branch tripping of a distribution network based on dual similarity according to claim 1 is characterized in that: The step 1) is specifically as follows: (a) Each substation selects a state matrix based on its own historical data. The number of substations is determined by the substation itself. The data source is the current, voltage, load factor, and power measurement data sampled every 15 minutes. The state matrix is ​​a sample selected from the historical data. (b) In the case of missing data, if the missing period is less than 1 hour, linear interpolation is used to fill it in; if the missing period is greater than 1 hour, the data for that day is excluded; (c) In the case of data anomalies, the obviously abnormal moment data are directly eliminated.

3. The intelligent prediction and analysis method for low-voltage branch tripping of a distribution network based on dual similarity according to claim 2 is characterized in that: The selected state matrix in step (a) is specifically: Select samples from the given historical data X and form a state matrix D. That is, a process or equipment has n related features. Suppose it is sampled at a certain time i, and the n collected features are selected as a working condition; X(i)=[x1,x2,…,x n ] T Combined with the changes in working conditions during the set time, m working conditions are selected and the state matrix D is formed. Each column vector in the state matrix represents an operating condition of the equipment.

4. The intelligent prediction and analysis method for low-voltage branch tripping of a distribution network based on dual similarity according to claim 1 is characterized in that: The step 2) is specifically as follows: (a) Select phase A current, phase A voltage, phase B current, phase B voltage, phase C current, phase C voltage, load factor, active power, and reactive power as features and sort them in time series; (b) Subtract the features of the previous moment from the features of the next moment to find the corresponding changes of each feature. The features and their changes are combined into a total of 18 feature dimensions to form an n*18 matrix, where n is the number of moments.

5. The intelligent prediction and analysis method for low-voltage branch tripping of a distribution network based on dual similarity according to claim 1 is characterized in that: The step 3) is specifically as follows: Based on the k-means clustering method, m cluster centers are selected from the normal operation data of the past month to form an m*18 initial state matrix, and the value of m ranges from (30, 50).

6. The intelligent prediction and analysis method for low-voltage branch tripping of a distribution network based on dual similarity according to claim 1 is characterized in that: In step 3), the final state matrix is ​​obtained by iteratively optimizing the initial state matrix; at the same time, the LSTM model is trained based on the training set to obtain a station feature prediction model that can predict the next working condition based on the working condition data of the previous three hours, specifically: (a) Select historical data from 96 time points every day for the past month, with a 15-minute interval between each time point, and divide it into a training set and a test set. The test set is the feature set of tripping days when tripping is known to occur, and the training set is the feature set of normal operation. (b) The initial state matrix is ​​input into the pattern similarity operator to calculate the reconstructed estimated value. Based on the grid operation mechanism, eight feature dimensions are analyzed: phase A current, phase A voltage, phase B current, phase B voltage, phase C current, phase C voltage, load factor, and active power. The relative error between the estimated value and the true state value of each feature dimension is calculated, and the error of each feature dimension is obtained by weighted average. (c) Errors with an absolute value greater than 10 in the average positioning error are correspondingly eliminated from the original data and the initial state matrix and regarded as abnormal values ​​mixed into the normal working conditions; (d) When the mean of the error is greater than 0.1 or the variance is greater than 1, and the length of the state matrix is ​​less than 80, add features to the initial state matrix. The calculation method is as follows: calculate the predicted value obtained by the original data and the state matrix through the pattern similarity operator, calculate the error between the predicted value and the true value, locate the feature with the largest error, add this feature to the state matrix, and delete it from the original data; (e) Based on the final state matrix, calculate the mean and variance of the relative error between the estimated value and the true state value, save the mean, variance and the final state matrix, and calculate the error threshold. The calculation formula is as follows: Where G is the error threshold, error is the average error of the nth normal working condition, and std(·) is the variance calculation formula; (f) Use the historical data of the most recent month as training data without excluding tripping data, perform max-min linear normalization, and use the LSTM model for training. Use the data of the first three hours to predict the data of the next 15 minutes, and use 12 data points to predict one data point in the future. Then, denormalize the predicted value, calculate the variance and mean of the error of each feature, and calculate the error threshold according to the error threshold calculation formula and save it.

7. The intelligent prediction and analysis method for low-voltage branch tripping in a distribution network based on dual similarity according to claim 6 is characterized in that: The reconstructed estimated value obtained by inputting the initial state matrix into the pattern similarity operator is specifically: Input condition x at a certain moment in It is composed of 1*18 feature data, where the first 9 features are phase A current, phase A voltage, phase B current, phase B voltage, phase C current, phase C voltage, load factor, active power, and reactive power. The 10th to 18th features are the changes of the first 9 features compared to the previous moment. The specific form is as follows: x in =[x1x2,x3,x4,x5,x6,x7,x8,x9,x 1in x 2in ,x 3in ,x 4in x 5in ,x 6in ,x 7in ,x 8in ,x 9in ] T Input working condition x in Comparing the similarity of each operating condition with the state matrix (D) will produce a similarity vector a, which contains the same number of elements as the training matrix elements stored in the state matrix; (b) Convert the similarity vector representing the similarity between the current working condition and the historical normal working condition into a weight vector (w): w0=G -1 ·to Where: is a nonlinear operator, and is chosen as the Euclidean distance between two vectors, that is: (c) Producing an estimate through a linear combination of samples and weights x out =D·w Right now:

Images