A salient feature of a renewable energy power supply system (REPSS) on islands is the high level of uncertainties caused by high penetration of volatile power sources, such as wind and solar photovoltaic farms. This creates large forecast errors under some conditions and makes the fixed time-scale dispatch impotent in maintaining system reliability. To tackle this challenge, a time-scale adaptive dispatch method is developed in this paper. The time-scale for dispatch REPSS on islands is adjusted online according to the confidence interval of forecast error predicted by neural network. Extensive tests have validated the effectiveness of the presented method in offsetting the uncertainties in the system and improving the system reliability.