With China's social and economic rapid development, the conflict between the water supply and demand had become increasingly conspicuous, and would have the further deterioration tendency. Therefore, to forecast water demand was very important for the water resources reasonable plan and the optimized allocation. There were a variety of methods to predict the water demand; however, for different cities or regions, it was necessary to choose the right forecasting method to predict the water demand. Take Changzhi city in Shanxi Province as the modeling object, and then use the gray model GM(1,1) and the RBF (radial basis function) neural network to predict the water demand of Changzhi city in 1998、1999 and 2000 years. Two prediction models have high prediction accuracy. The maximum relative error of the gray model GM (1, 1) was 8.16%, the average absolute relative error was 4.08%; the maximum relative error of the RBF neural network model was 6.89%, the average absolute relative error was 2.74%. The results shown that the prediction accuracy of the RBF neural network model was better than the gray model GM (1, 1), and the forecast for years was longer than the gray model GM (1, 1), so the RBF neural network was more suitable to forecast water demand for Changzhi city.