Mountain area power distribution network fault positioning interpretability method based on recurrent neural network
By constructing a fault location model based on a recurrent neural network and training it using the gradient descent algorithm, the problems of low efficiency and insufficient transparency in fault location in mountainous power distribution networks were solved, achieving interpretability and transparency in fault location and reducing the cost of misjudgment.
Patent Information
- Application Number
- CN202510939479.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-08
- Publication Date
- 2025-12-09
AI Technical Summary
Fault location in mountainous power distribution networks relies on manual investigation, which is inefficient and costly. Neural network location results lack theoretical support and transparency, and the cost of misjudgment is high, making it difficult to promote in complex mountainous areas.
A fault location model based on a recurrent neural network is constructed and trained using the gradient descent algorithm. The convergence, stability, monotonicity, and process transparency are verified through interpretability methods to achieve interpretability of the fault location results.
It improves the interpretability and credibility of fault location, ensures the effectiveness, safety and accuracy of the model, reduces the cost of misjudgment, and realizes the transparency and visualization of fault location.
Smart Images
Figure CN121090974A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of power distribution network fault positioning, and particularly relates to an explainability method for mountainous power distribution network fault positioning based on a recurrent neural network. BACKGROUND
[0002] Due to complex terrain and wide distribution of lines, fault positioning of a mountainous power distribution network highly depends on manual investigation, and is low in efficiency and high in cost. In recent years, a neural network technology is introduced into the field of fault positioning, but the "black box" characteristics of the neural network technology lead to a lack of user trust in positioning results. The existing method has the following problems: the neural network positioning result lacks theoretical support, is prone to misjudgment, and is high in trial and error cost; the training process is not transparent, and the rationality of weight value updating and error change cannot be verified; mathematical proof of model convergence, stability and monotonicity is lacked, and it is difficult to popularize to a complex mountainous scene. Once positioning is wrong, the whole line needs to be re-investigated, and the trial and error cost is huge. SUMMARY
[0003] In view of the defects of the prior art, the application provides an explainability method for mountainous power distribution network fault positioning based on a recurrent neural network, which has the advantages of positioning faults and the like, and solves the above technical problems.
[0004] To achieve the above purpose, the application provides the following technical scheme: an explainability method for mountainous power distribution network fault positioning based on a recurrent neural network, comprising the following steps: S1: constructing a fault positioning model based on a recurrent neural network; S2: inputting line real-time data into the fault positioning model, and outputting a fault positioning result; S3: verifying the fault positioning result by using an explainability method.
[0005] As a preferred technical scheme of the application, the fault positioning model based on the recurrent neural network comprises an input layer, a first hidden layer, a second hidden layer, a third hidden layer, an output layer and a recursive link; The output of the first hidden layer is: wherein, is a weight value vector between the nth node of the first hidden layer and the input layer, represents the number of input elements (real-time data related to power distribution network line faults), represents the number of time iterations, represents a network input value, which contains normal input values and output values of the last iteration, This represents the number of nodes in the inputlayer.
[0006] The network connection combination between the first hidden layer and the second hidden layer is specifically as follows: Among them, record This is the set of connections between the first hidden layer and the q-th neuron in the second hidden layer r. This is the set of connections between the second hidden layer and the nth neuron in the first hidden layer. and Represent layer and The number of nodes in a layer. For a random set. ,remember for The number of elements, , , , They represent The number of elements, This is the set of connections between the first hidden layer and the q-th neuron in the second hidden layer. This is the set of connections between the nth neuron in the first hidden layer and the entire second hidden layer.
[0007] The output structure of the second hidden layer is as follows: in, This refers to the output structure of the network. The network output of the third hidden layer is: in, This is the activation function.
[0008] As a preferred embodiment of the present invention, S1 further includes a loop-based method. The training of the neural network fault location model involves using the gradient descent algorithm, with the following update iteration rules: in, It is the squared error function. This represents the change in weights at the k-th iteration. This indicates that the error function is effective for the k-th iteration. The change in weight, This indicates that the error function is effective for the k-th iteration. The change in weight, The initial weights represent the weights. This refers to the learning rate during the training process.
[0009] As a preferred technical solution of the present invention, the interpretability verification method for the fault location result in S3 includes convergence interpretability, stability interpretability, monotonicity interpretability, process interpretability, and result solvability.
[0010] Compared with existing technologies, this invention provides an interpretable method for fault location in mountainous power distribution networks based on recurrent neural networks, which has the following advantages: This invention improves the acceptability of fault location based on neural network methods by making the location model, location process, and location results interpretable. It makes the convergence of the location model interpretable, proving that the fault location model is effective. It makes the stability of the location model interpretable, proving that the fault location model is safe. It makes the monotonicity of the location model interpretable, proving that the fault location model can definitely find the fault location. It makes the process of the location model interpretable, realizes the transparent display of the entire fault location process, and realizes the display of the calculation results of the fault location, breaking the black box phenomenon. Attached Figure Description
[0011] Fig. 1 For the cycle of this invention Neural network diagram; Fig. 2 This is a visual diagram illustrating the entire fault location process of the present invention. Detailed Implementation
[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0013] Please see Figs. 1-2 A loop-based An interpretable method for fault location in mountainous power distribution networks using neural networks first constructs a recurrent... A neural network fault location model takes real-time line data related to the fault as its input. The network structure includes an input layer, a first hidden layer (summation layer), a second hidden layer (productive layer), a third hidden layer (summation layer), an output layer, and a recursive link. The model is then evaluated for feature interpretability: convergence demonstrates its effectiveness, stability ensures its safety, and monotonicity guarantees its ability to locate the fault. Furthermore, the process is made interpretable through visualization technology, providing a transparent view of the entire fault location process, including data input, weight updates, hidden layer calculations, and result output. Finally, the result is interpreted by using the network output formula, starting from iteration t and continuing until the end, to demonstrate the interpretability of the result. Build based on loop The neural network fault location model consists of six parts: the first part is the input layer, the second part is the first hidden layer, the third part is the second hidden layer, the fourth part is the third hidden layer, the fifth part is the output layer, and the sixth part is the recursive link (i.e., the dotted line part in the diagram), which loops... The output of the neural network is in the form of The sum of, i.e. ,in, Represents input element The number of (real-time data related to power distribution line faults) Represents the number of time iterations. This means that the network input values include both normal input values. It also includes the output value of the previous iteration. . , and They represent inputlayer, layer and The number of nodes in the layer, and The activation function representing the network being trained, and the network structure is as follows: Fig. 1 As shown: Therefore, the overall input value of the network can be expressed as: ,in Obviously, when hour, In other words, during the first iteration, the network's input elements only contain Starting from the second iteration, recursively process the elements. Newly joined .
[0014] remember Let be the weight vector between the third hidden layer and the second hidden layer, where This represents the number of nodes in the second hidden layer. This is the weight vector between the nth node of the first hidden layer and the input layer. The connection weight between the first and second hidden layers is fixed at 1. For the first hidden layer, let's call it... For variables in the first hidden layer, Let be the number of neurons in the input layer. Therefore, the output of the first hidden layer can be expressed as: The connection state between the first and second hidden layers is dynamically changing. Neurons can achieve full or partial connections. In the case of full connection, the number of connection combinations between neurons is... When partial connections are present, the number of connection combinations of neurons is less than [a certain value]. .remember This is the set of connections between the first hidden layer and the q-th neuron in the second hidden layer r. This is the set of connections between the second hidden layer and the nth neuron in the first hidden layer. For a random set... ,remember for The number of elements allows us to determine the network connection combination between the first and second hidden layers. For the second hidden layer, let's call it... Given the output structure of the network, then: For the third hidden layer, its output value can be expressed as: in, Activation function .
[0015] cycle The entire training and weight update of the neural network uses the gradient descent algorithm, targeting the weight sequence. The update and iteration rules are as follows: in, It is the squared error function. This represents the change in weights at the k-th iteration. This indicates that the error function is effective for the k-th iteration. The change in weight, This indicates that the error function is effective for the k-th iteration. The change in weight, The initial weights represent the weights. The learning rate during the training process, in the loop In neural network models, It is a fixed value.
[0016] Features can be explained: Then, for a fault location model based on a neural network, to assess whether the model is reasonable, effective, feasible, and has generalization value, it is necessary to provide interpretable proofs of the fault location model from aspects such as convergence, stability, and monotonicity, so as to improve the acceptability of the model.
[0017] (1) Convergence demonstrates that the fault location model is effective (convergence proof) (a) The gradient of the error function converges infinitely close to 0: loop When training a neural network using the gradient algorithm, the gradient of the weights needs to be calculated during the iteration process. When the number of iterations is large enough, the gradient value is expected to approach 0.
[0018] Based on certain valid assumptions and a series of formula derivations, we hope to arrive at the following conclusion: (b) Weights converge to the optimal value: Use gradient algorithm to train the loop. The ultimate goal of neural networks is to find an optimal weight that minimizes the error function. In other words, when the number of iterations is large enough, the weight updates are expected to eventually converge to the optimal weight.
[0019] Based on certain valid assumptions and a series of formula derivations, we hope to derive the following conclusion: (2) Stability ensures that the fault location model is safe (stability proof). cycle Neural networks are recurrent neural networks, possessing the property of time sharing. During training, from the perspective of dynamic stability, it is necessary to construct the energy function of the network, hoping that the derivative of this energy function with respect to time is equal to 0, i.e., cyclical. Given the energy function EN(t) of the neural network model, and combining it with the dynamic system equations, we aim to derive the following by differentiating the energy function at t: When t approaches infinity ; (3) Monotonicity ensures that the fault location model can find the fault location (proof of monotonicity). cycle When training a neural network using the gradient algorithm, an error function needs to be constructed, and during the iteration process, it is expected that the error function is monotonically decreasing.
[0020] Based on certain assumptions, and using the expansion of Taylor's formula and the Cauchy-Schwartz inequality, we hope to derive the following conclusion: The process can be explained: Throughout the fault localization process, visualization is used to transparently display everything from data input, weight updates, hidden layer calculations to the final network output and fault location determination. Fig. 2 As shown; The result can be explained as follows: The process of obtaining the result by outputting the formula through the network, with the iteration number t starting from 1 and ending at the end, can be explained as follows: in, Represents the number of time iterations. Represents network input values (including normal input values) It also includes the output value of the previous iteration. ), , and They represent inputlayer, layer and The number of nodes in the layer, and Activation function representing the training network .
[0021] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. An interpretable method for fault location in mountainous power distribution networks based on recurrent neural networks, characterized in that: Includes the following steps: S1: Constructing a loop-based system A fault location model using neural networks; S2: Input the real-time data of the line into the fault location model and output the fault location result; S3: Verify the interpretability of the fault location results using a method.
2. The interpretability method for fault location in mountainous power distribution networks based on recurrent neural networks according to claim 1, characterized in that: The cycle-based The fault localization model of a neural network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, an output layer, and a recursive path; The output of the first hidden layer is: in, Let be the weight vector between the nth node of the first hidden layer and the input layer. Represents input element The number of Represents the number of time iterations. Represents the network input value, i.e. , Represents the number of nodes in the inputlayer; The network connection combination between the first hidden layer and the second hidden layer is specifically as follows: Among them, record This is the set of connections between the first hidden layer and the q-th neuron in the second hidden layer r. This is the set of connections between the second hidden layer and the nth neuron in the first hidden layer. and Represent layer and The number of nodes in a layer and These represent the network connection combinations between the first hidden layer and the second hidden layer, and the network connection combinations between the second hidden layer and the first hidden layer, respectively. , They represent , The number of elements, This is the set of connections between the first hidden layer and the q-th neuron in the second hidden layer. This is the set of connections between the entire second hidden layer and the nth neuron in the first hidden layer; The output structure of the second hidden layer is as follows: In this configuration, the connection weight between the first hidden layer and the second hidden layer is fixed at 1. This is the output structure of the second hidden layer. This represents the output of the q-th neuron in the second hidden layer. This represents the value of the i-th node in the first hidden layer. This means when i belongs to When the set is executed, the output of the first hidden layer is multiplied together to obtain the value. The network output of the third hidden layer is: in, Indicates activation function , This represents the network output of the third hidden layer at the t-th iteration. This is the weight vector between the third hidden layer and the second hidden layer. , This represents the output value of the second hidden layer, specifically... , This represents the weight between the q-th node of the third hidden layer and the second hidden layer, where q∈[1,Q]. This represents the output value of the q-th node in the second hidden layer. This represents the output value of the third hidden layer when it has not undergone an activation function.
3. The interpretability method for fault location in mountainous power distribution networks based on recurrent neural networks according to claim 1, characterized in that: S1 also includes loop-based The training of the neural network fault location model involves using the gradient descent algorithm, with the following update iteration rules: in, The squared error function, This represents the change in weights at the k-th iteration. This indicates that the error function is effective for the k-th iteration. The change in weight, This indicates that the error function is effective for the k-th iteration. The change in weight, This represents the initial weight values before any iterative calculations occur. This refers to the learning rate during the training process.
4. The interpretability method for fault location in mountainous power distribution networks based on recurrent neural networks according to claim 1, characterized in that: The interpretability verification method for fault location results in S3 includes convergence interpretability, stability interpretability, monotonicity interpretability, process interpretability, and result solvability.