The focus of this paper is on VLIW instruction scheduling that minimizes the variation of power consumed by the processor during the execution of a target program. We use rough set theory to characterize the imprecision inherent in the instruction-level power model that is obtained through empirical measurements. The optimal instruction scheduling problem based on such a power model is formulated as a chance-constrained rough program which is solved by a problem-specific genetic algorithm. Efficiency of the algorithm is greatly improved through a novel rule-based approach to rank the intermediate candidate schedules. Experimental results using the MediaBench and Trimaran benchmarks show that the near-optimal schedules obtained are significantly better than those obtained through the mixed-integer programming approach. Computational requirements are low enough for the technique to be adopted by practical compilers.