Owing to the fluctuation of water quality in urban river which polluted by drainage along river, one-dimension uncertain water quality model embeded neural network is established. Genetic algorithms and a modified fitness function are used to optimize parameters of the uncertain model. Examples illustrate that the uncertain model has higher prediction accuracy with the average accuracy over 80% than the certain model, and is more sensitive to the fluctuation of pollutants discharged into the river. The uncertain model has a significant advantage of prediction and could better adapt to the changing urban water environment, especially at points close to the pollution sources.