A short-term load forecasting neural network model based on rough sets is established to predict future power demand one day to one week ahead. The model is validated by a set of data including both weather factors from Central Weather Bureau and history load from Taiwan Power Company. Rough set is applied to modify the model by eliminating redundant attributes. The impact of preprocessing increases the reliability of the neural network model, which successfully avoids the complexity of the system; hence it performs better than the traditional neural network model. The model for forecasting is simulated in MATLAB and the results can increase the reliability of power system to further help energy economic and strategy planning.