Geostatistical modelling techniques are capable of generating high-resolution reservoir models. Since a limited amount of information is available to model the reservoir, geological uncertainty is represented through a suite of equally probable models. Unfortunately, these high-resolution models are often too large to process through numerical flow simulators. Upscaling methods are required that reduce the size of detailed models while preserving the important geological characteristics of the reservoir. Most upscaling methods currently in vogue work with uniform grids. A new upscaling approach based on non-uniform coarsening with optimum power average is presented. The proposed algorithm identifies likely high connectivity regions using streamline simulations, then constructs a non-uniform coarse-scale grid preserving the areas with probable high connectivity and assign equivalent permeability to the coarse grid blocks using an optimum power average technique. The power average exponent is calibrated using the data from a series of single-phase flow simulations.