Você está na página 1de 10

The Application of Neural Networks on Three Phases Shunt Active Power Filter

Hanny H. Tumbelaka*, Lawrence J. Borle** and Chem V. Nayar*


* Department of Electrical and Computer Engineering Curtin University of Technology, Western Australia ** School of Electrical, Electronic and Computer Engineering. University of Western Australia, Western Australia

Abstract Increasing number in operating non-linear equipment loads will cause the harmonics on power system and reduce the quality of electric power. Harmonics on power systems can be overcome by using a Shunt Active Power Filter. Shunt Active Power Filter is three phases Current Controlled Voltage Source (CC-VSI) that connected to Point of Common Coupling (PCC). Neural networks compensation method is used as a control series to generate the reference current for Shunt Active Power Filter which is obtained from the load current. Neural networks determine the amplitude of sine and cosine components of the reference current. The simulation results showed that three phases Shunt Active Power Filter by neural networks compensation method is effectively able to eliminate harmonics and generating sinusoidal current source/grid either in steady state or dynamic conditions. Key words: Active Power Filters, Harmonics, Neural Network

1. Introduction The increasing in using power electronics on electric power systems caused a lot of nonlinear loads are installed. Operation of non-linear loads will cause harmonic in electrical distribution systems. It will reduce the quality of electric power due to the current non-linear loads did not form sinusoidal waves purely, although voltage source that used at that time formed the sinusoidal wave purely. The solution to overcome harmonic problems can be done by installation the harmonic filters. Generally, harmonic filters can be classified as Passive Power Filter and Active Power Filter. The use of Passive Filter to overcome harmonic problems has some weakness, among others, will only be used to filter one harmonic frequency (single tune). This means that needs some filters to overcome a number of harmonic currents. In addition, this filter is not flexible to load changes, will cause a parallel resonance in power systems, has design size and weight of the inductor L and C are large enough to filter the harmonics, as well as having LC filter characteristics that strongly influenced by system impedance which difficult to know surely because changing to the networks configuration [1]. To overcome the weakness from Passive Power Filter, then Active Power Filter is used. Thus, the use of Active Power Filter will be investigated to eliminate the harmonics. Over the years, compensation method has been developed to control Active Power Filter [2] [3]. Each method has its uniqueness, excellence and weaknesses. Current developments, artificial intelligence techniques in particular of artificial neural network techniques [4] was started to be applied to detect harmonics. Therefore, this paper will discuss the application of three phases Active Power Filter that works by injecting anti-harmonics current to relieve harmonics on power systems by artificial neural networks method.

2. Shunt Active Power Filter Configuration Shunt Active Power Filter is the Voltage Source Inverter (VSI) which consists of six pieces of IGBT with anti-parallel diode. On DC side consist DC capacitors and inductors installed on the AC side that connected to Point of Common Coupling (PCC), parallel with the load. VSI is operated by control the inductor current such that according to the desired reference current. Thus, current sensor is installed on output side (AC) of VSI. The output of current sensor is compared to the reference current which comparison results produce Pulse Width Modulation (PWM) signal to trigger IGBT. This operation mode is called the Current-controlled Voltage Source Inverter (CC-VSI). Reference current is obtained by detecting load current which contain harmonic by using current sensor. The load current is then processed to separate the fundamental component and harmonic components. In this case, the process of separation is performed by using a artificial neural networks method. Harmonic components are used as reference current to control inductor current of the CC-VSI. Thus, CC-VSI produces anti-harmonics current against harmonics current caused by the load, so that source current become sinusoidal. Block diagram Shunt Active Power Filter is shown in figure 1.

Figure 1. Shunt Active Power Filter Configuration In order to series can work stable and current can flow, then CC-VSI voltage on DC side should be greater than peak-to-peak voltage on AC side [5]. Voltage on DC side must be maintained constant by using PI (Proportional Integral) control. If not, then the losses in the series will cause the DC-bus voltage decrease.

3. Method of Neural Networks 3.1. Basic Principles of Shunt Active Power Filter with Neural Network Control Some of the periodic signal can be described as sum of sine and cosine components. This concept is the basis of designing artificial neural network architecture in estimating the harmonic components in non-linear load. Load current iL(t) can be described as:

where Wan and Wbn is the amplitude of sine and cosine component of the load current iL(t). Load current iL (t) can be simplified into the equation: iL (t) = WT X(t) where the weighting matrix: WT = [wal wblwan wbn] and vector input of sine/cosine: X(t)T = [sin(t) cos(t)sin(nt) cos (nt)] (4) (3) (2)

Based on above equation, the architecture of neural networks compensation can be described as follows:

Figure 2. Neural Network Architecture

As shown in Figure 2, neural network have a constant current input of N pairs in which each pairs of input consists sine and cosine inputs and represents fundamental component and harmonic component of load current. Weight of each input from artificial neural network represents the amplitude of fundamental or harmonic current. Output of neural networks is estimate load current (iest) and estimate fundamental load current (ip). The process of training is done by changing the weights of WT so that e(k) be zero. e(k) is the error between actual load current (iL) and estimated load current (iest). Thus, if e(k) has a zero or near zero, then load current iL and fundamental current ip can estimated exactly. So that, reference current i*C(k) for Active Power Filter can be calculated by deducting actual load current iL and estimation results of fundamental current ip. 3.2. Training Widrow-Hoff Algorithm for Neural Network on Active Power Filter The process of training is done by using Widrow-Hoff algorithm with purpose to minimize the amount of error e(k) between load signal iL(t) and estimated signal iest (k), where e(k) can be expressed by the equation below: e(k) = iL - iest (k) (5)

Reference current i*C(t) of Active Power Filter will be obtained based on the results of weights calculation. The basic principle of weight calculation by artificial neural networks algorithm can be described as follows:

Figure 3. Calculation of Artificial Neural Networks Weights Artificial neural networks in figure 3 consists a neuron, where the weights of neurons expressed as W. The input of system is source voltage VS(t) that raised, while the output of fundamental active current iP(t) and the output of system is reference current i*C(t) of Active Power Filter. Based on the Widrow-Hoff algorithm, the equation to update the weights can be expressed as follows: W(n) = W (n-1) + icr Vs (n-1) (6)

where: icr Vs (n-1) = load changes = learning rate, 0 < < 1 icr (n) = iL (n) W (n) Vs (n) Based on equation 6, the control design by Widrow-Hoff algorithm can be described as shown in Figure 4. Training process of artificial neural networks by learning Widrow-Hoff methods are used to obtain the value of weight in a certain learning rate. At the end, the reference current Active Power Filter can be calculated by using below equation: ic (t) = iL (t) WVs (t) (7)

Figure 4. Compensation circuit i*C(t) Based on equations 7, the modeling of control series in Matlab Simulink to obtain the value of weight W by Widrow-Hoff algorithm can be described as follows:

Figure 5. The Series of Neural Networks Compensation in Matlab Simulink

4. Simulation Results The concept of Shunt Active Power Filter by neural networks compensation method has been proved by computer simulation using the combination of PSIM and Matlab Simulink. CCVSI series is modeled in PSIM, while the neural network control is designed in Matlab Simulink. Figure 6 through 9 is one of the simulation results that has been done by Matlab Simulink from the series of artificial neural networks compensation that consisting of the input current (load current), the error e(k), the estimated current iest(k) and sine/cosine current of fundamental component iP, with learning rate 0.4. Input current in figure 6 is a non-linear load current which has been distorted by the harmonics. Estimated current is generated by compensation series which depends on the errors value. Estimated current is the difference between input current and errors. The smaller value of the error, then estimates current will be more similar with input current.

Figure 6. Input current expenses

Figure 7. Error e (k)

Figure 8. Estimated current iest(K)

Figure 9. Fundamental components of ip Figure 7 shows the error values are very small at steady state conditions so that the estimated current obtained is almost same as input current. It means reference current for Active Power Filter is obtained by calculating the difference between load current iL and fundamental component current iP of compensation series. Figure 10 shows the three-phases load current that contain harmonics. By using the network neural, Thress-phases reference current is obtained for CC-VSI as shown in figure 11. The operation of CC-VSI as Shunt Active Power Filter produces sinusoidal three-phases source current and balanced (figure 12). The THD comparison of load and source current (after compensation) can be seen in table 1. From figure 10 to 12 also illustrated the load changes. It appears that neural network is able to adjust reference current to the load changes. Source/grid current remains at sinusoidal shapes.

Figure 10. Three-phases Load Current

Figure 11. Reference Current

Figure 12. Source Current after Compensation

THD current (%) Load Source/grid

Phase R 25.5 1.8

Phase S 26.1 2.1

Phase T 25.1 1.9

Table 1. THD of load current and source current

5. Conclusion Based on the analysis of simulation results obtained, we can conclude the following: 1. By adjusting the learning rate, artificial neural networks is able to adjust weights W, hence references current that used in CC-VSI is formed. The desired reference current is obtained when the error e (k) equal to zero. In this case, the estimated current produced resembles the load input current. 2. The simulation results showed that three phases Shunt Active Power Filter series by using neural networks compensation methods is effectively able to overcome load harmonic current. 3. The simulation results also showed that in dynamic conditions, series of three phases Shunt Active Power Filter by neural networks compensation method is effectively able to provide quick response during load changes, so harmonic current can be compensated properly.

Reference [1] Jou, H.-L., J.-C. Wu, and K.-D. Wu, Parallel Operation of Passive and Hybrid Power Filter Power Filter for Harmonic Suppression. IEE Proc. Generation Transmission Distribution, 2001. 148(1): p. 8-14 [2] El-Habrouk, M., M.K. Darwish, and P. Mehta, Active power filters: a review. Electric Power Applications, IEE Proceedings-, 2000. 147(5): p. 403-413. Green, T.C and J.H Marks, Control techniques for active power filters. Electric Power Applications, IEE Proceedings-, 2005. 152(2): p. 369-381. Temurtas, F., et al., Harmonic detection using feed forward and recurrent neural networks for active filters. Electric Power Systems Research, 2004. 72(1): p. 33-40. Tumbelaka, H.H., L.J. Borle, and C.V. Nayar. A New Approach to Stability Limit Analysis of A Shunt Active Power Filter with Mixed Non-linear Loads. in Australasian Universities Power Engineering Conference (AUPEC). 2004. Brisbane, Australia: ACPE. p. ID: 121

[3]

[4]

[5]

Você também pode gostar