A power distribution network fault discrimination method based on a PCCNN model
By directly generating full-order polynomial product terms using the PCCNN model and performing weighted summation, the problems of slow speed and low accuracy in distribution network fault location are solved, achieving fast and accurate fault identification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-30
- Publication Date
- 2026-07-10
AI Technical Summary
Existing technologies for fault location in power distribution networks suffer from slow fault location speed and low accuracy, failing to meet the requirements for both speed and high accuracy.
A fault discrimination method based on the PCCNN model is adopted. By constructing a neural network containing an input layer, a hidden layer with a π∑∑ architecture, and an output layer, full-order polynomial product terms are directly generated. Weighted summation and nonlinear mapping are then performed to capture the nonlinear high-order correlation between electrical data, thereby achieving accurate screening of fault features and suppression of noise interference.
It improves the sensitivity and stability of fault identification, enhances the generalization ability of the model, ensures the integrity and accuracy of fault features, reduces noise interference, and improves the speed and accuracy of fault location.
Smart Images
Figure CN122365046A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of distribution network fault identification technology, and in particular to a distribution network fault identification method based on a PCCNN model. Background Technology
[0002] Power distribution networks are closely related to people's lives. When a fault occurs in the distribution network, if the fault location cannot be quickly and accurately located and eliminated, it will not only cause huge economic losses but also personal injury accidents. Faced with a fault, the options are either to cut off a large area of power to eliminate the fault, or to use artificial intelligence to accurately and quickly locate and eliminate the fault. As the requirements for power supply reliability become increasingly stringent, the former method is no longer feasible, and the latter is becoming a trend.
[0003] Traditional neural networks have been widely used in distribution network fault location applications, but some problems have gradually emerged during application. For example, deep neural networks have extremely strong learning capabilities and can guarantee high fault location accuracy, but the network itself requires a large amount of memory resources, resulting in slow convergence speed and failing to meet the requirements for rapid fault location. Traditional BP neural networks have a simple structure and fast convergence speed, but their nonlinear processing capabilities are poor, leading to low fault location accuracy. Therefore, in the application of distribution network fault location, finding a new neural network structure that meets the requirements of speed and high accuracy has become an inevitable task. Summary of the Invention
[0004] In view of the above-mentioned prior art, the present invention provides a method for power distribution network fault identification based on a PCCNN model, which mainly solves the technical problems existing in the background art.
[0005] To achieve the above objectives, the technical solution of this invention is implemented as follows: This invention discloses a method for power distribution network fault identification based on a PCCNN model, the method comprising the following steps: Fault labeling is performed on historical data of the power distribution network, a PCCNN model is established and trained based on the historical data; Real-time data is input into the trained PCCNN model, which outputs fault prediction results. The PCCNN model includes an input layer connected in sequence, layer, layer, The input layer is used to receive the electrical characteristic input vector of the distribution network, and the output layer is used to receive the input vector of the distribution network. The layer is used to generate and output full-order polynomial product terms from the input features of the input layer. Layers are used for The polynomial product terms of the same degree output by the layer are subjected to weighted summation. Used for The full-order output of the layer is globally linearly weighted and a bias term is added. The output layer has a built-in activation function used to... The output of the layer is mapped and the final fault diagnosis result is output.
[0006] Optionally, fault labeling is performed on historical data of the distribution network, a PCCNN model is built and trained based on the historical data, specifically including: Historical data that experienced failures were labeled as 1, and historical data that did not experience failures were labeled as 0. The training set and the test set were then divided based on the labeled historical data. A PCCNN model is constructed and trained based on the training set. The trainable parameters of the model are optimized through the gradient algorithm, the mean squared error loss function is minimized, the optimal weight parameters of the model are obtained, and the model training is completed by fixing the optimal weight parameters.
[0007] Optionally, the power distribution network electrical feature input vector received by the input layer includes the current data and voltage data of the three phases A, B, and C of the power distribution network, and the dimension of the input vector is 6-dimensional.
[0008] Optionally, the When generating full-order polynomial product terms, for the N-dimensional input features of the input layer, polynomial product terms of order 1 to N are generated sequentially. Each order polynomial product term is a combination product of all input features under that order. The number of product terms of order j is the binomial coefficient of selecting j elements from the N input features, where the value of j ranges from 1 to N.
[0009] Optionally, the The layer's output vector is:
[0010] The q-th element of the j-th order subvector is:
[0011] in, For the i-th input vector, For the input layer metrics, for The q-th subset, For indicator set The set consisting of all j-element subsets in the set.
[0012] Optionally, the Layers When performing a weighted summation of polynomial product terms of the same order output by the layer, for Each polynomial product term of the layer output is assigned a corresponding trainable weight vector. The output of each order is the weighted sum of all polynomial product terms of that order and their corresponding weight vectors. The final output of the layer is a one-dimensional vector with the same order as the input features.
[0013] Optionally, the The output of the j-th node in the layer is:
[0014] in, To select j elements from N input nodes, the binomial coefficients are... To and Dimensionality-matched trainable weight vectors; The output vector of the layer is .
[0015] Optionally, the Layers When processing the output of the layer, it is Each order result in the layer output is assigned a corresponding constraint weight. The global linear weighted result is the sum of the products of all order results and their corresponding constraint weights, which are then superimposed with the model's bias term before being output to the output layer.
[0016] Optionally, the The layer output is:
[0017] in, for Layers and Trainable weight vectors between layers For the corresponding The constraint weight values, As the bias term of the model, the final output of the PCCNN model is ),in This is the activation function built into the output layer.
[0018] The beneficial effects of this invention are: it employs a generation logic that directly constructs higher-order product features, through... The layer directly performs full subset multinomial product operations on the original electrical feature vectors of the input layer, skipping the initial linear combination step and directly capturing the nonlinear high-order correlations between electrical data in the distribution network. This design directly preserves the fault feature correlations of the original data, completely avoiding the fault feature distortion problem caused by multiplication after linear operations; at the same time, it retains single-dimensional electrical features through univariate combination, captures two-parameter interaction features through binary combination, and captures multi-parameter coupling features through multivariate combination, fully covering the full-dimensional electrical feature changes caused by distribution network faults. It can effectively amplify the feature differences of minor faults and hidden faults, ensure the integrity of fault features, and fundamentally improve the model's sensitivity to fault data. This solves the industry pain point of existing technologies being insufficient in capturing weak and hidden fault features and prone to missed detection. pass Layers The output polynomial terms of the same order from the layer are grouped and weighted to achieve unified screening and weighting of features of the same order; through Each layer assigns independent constraint weights to the weight vector of each order group, enabling overall control of feature groups of different orders. This design achieves precise screening of fault features through order classification, strengthens the weighting of fault-sensitive features in the same order group, and precisely suppresses redundant noise features in the same order group. From an architectural perspective, it avoids invalid interference from features of different orders, allowing the model to accurately focus on the core fault features. At the same time, the order-independent constraint weights can adaptively adjust parameters from the effective solution space of the corresponding order through training, rather than being limited to a single optimal solution. This significantly enhances the model's generalization ability, effectively prevents overfitting, significantly reduces noise interference in the data collected from the distribution network field, and improves the model's fault discrimination stability under complex operating conditions. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a flowchart illustrating the power distribution network fault identification method based on the PCCNN model in the embodiments of this application; Figure 2 This is a schematic diagram of the architecture of the PCCNN model in the embodiments of this application; Figure 3 A comparison chart of the error curves of the two models built; Figure 4 A comparison chart showing the changes in test set accuracy for the two models. Figure 5This is a bar chart showing the constraint weights of the constructed PCCNN model. Detailed Implementation
[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. In the following description, the expression "some embodiments" refers to a subset of all possible embodiments; however, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.
[0022] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.
[0023] It should be understood that the present invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Furthermore, the terminology used herein is intended only to describe particular embodiments and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “compose” and / or “comprising,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.
[0024] It should also be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "inner," "outer," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.
[0025] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.
[0026] Please refer to the attached document. Figures 1 to 2 This invention discloses a method for fault identification in power distribution networks based on a PCCNN model. The method includes the following steps: S1. Mark faults in historical data of the distribution network, establish a PCCNN model and train it based on historical data; In some optional implementations, historical operating data of the distribution network is extracted from the distribution network master station. The extracted historical operating data includes six core electrical features of the distribution network: A-phase current, B-phase current, C-phase current, A-phase voltage, B-phase voltage, and C-phase voltage. This constitutes the basic data source for model input. The extracted historical operating data undergoes data cleaning and normalization preprocessing to eliminate the interference of differences in the dimensions of different electrical features on the model training process, thereby improving the convergence speed and stability of model training. Fault labeling is performed on the preprocessed historical operating data. Sample data of distribution network faults in the historical operating data are labeled as the target output value 1, and sample data of the distribution network operating normally without faults in the historical operating data are labeled as the target output value 0. After labeling, all labeled samples are divided into training set and test set according to a preset ratio. The training set is used for iterative optimization and fitting of model parameters, and the test set is used to verify the fault discrimination accuracy of the model during training and to select the optimal model parameters.
[0027] The PCCNN model established in this application comprises one input layer, three hidden layers, and one output layer. The three hidden layers adopt a ∏∑∑ architecture, which is composed of… layer, layer, Layer composition, through The layer constructs a polynomial output from the input layer. layer, Polynomial terms of the same order are grouped into a single-layer unit. After applying weighted constraints to each unit, the result is... The layer generates an output, which ultimately produces a single output, obtained through the activation function g().
[0028] The input layer is used to receive the electrical characteristic input vector of the distribution network, the input vector being... It means that, among them This represents the electrical characteristic value of the i-th input node in the input layer, and also sets... This represents the set of indicators for the input layer.
[0029] The first hidden layer of the PCCNN model uses fixed weights and has no parameters to be trained. Its core function is to directly perform a full subset polynomial product operation on the electrical feature input vector output by the input layer, skipping the initial linear combination step in existing technologies, directly capturing the nonlinear high-order correlations between data, and generating full-order polynomial product terms from order 1 to 6. This fully preserves the nonlinear coupling correlations between the original electrical data, avoids the distortion of fault features caused by the linear combination followed by multiplication, amplifies the feature differences between minor and hidden faults, and ensures the integrity of fault features. The layer's output vector is:
[0030] Where the j-th order subvector It consists of the product of all j-ary input features. The formula for calculating the q-th element is:
[0031] in, For the i-th input vector, For the input layer metrics, for The q-th subset, For indicator set The set consisting of all j-element subsets in the set.
[0032] The dimension of the subvector is consistent with the binomial coefficient C(N,j), where the first-order subvector corresponds to the six single-dimensional electrical features themselves, the second-order subvector corresponds to the product of all pairwise electrical features, and so on, with the sixth-order subvector corresponding to the total product of all six electrical features. The subvectors of all orders are concatenated in order to form the subvector. The complete output vector of the layer, the output result is transmitted to layer.
[0033] The This layer is the second hidden layer of the PCCNN model, and its core function is to... The output of each layer performs grouped and weighted summation of polynomial products of the same order, enabling adaptive filtering and weighting of fault features of different orders. By classifying by order, precise feature selection is achieved, avoiding invalid interference from features of different orders and allowing the model to focus on the core fault features. Each polynomial product term of order in the layer output is assigned a trainable weight vector with corresponding dimension matching. The output of the j-th node in the layer is:
[0034] in, To select j elements from N input nodes, the binomial coefficients are... To and The dimensionality-matched trainable weight vector can adaptively adjust weight values during training to strengthen features sensitive to fault detection and suppress irrelevant noise features. The output vector of the layer is The dimension of the output vector is the same as the total order of the input features, and its output is transmitted to... layer.
[0035] The This layer is the third hidden layer of the PCCNN model, and its core function is to... The full-order outputs of the layer are subjected to global linear weighting and the model's bias terms are added to them, resulting in... Each order of the layer output is assigned an independent constraint weight, enabling overall control of feature groups of different orders. Layers and The trainable weight vector between layers is The The layer output is:
[0036] in, for Layers and Trainable weight vectors between layers For the corresponding The constraint weight values, As the bias term of the model, the values of the constraint weights and bias term can be adaptively adjusted during training. This enhances the weights of features of effective orders and suppresses the weights of features of irrelevant orders, further improving the model's fault detection accuracy and generalization ability. The output of the layer is transmitted to the output layer, which has a built-in activation function for... The output results are nonlinearly mapped to produce the final fault diagnosis result. The final output formula of the model is: ),in The activation function built into the output layer is the Sigmoid function, which is used here. The output result after the activation function mapping takes values between 0 and 1, thereby realizing the binary classification of distribution network faults.
[0037] After completing the PCCNN model structure, the model is trained based on the aforementioned training set. During training, mean squared error is used as the loss function to measure the error between the model's actual output and the target output. The formula for the loss function is:
[0038] in, This represents the total number of training samples in the training set. This represents the model's actual output for the a-th input sample. Let be the target output corresponding to the a-th input sample, and let the error term for a single sample be . The core objective of training is to adjust the trainable weight vector of the model. To minimize the loss function The optimal weight parameters of the model are obtained. During the training process, the gradient algorithm is used to iteratively optimize the trainable parameters of the model. The gradient of the loss function with respect to each trainable parameter is derived through the chain rule, thereby determining the update direction and update step size of the parameters.
[0039] During iterative training, after each preset number of iterations, the fault identification accuracy of the current model is verified using a test set. The corresponding model parameters and accuracy results are recorded. When the loss function converges to a preset threshold or the number of iterations reaches the preset maximum number of iterations, iterative training is stopped. The model parameters with the highest fault identification accuracy on the test set are selected as the optimal weight parameters. The optimal weight parameters are fixed to complete the training process of the PCCNN model.
[0040] S2. Input the real-time data into the trained PCCNN model. The output value after activation function mapping is in the range of 0 to 1. Based on the output, perform fault judgment of the real-time operating condition of the distribution network. When the output result is 1, it is judged that a fault has occurred in the distribution network and further fault investigation and elimination operations are required. When the output result is 0, it is judged that the distribution network is in a normal fault-free operating condition and no additional operations are required. Continuously monitor and judge the real-time operating data of the distribution network to complete the entire process of outputting the fault prediction results of this real-time data.
[0041] To verify the performance of the PCCNN model in this application, a Recurrent Sigma-Pi-Sigma Neural Network (RSPSNN) and the PCCNN model of this application were constructed for comparison and verification.
[0042] The dataset used consists of 1000 historical operational data entries from a single node of the distribution network, including 500 normal data entries and 500 fault data entries. Phase A current, phase B current, phase C current, phase A voltage, phase B voltage, and phase C voltage are extracted as core discriminant features, normalized, and used as the input vector for the PCCNN. The training and test sets were divided in an 8:2 ratio. The experiment was conducted in a Python 13 environment.
[0043] To ensure fairness in the experiment, the parameter scale and environmental parameter settings of the two models were kept consistent. The specific configuration is shown in Table 1.
[0044] Table 1 Parameter Settings
[0045] Training employs batch gradient descent, using mean squared error (MSE) as the loss function. Figure 3 It can be seen that the MSE of both models gradually decreases with the increase of training epochs, but PCCNN can escape local optima more quickly and trains faster with the same number of iterations.
[0046] from Figure 4 It can be seen that after 180 iterations, PCCNN achieved an accuracy of 99.9% on the test set, while the recursive Sigma-Pi-Sigma neural network achieved 96.5% accuracy after 1000 iterations. Combining the error curves, PCCNN demonstrates higher accuracy in fault detection on the test set and stronger generalization ability, even with higher training set errors. Figure 5 As can be seen, the constraint weights of PCCNN can suppress the feature weights of irrelevant order terms as a whole, and enhance the feature weights of effective order terms as a whole. The experimental results are shown in Table 2, indicating that PCCNN outperforms the recurrent Sigma-Pi-Sigma neural network on this dataset.
[0047] Table 2 Experimental Results
[0048] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for fault identification in distribution networks based on a PCCNN model, characterized in that, The aforementioned method includes the following steps: Fault labeling is performed on historical data of the power distribution network, a PCCNN model is established and trained based on the historical data; Real-time data is input into the trained PCCNN model, which outputs fault prediction results. The PCCNN model includes an input layer connected in sequence, layer, layer, The input layer is used to receive the electrical characteristic input vector of the distribution network, and the output layer is used to receive the input vector of the distribution network. The layer is used to generate and output full-order polynomial product terms from the input features of the input layer. Layers are used for The polynomial product terms of the same degree output by the layer are subjected to weighted summation. Used for The full-order output of the layer is globally linearly weighted and a bias term is added. The output layer has a built-in activation function used to... The output of the layer is mapped and the final fault diagnosis result is output.
2. The method for power distribution network fault identification based on a PCCNN model according to claim 1, characterized in that, Fault labeling is performed on historical data of the power distribution network, a PCCNN model is built and trained based on the historical data, specifically including: Historical data that experienced failures were labeled as 1, and historical data that did not experience failures were labeled as 0. The training set and the test set were then divided based on the labeled historical data. A PCCNN model is constructed and trained based on the training set. The trainable parameters of the model are optimized through the gradient algorithm, the mean squared error loss function is minimized, the optimal weight parameters of the model are obtained, and the model training is completed by fixing the optimal weight parameters.
3. The method for fault identification in a distribution network based on a PCCNN model according to claim 1, characterized in that, The input layer receives the power distribution network electrical feature input vector, which includes the current data and voltage data of the three phases A, B, and C of the power distribution network. The input vector has a dimension of 6.
4. The method for power distribution network fault identification based on a PCCNN model according to claim 3, characterized in that, The When generating full-order polynomial product terms, for the N-dimensional input features of the input layer, polynomial product terms of order 1 to N are generated sequentially. Each order polynomial product term is a combination product of all input features under that order. The number of product terms of order j is the binomial coefficient of selecting j elements from the N input features, where the value of j ranges from 1 to N.
5. The method for power distribution network fault identification based on a PCCNN model according to claim 4, characterized in that, The The layer's output vector is: The q-th element of the j-th order subvector is: in, For the i-th input vector, For the input layer metrics, for The q-th subset, For indicator set The set consisting of all j-element subsets in the set.
6. The method for power distribution network fault identification based on a PCCNN model according to claim 5, characterized in that, The Layers When performing a weighted summation of polynomial product terms of the same order output by the layer, for Each polynomial product term in the layer output is assigned a corresponding trainable weight vector. The output of each order is the weighted sum of all polynomial product terms of that order and their corresponding weight vectors. The final output of the layer is a one-dimensional vector with the same order as the input features.
7. The method for fault identification in a distribution network based on a PCCNN model according to claim 6, characterized in that, The The output of the j-th node in the layer is: in, To select j elements from N input nodes, the binomial coefficients are... To and Dimensionality-matched trainable weight vectors; The output vector of the layer is .
8. The method for fault identification in a distribution network based on a PCCNN model according to claim 7, characterized in that, The Layers When processing the output of the layer, it is Each order result in the layer output is assigned a corresponding constraint weight. The global linear weighted result is the sum of the products of all order results and their corresponding constraint weights, which are then superimposed with the model's bias term before being output to the output layer.
9. The method for power distribution network fault identification based on a PCCNN model according to claim 8, characterized in that, The The layer output is: in, for Layers and Trainable weight vectors between layers For the corresponding The constraint weight values, As the bias term of the model, the final output of the PCCNN model is ),in This is the activation function built into the output layer.