Artificial neural network of relay protection
Abstract According to modern control technology of artificial neural networks to propose a plan to protect principles of composition, and analyzes the principles of feasibility and technical difficulty of achieving.
Artificial neural network (Aartificial NeuralNetwork, hereafter ANN) is a simulation of the structure of biological neurons proposed an information processing method. As early as 1943, by psychologists and mathematicians WalthH.Pitts WarrenS.Mcculloch proposed neuron model, after being ignored for some time, the rapid rise of another 80 years [1]. ANN reason for being a great concern, because it has the essential characteristics of the non-linear, parallel processing, robustness and self-organization self-learning capability. Which studied the most mature is the error back propagation model algorithm (BP algorithm, BackPropagation), its network structure and algorithm intuitive, simple, application in the industry more.
Trained ANN for analysis of vibration data on the use of the machine monitoring and fault detection, some parts of the fatigue life prediction [2]. Non-linear neural network compensation and the application of robust control synthesis method (the robust control using variable structure control or sliding mode control), in real-time industrial control implementation process more effective [3]. Artificial neural network (ANN) and fuzzy logic (FuzzyLogic) integrated to achieve a motor fault detection heuristic reasoning. Non-linear problem, the BP algorithm by ANN case study running to accurately adjust the value of solving the internal power [4].
Therefore, there is a lot of power system in this complex system of nonlinear terms, ANN theory in power system has great potential, is currently involved in such transient, dynamic stability analysis , load forecasting, optimal combination unit, alarm processing and fault diagnosis, distribution network loss calculation, generation planning, economic operation and power system control, etc. [5].
This paper, a neural network-based (ANN) theory of the protection principle.
1 Overview of artificial neural network
BP algorithm is a monitoring study skills, it is true by comparing the output unit outputs the difference between the values and hope to adjust the weights of the network path, so the next time the same input, the network the output value close to hope. Figure 1 is a structural model of artificial neural Ui, Ui diagram for the internal state of neurons, Qi is the threshold value, Yi is the output signal, Xi (i = 1,2, …, n) for the neurons receive signals. The model can be expressed as:
Type of Wji?? Connection weights. BP algorithm of neural network graph shown in Figure 2, set the network input module p, ordered network output unit j under the action of the output Opj. If you want the output value is Tpj, then the error is Dpj = Tpj-Opj. If the input module of the first i-unit input for the Ipi, p is the input module, the input and output contacts j I contacts between the amount of weight change:
Wpji = zDpjIpi The formula, z is a constant. When the repeated iteration of the type, they can converge to the actual value of the target value [6]. Existing transmission network in which the hidden layer lines, but also output cable, every arrow has a certain weight.
Put into operation in the neural network before the application of large amounts of data, including normal operation, abnormal operation, as part of its training content, to a certain degree of input and desired output by BP algorithm to constantly modify the network weight. In the put into operation, also under the particular circumstances of the scene-site learning, to expand the knowledge of the amount of memory ANN. View from the algorithm theory, parallel processing and nonlinear function of BP algorithm is a major advantage.
2 neural network based protection Neural network, the protection device can determine a more complex model, and its causal relationship is more complex, nonlinear, ambiguous, the dynamic and non-stationary random. It is the neural network (ANN) and expert system (ES) integration of the neural network expert system, which, ANN is the value, Lenovo, and self-organization, the bionic approach, ES is a cognitive and heuristic.
Shown in Figure 3, the device can directly access routes in and around the analog, digital, feature transformation by the model input to the neural network, learned under the previous training materials, data, reasoning, analysis, evaluation of the output. Expert system for process control and training operation, according to the best way to collect data or collected by the analysis process and then control the output results of the assessment, determine its accuracy, consistency, made a final ruling by the transformation output to the executing agency. Even the new protection, there will be some functional modules may be incorrect action, then human intervention can be extended after the expert system database or by the expert system to determine, as a training sample of this part of the function of ANN training module, a change Some of the wire weight, so that the same situation next time to reduce the possibility of an incorrect action.
ANN Here is a simple example of line protection. When the power system fault, the transmission lines of each phase, the sequence voltage and current are changing, too, especially after the failure phase fault phase voltage and phase current, and grounding system ground fault zero sequence current changes are obvious representative. Such as the number of neurons selected for the input layer 14, respectively Uar, Uai, Ubr, Ubi, UcrUci, Iai, Ibr, Ibi, Icr, Ici, Ior, Ioi (subscript r and i represent the real and imaginary parts ), the selected number of output layer neurons is 5: YA (A phase), YB (B phase), YC (C phase), YO (ground), YF (direction), the output value of 1, on behalf of the election in; output value of 0 represents no select (YF is 0 for reverse). This five output to satisfy demand line direction protection (with no forward beyond), the number of hidden layer neurons 2N +1 (N is the number of input layer neurons). Training set contains 14 input variables and five output variables, a sample of the test sample set, only 14 input variables. Figure 4 Bilateral selected for study power system, transmission lines, the system is equivalent, zero sequence parameters shown in Figure 4.
Considered, including single-phase ground fault type (K1), two-phase short circuit (K2), two phase to ground (K1? 1), three-phase short circuit (K3).
Shown on Figure 4 Bilateral 500 kV power system fault conditions of various operating modes and the establishment of training samples.
In the normal state, so that h = (EM) / (EN), h = 1,
With the load change, taken as -60 °, -50 °, -40 °, -30 °, -20 °, -10 °, 0 °, 10 °, 20 °, 30 °, 40 ° 50
I am a professional writer from Mp3 Player Manufacturers, which contains a great deal of information about medical plastic tubing , double foam tape, welcome to visit!
Processing your request, Please wait....
