Back to overview

Calibrating stochastic hydrological models to signatures

Applicant Albert Carlo
Number 169295
Funding scheme Project funding
Research institution Swiss Federal Institute of Aquatic Science and Technology (EAWAG)
Institution of higher education Swiss Federal Institute of Aquatic Science and Technology - EAWAG
Main discipline Hydrology, Limnology, Glaciology
Start/End 01.04.2017 - 30.06.2022
Approved amount 174'108.00
Show all

All Disciplines (2)

Hydrology, Limnology, Glaciology

Keywords (4)

Hydrological signatures; Stochastic Models; multi-model ; Approximate Bayes Computations

Lay Summary (German)

Dieses Forschungsprojekt setzt sich zum Ziel, mit Hilfe von verbesserten Regen-Abfluss-Modellen und Inferenzverfahren zu zuverlässigeren Vorhersagen von Abflusssignaturen (Spitzen, Ganglinien, etc.) in natürlichen Einzugsgebieten zu gelangen.
Lay summary

Ziele des Forschungsprojektes:

Es ist seit langem bekannt, dass auf grossen Zeitskalen sowohl Regen- als auch Abflusszeitreihen bemerkenswert universellen Skalengesetzen folgen. Dies deutet darauf hin, dass viele der klein-skaligen Prozesse im Zustandekommen dieser gross-skaligen Gesetze keine Rolle spielen, und letztere daher mit relativ einfachen stochastischen Modellen beschrieben werden können. Dies wollen wir ausnützen um zu zuverlässigeren Abflussprognosen insbesondere von Extremereignissen zu gelangen.

Stochastische Modelle an gemessenen Daten zu kalibrieren und die resultierende Parameterunsicherheit abzuschätzen ist eine sehr rechenintensive Aufgabe, welche wir mit unseren neu entwickelten Algorithmen angehen werden. Dies wird uns erlauben die Unsicherheiten in den geschätzten Modell-Parametern zuverlässiger zu schätzen und damit auch die Zuverlässigkeit der Prognosen zu erhöhen.

Eine einfache Art und Weise stochastische Modelle zu kalibrieren besteht darin, bloss gewisse Signaturen von Regen und Abflusszeitreihen mit entsprechenden simulierten Signaturen zu vergleichen. Dies ist insbesondere für Einzugsgebiete ohne gemessene Ganglinien interessant. Oft kann man nämlich aufgrund klimatischer und geologischer Eigenschaften des Einzugsgebietes dennoch Aussagen über gewisse Regen-Abfluss-Signaturen machen. Skalengesetze sind bisher in der Hydrologie im Zusammenhang mit Signaturen kaum berücksichtigt worden. Diese dürften aber wichtige Informationen über das zugrunde liegende Einzugsgebiet enthalten.

Wissenschaftlicher und gesellschaftlicher Kontext:

Zuverlässigere Prognosen von Abflusscharakteristika sind sowohl für den Hochwasserschutz als auch für die Prognose der Gewässerqualität sehr wichtig. Neben dieser Bedeutung für die Hydrologie werden die Einsichten aus diesem Projekt auch für die nichtlineare Zeitreihenanalyse im allgemeinen nützlich sein.

Direct link to Lay Summary Last update: 01.11.2016

Responsible applicant and co-applicants


Scientific events

Active participation

Title Type of contribution Title of article or contribution Date Place Persons involved
EGU Talk given at a conference Bayesian parameter inference in hydrological modelling using a Hamiltonian Monte Carlo approach with a stochastic rain model 23.05.2022 Vienna, Austria Albert Carlo; Ulzega Simone;
EGU Poster Scaling Laws in River Runoff 04.04.2018 Vienna, Austria Coutandin Thomas; Albert Carlo;

Associated projects

Number Title Start Funding scheme
152824 Using Commercial Microwave Links and Computer Model Emulation to Reduce Uncertainties in Urban Drainage Simulations (COMCORDE) 01.09.2014 Interdisciplinary projects


Input uncertainty, model deficits and randomness due to the aggregated description of the system lead to problems in the calibration of hydrological models.If they are not carefully accounted for, they lead to incorrect uncertainty estimates of predictions.Large research efforts over the past 20 years led to many important insights, but the basic underlying problems are still not adequately considered in current hydrological modeling efforts.Indications of these problems are (i) systematic deviations of model results from observations, (ii) incorrect signatures of model results, such as peak heights, discharge quantiles, and shapes of recession curves, and (iii) poor quantification of the prediction error.The objective of this proposal is to make a further step in improving hydrological models, their calibration, and the quantification of their prediction uncertainty.This is intended to be done by bringing together knowledge and methodologies from hydrology, statistics and numerics.In particular, a multi-model approach, facilitated by the framework implemented by one of the co-PIs, will be used to learn as much as possible about hydrological mechanisms and transfer this knowledge into appropriate model structures.Statistical knowledge will be used to formulate rigorous stochastic models to appropriately consider intrinsic randomness caused by incomplete observations and aggregated model states (the same observed input can lead to different outputs as it can be associated with different real inputs).Finally, modern statistical techniques, in particular an Approximate Bayesian Computation (ABC) algorithm developed by the PI, will be applied where needed to deal with the complexity of likelihood functions of stochastic models.This approach guarantees an approximate, but statistically rigorous approach to Bayesian inference using (hydrology-inspired and automatically generated) signatures of hydrological time series.Some of these approaches have been applied to hydrology already, but testing them jointly in a unified framework is expected to lead to new and innovative results.In particular, we expect that calibration based on time series or based on signatures will deviate less from each other when applied to the most appropriate, stochastic hydrological model compared to the experience from just applying them to deterministic simulation models as it has been demonstrated in the literature (this led to quite different calibration results and to a discussion about which kind of calibration, likelihood-based or ``multi-objective'' calibration, is more appropriate).This would confirm that the suggested approach is more robust regarding the calibration endpoints and would thus be an indication of a better representation of the real system by the model.This hypothesis will be explored by testing the suggested techniques with a variety of models that differ in the structure of the deterministic model and the formulation of stochasticity, and by applying these models to data from catchments of different complexity.The expected results of the study are (i) to gain more insight into the causes of the calibration problems of hydrological models, (ii) to suggest better model structures and inference techniques to decrease the severity of these problems, and (iii) to get improved estimates of prediction uncertainty of hydrological models.All three points are of high scientific interest, points (ii) and (iii) are also very important for the practice of hydrological forecasting for flood protection, water quality simulation, etc.