Machine learning for power system impedance estimation
Kamyab Gibaki, Edinburgh Napier University
The increasing penetration of power electronic converter interfaced distributed energy sources into the power grid (e.g. Wind), introduced new challenges to the monitoring, protection and control of the system. The high amount of converter interfaced sources impacts the stability of the system, especially in a weak grid with high inductive impedance. The knowledge of the grid impedance at the connection point of the converter to the grid (PCC) is essential for improving the control strategy and overall grid stability by either changing the control actionor re-tuning the controller. In micro-grids, a variation of impedance is an indication for islanding or grid connection mode operations. Furthermore, the knowledge of grid impedance will be useful to improve the power quality, detection of the fault location, ground faults andgrid unbalanced operation. Various methods are proposed to estimate impedance of power network, though all of the methods can be classified into passive and active methods or combination of these two. The passive methods are known to be ‘non-invasive’ and active methods are ‘invasive’. Generally, the active methods are invasive as a disturbance signal is injected to the grid. Then the signal processing techniques are used to estimate the impedance of the grid. However, passive methods do not need any disturbance injection, and the available information of the non characteristic current and voltage at the PCC is used to estimate the grid impedance. Hence, the performance of the power system is not degraded using passive methods. Both the passive and active impedance estimation methods have shortcomings for power system applications.The passive methods can become inaccurate, and the active methods impact power quality. In this research project, to avoid the shortcomings of the other estimation methods, a passive machine learning based technique to estimate the impedance of the power grid at the point ofcommon coupling of a converter interfaced distributed generation source is proposed. The proposed method is based on supervised learning and provides a fast and accurate estimation of the grid impedance without adversely impacting the power quality of the system. This method does not need an injection of additional signals to the grid and provides an accurate estimation of the grid impedance. Multi-objective NSGA-II algorithm is used for optimisation and tuning the random forest model for accurate estimation of both R and X. The resistive and inductive reactance of grid is estimated using Random Forest model due to its capability in the prediction of multiple output values simultaneously.