Combining large-scale and regional hydrological forecasts using simple methods

The development and expanded application of large-scale hydrological models has produced forecasts that often overlap with more targeted, regional hydrological forecasts. Here the possibility is explored for using simple methods to combine forecasts from a large-scale model, the Great Lakes portion of the National Surface and River Prediction System (NSRPS), and a regional system, the Systeme de Prevision Hydrologique (SPH) which covers southern Quebec, to improve regional forecasts. Outputs from the two forecasting systems are combined using multiple methods, including the simple mean, a weighted average in which the weights are optimized using the Kling-Gupta Efficiency (KGE), the Reduced Continuous Ranked Probability Score (RCRPS), and Ignorance Score (IGN) as cost functions, and weights calculated from the residual errors of the models. Bayesian Model Averaging (BMA) is also used to combine the probabilistic forecasts from both systems. The results show that it is possible to improve regional hydrological forecasts by using simple weighted combinations with forecasts from the large-scale system, even though the regional system performs clearly better. Performance is assessed via many well-known metrics, such as Nash-Sutcliffe Efficiency (NSE), KGE, RCRPS, and IGN. Results are averaged over 40 gauging stations and analyzed at lead times from 3 to 120 h. Improvements in all criteria for lead times over 60 h are observed, and there is no loss in performance at any lead times. Finally, the methods are used in a leave-one-out setup containing 29 validation basins to simulate performance on ungauged basins. The performance gain for ungauged basins is similar to that of the gauged basins, demonstrating that these simple methods can also improve forecasts in more remote territories where no gauging is available. Le developpement et l'application etendue des modeles hydrologiques large-echelle ont produit des previsions qui chevauchent regulierement des previsions hydrologiques regionales, plus specifiques a un territoire. Cette etude porte sur l'utilisation de methodes simples permettant la combinaison des previsions large-echelle de la partie des Grands Lacs du National Surface and River Prediction System (NSRPS) aux previsions regionales du Systeme de Prevision Hydrologique (SPH) qui couvre le sud du Quebec, afin d'ameliorer ces previsions regionales. Les sorties de ces deux systemes sont combinees selon plusieurs methodes incluant la moyenne simple, des moyennes ponderees qui utilisent comme fonction de cout le Kling-Gupta Efficiency (KGE), le Reduced Continuous Ranked Probability Score (RCRPS) et l'Ignorance Score (IGN), en plus de poids calcules selon les residus des modeles. Le Bayesian Model Averaging (BMA) est aussi utilise pour combiner les previsions des systemes. Les resultats montrent qu'il est possible d'ameliorer les previsions hydrologiques regionales en utilisant de simples combinaisons ponderees avec les previsions du systeme large-echelle, et ce meme si le systeme regional performe mieux que le systeme large-echelle. La performance est evaluee selon plusieurs metriques bien connues, telles que le Nash-Sutcliffe Efficiency (NSE), KGE, RCRPS et IGN. Les resultats sont moyennes sur 40 stations de jaugeage et analyses pour des horizons de 3 a 120 h. Des ameliorations sont observees sur tous les criteres pour des horizons de 60 h et plus, et aucune perte de performance n'est observee sur tous les horizons. Finalement, les methodes de combinaison sont utilisees dans une configuration << leave-one-out >> contenant 29 bassins de validation afin de simuler la performance des combinaisons sur les bassins non-jauges. Le gain en performance pour les bassins non-jauges est similaire au gain des bassins jauges, demontrant que ces methodes simples pourraient aussi ameliorer les previsions des territoires plus eloignes ou des stations de jaugeage ne sont pas disponibles.