Anti-islanding detection in grid-connected inverter system using active frequency drift technique with random forest

The increase in penetration levels of distributed generation (DG) into the grid has raised concern about undetected islanding operations. Islanding is a phenomenon in which the grid-tied inverter of a distributed generation system, and some of the local loads are disconnected from the grid. If this condition is not detected and the generation (e.g. from a photovoltaic energy source) remains operative, the isolated DG system will stay energised by the inverter. The phase mismatch between the grid and the inverter voltage makes this scenario undesirable since it could be hazardous for the maintenance operator and could harm the inverter and loads in the event of an unsynchronised reconnection of the grid. Consequently, the article presented a novel hybrid active anti-islanding approach for fast and reliably detecting unintended islanding. For the modelling and experimental setup, a multiphase grid-tied photovoltaic distributed generating system was employed, and it was regarded as a viable application. Initially, the study introduces a fault-tolerant control (FTC) technique of data-driven predictive control to limit the impact of grid faults on inverters. Furthermore, the article suggested the Sandia frequency and voltage shift (SFVS) approach for inverter-based distributed generation to identify an islanding state. The approach employs a positive feedback gain to minimise NDZ and THD; moreover, the system does not affect the power quality. To minimise system damage, this scenario necessitates the use of effective islanding detection algorithms. This study suggests using empirical mode decomposition (EMD) to enhance power quality to extract detailed coefficients, which are subsequently processed to detect common transient fluctuations during islanding. In the grid-tied inverter, random forest (RF) is also utilised to categorise the condition as islanding or non-islanding. This exhibits an acceptable trade-off between output power quality and islanding detection effectiveness. The quality of the data and the selection of hyperparameters impact the performance of ML models. Following that, hyperparameter adjustment of the random forest model is done using the memetic firefly algorithm (MFA), which has a considerable impact on overall classification performance. The hyperparameters of the suggested RF model have been successfully calculated to attain the maximum accuracy and the least loss. Results demonstrate that the suggested method produces less THD (2.44)