Abstract The quality of power in modern‐day power system is polluted with increased penetration of converter‐based distributed generations such as wind farm, solar PV system. In such scenarios detection of islanding and power quality disturbances as well as the removal of these from the system is quite crucial for equipment and maintenance personnel safety. Here, a down‐sampling empirical mode decomposition (DEMD) and optimized random forest (RF) machine learning approach are hybridized to detect islanding conditions with reduced non‐detection zone (NDZ) and classify non‐islanding power quality events in a highly wind energy penetrated distribution generation system. DEMD has a special ability to filter out the fundamental signal from the polluted signal and random forest is quite an unbiased non‐linear machine learning approach. Moreover, an improved grey wolf optimization technique is proposed to optimize the parameter of RF. The proposed technique is simulated in MATLAB/Simulink with IEEE 13‐Bus test grid. The efficacy of the proposed method is evaluated through comparative analysis with existing machine learning techniques under normal and noisy environments as well as validated in narrow NDZ with lesser detection time.