Clustering as Approximation Method to Optimize Hydrological Simulations

Accurate water-related predictions and decision-making require a simulation of hydrological systems in high spatio-temporal resolution. However, the simulation of such a large-scale dynamical system is compute-intensive. One approach to circumvent this issue, is to use landscape properties to reduce model redundancies and computation complexities. In this paper, we extend this approach by applying machine learning methods to cluster functionally similar model units and by running the model only on a small yet representative subset of each cluster. Our proposed approach consists of several steps, in particular the reduction of dimensionality of the hydrological time series, application of clustering methods, choice of a cluster representative, and study of the balance between the uncertainty of the simulation output of the representative model unit and the computational effort. For this purpose, three different clustering methods namely, K-Means, K-Medoids and DBSCAN are applied to the data set. For our test application, the K-means clustering achieved the best trade-off between decreasing computation time and increasing simulation uncertainty.