A single-phase ground fault intelligent wire selection method based on an ant colony algorithm

For the single-phase ground fault, which has the highest probability of occurrence in a 10 kV distribution network, the ant colony algorithm is combined with a neural network for fault line selection, and the algorithm is optimized. The study builds a 10 kV distribution network model in MATLAB and simulates the neutral ground fault. It compares the results of the optimized line selection algorithm, and the traditional line selection algorithm proves that the accuracy of the method in identifying faulty lines is significantly better than that of the traditional method.


Introduction
The single-phase grounding fault selection problem of a small current grounding system has not been effectively solved because of the weak fault current signal, arc instability, and the influence of random factors during small current grounding faults.Over the years, experts and scholars at home and abroad have conducted a lot of research and proposed many wire selection methods based on different principles [1] .However, there is a general problem of low accuracy, limited applicability, and susceptibility to factors such as transition resistance and fault phase angle of the single criterion wire selection method.
With the development of artificial intelligence theory, artificial intelligence algorithms are gradually applied to fault line selection.The neural network, with its good parallel processing capability, adaptive learning capability, and associative memory capability, intelligently mix various single-criteria routing methods, is used to select fault lines.However, it also has the disadvantages, such as the tendency to fall into local minima when solving, which leads to a long time and low reliability of line selection [2][3][4] .The ant colony algorithm has strong global optimization capability and distributed computing capability.Combining the ant colony algorithm with neural networks for fault line selection makes up for the shortcomings of neural networks and improves the reliability of neural network routing methods.

Design of parameters of the neural network
The system has five 10 kV feeders and one busbar.The zero-sequence voltage and zero-sequence current of each line are extracted using the energy method, fifth harmonic method, and wavelet packet analysis, respectively, as the energy eigenvolume, fifth harmonic eigenvolume, and transient eigenvolume.This is used as the input vector to the neural network denoted as X.Then:  x , x , x , x , x , x , x , x , x , x , x , x , x , x , x , x , x , x The neural network uses one hidden layer, but there has no effective method to decide the number of nodes in this layer.If the number of nodes is too many, then it will increase the training time of the network [5] ; if the number of nodes is too few, it will lead to poor fault tolerance and weak generalization ability of the system.Therefore, a number of hidden layer nodes can be selected by referring to the empirical Formula (3): in the formula, i n is number of input layer nodes [6] , o n is the number of output layer nodes, and a is a constant between 1 and 10.

Principle of the line selection method based on an ant colony algorithm optimized neural network
First, the zero sequence voltage and current of each line are extracted, multiplied, and integrated by the energy method.The energy wave line amplitude and phase of each line are obtained.The fifth harmonic method is used to extract the phase and amplitude of the fifth harmonic in the zero sequence current through a fast Fourier transform.The maximum value and direction of each line are obtained through wavelet packet transform for zero-sequence currents.Then, their fault measures are calculated using the corresponding fault measure functions as the input quantity of the BP neural network.Finally, this input is fed into the ant colony neural network model obtained after training, and then the fault routing results are output.The structure of the system based on the ant colony algorithm optimization neural network is shown in Figure 1.The steps in the line selection method based on an ant colony algorithm-optimized neural network are as follows:

Wavelet packet analysis
(1) The fault characteristics of the transient energy, transient component, and fifth harmonic component are extracted from the zero sequence voltage and zero sequence current signals using the energy method, fifth harmonic method, and wavelet packet analysis, respectively.
(2) The fault measures for each of the three fault characteristics are calculated separately to form the sample data.
(3) One part of the sample data is used as the training sample of the neural network, and the other part is used as the test sample.
(4) The ant colony algorithm combined with the LM-BP algorithm is used to train the thresholds and weights of the neural network and obtain a new fault line selection model with optimal parameters after the training is completed.
(5) The trained neural network line selection model is validated using test samples.The final neural network will output the results of the line selection in [-1,1], and the faulty line can be identified according to the following rules: if the fault measurement of the line is greater than 0.5, the line is considered faulty; if the fault measurement of the line is less than -0.5, the line is considered non-faulty; if the fault measurement is in [-0.5,0.5], the line cannot be judged.If the fault measurement of all lines is less than 0, then the line is considered to be a bus fault.

Simulation validation
In order to compare the optimized ant colony algorithm neural network with the traditional simple BP neural network, training weights, and thresholds for the two methods separately, we used simulations with the sample data and compared the mean error and mean squared error [7][8][9] .
This paper uses the SimPowerSystems module library in MATLAB/Simulink to build a model of a 10 kV distribution line.110 kV distribution system generally uses the neutral point that is not grounded and the neutral point through the arc extinguishing coil grounding two ways.The system structure schematic is shown in Figure 2.  The system is a radial network of five feeders, and the system power supply is an infinity.The main transformer is a three-phase double-winding step-down transformer with a voltage rating of 100/10 kV.The capacity is 100 MVA, and the frequency is 50 Hz.
The initial thresholds and weights of the simple BP neural network are generated randomly, using the BP algorithm to train the samples to obtain the trained thresholds and weights [10] .The weights and thresholds of the ant colony neural network have a set of 40 random numbers, training one number at a time.Optimal weights and thresholds are obtained by searching and filtering.The training error was trained to a minimum using the two algorithms respectively, and the mean squared error obtained is shown in Figures 3-4.A set of sample data with inconsistent results of the three line selection methods is used as a test sample to compare the differences in the line selection results obtained from the simulation of the test sample by the two neural networks.This is shown in Table 1.
Table 1.The comparison of test errors of two kinds of neural network L1 L2 L3 L4 L5 Busbar Average error BP Neural Network -0.9993 -0.9999 -1.0000 -0.8162 0.4157 -1.0002 0.1282 Ant colony neural network -1.0000 -1.0000 -1.0000 -1.0000 1.0000 -1.0000 0.001 As can be seen from Table 1, when one of the three line selection methods is a misjudgment, the simple BP neural network training yields a result of 0.4157<0.5, which fails to identify the fault line.In contrast, the ant colony neural network training obtained a result of 1, accurately identifying line L5 as a fault line.
For different single-phase grounding faults in neutral via arc extinguishing coil grounding system, we use different fault lines, fault locations, fault phase angles, and fault grounding resistance of the combination of different cases of ground fault simulation experiments.Table 2 shows the results of line L5, in which 90% of the single-phase ground fault occurred away from the bus, and the grounding resistance is 2000 Ω.

Conclusion
This design combines the ant colony algorithm and BP neural network algorithm.By optimizing them to design an intelligent line selection method for wire selection method for single-phase ground faults distribution networks, it effectively solves the problem of long training time caused by the high dimensional input of the network.By building a 10 kV simulation model in MATLAB for validation, when one of the selection methods has a selection error, the fault line can still be selected correctly.It is shown that the selection method based on an ant colony algorithm optimized neural network is effective for fault selection of neutral point via arc extinguishing coil grounding system and has a high selection accuracy.

Figure 1 .
Figure 1.Structure diagram of fault line detection system based on ant colony algorithm to optimize neural network

Figure 2 .
Figure 2. The diagram of the 10 kV distribution network system

Figure 3 .
Figure 3.The training error curve of the BP neural network

Figure 4 .
Figure 4. Training error profile of optimized neural network based on ant colony algorithm in the formula x 1 , x 2 , x 3 , x 4 , x 5 , x 6 are energy characteristic quantities; x 7 , x 8 , x 9 , x 10 , x 11 , x 12 are fifth harmonic characteristic quantities; x 13 , x 14 , x 15 , x 16 , x 17 , x 18 are transient characteristic quantities.Thus, the neural network input layer has 18 nodes.And the network has to output the fault state

Table 2 .
Simulation results of test samples of arc suppression coil grounded system