Selective Harmonic Elimination in Single Phase Inverter with LC Filter using Artificial Neural Network

In this paper presents the using of selective harmonic elimination pulse width modulation (SHEPWM) switching technique in the unipolar single phase inverter with LC type of filter circuit and concept of using artificial neural network (ANN) to simplify the designs of components value to the result. The need of harmonics minimization is required in the current inverter. The SHEPWM is one of technique that widely used which has an advantage on selectable the harmonic order to be eliminated and operates on lower switching frequency scheme, leading benefit to decrease switching loss. The low pass filter is still highly required to minimize harmonics from SHEPWM inverter output. The LC low pass filter is selected which is highly efficient on harmonic elimination and improve the quality of the output. The SHEPWM and LC filter are performed simulation in PSIM. To avoid the difficulty of solving the SHEPWM non-linear equation and reduce the step to find the filter components, the ANN is proposed to simplify the process of design and simulation which provide the LC filter value from the desired parameters by using MATALB Simulink. The 3^rd with 5^th, 3^rd with 7^th and 3^rd with 9^th harmonics order are proposed in simulation scope with fixed value of capacitor 100 uF, 220 uF and 330uF and total harmonic distortion (THD) from 2% to 5% to be totally 270 ANN training data. The simulation result of 3^rd and 5^th elimination of harmonic order is lower THD than other for the output before filter and the 3^rd and 9^th elimination of harmonic order are required less LC filter value than other to achieve same THD. The parameter value from ANN model comparing with simulation in PSIM are analyzed and result in the lowest error at 0.27% and the highest at 1.51% from assigned testing value. The concept of using ANN is presented which be possible solution to apply. However, the result has such a bit of error that indicating a greater number of training data is required to achieve more accuracy in ANN model.