This paper presents two heuristic self-organization methods for construction of a non-linear prediction model of river flows. The self-organization algorithms have multi-layered structures of the perceptron type and provide the optimally complex non-linear equation of the input-output relation. The algorithms are applied to the river-flow prediction of Karasu JRiver and Katsura River in Japan. The performance of the prediction models by the self-organization methods is compared with that of the hydrological models. The numerical comparison shows that without any hydrological and geographical knowledge the prediction models presented here aro superior to the elaborate hydrological models.