(Department of Electrical Engineering, Southeast University, Nanjing 2 10096)
(Department of Electrical and Electronic Information Engineering, City University, UK EC 1VOHB)? In this paper, a neural network differential protection method based on conventional protection principle is proposed. Neural network has excellent pattern recognition ability. Firstly, it is analyzed theoretically that the conventional ratio braking characteristics in differential protection can be realized by a single neuron, and the ratio braking characteristics differential protection composed of a single neuron sensor can be realized. On this basis, a multi-layer neural network differential protection model with nonlinear braking characteristics is further proposed. This method corresponds the setting value in traditional protection with the weight coefficient in artificial neural network structure, and obtains the best protection scheme by using learning algorithm. Operators can choose training results according to their experience, which completely solves the problem of poor reliability after applying ANN technology to relay protection engineering practice and is easy to be accepted by field operators. This method has important engineering and theoretical significance. ? Generator transformer differential protection; Artificial neural network; Perceptron; Digital protection; Internal fault; 1 Introduction Power system relay protection is a comprehensive subject. With the development of digital relay protection technology, various new principles and technologies have been widely used in digital protection. Artificial neural network has the ability of self-learning and self-adaptation, and the application of artificial neural network in relay protection has become a research hotspot. If neural network technology is applied to real-time control fields, such as relay protection, it requires high reliability and is difficult to put into practical application. The difficulty mainly lies in the lack of normalization ability of neural network, that is, the trained network cannot guarantee the reliability of 100%. However, from the papers published by scholars at home and abroad, a large number of research results have emerged, such as distance protection, fault classification, fault location using artificial neural network, etc. [1 2], and achieved good results; It is rare to apply neural network to generator protection. For example, reference [3] puts forward generator stator winding protection based on differential principle, which adopts two neural networks. The first one is used for fault detection, which detects whether the generator is in normal, external or internal fault state. Another neural network is used for fault phase selection. The input of neural network is five continuous sampling values of generator terminal, neutral point and rotor winding current. Reference [4] puts forward that the calculated sampling values of action current and braking current are used as the input of neural network, and the training time of the network is shortened by genetic algorithm, which has achieved good results. ? The generator differential protection realized by different principles shows different behaviors when the power system fails, especially when the power system is short-circuited, the system is transient and CT is saturated. Even if the same principle is adopted, such as the differential protection of the ratio braking generator, if different digital filters and different algorithms are adopted, the performance of the algorithm will be different when it fails because of the different suppression effects on harmonic and DC components, which is why the correct operation rate of power system protection should be 100% in theory, but it is not high in practice [5]. Because the basic requirement of relay protection for protection devices is reliability, with the increasing installed capacity of power system, the reliability requirements for devices are getting higher and higher. Because the operation of power system is ever-changing, the existing experience can not cover all possible faults, and artificial neural network is based on the principle of self-learning and self-adaptation on the basis of previous experience learning, so the operators of power system often dare not put it into actual system operation, and the setting method after applying artificial neural network technology to protection is different from previous experience. It is impossible to complete the setting process of protection by training a neural network with arbitrary structure. How to ensure that the trained network must be reliable? Therefore, the key problem of applying neural network to practical protection is how to ensure its reliability. Based on the above problems, this paper puts forward an artificial neural network differential protection scheme based on the conventional protection principle, which fully combines the conventional ratio differential protection principle with the artificial neural network principle, and completely solves the concern of field personnel about the reliability of neural network. Because the protection is based on the traditional differential protection principle, the trained network will not be worse than the traditional protection in reliability, and it is easy to realize in engineering. ? Realization and analysis of single neuron ratio braking characteristics? 2. 1 ratio braking characteristic differential protection and single neuron realization principle The ratio braking characteristic principle is an improvement on the traditional protection principle in digital protection. Its operating current is not fixed, but increases with the increase of external short-circuit current, so it can ensure the protection from misoperation when external short-circuit occurs, and has high sensitivity to internal short-circuit [6, 7]. Its action characteristics are shown in figure 1. ? As can be seen from the figure (1), it consists of three parts: no braking area, ratio braking area and quick break area. When the braking current is less than the inflection point current ig, the action current is a constant Iq starting current; When the braking current is greater than the inflection point current, the action current increases along a straight line with the increase of braking current; When the action current is greater than the differential quick-break current, it reflects that the fault situation is serious and the protection will not delay the action. The action equation is as follows: when Iz≤Ig (1), no braking zone means Id≥Iq? Ratio braking area, that is, Id≥Ks(Iz-Ig)+Iq, when Iz≥Ig (2)? Fast break zone Id≥Isd? (3)? Take the direction flowing to the generator as the positive current direction, the differential current is ID = | in+it |, and the braking current is in-it |, in which the secondary CT current at the generator neutral point is IN, the secondary CT current at the machine end is it, the curve inflection point current is Ig, the curve starting current is Iq, and the curve slope is KS. Relative to the positive direction of the current flowing into the generator. From the analysis of the above action equation, it can be concluded that the equation actually realizes the function of a pattern classifier, and divides the two-dimensional plane space composed of Id and (Iz-Ig) into two categories: action area and non-action area. Because of the strong pattern recognition ability of artificial neural network, the above classification characteristic curve can be realized by a single neuron perceptron. In the actual protection setting, the slope KS of the ratio braking curve is generally determined according to the experience of protection operation. After introducing artificial neural network technology into differential protection, the previous setting experience will continue to play a role. For example, the value of empirical KS can be used as the initial value of the weight coefficient of a single neuron. In this example, the classifier realized by a single neuron is realized and trained, and the ratio braking part of differential characteristics is obtained. ? According to the setting principle of generator differential protection, it is necessary to set the curve inflection point current Ig and curve starting current Iq of differential protection. In this experiment, the setting principle of conventional differential protection is adopted to fix two parameters, Ig and Iq, and the action current and the connection weight of neurons are set to be constant 1. At this time, the parameter of a single neuron is only the curve slope Ks of the ratio braking part, and then the optimal slope Ks of the ratio braking curve is found according to the training method of artificial neural network. As shown in Figure 2, the output equation of the sensor is the equation of the ratio braking part of the differential protection. Further research should also be that Ig and Iq are also obtained through the training process of artificial neural network. The network adopts multi-layer structure and the transfer function is nonlinear, so as to get the best approximation of the unbalanced current characteristic curve caused by the error of secondary circuit equipment such as CT and get the nonlinear braking performance. ? According to the actual field data and simulation data, the training samples are formed, and the single neuron network is trained by a certain learning algorithm, so that the network weight Ks converges to an appropriate value. From the field measurement data and simulation data, we can get I? D, (Iz-Ig) (denoted as Iz 1) and the output action signal are used as sample pairs, which constitute the sample space of a single neural network. ? 2.2 Perceptron Differential Protection Training Algorithm Based on Conventional Protection Principle As can be seen from Figure 2, the weighted sum of the inputs of neurons is:? The transfer function f is taken as a step function, and the output value of neurons is:? The error function of a single neuron is defined as: where n is the number of samples; K stands for sample. ? It can be seen from the above formula that the error of a single neural network is only a function of the ratio braking coefficient when the sample set is determined. The process of neural network training is to adjust the weight coefficient Ks through sample learning to minimize the sum of squares function of errors. ? According to the learning rules of perceptron network, we can get [8]:? 1) Given the initial k? S value, according to the setting experience of conventional protection, the initial Ks can be 0.5; ? 2) input a sample X=(IdIz 1) and its expected output t (tutor signal, if X∈ internal fault, t= 1, if X∈ external fault or normal situation, t = 0); ? 3) Calculate the actual output of the neuron according to formulas (4) and (5); ? 4) according to formulas (7) and (8), correct k? The value of s, where η is the learning rate; ? 5) Go to 2) Until K? S of all samples are stable. ? Training and simulation of a single neural network? 3. 1 learning sample acquisition? Taking the Three Gorges Unit as an example, an actual generator-transformer system is constructed with the following parameters. ? Rated capacity? Sn？ =777.8 ? MVA？ ; ? The rated power factor cosφ= 0.9；; ? Rated line voltage UN=20 kV? ; ? The wiring diagram of the direct-axis unsaturated ultra-instantaneous reactance system with rated current IN=22 453.2A is shown in Figure 3. ? Simulate internal phase-to-phase short-circuit faults of various generators and external phase-to-phase short-circuit faults with different short-circuit reactances respectively. Fault current passes through twice? CT？ After transmission, the currents at neutral point and machine end are obtained, and then the data values of braking current and action current are calculated. The sample should also include the current data during the normal operation of the generator; Considering the practical operation experience, the data used for setting differential protection in practical engineering should also be put into the sample set as samples; In addition, the current data of the generator's previous fault records should be added to the sample set. ? Simulation of (1) external interphase short circuit? Use it? Matlab？ When the software simulates the external interphase short circuit fault, considering the discreteness of CT ratio, the setting values of the ratio can be roughly equal, but slightly different, so that the steady-state unbalanced process of the secondary circuit can be simulated. Considering the transient unbalanced current, the time constant of setting CT can be slightly different, which can simulate the transient unbalanced process of secondary circuit. At the same time, considering that CT will be saturated when the short-circuit current is too large, CT adopts the magnetization characteristics of two broken lines to simulate CT saturation. ? (2) Collection of internal turn-to-turn short circuit fault data? Internal short-circuit fault data are obtained by multi-loop analysis method compiled by Southeast University. ? Fault condition: the generator is not connected to the system during no-load operation, that is, when the generator is internally short-circuited, the system does not provide short-circuit current to the generator; At the same time, it should also include the internal short-circuit fault data under load condition when the generator is connected to the system; Short circuit fault data of transition resistance. The following is a list of fault conditions in different situations:? 1) Phase A 1 branch and Phase B1branch are short-circuited. The short-circuit point shifts from 1 to 36; Obtaining 36 groups of sample data; ? 2) Short circuit between phase A 1 branch and phase B 1 branch, where A is fixed at neutral point and B moves, and 36 sets of data are obtained; ? 3) With 30% load, there is a short circuit between phase A 1 branch and phase B 1 branch, and the short circuit point is moved from 1 to 36, and 36 sets of data are obtained; ? 4) With 60% load, there is a short circuit between the A-phase 1 branch and the B-phase 1 branch, and the short circuit point is moved from 1 to 36, and 36 sets of data are obtained; ? 5) With 100% load, phase A 1 branch is short-circuited with phase B 1 branch, and the short-circuit point is moved from 1 to 36, and 36 sets of data are obtained; ? 6) When the load is 30%, the resistances of phase A 1 branch and phase B 1 branch near the neutral point are short-circuited, and the resistance is 0.05 j ohm (j= 1… 17). ? Draw external fault and internal fault data on the decision plane, as shown in Figure 4, where the horizontal axis is braking current and the vertical axis is action current. O stands for external fault and x stands for internal fault. ?
From the distribution of sample points in Figure 4, the boundary between internal fault and external fault is obvious, and the internal fault and external fault are linearly separable, and the two areas can be separated by a straight line between the two modes. Judging from the trend of error changing with Ks, the value of Ks is about 0.23. It can be concluded from the training process in Figure 6 that the network converges after a limited number of iterations, and the final optimal value of Ks is 0.234 6. ? 3.2 Selection and analysis of training results? When the neuron training is completed, because the weight coefficient of the network has obvious physical significance, the action characteristics of the differential relay can be drawn according to the training results, and the quality of the network training results can be judged manually according to the obtained action characteristics. If the training results have obvious problems, that is, they are inconsistent with the usual operating experience, they will be rejected. Then carefully analyze the training samples to find out whether there are obvious bad data samples in the samples, so that the samples can be integrated into a linear inseparable problem and the bad samples can be eliminated from the sample set. The weight of neural network corresponds to the setting value of protection, and people can analyze its training results to decide whether to accept or reject it. Only by organically combining the principles of artificial neural network with practical engineering problems can the theoretical knowledge of neural network be applied to practical engineering. In principle, artificial neural network is an idea of structural interconnection, which is dominated by some mathematical optimization methods. We can't just get the nonlinear mapping of input and output through a neural network with arbitrary structure, and then we can get the solution of this problem. That is to say, we should untie the internal structure of neural network, give the weights of neural network a certain physical meaning, and correspond the parameters in our usual problem-solving methods with the weights of neural network, so that the acquisition of these parameters can be attributed to a training process and can be expanded and extended, so that we can make full use of the structure of neural network and its learning and training methods. ? The transfer function is a step function, a linear function or a nonlinear function, but different braking curves can be obtained by using different transfer functions, which can be obtained through mathematical analysis. The braking curve of single neuron of step function is linear. After adopting the neural network with the above structure, the weights and thresholds of the network are endowed with certain physical significance, which makes the differential protection architecture based on the neural network principle of conventional differential protection easy to be accepted by field personnel. It can be seen that even if neural network is not trained, the initial value set by experience can ensure the normal operation of neural network differential protection, because it is actually a conventional differential protection with proportional braking characteristics. Once the neural network training converges, the value of Ks will be more in line with the actual situation. ? If the structure of neural network is further improved and the number of hidden nodes is increased, the nonlinear braking curve can be obtained, which makes the braking curve closer to the unbalanced current of differential circuit. As we all know, neurons in the hidden layer of neural network represent a decision boundary. If we increase the number of hidden layer connections, it is similar to generating a piecewise linear ratio braking curve. The decision boundary of the trained network is closer to the error curve of CT secondary loop. In engineering, differential protection with proportional braking characteristics generally adopts two-stage characteristic curve, but some of them adopt three-stage characteristic curve. The third disconnection has a certain effect on preventing differential protection misoperation caused by ct saturation, and various braking curves can prevent CT? The misoperation effect of differential protection caused by saturation is different. According to the above conclusions, it can be seen that if the nonlinear braking curve is realized by artificial neural network, the performance of differential protection will be improved more obviously. ? Multi-layer neural network realizes differential protection of generator with nonlinear braking characteristics. The above single neuron scheme only uses artificial neural network to realize differential protection of ratio braking characteristics, and the neural network has very good nonlinear mapping ability. If neural network is only used in this way, it will lose its practical engineering significance, and at best, it will only use the learning algorithm of artificial neural network to obtain the best protection of ratio braking characteristics. It is proved that the combination of artificial neural network and differential protection is feasible, and the excellent characteristics of neural network are not fully utilized, so further research is to use multi-layer artificial neural network to realize differential protection of nonlinear braking characteristics. ? From the analysis of a single neural network, it can be seen that adding a hidden neuron is equivalent to adding a classification boundary, and changing the nodes of neurons into nonlinear transfer functions is equivalent to changing classification straight lines into classification curves, which can enhance the expression ability of neural networks. As shown in figure 7, a three-layer neural network differential protection is realized, in which the first neuron mainly represents the non-braking area; The second intermediate neuron mainly represents the corresponding braking area; The third neuron mainly represents the corresponding fast break zone. Give the weight the initial value as shown in the figure, and add inequality constraints to limit the weight to a certain area. From the initial values set in Figure 7, a nonlinear braking curve can be obtained. Different from the ratio braking curve, it is smooth and derivable everywhere, and more practical nonlinear braking characteristics can be obtained through actual sample training. ? The learning and training algorithm of the network adopts BP algorithm, but because some weights in the network represent certain physical meanings, some weights are fixed, and the corresponding weights have a certain range of values, they are constrained by practical experience. How to add these inequality constraints to the defined error function and the selection of network transfer function need further study, so the general BP algorithm needs to be improved before it can be used. In the future, further research will be mainly carried out in these areas. ? 5 conclusion? Firstly, this paper analyzes the feasibility of using a single neural network to realize longitudinal differential protection with proportional braking characteristics, and carries out simulation experiments. From the simulation results, it can be seen that it is feasible to realize the ratio braking characteristics of differential protection by using single neuron sensor, and the best ratio braking characteristics can be obtained. The analysis shows that the specific parameters in the traditional protection principle correspond to the weights in the artificial neural network, and the weights of the neural network are endowed with certain physical significance. The excellent training algorithm of neural network is used to obtain the best weight coefficient and the setting parameters in traditional protection. Protection operators can decide the selection of parameters according to their own experience, which completely solves the problem that the reliability of artificial neural network in protection can not be guaranteed. On this basis, the principle of differential protection with nonlinear braking characteristics based on multi-layer neural network is proposed and preliminarily analyzed. Neural network differential protection based on traditional protection principle organically combines artificial neural network technology with practical protection, and its principle is simple to realize and easy to be accepted by field operators. This scheme has important engineering and theoretical significance.