Animal fault analysis method in overhead distribution system based on wavelet decomposition and neural network

Abstract

An animal fault analysis method for overhead distribution systems based on wavelet decomposition and neural networks includes identifying features of animal-caused outages, including animal life cycle, weather clear conditions; using discrete wavelet transform to decompose the time series of animal-caused outages each week into two components, and constructing a separate neural network for each decomposed coefficient series; combining the outputs of the neural networks to obtain an estimate of animal-caused outages each week according to wavelet reconstruction techniques.

Description

TECHNICAL FIELD

[0001] The present application relates to an animal fault analysis method in overhead distribution system based on wavelet decomposition and neural networks, belonging to the field of distribution system. BACKGROUND

[0002] The reliability of distribution systems is largely affected by different environmental factors, and since animals cause a large number of outages in overhead distribution systems, it is very important to investigate these outages. The objective is to estimate the animal-caused outages in a specific year using historical data from previous years. Although animals cause serious faults, there are few studies and the existing techniques have methods based on neural networks that identify animal-caused outages from a series of illnesses caused by different factors, but their methods do not solve the problems specific to our objective. SUMMARY

[0003] In order to overcome the deficiencies of the prior art described above, the present application proposes an animal fault analysis method in overhead distribution system based on wavelet decomposition and neural networks. The time series of animal-caused outages per week are decomposed into two components using discrete wavelet transform, and a separate neural network is constructed for each decomposed coefficient series. The outputs of the neural networks are combined to obtain the estimate of animal-caused outages per week according to wavelet reconstruction techniques.

[0004] The technical solution of the present application is as follows:

[0005] An animal fault analysis method in overhead distribution system based on wavelet decomposition and neural networks, comprising the following steps:

[0006] Step 1) Identify the characteristics of animal-caused outages, including the life cycle of animals, sunny weather conditions;

[0007] The examination of outage data from utility companies shows that animal-related outages occur randomly, with an increased likelihood of outages in certain months due to the life cycle of animals, and an increased likelihood of outages in sunny weather conditions due to increased animal activity. Due to the randomness of outages, it is not practical to accurately predict the number of animal-related outages in a short duration (e.g. specific hours or a day) or in a specific overhead line. Spatial or temporal aggregation of data is required to obtain meaningful patterns. A very short duration or a very small area will result in a larger statistical range, resulting in lower confidence in the results. Aggregation over very long durations can result in loss of data information. Similarly, spatial aggregation over very large territories can introduce errors due to spatial variations in weather. Almost all animal-caused outages occur on sunny days [days when the temperature is between 40 and 85 degrees and there are no other weather-related activities. The number of sunny days per week is used as another input to the model.

[0008] Step 2) The time series of weekly animal-induced interruptions is decomposed into two components using the discrete wavelet transform and a separate neural network is constructed for each of the decomposed coefficient series

[0009] The formula is as follows:

[0010] ψ j，k = 2 -j / 2 ψ(2 -j t k)j, k e Z (6)

[0011] where the normalizing factor makes the basis and mother functions have the same energy.

[0012] Step 3) The decomposition formula is as follows:

[0013]

[0014] Step 4) The wavelet coefficient neural networks are proposed, including low frequency information and high frequency information;

[0015] Step 5) A hybrid approach based on artificial immune system is adopted

[0016] At each time step in the prediction phase, the neural networks in the library are ranked according to their affinities. Then a number of clones of each network is obtained according to the ranking of the networks. The neural network with the highest affinity produces four clones, while the networks ranked second, third and fourth produce three, two and one clone, respectively. The fifth network produces two clones, and the rest of the clones only one each. Each clone has the same weights as the original network.

[0017] The clones are then subjected to hypermutation. In this step, the weights are altered by adding small random perturbations. If the set of weights of any clone is denoted by a matrix W, then hypermutation is implemented as

[0018] W = W + δ · U(-1,1) (10)

[0019] In the above equation, the quantity δ is a constant related to hypermutation and U(-1,1) is a matrix of the same size as the random numbers uniformly distributed in the interval [1,1]. To further improve the clones, the weights are trained using the regular gradient descent algorithm as follows

[0020]

[0021] where a constant (called the learning rate) is the gradient operator with respect to the weight matrix, and E is the squared error. For each clone, the standard backpropagation algorithm is applied iteratively ten times, after the hypermutation and backpropagation steps, the affinity of the cloned neural network is evaluated and inserted in the library. Only the best ten networks are kept for the next step, while the rest are discarded.

[0022] The present invention has the advantage over the prior art that the time series of animal-caused outages per week is decomposed into two components using discrete wavelet transform and a separate neural network is constructed for each of the decomposed coefficient series. The outputs of the neural networks are combined according to wavelet reconstruction techniques to obtain an estimate of the animal-caused outages per week. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is the neural network graph for low frequency information and b is the neural network graph for high frequency information; DETAILED DESCRIPTION

[0024] The present invention will be described in detail below with reference to the attached drawings, which should be considered in conjunction with the following detailed description. It should be emphasized that the following description is merely exemplary and is not intended to limit the scope of the invention or its applications.

[0025] A method for animal fault analysis in overhead distribution systems based on wavelet decomposition and neural networks, comprising the following steps:

[0026] Step 1) Identify the characteristics of animal-caused outages, including the life cycle of the animal, the weather conditions;

[0027] An examination of outage data for a utility company shows that animal-related outages are random, with an increased likelihood of outages in certain months due to the life cycle of the animal and an increased likelihood of outages in sunny weather conditions due to increased animal activity. Due to the randomness of outages, it is not practical to accurately predict the number of animal-related outages in a short duration (e.g., a specific hour or day) or in a specific overhead line. Spatial or temporal aggregation of data is required to obtain meaningful patterns. Aggregation over very short durations or very small areas will result in a large statistical range, leading to lower confidence in the results. Aggregation over very long durations can result in loss of data information. Similarly, spatial aggregation over very large territories can introduce errors due to spatial variations in weather. Almost all animal-caused outages occur on sunny days [days when the temperature is between 40 and 85 degrees and there are no other weather-related activities. The number of sunny days per week is used as another input to the model.

[0028] Step 2) Decompose the time series of animal-caused outages per week into two components using discrete wavelet transform and construct a separate neural network for each of the decomposed coefficient series, according to the following formulas:

[0029] ψ j，k = 2 -j / 2 ψ(2 -j t k)j, k e Z (6)

[0030] where the normalization factor makes the basis and mother functions have the same energy.

[0031] Step 3) The decomposition filter formula is as follows:

[0032]

[0033] Step 4) Wavelet coefficient neural networks are proposed, including low frequency information and high frequency information;

[0034] Step 5) A hybrid method based on artificial immune system is adopted

[0035] In each time step of the prediction phase, the neural networks in the library are ranked according to their affinities. Then, a number of clones of each network is obtained according to the ranking of the networks. The neural network with the highest affinity generates four clones, while the second, third and fourth ranked networks generate three, two and one clone, respectively. The fifth network generates two clones, and the rest of the clones only one each. The weights of each clone are the same as the original network.

[0036] Then, the clones are subjected to hypermutation. In this step, the weights are changed by adding a small random perturbation. If the set of weights of any clone is denoted as a matrix, then the hypermutation is implemented as

[0037] W = W + δ · U(-1, 1). (10)

[0038] In the above equation, the quantity is a constant related to the hypermutation, and U is a matrix of the same size as the random numbers uniformly distributed in the interval [1, 1]. To further improve the clones, the weights are trained using the conventional gradient descent algorithm as follows

[0039]

[0040] where a constant (called the learning rate) is the gradient operator with respect to the weight matrix, and E is the squared error. For each clone, the standard backpropagation algorithm is applied iteratively ten times, and after the hypermutation and backpropagation steps, the affinity of the neural network of the clone is evaluated and inserted into the library. Only the best ten networks are kept for the next step, while the rest are discarded.

Claims

1. A method for animal-caused outage analysis in overhead distribution system based on wavelet decomposition and neural networks, comprising the following steps: Step 1) identifying the characteristics of animal-caused outages, including the life cycle of animals, sunny weather conditions; Step 2) decompose the time series of each week animal caused interruption into low frequency components and high frequency components by using discrete wavelet transform, and construct low frequency information neural network and high frequency information neural network for each decomposed coefficient sequence respectively; wherein, The decomposition formula is: The transformation formula is: ψ j，k = 2 -j / 2 ψ(2 -j t k)j,k∈Z (6) where the normalization factor makes the basis and mother functions have the same energy; Step 3) neural network optimization method based on artificial immune system: 3.1 In each time step in the prediction phase, sort the neural networks in the library according to their affinities, then obtain many clones of each network according to the ranking of the networks: The top-ranked neural network with the highest affinity produces four clones; The second, third and fourth-ranked networks produce three, two and one clone, respectively; The fifth-ranked network produces two clones; The remaining networks each generate one clone; Each clone has the same weights as the original network; 3.2 Perform hypermutation on the clones: Change the weights by adding a small random perturbation, if any clone's weight set is denoted as a matrix, then hypermutation is implemented as W = W + δ · U(-1, 1). (10) where the quantity is a constant related to hypermutation, and is a matrix of the same size as the random number uniformly distributed in the interval [1, 1]; 3.3 Further improve the clones, train the weights using the conventional gradient descent algorithm as follows: where a constant, called the learning rate, is the gradient operator with respect to the weight matrix, and E is the squared error; 3.4 For each clone, apply the standard backpropagation algorithm iteratively ten times, after the hypermutation and backpropagation steps, evaluate the affinity of the neural network of the clone and insert it into the library, only the best ten networks are retained for the next step, the rest are discarded; Step 4) combine the outputs of the optimized low-frequency information neural network and high-frequency information neural network through wavelet reconstruction, output the estimated results of the weekly outages caused by animals.

Images