A Mixture-Density-Network Based Approach for Finding Rating Curves: Facing Multi-Modality and Unbalanced Data Distribution

Abstract

In this paper, the use of MDNs (Mixture Density Networks) is proposed for deciding rating curves. This method is beneficial especially when a single curve is developed when the relation between stage and discharge exhibits hysteresis. The computational analyses performed for the Han River and Mokkye stations showed that the MDN-based method yields more meaningful results than the conventional least squares approach. Of particular significance was the possible identification of the bi-modal characteristics of rating curves under the proposed method.