Predictive Control of Unified Power Quality Conditioner (UPQC) using Random Forest (RF) Algorithm

Abstract

Power quality is an essential problem that is becoming more significant to energy consumers because non-linear load has been extensively employed in power sectors. Harmonics are produced when nonlinear loads convert AC line voltage to DC, which has detrimental effects on electrical equipment. Conductors and distribution transformers of a power system may overheat if distortion causes an increase in current flow. Temperature increases will reduce the lifespan of electronic gadgets and disrupt power grids if not addressed. Using the Random Forest (RF) algorithm as a machine learning technique, this research predicts the correct value of proportional gain (P) and integral gain (I) to minimize Total Harmonic Distortion (THD) of the source current on the Unified Power Quality Conditioner (UPQC) circuit in the state of voltage sag reduction. There are three case studies on voltage sag reduction at 40%, 30%, and 10%. RMS of source voltage and RMS error in the DC link capacitor voltage regulation circuit that is in the Shunt Controller in the UPQC circuit are recorded. The total number of recorded data is 2,000,000.  These data will undergo the multiple time lag process to enhance the data, principal component analysis to reduce the data's dimensionality, and a random forest technique to predict the optimal values of P and I.  Using regression analysis and root mean square error (RMSE), the performance of the utilized model will be evaluated. According to the observed results, multiple time lag with a lag value of 24, principal component analysis with a value of 0.000001, tree and leaf values of 100 and 2 for P, and 10 and 5 for I, will create more accurate results. For testing and validating P, the chosen RMSE values are 6.0392e-10 and 7.8365e-13. Meanwhile, the testing and validation RMSE values for I are 1.4738-10 and 1.1427e-13, respectively. The regression result for both instances P and I is equal to 1 considering the optimal setup. Finally, the THD value of the source current was reduced successfully.