The efficiency of nonparametric methods based on residual analizes and parametric method to estimate hydrological model uncertainty

Despite modern scientific knowledge and computational power in hydrology, the key to properly addressing hydrologic uncertainty remains a critical and challenging one. Here, we applied lumped HBV hydrological model to describe the uncertainty in runoff prediction in Chehl-Chay watershed in Golestan province. We applied a new framework for uncertainty analysis that is rooted on ideas from predicting model residual uncertainty. The uncertainty calculated by local Errors and Clustering (EEC) is compared with estimates from two non parametric methods (quantile regression (QR) and random forest (RF)) and a parametric method (GLUE). Firstly, the model parameters were optimized by Shuffled Complex Evolution approach and model residuals of test data were computed. Fuzzy clustering in EEC is carried out by the fuzzy c-means method and employs four clusters, predictive discharges, observed discharges, rainfall values and residuals in study basin. The results of this case study show that the uncertainty estimates obtained by EES which is trained by SVM gives wider uncertainty band and RF gives narrower uncertainty band. The best overall uncertainty estimates according to the PICP, MPI and ARIL indices were obtained with QR and then EEC. In comparison with non-parametric, with respct to all indices nonparametric methods had better performance than GLUE method.