There are too many impact factors of water demand in the urban water demand prediction model and most of the factors are multicollinear. Besides, the BP neural network has slow convergence rate and easily gets into a local optimum. To tackle these problems, we proposed an improved prediction model by combining the principal component analysis (PCA), genetic algorithm (GA), and back propagation neural network (BPNN). Taizhou city was taken as a case for study. We established a water demand prediction model that selects the main impact factors of water demand by principal component analysis and optimizes the connection weights and thresholds of the BP neural network by genetic algorithm. The BP neural network prediction model was set up as the contrast model. The results showed that the average relative error and the maximum relative error of water demand prediction by the improved model in 2003-2014 in Taizhou city were 0.564% and 1.681% respectively. The precision was superior to that of the BP neural network prediction model. The results predicted by the GA-BP model matched with the actual water demand data of Taizhou city, and the model had faster calculation speed and higher precision. It can be used as an effective method for water demand prediction.