Multi-parameter approaches for improved ensemble prediction accuracy in hydrology and water quality modeling

Abstract

Ensemble approaches can be a quick way to improve hydrological prediction accuracy by considering multiple plausible modeling representations of a system. Previous studies showed that the use of multiple types of models could increase prediction accuracy, but model selection subjectivity and parameter equifinality issues remain unsolved. This study proposes to consider multiple parameter sets identified in the model calibration process and information on their performance for improved accuracy of ensemble hydrological prediction. In this study, we investigate whether multi-parameter ensemble (MP) compared to multi-model ensemble (MM) approaches can improve the accuracy of hydrology and water quality prediction. A variety of ensemble methods, including simple averaging, median filtering, weighted averaging, Bayesian model averaging (BMA), and their variants, were tested in watershed modeling implemented to predict daily water discharge and total phosphorus (TP) loads of a study watershed. The parameter spaces of two watershed loading models, Soil and Water Assessment Tool (SWAT) and Hydrological Simulation Program-FORTRAN (HSPF), were sampled using a multi-objective heuristic optimization algorithm, namely A Multi-Algorithm Genetically Adaptive Multiobjective (AMALGAM). The modeling experiment showed that the MP approach more effectively improved modeling accuracy statistics compared to the MM of both models as well as single-model and single-parameter approaches. The ensemble approaches did not always improve modeling accuracy, and their efficiency depended on ways to determine weights for ensemble members. The BMA method outperformed the other weighting schemes tested in this study, and its performance was responsive to the number and diversity (or variance) of ensemble member candidates and the computational cost to create the candidates. The findings demonstrate how the information of parameter sets sampled in model calibration and their performance statistics could help improve hydrological prediction accuracy.