【導(dǎo)讀】摘要--經(jīng)過(guò)調(diào)試的前饋神經(jīng)網(wǎng)絡(luò)區(qū)分電力變壓器勵(lì)磁涌流和故障電流。的調(diào)試算法是反向傳播，假定最初的S型傳遞函數(shù)為網(wǎng)絡(luò)處理單元。網(wǎng)絡(luò)進(jìn)行訓(xùn)練單位的傳遞函數(shù)改為硬限幅器的閾值等于在調(diào)試中的S形偏差。一個(gè)FFNN可被視為一種替代方法，使數(shù)字繼電器在浪涌和故障電流之間實(shí)現(xiàn)判別。延遲，2）根據(jù)所測(cè)量的電流的諧波含量,抑制或阻斷繼電器動(dòng)作。第一個(gè)解決方案已。被用于初級(jí)過(guò)電流保護(hù)和在差分格式中。然而，這是不理想的，因?yàn)檠舆t內(nèi)部故障跳閘。時(shí)間存在潛在危險(xiǎn)。故障，那么所述的第二或第五諧波檢測(cè)是不充分的指標(biāo)。該論文中，在提出的方法的基礎(chǔ)上使用一個(gè)主相電壓作為控制信號(hào)。適于在數(shù)字式保護(hù)繼電器實(shí)現(xiàn)該目標(biāo)。一個(gè)前饋神經(jīng)網(wǎng)絡(luò)的一般結(jié)構(gòu)在圖1中示出。涌檢測(cè)和保護(hù)），可能會(huì)使用相同的采樣率。3）事實(shí)上上述界限的問(wèn)題作為設(shè)計(jì)FFNN之一，給定的變壓器電流的樣本的序列，這意味著，一個(gè)序列有112個(gè)樣本。第一矢量p1的元素對(duì)應(yīng)于第12個(gè)樣品的i，矢量

　　

【正文】 in the output layer) NN must be 1, necessary to indicate “true” and “false,” or “0” and “1”, as indicated in section . In this case it was chosen that the work’s output be 0 when the applied current is an inrush, and 1, when it is not an inrush . Basic Architecture A functional block representation of the scheme used to achieve the goals stated in the former sections is shown in Figure 3. This kind of work is sometimes called timedelay neural work [16] and has been used recently in another power system application [18]. Feed forward works have also been applied to the detection of high impedance faults with remarkable results [19,20]. Notice that the scheme given in Figure 3 is equivalent to the scheme of Figure 1, if the input xk is equal to the thk sample of input x(t). In other words, the work will receive the 12 samples of each window every sample period, and must make a decision based on these 12 samples (. one cycle), which implies the number of inputs must be 12. The number of layers and the number of units on each layer is determined by the heuristic process described in section (the only layer pletely defined at this point is the output layer) Figure 3. Timedelay work. 3. TRAINING STRATEGY . Training Examples Given that the work has to distinguish between two kind of signals, two sets of examples signals (cases) were prepared for that purpose: the inrush cases and the fault cases. The inrush cases were measured in the laboratory energizing at random a small power transformer of 50 VA, 1202040 V. The cases taken to train the work were chosen such that they represented an acceptable range of inrush current shapes (the six cases described sampling rate at which the inrush signal examples were measured was originally 84 samples per 60 Hz cycle, and then resampled (at a rate 7 times lower) to get the desired 12 samples/cycle. Figure 4 show some of the inrush training signals. The fault cases were generated by puter, some of these using an electromagic transients program [23], others by simple generation of faultlike signals (response of an RL circuit), and a third group of faultlike currents infected with second, third and fifth harmonics Figure 4. Four inrush training cases. Two special cases grouped with the inrush cases were the zero current (indicating the transformer is deenergized) and the load current (a simple sinusoidal wave with a magnitude equal to the transformer load current). For each case, signals were sampled at 12 samples per cycle (with a window length equal to one cycle) and the total of samples was limited to have about 100 windows (. seconds per case). . Data Windowing and Training Matrices The training matrices were built in such a way that the work was trained to produce a 0 output when the presented signal was an inrush, a zero current or a full load condition。 and a I output for the other cases (the authors do not insinuate the method can be used to do all the protection function, since additional research will be needed to prove it). Vectors pj were built from the cases in the following way (see Appendix 2 for the definition of these vectors). Suppose the first of the cases’ signals is characterized by the sequence: ? ?1 1 221 , iiiii k ??????? （ 1） this means, a sequence of 112 samples. The first vector p 1 elements corresponds to the first 12 samples of i, the vector p2 entries are the 12 samples from i2 to i13 and so on, until pleting the first 100 columns of example matrix P defined in eqn. (A11) and (A13). The same process was repeated for each cases’ signal until having an example matrix of 936 columns (the number of examples, ne). This matrix was built such that the first 600 examples correspond to inrush examples (desired output=O), and the other 336 examples correspond to faults (desired output=l). Since the FFNN has only one output, matrix D became a horizontal vector of 936 elements, the first 600 entries are 0’s and the other 336 are 1’s. In practice, a target tolerance of was used, meaning that the work was trained to produce a response of or greater to represent one class and or less to represent the other class. This is necessary because the nature of the nonlinear squashing function is such that it can never assume the precise values of or . . Training Process Once the training matrices P and D were defined, the backpropagation algorithm was applied to the problem. The MATLAB Neural Network Toolbox [24] was used for such a purpose. Function TRAINBP was used with sigmoid transfer functions and a learning rate variable between and . The admissible error (sum of square errors in an epoch) was , for which it was necessary to present the training matrices between 200 and 600,000 times, depending on the work size (number of units) and the learning rate employed. 4. IMPROVING THE NETWORK . Network Testing and Pruning As mentioned in section it is desired that the work be applied in a digital relay implementation. This implies a promise between speed and accuracy. As is well known, the FFNN classification time (the required time to produce an output given an input) depends on the number of units in the work, so it is very important to have the lowest number of units but without jeopardizing the quality of the classification. Figure 5 shows the architectures tested for this purpose.