Power system relay protection is a comprehensive discipline. Today, with the development of relay protection technology to digital, various new principles and technologies have been widely used in digital protection. Artificial neural network has self-learning and self-adaptive ability. Applying artificial neural network to relay protection field has become a research hotspot. If neural network technology is applied to real-time control field, such as relay protection, it requires relatively high reliability. Sexuality, so it has certain difficulty when it is put into practical application. The difficulty is mainly reflected in the lack of normalization ability of the neural network, that is, the trained network can not guarantee 100% reliability, but it is published by scholars at home and abroad. According to the paper, a lot of research results have emerged, such as artificial neural network to achieve distance protection, fault classification, fault location, etc., and have achieved good results; the application of neural network to generator protection, comparison Rarely, such as the generator stator winding protection based on the differential principle, it uses two neural networks, the first one is used for fault detection, detecting whether the generator is in a normal, external fault or internal fault state; another nerve The network is used for fault selection. The input of the neural network for the continuous 5 samples of the current of the generator and the neutral point and the rotor windings proposes the calculated values ​​of the action current and the braking current as the input of the neural network, and uses the genetic algorithm to Shorten the training time of the network and achieved good results.
Generator differential protection realized by different principles, the performance behavior is very different in the case of power system failure, especially in the power system short-circuit fault, the system is in transient and CT saturation flow is /d=IZv+Irl This is especially true for dynamic currents; even if the same principle is used for differential braking of differential braking generators, if different digital filters and different algorithms are used, the suppression of harmonics and DC components by the algorithm is not the same. The behavior of the fault is different when it is faulty. This is why the protection of the power system should theoretically be 100% correct, but the reason why the correct action rate is not high is that the basic requirement of the protection device for relay protection is reliability. With the increasing installed capacity of the power system, the reliability requirements of the device are also getting higher and higher; because the operation of the power system is ever-changing, the existing experience cannot cover all possible failures, and the artificial neural network is based on the past. Self-learning and adaptive principles based on empirical learning, so the operating personnel of the power system often dare not invest After going to the actual system to operate, and applying the artificial neural network technology to the protection, the setting method is different from the previous experience. The training process cannot be completed by training an arbitrary structure of the neural network. How to ensure the training after training The network must be reliable. Therefore, the key problem in applying neural network to actual protection is how to ensure reliability. Based on the above problems, this paper proposes an artificial neural network differential protection scheme based on the principle of conventional protection. The combination of the conventional ratio differential protection principle and the artificial neural network principle completely solves the problem of the field personnel's reliability to the neural network. Because the protection is based on the traditional differential protection principle, the trained network is Reliability is not worse than traditional protection, and it is easy to obtain from engineering. 2 Single neuron ratio braking characteristics are realized and analyzed. 2.1 Ratio braking characteristics difference protection and single neuron realization principle Ratio braking characteristics The principle is that the traditional protection principle is improved in digital protection, and its action current is not
Rate braking zone and quick-break zone, when the braking current is less than the inflection point current Ig, the operating current is a constant Iq starting current; when the braking current is greater than the inflection point current, the operating current is long along a straight line with the braking current When the operating current is greater than the differential quick-break current, the fault condition is severe, and the protection will operate the exit without delay. The action equation is as follows: the direction of the flow to the generator is the positive direction of the current, the secondary CT current of the generator neutral point is In, the secondary CT current of the machine is IT, the inflection point current of the curve I, the curve The starting current Iq, the slope of the curve Ks is analyzed by the above action equation in the positive direction of the current flowing into the generator. The equation actually implements the function of a mode classifier, which will be /d and (/z-/g) The two-dimensional plane space composed is divided into two categories: an action area and a non-action area. Because the artificial neural network has strong ability in pattern recognition, the above classification characteristic curve can be completely realized by a single neuron perceptron. In actual protection setting, it is generally determined according to the experience of protection operation. The slope Ks of the ratio braking curve, after introducing the artificial neural network technology into the differential protection, the previous tuning experience will continue to play a role, such as the empirical Ks value, as the initial value of the weight coefficient of the single neuron. This example implements and trains a classifier implemented by a single neuron to obtain a ratio braking portion of the differential characteristic.
According to the setting principle of generator differential protection, it is necessary to adjust the curve inflection point current /g of the differential protection, and the starting current /q of the curve. This test first fixes the two parameters of /g/q by the setting principle of conventional differential protection. The action current and the connection weight of the neuron are fixedly set to a constant value 1. At this time, the parameter of the single neuron only has the curve slope Ks of the ratio braking portion, and then the ratio braking is found according to the training method of the artificial neural network. The optimum slope Ks of the curve, as can be seen from the figure, is that the output equation of the perceptron is the equation of the ratio braking portion of the differential protection. Further research should also be that /g and /q are also acquired through the artificial neural network training process, and the network adopts a multi-layer structure, and the transfer function is nonlinear, thereby obtaining the error caused by the secondary circuit equipment such as CT. The best approximation of the unbalanced current characteristic curve, the nonlinear braking performance is obtained. The single neuron with the ratio braking characteristic constitutes the training sample according to the actual collected field data and the simulated data, and the training sample is trained by a certain learning algorithm. A single neuron network that converges the weight Ks of the network to a suitable value. From the field-measured data and the simulated data, a large number of Ig) (recorded as known) and the output action signal are used as the target pair of sample pairs. These sample pairs make up the sample space of the single neuron network.
2.2 Perceptron differential protection training based on the principle of conventional protection It can be seen that the input weighted sum of neurons is: transfer function / taken as a step function, the output value of the neuron: the error function defining a single neuron is: Where N is the number of samples; k represents a sample.
It can be seen from the above formula that when the sample set is determined, the error of the single neuron network is only a function of the ratio braking coefficient. The process of neural network training is to learn the sample, and adjust the weight coefficient Ks so that the squared error function reaches The smallest.
According to the learning rules of the perceptron network, it can be obtained: given the initial Ks value, the initial Ks can be taken as Q5 according to the conventional protection setting experience; output t (instructor signal, if internal fault, t= 1, if external fault or normal situation (t)); (5) Calculate the actual output of the neuron; (8) Correct the value of Ks. 5) Go to 2) until Ks is stable for all samples. Training and simulation of 3 single neural networks 3.1 Learning samples Take the Three Gorges unit as an example to construct an actual generator transformer system. The parameters are as follows.
The actual generator transformer group composition diagram simulates various internal phase-to-phase short-circuit faults of the generator, and the phase-to-phase short-circuit fault currents of different external short-circuit reactances are subjected to secondary CT transmission, and the currents at the neutral point side and the machine end are obtained, and then The calculated data values ​​of the braking current and the operating current should also include the current data during normal operation of the generator; considering the actual operating experience, the data used to adjust the differential protection in the actual project should also be used as a sample. Going to the sample set; further, the current data of the fault recordings of the generators should be added to the sample set.
(1) Simulation of external phase-to-phase short circuit When using Matlab software to simulate the external phase-to-phase short-circuit fault, considering the discreteness of the CT ratio, the ratio of the ratio can be roughly equal, but slightly different, so that the secondary loop can be simulated. The steady-state unbalanced process; considering the transient unbalanced current, the time constant of setting the CT can be slightly different, so that the secondary loop transient unbalance process can be simulated; at the same time, considering that the short-circuit current is too large, the CT will be saturated. CT uses the magnetization characteristics of the two-fold line characteristics to simulate the saturation of CT.
(2) Acquisition of internal inter-turn short-circuit fault data The internal short-circuit fault data is obtained by the multi-loop analysis method compiled by Southeast University: the generator is not connected to the system under no-load operation, and the obstacle is also relatively obvious in the internal External interception fault blis! The physical meaning of the fixed neural network provides the short-circuit current for the differential protection net machine; it should also include the internal short-circuit fault data under load with the generator connected to the system; the fault data with the transition resistance short-circuit is as follows The fault condition under different conditions is: the phase-to-phase short-circuit short-circuit point of the A-phase 1 branch and the B-phase 1 branch is shifted from 1 to 36; 36 sets of sample data are obtained; short-circuit between the A-phase 1 branch and the B-phase 1 branch, A is fixed at the neutral point, B moves, and 36 sets of data are obtained; short circuit, short circuit point is shifted from 1 to 36, and 36 sets of data are obtained; short circuit between, short circuit point is shifted from 1 to 36, and 36 sets of data are obtained; neutral The resistance near the point is short-circuited, the resistance is 0. The error function graph (the horizontal axis is Ks). From the distribution of the sample points, the internal fault and the external 3.2 training result selection and analysis, after the neuron training is completed, because of the network right The value coefficient has obvious physical meaning, so the action characteristics of the differential relay can be drawn according to the training result. According to the obtained action characteristics, the network training result can be manually judged if the training result is good. If there is an obvious problem, that is, it is inconsistent with the normal operating experience, it will refuse to accept; then carefully analyze the training samples to find out whether there are any unfavorable data samples in the sample, so that the sample set becomes a linear inseparable problem, and the bad sample is taken. Removed from the sample set. The weight of the neural network is consistent with the setting of the protection. People can analyze its training results to determine the trade-off. Only by combining the principle of the artificial neural network with the actual engineering problems can the theoretical knowledge of the neural network be applied to the actual situation. Going in the project. In principle, the artificial neural network is a kind of structural interconnection idea. Some optimization methods that govern it or not can not obtain the nonlinear mapping of input and output through a neural network with arbitrary structure, and obtain the problem. The solution, that is, to solve the internal structure of the neural network, give the weight of the neural network a certain physical meaning, and associate the parameters in the method we usually solve the problem with the weights of the neural network, and put these parameters The acquisition comes down to a training process, and can be extended and extended, so that the structure of the neural network and its learning training method can be fully utilized as a step function, or a linear function or a nonlinear function, but Different transfer functions can be obtained after different transfer functions, which can be mathematically analyzed. The braking curve of a single neuron using a step function is linear. After adopting the neural network with the above structure, the weight and threshold of the network are given to the structure above the principle of conventional differential protection, which is easy to be accepted by the field personnel. It can be seen that even if the neural network is not trained, it is set by the experience. The value also guarantees that the neural network differential protection works normally, because in fact it is a differential protection of the conventional ratio braking characteristics. Once the neural network training converges, the value of Ks will be more in line with the actual situation if further Improve the structure of the neural network, add the number of hidden layer nodes, you can get the nonlinear braking curve, make the braking curve closer to the unbalanced current of the differential circuit. We all know that a hidden layer of neurons in the neural network It represents a decision boundary. If we add the number of contacts in the hidden layer, it is similar to the generation of a piecewise linear ratio braking curve. After training, the decision boundary of the network is closer to the error curve engineering of the CT secondary loop. The differential protection of the ratio braking characteristic generally adopts a two-stage characteristic curve, but also adopts a three-stage characteristic curve. The third section of the fold line has a certain effect on preventing the differential protection of the CT saturation, and the different brake curves are different for preventing the differential protection caused by the CT saturation. It can be seen that if we adopt a nonlinear braking curve implemented by an artificial neural network, the performance of the differential protection will be significantly improved.
4 Generator differential protection with nonlinear braking characteristics realized by multi-layer neural network The above single neuron scheme only uses the artificial neural network to realize the differential protection of the ratio braking characteristics. The neural network has a very good nonlinear mapping. The ability to use the neural network only loses the practical engineering significance. At best, it only uses the artificial neural network learning algorithm to obtain the best ratio braking characteristics protection, which proves that the artificial neural network and differential protection are combined. It is feasible and does not use all the excellent characteristics of the neural network, so further research is the differential protection of the nonlinear braking characteristics realized by the multi-layer artificial neural network.
From the analysis of the above single neural network, it can be known that adding a hidden layer of neurons is equivalent to adding a classification boundary, and changing the node of the neuron to a nonlinear transfer function is equivalent to changing the classification line to a classification curve. The network's ability to express problems, such as a three-layer neural network differential protection, the first neuron mainly represents the no-braking zone; the second intermediate neuron mainly represents the corresponding braking zone; the third The neurons mainly represent the corresponding quick-break zone to give the weights the initial values ​​shown in the figure, and add the unequal constraint conditions to limit the weights to a certain area. The learning training of the nonlinear braking curve network of the multi-layer neural network is illustrated. The algorithm uses the BP algorithm, but because some weights represent a certain physical meaning in the network, some weights are fixed, and the corresponding weights have a certain range of values, subject to practical experience. How to add these inequality constraints to the defined error function, and the selection of the transfer function of the network needs further research, so it needs to be The improved BP algorithm can be used to further explore these aspects in the future. 5 Conclusion In this paper, the feasibility of the differential protection of the ratio braking characteristic realized by a single neuron network is analyzed and simulated. test. It can be seen from the results of the simulation test that the ratio braking characteristic of the differential protection using a single neuron sensor is feasible, and the best ratio braking characteristics can be obtained, indicating that the specific parameters and labor in the traditional protection principle are The weights in the neural network are matched, the weight of the neural network is given a certain physical meaning, and the best weight coefficient is obtained by using the excellent training algorithm of the neural network. It can be seen from Table 2 that 10kV is caused by the asymmetry of the load. The line current is asymmetrical. For example, the B-phase current of branch 1 is 19.385AA, the phase current is 12 673A, and the B-phase current is higher than the A-phase current of 53%. According to the symmetrical power flow (Table 4), the three-phase of the branch The current is 15.802A. At this time, the actual current of phase B may be higher than the calculated value. 22 It can be seen from Table 3 that the calculated value of the three-phase power flow is close to the actual measured value and the maximum deviation between the calculated value and the measured value is 3.6%. The deviation is 1.04*, and the average deviation is 3.28%. The degree of deviation can be corrected by the state estimation.
5 Conclusions The three-phase power flow calculation method proposed in this paper is simple in principle, stable in convergence, fast in calculation, and can meet the requirements of real-time calculation. The example shows that the current asymmetry in China's 10kV distribution line is serious, so the power distribution Real-time power flow calculation in automated SCADA should use three-phase power flow calculation method
Custom Camlock,Brass Industrial Hose Coupling,Aluminium Camlock Type Dc,Adaptor Hose Shank
SHIJIAZHUANG STRONGFLEX I/E CO.,LTD. , https://www.sjzhydraulichose.com