Shandong Provincial Natural Science Foundation (Q97FG8154) and Tsinghua University National Key. The function of the load-shedding t-neural network structure diagram is different from that of the load-cut decision-making part, so different structures should be used. The classification part not only needs to distinguish between stability and instability, but also distinguish whether it should cut the machine or load it, and implement it with a fuzzy competitive radial basis network. There are two outputs in the stable classification part, that is, yi and y2.yi are for system stability (the other output is not considered at this time), 1 means system instability; y2 is 1 means that the machine needs to be cut, and 0 means the load is required to be cut. At the same time, activate the next level of load shedding decision network. The load shedding decision network obtains the corresponding load shedding control law based on the input quantity, mainly realizes the function approximation function. To improve the approximation accuracy, a fuzzy competitive feedforward network is proposed, which is to use the fuzzy competition network (FCN) to perform fuzzy clustering on the sample. For each type of sample, a different sub-BP network is used for function approximation. See the stable classification network structure. The 21FCN training fuzzy competition network is a network without tutor learning. It can group according to the different characteristics of the sample, divide the sample space, and reduce the training burden of each sub-BP network. The input is connected, the number of output nodes is the number of clusters, and the output U is the membership degree of the input vector pair. Let the input be an N-dimensional vector X (XI, X2..., x*), and there are R output nodes. The connection weight Wi is defined as the center of each type, and U is defined as the fuzzy equality relationship between the input and the weight. Di Yin 0, the training steps of updating the connection right FCN according to the gradient descent rule are as follows: 1 assigning an initial value to the weight within a certain range; 2 randomly extracting a sample, calculating Awj and correcting wj4 to stop if the convergence condition is satisfied, otherwise returning Step 2. The convergence condition can be defined as a change in the weight that is less than a certain small positive value. Assuming that the FCN classifies the samples into R categories, there are R sub-BP networks corresponding to them. After the FCN is trained, the input of each sample is sequentially added to the input of the trained FCN to generate R outputs Ui~Ur. If Ui is the maximum value, the sample becomes the i-th sub-BP network. Training samples. It is important to note that for some samples, the FCN can produce several approximate outputs, and these samples should be trained in several corresponding sub-networks to improve the generalization ability. 2.2 BP network training and training BP network, that is, adjusting the weight and threshold vector w to make *n-dimensional row vector * training sample number output number * for the p-th sample neural network k-th output unit expected output value and actual The output value is the maximum relative rocking angle 3 (i=0,1,2) between the generator axes and its rate of change v, (=1,2). A total of 5 variables are used as inputs to the stable classification network; The network and the total load of each § and 5 load-shedding points at steady state have a total of 4 variables as inputs. Where § reflects the steady state operating conditions and grid structure before the system failure, § and § reflect the system's fault information, they contain a wealth of system stability information, V can reflect the maximum swing angle change trend, each load shedding The total load of the point can reflect the load level of the system. These quantities can basically map the transient stability of the first pendulum of the system. The simulation results also show that the neural network can better approximate the functional relationship between the four variables and the load shedding at different operating points, wiring modes and fault conditions. This gives an appropriate control decision. The output of the load shedding decision network is a percentage of the load shedding. At present, the baud rate of power communication is generally up to 4800 ~ 9600b / sPMU sampling period is 0.033 ~ 0.083s, in the United States Bonneville Electric Power Authority (BPA), the actual running PMU to the control center communication line per s 30 phasor information can be transmitted. The main protection generally operates at 0.1s after the fault. Therefore, in practical applications, the sampling period is taken as 0.05s. The above five variables can be simply calculated by the control master based on the measurement information transmitted from the PMUs installed in each generator node. obtain. 3.2 How to obtain a sample How to obtain a sample that can fully represent the problem itself has a great impact on the results of the neural network. Too few samples can not reflect all aspects of the problem, too much will add complexity to the problem. The method, this article is based on experience to obtain samples as follows. Considering one of the three parallel lines 145-132, single-phase, two-phase and three-phase grounding short-circuit occurs at 0s, and the faulty line is cut off by 0.1s. The sample set is generated as follows: the generator output of the 1st node 145 is 600MW, the active variation range is 70%~130%, and the generator node with the remaining output of more than 50MW is only changed once from the basic output to 80%. And only one generator output changes at a time; 2 each load fluctuates randomly between 90% and 110%; 3 the fault locations are at the ends and in the middle of the line. Considering the faults between lines 145-144, the sample set is also generated as described above, thus taking into account the situation of different faulty lines. In order to adapt to changes in the wiring mode of the system, in addition to consider 145-132 to disconnect a line, 145 * 144 failure. In this way, a total of 5,765 samples were obtained, of which 505 were unstable. Machine, load shedding, first cut 50% load at 5 load-cut points (5 nodes total base load is 547MW). If the system is stable, linear interpolation finds a minimum load-cut percentage (also can be based on 4 conclusions. Synchronous phasor measurement technology and artificial neural network have studied the transient stability region control decision. The actual system simulation results show that the proposed two-stage neural network structure is reasonable, and can better coordinate the cutting machine and load shedding. Stability classification Partial classification accuracy is high and training speed is fast; the fuzzy competitive feedforward network in the decision part can give a fairly accurate load-cutting decision. How to propose a systematic sample acquisition method, reduce the sample workload and how to extend it to the subsequent remote pendulum , are the goals of further research. Encrypted Oxford Cloth Laptop Backpack Encrypted Oxford Cloth Laptop,Custom Design Backpack,15-Inch Backpack,Encrypted Oxford Cloth Backpack SUZHOU INNEST IMP.&EXP.TRADING CO.,LTD , https://www.suzhouinnest.com