Fault position identification method for power distribution system based on artificial intelligence ammeter data visualization

Abstract

According to the artificial intelligence ammeter data visualization-based power distribution system fault position identification method provided by the invention, three basic learners, namely a random forest, a K-proximity algorithm and an artificial neural network, are combined, and an integration method of combining a plurality of different machine learning models into a prediction model can be adapted to different scenes, so that the accuracy of the prediction model is improved. For example, noise and data loss, fault resistance change, load or power distribution feeder structure change and the like can be realized. An integrated voting classifier is developed, and a random forest, kNN and an artificial neural network model are utilized to classify fault types and identify a limited number of data points of FL only using voltage measurement. According to the method, the key problem of insufficient multiple features is overcome, the fault position can be quickly identified, the method is not influenced by the initial angle of the fault and the resistance on the line, the anti-interference capability is high, the classification accuracy of multiple fault types is high, the detection precision is high, and the cost is not too high.

Description

Technical Field

[0001] This invention relates to the field of power distribution system technology, and in particular to a method for identifying the location of faults in power distribution systems by visualizing data from artificial intelligence electricity meters. Background Technology

[0002] In recent years, due to factors such as a lack of available data, unbalanced network topology, varying load conditions, and the integration of distributed energy resources into microgrids, rapid and accurate fault location in distribution systems containing microgrids has become a major challenge for power systems. Therefore, integrated voting fault classification and location identification methods for microgrid distribution systems are currently a hot research topic. However, existing technologies have too many limitations; single-feature methods struggle to handle complex faults, while multi-feature methods and deep learning methods, although offering better performance, suffer from high costs and poor real-time performance. To achieve efficient and effective integrated voting fault classification and location identification in microgrid distribution systems, an integrated voting fault classification and location identification method based on smart meter measurements is imperative.

[0003] Existing short-term load forecasting methods can be categorized into several types, including fault classification and location identification methods based on wavelet transform and deep learning; multi-feature fault detection and classification methods; fault classification and location identification methods based on deep learning and traveling wave ranging; and fault classification and location identification methods based on Bayesian networks. These methods lack efficiency and effectiveness, exhibiting several drawbacks: limitations of single-feature methods (e.g., threshold-based fault detection relies on preset thresholds for fault judgment, but improper threshold selection can easily lead to false alarms or misreporting, and they cannot adapt to complex fault modes); shortcomings of multi-feature methods (e.g., multi-feature fault detection requires constructing multiple input feature values, but how to effectively integrate these features remains a problem to be solved); and limitations of ensemble voting methods (e.g., the output results of different fault detection technologies may differ, and how to effectively fuse these results is a technical challenge). Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes a method for identifying the location of faults in a power distribution system using artificial intelligence-based visualization of electricity meter data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for identifying fault locations in a power distribution system using AI-powered electricity meter data visualization includes the following steps:

[0007] Includes the following steps:

[0008] S1. Analyze the ensemble learning for fault location identification. Use the capacitor voltage measurement values ​​of the end of the branch and the load point on the photovoltaic bus after the fault to identify the fault location. The ensemble learning combines multiple machine learning algorithms to build a set of hypotheses from the base learner and combine the results of the base learner through unweighted or weighted voting.

[0009] S2, Analysis and learning of artificial neural network learning algorithms, K-nearest neighbor learning algorithm, random forest algorithm;

[0010] S3. Design an integrated voting classifier, combining the three basic learners from step S2, train the base learners in each step, and then complete the final prediction through an integrated hard voting and soft voting method.

[0011] S4. Perform fault type classification and fault location identification. Fault classification is accomplished using faults on the selected bus voltage. After the faults are classified, the base learner and the ensemble model are trained using the relevant voltage vectors for each fault type.

[0012] Preferably, in step S1, let B be the set of all buses in the distributed network, b be the set of buses with intelligent modules, and the cardinality of b be |b| = ns. The set of voltage measurement values ​​after a fault is a row vector, x has a cardinality |x| = n, ns * ps, where ps is the phase associated with the corresponding bus. The fault location indicator identification task is described as a classification problem, where fc is the fault classification ensemble model, FL = f1, f…, fn are the FLI ensemble models for each fault type, yc = fc(x) predicts the fault type, and then yl = fl(xp) predicts the FL, where xp ∈ x is a subset of x, containing only the faulty phase measurement.

[0013] Preferably, in step S1, the ensemble learning process of the base model is as follows: There are T classifiers h1, h2…h… n By combining h i A classifier is used to predict the possible class labels b1, b2, ..., b in a set of "k". k ∈B, suppose that for an example x, the output of classifier his is a k-dimensional label vector (h1(x),…,h n h(x)), where h(x) is the output of class label bj.

[0014] Preferably, in step S1, the combination method has two types: hard voting or majority voting and soft voting. In majority voting, the ensemble model H... i The final prediction result is:

[0015]

[0016] In hard voting, the final output is the result produced by the majority of base models. This method statistically analyzes the results of all individual base models and selects the result with the highest frequency. In soft voting, the individual probability outputs of all base classifiers are averaged, and the final output is selected based on this probability. In soft voting, the final output for class b is as follows:

[0017]

[0018] Preferably, in step S2, the artificial neural network learning algorithm model has an input layer and an output layer, with the layers between them called hidden layers. Neurons use the weighted outputs of other neurons in the previous layer as input information for each layer. Each neuron calculates the result as shown in the following formula and passes it to the next neuron.

[0019]

[0020] x pq The signal value w from input q to neuron p. pq It is the association weight between neuron IP and input q, while b p This is the association bias of neuron p, and then the output of each neuron is fed into the activation function, as follows:

[0021]

[0022] φ is the activation function, y p It is the output of each neuron. In each layer, neurons receive outputs from other connected units, add weighted inputs, and pass the output through an activation function.

[0023] Preferably, in step S2, the K-nearest neighbor learning algorithm assigns weights to the contributions of neighbors based on a similarity criterion, such that the contribution of the nearest neighbor to the average value is greater than that of the farthest neighbor. This includes two hyperparameters: distance metric and the number of neighbors k. The distance metric is expressed as follows:

[0024]

[0025] x and y are two inputs with n characteristics, x i and y i d represents the i-th attribute of x and y respectively, the smaller value of d indicates the similarity between the two samples, and the K-nearest neighbors of sample x determine the final output through majority voting.

[0026] Preferably, in step S2, the parameters of the random forest algorithm include: the number of trees in the forest (b), the number of features in each branch (m), and the minimum number of samples or leaf size. The method for implementing the random forest model is as follows:

[0027] S21. Generate a sample dataset b of size N;

[0028] S22. For each sample dataset, grow the RF tree by iterating through the following process until the leaf size is reached: select m random features from the feature set; for each split point, select the best feature; and divide each node into two branch nodes according to the decision criteria.

[0029] S23. A set of spanning trees. Calculate the classifier for a specific sample x based on the average output of each tree.

[0030] Preferably, in step S4, in the fault type classification step, the number of features for each fault event is the number of substations n multiplied by the corresponding phase number. All voltage measurement data are used as input, and one-hot encoding is used to convert the output of the fault type from classification to numerical value. Each classifier generates an output vector C with different faults. The final output is predicted by majority voting, using the prediction of each classifier.

[0031] Preferably, in step S4, during the fault location and identification process, voltage, current, and power data are recorded in real-time using the DNP3 protocol. If a predefined event occurs, the substation communicates with the master station to set this event as a voltage threshold violation.

[0032] 0.95pu≤V≤1.05pu

[0033] If any SMS message is detected to exceed the voltage threshold, all SMS data will be collected via DNP3 communication in the form of a voltage array and sent to the utility data center: V1V2,V3……V n The integrated algorithm predicts the type and location of the fault.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] This invention utilizes event-driven smart meter voltage data for fault classification and localization. The integrated voting classifier consists of three base learners: random forest, k-nearest neighbors, and artificial neural network. The fault localization problem is transformed into a classification problem; first, fault types are classified, and then the faulty bus is identified based on the fault type. This method overcomes the key problem of insufficient multi-feature coverage, enabling rapid fault location identification, unaffected by the initial fault angle and line resistance, exhibiting strong anti-interference capabilities, high classification accuracy for multiple fault types, high detection precision, and reasonable cost. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the fault types and fault location process of the present invention.

[0037] Figure 2A diagram of the ensemble learning process;

[0038] Figure 3 Flowchart for fault location identification;

[0039] Figure 4 This is a flowchart of the DNP3 communication process. Detailed Implementation

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0041] like Figure 1 As shown in the figure, this embodiment proposes a method for identifying the location of faults in a power distribution system using artificial intelligence-based electricity meter data visualization, including the following steps:

[0042] Includes the following steps:

[0043] S1. Analyze the ensemble learning for fault location identification, with the aim of improving the robustness and overall performance of the model.

[0044] Fault location is identified by measuring the capacitance voltage at the end of the branch and the load point on the photovoltaic bus after a fault. Let B be the set of all buses in the distributed network, b be the set of buses with smart modules, and the cardinality of b be |b| = ns. The set of voltage measurements after a fault is a row vector, x has a cardinality |x| = n, ns * ps, where ps is the phase associated with the corresponding bus. The fault location indicator identification task is expressed as a classification problem, where fc is the fault classification ensemble model, FL = f1, f…, fn are the FLI ensemble models for each fault type, yc = fc(x) predicts the fault type, and then yl = fl(xp) predicts the FL, where xp ∈ x is a subset of x, containing only the faulty phase measurement.

[0045] Ensemble learning combines multiple machine learning algorithms to avoid poor decision probabilities and improve the robustness and overall performance of the model. It constructs a set of hypotheses from base learners and combines the results of the base learners through unweighted or weighted voting to reduce the variance and bias of the base models and improve overall performance. Ensemble methods introduce diversity, and the weaknesses of one algorithm are often compensated for by another, thus improving the accuracy of the ensemble or at least outperforming the worst-case scenario. The motivation for using ensemble models is to improve accuracy with limited data. Every model has its shortcomings, but different models have different strengths and weaknesses. Data sets that are too large or too small will affect the classification performance of a single learning algorithm. The decision boundaries separating different classes may be too complex or may fall outside the function space of the chosen classifier. Combining multiple models can improve overall accuracy and reduce the likelihood of poor performance under specific conditions. If individual classifiers have sufficient diversity and accuracy, ensemble methods will produce better results. An accurate classifier produces a lower error rate on new samples than random guessing, while multiple classifiers produce different errors on new samples, such as... Figure 2 The ensemble learning process of the basic model is as follows: Given T classifiers h1, h2…h… n By combining h i A classifier is used to predict the possible class labels b1, b2, ..., b in a set of "k". k ∈B, suppose that for an example x, the output of classifier his is a k-dimensional label vector (h1(x),…,h n h(x)), where h(x) is the output of class label bj.

[0046] There are two types of ensemble methods: hard voting or majority voting and soft voting. In majority voting, the ensemble model H... i The final prediction result is:

[0047]

[0048] In hard voting, the final output is the result produced by the majority of base models. This method statistically analyzes the results of all individual base models and selects the result with the highest frequency. In soft voting, the individual probability outputs of all base classifiers are averaged, and the final output is selected based on this probability. In soft voting, the final output for class b is as follows:

[0049]

[0050] Individual classifiers have been selected to ensure diverse outputs and sufficient accuracy. For the overall accuracy of the ensemble method to improve, the base classifiers must possess both diversity and sufficient accuracy. If the base classifiers are inherently not diverse enough, they will make the same mistakes, and combining them will not produce better results. Similarly, if each selected model is not accurate enough, the overall result may worsen. In most cases, the number of base classifiers is chosen to be odd to avoid receiving the same number of votes in a hard-vote scenario. However, increasing the number of models is only recommended if the overall accuracy improves. In this embodiment, adding two base models did not bring significant improvement, so three base models were chosen. Diversity and accuracy are the most important factors to consider when selecting base models. The three base learners used in the proposed ensemble method are artificial neural networks, nearest neighbors, and random forest models because of their high accuracy and diversity. These models are inherently different. Artificial neural networks are powerful classifiers capable of learning complex patterns in datasets. K-nearest neighbors are considered a memoryless lazy learning algorithm that classifies each new sample based on a calculated distance. The K-nearest neighbor algorithm often performs poorly when dealing with large-scale datasets, especially at cluster boundaries. Random forests are a combination of decision trees that can easily learn less complex features in a dataset and perform well even on large datasets.

[0051] S2, analyzing and learning artificial neural network learning algorithms, K-nearest neighbor learning algorithm, random forest algorithm, among which,

[0052] In an artificial neural network learning algorithm model, there is an input layer and an output layer. The layers between these two layers are called hidden layers. Neurons take the weighted outputs of other corresponding neurons in the previous layer as input information for each layer. Each neuron calculates the result shown in the following formula and passes it to the next neuron.

[0053]

[0054] x pq The signal value w from input q to neuron p. pq It is the association weight between neuron IP and input q, while b p This is the association bias of neuron p, and then the output of each neuron is fed into the activation function, as follows:

[0055]

[0056] φ is the activation function, y p This represents the output of each neuron. In each layer, neurons receive outputs from other connected units, add weighted inputs, and pass the output through an activation function. Weights and biases are updated to minimize the error.

[0057] The K-Nearest Neighbors learning algorithm assigns weights to the contributions of neighbors based on a similarity criterion, ensuring that the nearest neighbor contributes more to the average than the farthest neighbor. It includes two hyperparameters: a distance metric and the number of neighbors, k. The distance metric is expressed as follows:

[0058]

[0059] x and y are two inputs with n characteristics, x i and y i Let x and y be the i-th attributes of x and y, respectively. The smaller value of d represents the similarity between the two samples. The final output is determined by the K-nearest neighbors of sample x through majority voting. This approach combines competitive results with a lazy learning method.

[0060] The parameters of the random forest algorithm include: the number of trees in the forest (b), the number of features in each branch (m), and the minimum number of samples or leaf size. The method for implementing the random forest model is as follows:

[0061] S21. Generate a sample dataset b of size N;

[0062] S22. For each sample dataset, grow the RF tree by iterating through the following process until the leaf size is reached: select m random features from the feature set; for each split point, select the best feature; and divide each node into two branch nodes according to the decision criteria.

[0063] S23. A set of spanning trees; calculate the average value of a specific sample based on the output of each individual tree.

[0064] A classifier for x.

[0065] S3. Design an ensemble voting classifier, combining the three basic learners from step S2. Train the base learners in each step, and then complete the final prediction using an ensemble method of hard voting and soft voting. The entire process is as follows: Figure 3 As shown;

[0066] S4. Fault type classification and fault location identification are performed. Fault classification is completed using faults on the selected bus voltage. The SMS (Smart Meter) in the AMI (Advanced Metering Infrastructure) is used to collect event-driven data during and after the fault. In the proposed method, only the voltage amplitude in the SMS is used for fault classification. All 11 possible fault types are placed on different buses. As in DN, three-phase faults are rare, and three-phase ground faults and three-phase unground faults are combined into a three-phase fault (3P). Therefore, a total of 10 labels (Ag, Bg, Cg, AB-g, BC-g, CA-g, AB, BC, CA, 3P) are used for classification in the fault classification task, as shown in Table 1. In radial topology DN, the SMs at the branch endpoints provide the most relevant information for FLI. The number of features for each fault event is the number of substations n multiplied by the corresponding phase number. In the fault type classification mode, all 67 voltage measurement data are used as input. One-hot encoding is used to convert the fault type output from classification to numerical value. Each classifier produces an output vector C with different faults, as shown in Table 1. The final output is predicted by majority voting, using the prediction of each classifier.

[0067] Table 1 Fault Types

[0068]

[0069] After classifying the faults, the base learner and ensemble model are trained using the relevant voltage vectors for each fault type. The voltage profile after a fault shows that the affected phase voltages have the largest variance during a fault; therefore, only the affected phase voltages are used for the FL task. The size of the input vector fed into the model is the number of affected phases *p* multiplied by the corresponding number of short messages. For a single-line-to-ground (SLG) fault (i.e., Ag), the vector size is 1n, where 1 is the number of affected phases and n is the number of short messages with *a* phase readings. Similarly, for a two-line fault, the vector size is 2n, and for a three-phase fault, the vector size is 3n.

[0070] Microgrids consist of distributed energy resources (DERs), such as photovoltaic power plants, energy storage systems, and locally operating loads that can be handled as single entities. Microgrids typically operate at low and medium voltage levels, ranging from 400V to 69KV. Protecting and identifying faults (FLs) in microgrid scenarios is becoming a complex task with increasing DER penetration. Rapid fault clearance and power restoration require three steps: fault detection, fault classification, and fault line identification. Fault detection is not discussed in this embodiment, as relevant technologies already exist in outage management software. The integrated method proposed in this embodiment can be combined with existing detection methods to ensure that the microgrid can isolate the smallest possible portion of the fault. Advanced metering infrastructure enables grid monitoring and reporting capabilities. In the proposed method, SM voltage measurement capabilities have been used to obtain post-fault voltage measurements from metering points. Figure 4 The flowchart shows the use of DNP3 communication to utilize the SM block and modified AMI infrastructure for post-fault voltage data.

[0071] SMS in AMI can record voltage, current, and power data at near real-time resolution. However, constant data recording requires enormous bandwidth and data storage capacity. This embodiment proposes an improved function based on DNP3 communication and SM. During the fault location and identification process, voltage, current, and power data are recorded at real-time resolution using the DNP3 communication protocol. If a predefined event occurs, the substation communicates with the master station to set this event as a voltage threshold violation.

[0072] 0.95pu≤V≤1.05pu

[0073] pu is the per-unit value of voltage. If any SMS message detects a voltage threshold exceeding the limit, all SMS data will be collected via DNP3 communication in the form of a voltage array and sent to the utility data center: V1V2,V3……V n The integrated algorithm predicts the type and location of faults. This significantly reduces the bandwidth requirements of AMI infrastructure.

[0074] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying the location of faults in a power distribution system using artificial intelligence-based visualization of electricity meter data, characterized in that, Includes the following steps: S1. Analyze the ensemble learning for fault location identification. Use the capacitor voltage measurement values ​​of the end of the branch and the load point on the photovoltaic bus after the fault to identify the fault location. The ensemble learning combines multiple machine learning algorithms to build a set of hypotheses from the base learner and combine the results of the base learner through unweighted or weighted voting. S2, Analysis and learning of artificial neural network learning algorithms, K-nearest neighbor learning algorithm, random forest algorithm; S3. Design an integrated voting classifier, combining the three basic learners from step S2, train the base learners in each step, and then complete the final prediction through an integrated hard voting and soft voting method. S4. Perform fault type classification and fault location identification. Fault classification is accomplished using faults on the selected bus voltage. After the faults are classified, the base learner and the ensemble model are trained using the relevant voltage vectors for each fault type.

2. The method for identifying the location of faults in a power distribution system based on visualization of artificial intelligence meter data according to claim 1, characterized in that, In step S1, let B be the set of all buses in the distributed network, b be the set of buses with intelligent modules, and the cardinality of b be |b| = ns. The set of voltage measurement values ​​after a fault is a row vector, x has a cardinality |x| = n, ns * ps, where ps is the phase associated with the corresponding bus. The fault location indicator identification task is described as a classification problem, where fc is the fault classification ensemble model, FL = f1, f…, fn are the FLI ensemble models for each fault type, yc = fc(x) predicts the fault type, and then yl = fl(xp) predicts the FL, where xp ∈ x is a subset of x, containing only the faulty phase measurement.

3. The method for identifying the location of faults in a power distribution system based on visualization of artificial intelligence meter data according to claim 1, characterized in that, In step S1, the ensemble learning process of the base model is as follows: There are T classifiers h1, h2…h… n By combining h i A classifier is used to predict the possible class labels b1, b2, ..., b in a set of "k". k ∈B, suppose that for an example x, the output of classifier his is a k-dimensional label vector (h1(x),…,h n h(x)), where h(x) is the output of class label bj.

4. The method for identifying the location of faults in a power distribution system based on visualization of artificial intelligence meter data according to claim 1, characterized in that, In step S1, there are two types of combination methods: hard voting or majority voting and soft voting. In majority voting, the ensemble model H... i The final prediction result is: In hard voting, the final output is the result produced by the majority of the base models. This method statistically analyzes the results of all individual base models and selects the result that occurs most frequently. In soft voting, the individual probability outputs of all base classifiers are averaged, and the final output is selected based on this probability. In soft voting, the final output for class b is as follows:

5. The method for identifying the location of faults in a power distribution system based on artificial intelligence meter data visualization according to claim 1, characterized in that, In step S2, the artificial neural network learning algorithm model has an input layer and an output layer. The layers between these layers are called hidden layers. Neurons take the weighted outputs of other neurons in the previous layer as input information for each layer. Each neuron calculates the result as shown in the following formula and passes it to the next neuron. x pq The signal value w from input q to neuron p. pq It is the association weight between neuron IP and input q, while b p This is the association bias of neuron p, and then the output of each neuron is fed into the activation function, as follows: φ is the activation function, y p It is the output of each neuron. In each layer, neurons receive outputs from other connected units, add weighted inputs, and pass the output through an activation function.

6. The method for identifying the location of faults in a power distribution system using artificial intelligence-based meter data visualization as described in claim 1, characterized in that, In step S2, the K-nearest neighbor learning algorithm assigns weights to the contributions of neighbors based on a similarity criterion, ensuring that the contribution of the nearest neighbor to the average value is greater than that of the farthest neighbor. This includes two hyperparameters: distance metric and the number of neighbors k. The distance metric is expressed as follows: x and y are two inputs with n characteristics, x i and y i d represents the i-th attribute of x and y respectively, the smaller value of d indicates the similarity between the two samples, and the K-nearest neighbors of sample x determine the final output through majority voting.

7. The method for identifying the location of faults in a power distribution system based on artificial intelligence meter data visualization according to claim 1, characterized in that, In step S2, the parameters of the random forest algorithm include: the number of trees in the forest (b), the number of features in each branch (m), and the minimum number of samples or leaf size. The method for implementing the random forest model is as follows: S21. Generate a sample dataset b of size N; S22. For each sample dataset, grow the RF tree by iterating through the following process until the leaf size is reached: select m random features from the feature set; for each split point, select the best feature; and divide each node into two branch nodes according to the decision criteria. S23. A set of spanning trees. Calculate the classifier for a specific sample x based on the average output of each tree.

8. The method for identifying the location of faults in a power distribution system based on visualization of artificial intelligence meter data according to claim 1, characterized in that, In step S4, the fault type classification step, the number of features for each fault event is the number of substations n multiplied by the corresponding phase number. All voltage measurement data are used as input, and one-hot encoding is used to convert the fault type output from classification to numerical value. Each classifier generates an output vector C with different faults. The final output is predicted by majority voting, using the prediction of each classifier.

9. The method for identifying the location of a power distribution system based on visualization of artificial intelligence meter data according to claim 1, characterized in that, In step S4, during the fault location and identification process, voltage, current, and power data are recorded in real-time using the DNP3 protocol. If a predefined event occurs, the substation communicates with the master station to set this event as a voltage threshold violation. 0.95pu≤V≤1.05pu If any SMS message is detected to exceed the voltage threshold, all SMS data will be collected via DNP3 communication in the form of a voltage array and sent to the utility data center: V1V2,V3……V n The integrated algorithm predicts the type and location of the fault.

Images