WANG Zhi-wen,GONG Mao-fa,AN Bin,LI Lan-bing,LIU Tao

(College of Electrical Engineering and Automation,Shandong University of Science and Technology,Qingdao 266590,China)

0 Introduction

The transformer is a very important electrical equipment and plays a key role in long-distance transmission in the power system,and it can reduce the voltage level of transmission power load. Correctness of the transformer operation is directly related to reliability and safety of the entire power system. In order to ensure the transformer’s selectivity and sensitivity,it is necessary to correctly identify the inrush current and internal fault current of the transformer differential protection of main protection. In this paper,we use FPGA+DSP dual-CPU structure. In the structure,FPGA plays roles of digital filtering,feature extraction and algorithm implementation function as the chip of wavelet transform algorithm. DSP is used for parameter tuning,judgment of action,alarm,communication and other functions. The structure can be used in the occasions of high sampling frequency,large amount of data and high processing speed. It will greatly improve sensitivity and reliability of the power system[1].

1 Analysis of inrush current theory

In the transformer differential protection,the inrush current flows through only one side of the transformer. The inrush current is very small in the normal operation condition or an external fault excitation,and is usually less than 1% of the rated current. Its magnetization curve is non-linear,but the transformer operates normally in linear section of the magnetization curve. When an external fault occurs,its excitation impedance is very large and the excitation current is very small. But when the transformer no-load input voltage is restored after the external fault is cleared,the core of the magnetic flux cannot mutate,and aperiodic transient component appears in the magnetic flux with combined effect of the residual magnetic flux in the transformer core,which leads to the transformer core saturation. Since the voltage is alternating,thus the transformer core goes into and out of the saturated zone periodically. When it goes into the saturation zone,the excitation current will increase sharply,and its maximum will reach 6-8 times of the rated current. Inrush current’s characteristics are as follows:

1) The magnetizing inrush current contains a large number of high-order harmonic,specifically the second harmonic. The magnetizing inrush current often contains a lot of non-periodic components,so that the inrush current waveform tends to one side of the timeline.

2) Three-phase transformer inrush current is not only related to closing angle,residual saturation flux and system impedance,but also influenced by the three-phase winding connection mode and the magnetic structure.

3) The magnetizing inrush current waveform shows intermittent angle within the first few cycles. Each cycle has intermittent angle about 80°-100°. The higher the degree of core saturation is,the larger the intermittent angle gets[2].

2 Analysis of protection algorithm

The wavelet neural networks are neural networks model composed of the new wavelet transform theory and neural network theory. It takes advantage of wavelet theory that is frequency localization features and multi-resolution analysis (MRA),and makes full use of self-learning ability of the neural networks.

From theoretical analysis,wavelet coefficients of the magnetizing inrush current and fault current show a clear distinction. Inrush current waveform has many mutations and wavelet coefficients are really large. The waveform is accompanied by the whole process of magnetizing inrush current attenuation. Short-circuit current has larger wavelet coefficients in the initial moment of fault occurring. Then it rapidly attenuates and can be seen as a straight line even. So it is important to distinguish between magnetizing inrush current and fault current. Firstly,it needs to design pre-treats signal using wavelet transform method and to extract its characteristics of wavelet space,and makes the wavelet space as the feature space of neural networks pattern recognition,then extracts the neural networks of wavelet transform features to complete classification and function approximation. Finally,we can ultimately determine the magnetizing inrush or fault current. Fig.1 shows a block diagram of the method[3-6]above.

Fig.1 Block diagram of the wavelet neural networks identification method

In the paper,we useABphase as an example to illustrate that how the algorithm distinguish between magnetizing inrush current and fault current. Using the db5 wavelet to dealABdifferential current with 5 times the scale of one-dimensional resolution analysis can get the six band wavelet coefficients; then we rebuild a multi-scale wavelet by db5 wavelet,and the reconstructed signalScan be expressed as

S=a5+d5+d4+d3+d2+d1.

(1)

According to Eq.(1),one-dimensional wavelet approximation coefficients and detail coefficients are extracted. Then according to Eq.(2),the energy valueEsand the energy values of the five layers of detail coefficientsEWj(j=1,2,3,4,5) within the original signalSof one cycle are calculated.

(2)

According to Eq.(3),the wavelet decomposition feature vectorTof theABdifferential current in the from ofEsandEWjis obtained. Every two differential current feature vectors are consisted of five elements by wavelet decomposition. As a three-phase system,a total of 15 element feature vectorsTare needed to be the BP neural networks input vectors.

T=[T5,T4,T3,T2,T1]=

(3)

In order to make value of the feature vector falls in the range of (0,1),wavelet decomposition feature vectorTof theABdifferential current is made by the linear normalization pretreatment of Eq.(4),as BP neural networks input samples.

(4)

In order to correctly distinguish between magnetizing inrush current and fault current,it needs a one-variable of the BP neural networks output samples. The output layer is made ofS-type logarithmic function logsig. The output is limited in interval (0,1),and the ideal output can only be closed to 0 or 1,so as to set up a network ideal output value of 0.1 or 0.9. All sample biases are less than 0.3 after a series of training and verification. Therefore the following criterions are required.

When the neural networks outputysatisfies the relationship of the Eq.(5),we have

|y-0.1|<0.3.

(5)

That is to say,whenyis in the interval (0,0.4),it is determined as the magnetizing inrush current.

Similarly,when the outputysatisfies the relationship of Eq.(6),we have

|y-0.9|<0.3.

(6)

So whenyis in the interval (0.6,1),it is determined as the fault current.

In this paper,results obtained with 100 samples are shown in Figs.2 and 3. It can be be seen that the approach of Levenberg-Marquardt has a higher convergence speed and a smaller network training error.

Fig.2 Training samples’ approximate expected values

Fig.3 Error distributions of 100 samples’ output values and expected values

3 Design of hardware architecture

The design includes six modules which are AC input filter,analog-to-digital conversion,FPGA + DSP,display,trip alarm and communication. Hardware schematic diagram of the design is shown in Fig.4.

Fig.4 Diagram of hardware schematic

3.1 A/D conversion module

A/D converter module applies MAX125,which is MAXIM's production of high-speed,multi-channel,14-bit analog to digital conversion chip.

Operating mode,chip selecting,reading and writing and the terminal of MAX125 can be controlled by FPGA. According to the setting work mode,MAX125 can convert the analog input signal of each channel successively,and the converted valus are stored in the 4x14-bit buffer of the chip.

3.2 Communication module

On the basis of communication module,the design of TMS320X2812 DSP has RS232 and CAN bus interface,serial communication interface (SCI) module and eCAN module for Internet communication. Based on existing resources of the DSP,we select MAX3232 chip for serial communication and SN65HVD232 chip for CAN communication.

3.3 FPGA module

FPGA module is mainly used to achieve wavelet neural networks recognition algorithm. Wavelet MRA uses a distributed algorithm to achieve Mallat algorithm. The paper describes the FPGA implementation of wavelet MRA of a one-dimensional layer,while the multi-layer wavelet transform is repeatable and it can be obtained through wavelet transform cascade of a one-dimensional layer[7-9].

4 Design of software architecture

The protection device software is modular in design. The software module includes initialization module,fault identification module,fault processing module and display module. Software flow chart is shown in Fig.5.

Fig.5 Software flow chart

5 Conclusion

The paper presents a transgormer protection scheme of FPGA+DSP,and uses FPGA hardware language to achieve the wavelet neural networks algorithm. The scheme not only keeps a very good balance between the protection of speed and accuracy,but also enhances sensitivity and reliability of the protection device. The scheme of FPGA+DSP enhances real-time processing speed of transformer to distinguish between internal fault current and magnetizing inrush current. The protective device has a good reliability,sensitivity and a high practical value.

[1] ZHANG Xiao-ming. Study on a new method for digital transformer protection based on wavelet neuram network and FPGA. Qingdao: Shandong University of Science and Technology,2011.

[2] ZHANG Xiao,ZHANG Jian-wen. Research of magnetizing inrush current and fault current identification based on wavelet transform in transformer protection. Electrical Measurement and Instrumentation,2012,49(10): 73-77.

[3] LI Gui-cun,LIU Wan-shun. A algorithm of discrimination between inrush current and fault current of transformer based on self-correlation analysis. Automation of Electric Power Systems,2001,25(17): 25-28.

[4] ZHANG Xiu-chuan,HUANG Yi-zhuang. A fast algorithm of 2B-spline wavelet transform and its application to transformer protection. Relay,2004,32(1): 32-37.

[5] HU Xiao-peng,YI Li-gang. A new method of applying artificial neural network based on preprocess of wavelets to realize computer-based transformer protection. Relay,2004,32(6): 22-26.

[6] LIU Jian-li. Research on inrush current and its identification for transfomer. Zhenjiang: Jiangsu University,2011.

[7] Masud S,McCanny J V. Reusable silicon Ip cores for diserete wavelet transform applications. IEEE Transactions on Circuits and Systems: Fundamental Theory and Applieations,2004,51(6): 1114-1124.

[8] Saied M M. A study on the inrush current phenomena in transformer substations. Proeeedings of Industry Applieations Conference & Thirty-sixth IAS Annual Meeting,IEEE,Piscataway (NJ),2001,2: 1180-1187.

[9] Bronzeado H S,Yacamini R. Phenomenon of sympathetic interaction between transformer scaused by inrush transients. IEEE Proceedings-Science,Measurement and Technology,1995,142(4): 323-329.