Ensemble and probabilistic hydro-meteorological forecasts: predictive uncertainty, verification and decision making 

This session brings together scientists, forecasters, practitioners and stakeholders interested in exploring the use of ensemble hydro-meteorological forecast techniques in hydrological applications: e.g., flood control and warning, reservoir operation for hydropower and water supply, transportation, and agricultural management. It will address the understanding of sources of predictability and quantification and reduction of predictive uncertainty of hydrological extremes in deterministic and ensemble hydrological forecasting. Uncertainty estimation in operational forecasting systems is becoming a more common practice. However, a significant research challenge and central interest of this session is to understand the sources of predictability and development of approaches, methods and techniques to enhance predictability (e.g. accuracy, reliability etc.) and quantify and reduce predictive uncertainty in general. Ensemble data assimilation, NWP preprocessing, multi-model approaches or hydrological postprocessing can provide important ways of improving the quality (e.g. accuracy, reliability) and increasing the value (e.g. impact, usability) of deterministic and ensemble hydrological forecasts. The models involved with the methods for predictive uncertainty, data assimilation, post-processing and decision-making may include machine learning models, ANNs, catchment models, runoff routing models, groundwater models, coupled meteorological-hydrological models as well as combinations (multimodel) of these. Demonstrations of the sources of predictability and subsequent quantification and reduction in predictive uncertainty at different scales through improved representation of model process (physics, parameterization, numerical solution, data support and calibration) and error, forcing and initial state are of special interest to the session.

The session welcomes new experiments and practical applications showing successful experiences, as well as problems and failures encountered in the use of uncertain forecasts and ensemble hydro-meteorological forecasting systems. Case studies dealing with different users, temporal and spatial scales, forecast ranges, hydrological and climatic regimes are welcome.

The session is part of the HEPEX international initiative: www.hepex.org

Convener: Albrecht Weerts | Co-conveners: Trine Jahr Hegdahl, Schalk Jan van Andel, Fredrik Wetterhall
vPICO presentations
| Mon, 26 Apr, 15:30–17:00 (CEST)

vPICO presentations: Mon, 26 Apr

Chairpersons: Albrecht Weerts, Schalk Jan van Andel, Trine Jahr Hegdahl
Amina Msilini, Pierre Masselot, and Taha B.M.J. Ouarda

Hydrological processes and phenomena are naturally complex and nonlinear. Many physiographical variables such as those dealing with drainage network characteristics may influence streamflow characteristics and should be considered in regional frequency analysis (RFA). These variables have hence a significant impact on the effectiveness of flood quantile estimation techniques. Although many statistical tools are considered to estimate flood quantiles at ungauged sites in the hydrological literature, little attention has been given to the nonlinearity and to the high-dimensionality of physio-meteorological variable space. In this study, the multivariate adaptive regression splines (MARS) approach is introduced in RFA. This model allows to account simultaneously for non-linearity and interactions between variables hidden in high-dimensional data. MARS is hereby applied on two datasets of 151 hydrometric stations located in the southern part of the province of Quebec (Canada): a standard dataset (STA) including commonly used variables and an extended dataset (EXTD) combining STA with additional variables dealing with drainage network characteristics. It is then compared to generalized additive models (GAM), a state-of-the-art method for regional estimation. Numerical results show that MARS outperforms GAM, especially with the extensive database EXTD. The study suggests that MARS may be a promising tool to take into account the complexity of the hydrological phenomena involved and the increasing number of variables used in RFA.

How to cite: Msilini, A., Masselot, P., and Ouarda, T. B. M. J.: Accounting for high-dimensional predictors in RFA with MARS , EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-299, https://doi.org/10.5194/egusphere-egu21-299, 2021.

Rahim Barzegar, Jan Adamowski, and John Quilty

Hydrological time series modeling is an important task in water resources planning and management. However, time series may include noise, which can result in an inaccurate model. Therefore, removing noise from time series is valuable to obtain accurate predictions. The aims of this study are i) to develop and compare Long-Short Term Memory (LSTM) and Gated Recurring Units (GRU) Deep Learning (DL) models to predict hydrological time series and ii) to integrate a preprocessing method, Gaussian Filter (GF), to smooth out time series and couple it with DL to improve prediction accuracy. Moreover, the DL models are benchmarked against statistical time series models (e.g., Seasonal Autoregressive Integrated Moving Average (SARIMA)) to assess their added value for hydrological time series modeling. To establish predictive models, several monthly hydrological time series including water level (e.g., from the Great Lakes in North America, including Lakes Michigan, Ontario, and Erie (1918-2019)) and streamflow (e.g., gauging stations at Umfreville, along the English River, Ontario, Canada (1921-2019), Rapides Fryers, along the Richelieu River, Quebec, Canada (1937-2020) and near Lethbridge, along the Oldman River, Alberta, Canada (1957-2019)) were explored. For developing non-GF- and GF-DL models, time series were partitioned into training (70% of the data) and testing (the remaining 30% of the data) subsets and the time series’ past measurements up to 12 months (t-1, t-2, ..., t-12) were served to the DL models (LSTM and GRU) to predict the time series at time t. The structure of the DL models was tuned using Bayesian optimization. The SARIMA models (i.e., non-GF- and GF-SARIMA) were also implemented and tuned using pmdarima's auto-arima function. After calibrating the models, the testing step was implemented and the performance of the models was evaluated using statistical indicators including correlation coefficient, root mean square error, mean absolute error, the Nash-Sutcliffe efficiency coefficient, and Willmot’s index. The results of the developed DL models showed that the GRU outperforms the LSTM models. Moreover, both LSTM and GRU have superior performance when compared to the SARIMA models. It is observed that GF preprocessing significantly improves the accuracy of the developed DL and SARIMA models. It is concluded that coupling GF preprocessing with DL, due to capturing both linear and nonlinear features of the time series, represents a promising tool for obtaining accurate hydrological time series predictions.

How to cite: Barzegar, R., Adamowski, J., and Quilty, J.: Improving Deep Learning hydrological time series modeling using Gaussian Filter preprocessing, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-1644, https://doi.org/10.5194/egusphere-egu21-1644, 2021.

Mark Thyer, David McInerney, Dmitri Kavetski, Richard Laugesen, Narendra Tuteja, and George Kuczera

Sub-seasonal streamflow forecasts (with lead times of 1-30 days) provide valuable information for many consequential water resource management decisions, including reservoir operation to meet environmental flow and irrigation demands, issuance of early flood warnings, and others. A key aim is to produce “seamless” forecasts, with high quality performance across the full range of lead times and time scales.  

This presentation introduces the Multi-Temporal Hydrological Residual Error model (MuTHRE) to address the challenge of obtaining “seamless” sub-seasonal forecasts, i.e., daily forecasts with consistent high-quality performance over multiple lead times (1-30 days) and aggregation scales (daily to monthly).

The model is designed to overcome common errors in streamflow forecasts:

  • Seasonality
  • Dynamic biases due to hydrological non-stationarity
  • Extreme errors poorly represented by the common Gaussian distribution.

The model is evaluated comprehensively over 11 catchments in the Murray-Darling Basin, Australia, using multiple performance metrics to scrutinize forecast reliability, sharpness and bias, across a range of lead times, months and years, at daily and monthly time scales.

The MuTHRE model provides ”high” improvements, in terms of reliability for

  • Short lead times (up to 10 days), due to representing non-Gaussian errors
  • Stratified by month, due to representing seasonality in hydrological errors
  • Dry years, due to representing dynamic biases in hydrological errors.

Forecast performance also improved in terms of sharpness, volumetric bias and CRPS skill score; Importantly, improvements are consistent across multiple time scales (daily and monthly).

This study highlights the benefits of modelling multiple temporal characteristics of hydrological errors, and demonstrates the power of the MuTHRE model for producing seamless sub-seasonal streamflow forecasts that can be utilized for a wide range of applications.




How to cite: Thyer, M., McInerney, D., Kavetski, D., Laugesen, R., Tuteja, N., and Kuczera, G.: Do you want Seamless Subseasonal Streamflow Forecasts?    Ask MuTHRE!, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-3714, https://doi.org/10.5194/egusphere-egu21-3714, 2021.

José Quevedo, Daniel Firmo Kazay, Mariana Maria Werlang, Giovanni Gomes, Roberto Takahashi, Marcelo Zaicovski, Tannia Villanueva Aguero, and José Maria Fariña Jara

This work presents the development of the inflow ensemble forecasting system for Itaipu Dam. The system is based on combination of twelve Quantitative Precipitation Forecast (QPF) with three hydrological models and one hydrodynamic one-dimensional model. The QPF are provided by different meteorological institutions based in the results of the global Numerical Weather Prediction (NWP) models GFS and ECMWF, as well as of the regional NWP models WRF and COSMO executed by the Brazilian and Paraguay meteorological services (SIMEPAR, INMET and DINAC). The semi-distributed model MGB – Large Basin Model, the lumped models SMAP and HEC-HMS are considered as the hydrological models. Furthermore, the computed flows are propagated in a HEC-RAS scheme designed with extensive field data from bathymetry of Parana River and tributaries. The daily results are presented as fifteen inflow scenarios that are considered for the definition of a unique flow forecast. Finally, that forecast is used for the electric generation scheduling of the power plant. Each of these fifteen methods performance was evaluated as well as the suitability of the system for it purposes. For a short-term forecast horizon (less than 4 days), the performances of the hydrological models forced by the different rain forecast are quite similar. However, it is remarkable the difference between the results of the three hydrological models for the same horizon. On the other hand, for medium-term horizon (more than 4 days) both hydrological and meteorological models have diverse behavior and contribute for a wide representation of the possible scenarios. Overall, it has been showed that the simulations are complementary and provides to the forecaster a general overview of the hydrologic situation. Nevertheless, at this moment, further analysis of accuracy and reliability of the prediction have not been realized, so the forecaster needs to appeal to its own expertise to assure the consistence of the scenarios for decision-making.

How to cite: Quevedo, J., Firmo Kazay, D., Maria Werlang, M., Gomes, G., Takahashi, R., Zaicovski, M., Villanueva Aguero, T., and Fariña Jara, J. M.: Coupling meteorological forecasts with hydrologic and hydraulic models: The Itaipu Dam ensemble inflow forecasting system, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-6228, https://doi.org/10.5194/egusphere-egu21-6228, 2021.

Mohammed Amine Bessar, François Anctil, and Pascal Matte

The quality of water level predictions is highly dependent on the success of the flow forecasts that inform the hydraulic model. Ensemble predictions, by considering several sources of uncertainty, provide more accurate and reliable forecasts. In this project, we aim to evaluate a water level ensemble prediction system coupling a hydraulic model to an ensemble streamflow prediction system accounting for 3 sources of uncertainty: meteorological data, hydrological processing (multimodel) and data assimilation to update the initial conditions. The hydraulic model is previously calibrated and validated and the roughness coefficients are adapted as a function of flow according to predefined relationships developed for several river segments. The forecasts reliability and accuracy are then assessed at each layer of the forecasting system and the outcomes are illustrated comparing the ensembles skills and reliability for the considered events. Overall, the results show that accounting of the hydrometeorological uncertainty improves the performances of the water level forecasts for different lead times.

How to cite: Bessar, M. A., Anctil, F., and Matte, P.: Evaluation of a short-term ensemble water level forecasting system: case of the Chaudière River, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-8000, https://doi.org/10.5194/egusphere-egu21-8000, 2021.

Louise Arnal, Martyn Clark, Vincent Vionnet, Vincent Fortin, Alain Pietroniro, and Andy Wood

Sub-seasonal to seasonal streamflow forecasts represent critical operational inputs for many water sector applications of societal relevance, such as spring flood early warning, water supply, hydropower generation, and irrigation scheduling. However, the skill of such forecasts has not risen greatly in recent decades despite recognizable advances in many relevant capabilities, including hydrologic modeling and S2S climate prediction. In order to build a continental-scale forecasting system that has value at the local scale, the sources and nature of predictability in the forecasts should be quantified and communicated. This can additionally help to target science investments for tangible improvements in the sub-seasonal to seasonal streamflow forecasting skill.

As part of the Canada-based Global Water Futures (GWF) program, we are advancing capabilities for probabilistic sub-seasonal to seasonal streamflow forecasts over North America. The overall aim is to improve sub-seasonal to seasonal streamflow forecasts for a range of water sector applications. We are implementing an array of forecasting methods that integrate state-of-the-art mechanistic models and statistical methods. These include, for instance, a probabilistic sub-seasonal to seasonal streamflow forecasting system based on quantile regression of snow water equivalent observations, and a system based on the ESP approach (Day, 1985).

To guide forecast system developments over North America, we are currently quantifying streamflow predictability for different hydroclimatic regimes, forecast initialization times, and lead times, against both streamflow simulations and observations to quantify the effect of model errors. Building on the work from Wood et al. (2016) and Arnal et al. (2017), we are disentangling the dominant predictability sources (i.e., initial hydrological conditions and atmospheric forcings) of sub-seasonal to seasonal streamflow across North American watersheds. The results provide insights into the elasticity of predictability, i.e., the increase in streamflow forecast skill possible by improving a specific component of the forecast system, and will inform the forecasting system development.

Arnal Louise, Wood Andrew W., Stephens Elisabeth, Cloke Hannah L., Pappenberger Florian, 2017: An Efficient Approach for Estimating Streamflow Forecast Skill Elasticity. Journal of Hydrometeorology, doi: 10.1175/JHM-D-16-0259.1

Day, Gerald N., 1985: Extended streamflow forecasting using NWSRFS. Journal of Water Resources Planning and Management, doi:10.1061/(ASCE)0733-9496(1985)111:2(157)

Wood, Andrew W., Tom Hopson, Andy Newman, Levi Brekke, Jeff Arnold, and Martyn Clark, 2016: Quantifying streamflow forecast skill elasticity to initial condition and climate prediction skill. Journal of Hydrometeorology, doi: 10.1175/JHM-D-14-0213.1

How to cite: Arnal, L., Clark, M., Vionnet, V., Fortin, V., Pietroniro, A., and Wood, A.: Quantifying streamflow predictability across North America on sub-seasonal to seasonal timescales, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-8092, https://doi.org/10.5194/egusphere-egu21-8092, 2021.

Céline Cattoën, Stuart Moore, and Trevor Carey-Smith

Flooding is New Zealand’s most frequent natural disaster with an average annual cost of approximately NZ$51 million. Accurately forecasting convective and orographically enhanced precipitation for hydrometeorological ensemble prediction systems is challenging in Aotearoa New Zealand’s complex topographic regions with fast-responding and mostly ungauged catchments. Globally, designing convection-permitting ensemble flood forecasting chains is still a work in progress, with errors in the forecast rainfall amount and the location or timing of storm events a significant contributor to uncertainties in river flow forecasts. Given operational, computational and model representation constraints, compromises are often required on ensemble size, frequency of forecast issue times, NWP model resolution, domain size and data assimilation strategies. This research aims to design an optimal operational forecasting chain for convective-scale flood forecasting in New Zealand.  In doing so, our goal is to improve uncertainty representation in hydrometeorological forecasts during flood events by understanding the impact of convective-scale ensemble strategies.

The NWP model used is a local implementation of the UK Met Office-developed Unified Model.  The New Zealand Convective-Scale Model (NZCSM) is NIWA’s 1.5km high-resolution operational forecast model, configured such that convective processes develop explicitly. The New Zealand Ensemble (NZENS) is configured with similar convection-permitting model physics but operates with a 4.5km horizontal resolution and features up to 18 members.  Flood forecasts were produced by coupling several weather ensemble configurations with the semi-distributed hydrological model TopNet and its built-in statistical ensemble generation tool. TopNet is based on TOPMODEL concepts of runoff generation controlled by sub-surface water storage.

In this study, we evaluated three ensemble strategies for flood forecasting. The experiment design allowed for the effect of model horizontal resolution (and thus the representation of orography) to be investigated using ensemble forecasts from consecutive initialization times (a “lagged ensemble”), and from the same initialisation time (a “dynamical ensemble”). The third forecasting chain is a “statistical ensemble” generated by perturbing the deterministic 1.5km NWP model and hydrological states. For recent flood events across multiple case study catchments, we evaluated the impact of each approach on flood forecast performance. Flood forecasts were most sensitive to convective-scale forecasts with consecutive issue time initialisations (lagged ensemble) over other hydrometeorological ensemble configurations considered. Given dynamical ensembles are computationally expensive, the study suggests an optimal strategy might be to produce a small ensemble pool of dynamical forecasts at more frequent issue times combined with statistically post-processed ensembles rather than a larger ensemble pool generated less frequently.

How to cite: Cattoën, C., Moore, S., and Carey-Smith, T.: Designing an optimal flood forecasting chain using convective-scale ensembles: a sensitivity study., EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-8098, https://doi.org/10.5194/egusphere-egu21-8098, 2021.

Helen Titley, Hannah Cloke, Shaun Harrigan, Florian Pappenberger, Christel Prudhomme, Joanne Robbins, Elizabeth Stephens, and Ervin Zsoter

Global ensemble forecast models have been shown to have good skill in forecasting the track probabilities of tropical cyclones worldwide, but less well-studied is their ability to predict the hazards resulting from tropical cyclones, which in the case of fluvial flooding can extend far from the landfall location traditionally focussed on in operational tropical cyclone warnings. This work aims to investigate the key factors that influence the predictability of fluvial flood severity from tropical cyclones, using forecasts from the Global Flood Awareness System (GloFAS). GloFAS is jointly developed by the European Commission and the European Centre for Medium-Range Weather Forecasts (ECMWF) and is designed to provide a global overview of upcoming flood events to decision makers as part of the Copernicus Emergency Management Service, producing probabilistic river discharge forecasts driven by global ECMWF ensemble forecasts coupled to a hydrological model. This presentation will explore the chain of uncertainty through the forecasting process for several recent tropical cyclone flood events including Hurricane Iota and Cyclone Nivar. It investigates the influence on the overall predictability and uncertainty of the fluvial flood forecasts of various components of the forecasting chain, including the track, intensity, and precipitation forecasts for the tropical cyclone, and the hydrological catchment conditions and modelling.

How to cite: Titley, H., Cloke, H., Harrigan, S., Pappenberger, F., Prudhomme, C., Robbins, J., Stephens, E., and Zsoter, E.: Investigating the forecast predictability for fluvial flooding from tropical cyclones, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-8888, https://doi.org/10.5194/egusphere-egu21-8888, 2021.

Roya Narimani and Jun Changhyun

The quality and completeness of rainfall data have always played an important role in time series analysis and prediction for future water-related disasters. It requires to estimate missing data correctly for better results of rainfall prediction with high accuracy. In recent years, multilayer perceptron (MLP) neural networks have been applied to solve stochastic problems in data science. This study suggests a novel approach for estimating missing rainfall data with MLP neural networks. For this purpose, a mathematical model was created to analyze and predict the time series of daily rainfall data from 2003 to 2017 at six rain gauge stations in Seoul, Korea. Here, rainfall data with missing values during 20 days of time periods was considered for reconstruction of missing data at one specific rain gauge station from complete rainfall data records at five different stations. They were divided into training, validation, and testing datasets with a percentage of 70%, 15%, and 15%, respectively. This study investigates an effect of changes in data periods considered in MLP neural networks and it indicates that rainfall time series for a longer time period play a more effective role in rainfall data reconstruction.

How to cite: Narimani, R. and Changhyun, J.: Multilayer Perceptron Neural Networks for Estimating Missing Rainfall Data, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-10542, https://doi.org/10.5194/egusphere-egu21-10542, 2021.

Ankit Singh, Soubhik Mondal, and Sanjeev Kumar Jha

Short-term streamflow forecast is important for various hydrological applications such as, estimating inflow to reservoirs, sending alarms in case of extreme events like flood and flash floods etc. Flooding events in last few years in the Indian subcontinent emphasized the importance of more accurate streamflow forecasts and the possible benefit of high-resolution Numerical Weather Prediction (NWP) models has been confirmed. In India, National Center for Medium Range Weather Forecasting (NCMRWF) provides rainfall forecasts from its UK Met office Unified Model based deterministic model (NCUM), and ensemble prediction system (NEPS). The comparison of NCMRWF with the forecast from other agencies such as Japan Metrological Agency (JMA)and European Center for Medium Range Forecast (ECMWF) have been addressed in this work. Global NWP models developed by different international agencies applydifferent algorithms, initial and boundaries conditions.The usefulness of several forecasts in streamflow forecasting is still being investigated in India. Recent studies on streamflow forecasting by using different NWP models shows that the performance of streamflow forecasts directly depends on the skill of NWP models. Hydrological model also plays a vital role in stream flow forecasting, because different hydrological model have different structure, parameters and algorithms to simulate the flow.

            In this study we use the Soil and Water Assessment Tool (SWAT) a Hydrological Response Unit (HRU’s) based hydrological model. HRU is the area that contains similar type of soil, land use and slope properties in a subbasin. For comparison, the streamflow generated from the forecasted rainfall by NWP, we select three different NWP models namely JMA, ECMWF and NCMRWF for streamflow forecasting. Manot watershed part of Narmada River basin in central India is selected as the study area for this study. Streamflow is examined for monsoon (June to September) period of 2018 at multiple lead times i.e. 1 to 5 days. Rain-gauge based gridded Indian Meteorological Department (IMD) rainfall product is used as observed data in SWAT. All rainfall products are at 0.25*0.25-degree spatial resolution. The preliminary comparison between the simulated streamflow and the observation shows that the stream flow patterns produced by various forecast products are in good comparison with high peaks. Our results also indicate that the forecast accuracy of NCMRWF is closely comparable with other forecast products for all lead time. In addition, the setup of Variable Infiltration Capacity (VIC), the hydrological model for Streamflow forecasting is in progress. The VIC model is a grid-based model with variable infiltration soil layers and each of this layer characterizes the soil hydrological responses and heterogeneity in land cover classes. For routing, VIC model divides the whole basin into grides and water balance is calculated at the outlet of each and every grid and the flow simulate according to the flow direction. This model considers both the baseflow and the surface flow. The detailed results of ongoing work will be presented at the conference.

How to cite: Singh, A., Mondal, S., and Jha, S. K.: Developing semi-distributed hydrological models for streamflow forecasting in upper Narmada River basin, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-11770, https://doi.org/10.5194/egusphere-egu21-11770, 2021.

Ignacio Martin Santos, Mathew Herrnegger, and Hubert Holzmann

The skill of seasonal hydro-meteorological forecasts with a lead time of up to six months is currently limited, since they frequently exhibit random but also systematic errors. Bias correction algorithms can be applied and provide an effective approach in removing historical biases relative to observations. Systematic errors in hydrology model outputs can be consequence of different sources: i) errors in meteorological data used as input data, ii) errors in the hydrological model response to climate forcings, iii) unknown/unobservable internal states and iv) errors in the model parameterizations, also due to unresolved subgrid scale variability.

Normally, bias correction techniques are used to correct meteorological, e.g. precipitation data, provided by climate models. Only few studies are available applying these techniques to hydrological model outputs. Standard bias correction techniques used in literature can be classified into scaling-, and distributional-based methods. The former consists of using multiplicative or additive scaling factors to correct the modeled simulations, while the later methods are quantile mapping techniques that fit the distribution of the simulation to fit to the observations. In this study, the impact of different bias correction techniques on the seasonal discharge forecasts skill is assessed.

As a case study, a seasonal discharge forecasting system developed for the Danube basin upstream of Vienna, is used. The studied basin covers an area of around 100 000 km2 and is subdivided in 65 subbasins, 55 of them gauged with a long historical record of observed discharge. The forecast system uses the calibrated hydrological model, COSERO, which is fed with an ensemble of seasonal temperature and precipitation forecasts. The output of the model provides an ensemble of seasonal discharge forecasts for each of the (gauged) subbasins. Seasonal meteorological forecasts for the past (hindcast), together with historical discharge observations, allow to assess the quality of the seasonal discharge forecasting system, also including the effects of different bias correction methods. The corrections applied to the discharge simulations allow to eliminate potential systematic errors between the modeled and observed values.

Our findings generally suggest that the quality of the seasonal forecasts improve when applying bias correction. Compared to simpler methods, which use additive or multiplicative scaling factors, quantile mapping techniques tend to be more appropriate in removing errors in the ensemble seasonal forecasts.

How to cite: Martin Santos, I., Herrnegger, M., and Holzmann, H.: Impact of bias correction techniques on an ensemble of seasonal discharge forecast for the Danube upstream of Vienna, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-12145, https://doi.org/10.5194/egusphere-egu21-12145, 2021.

Rakesh Kumar Sinha and T.I. Eldho

The estimation of the impacts of climate change on hydrology at the local level comprises various sources of uncertainty. Especially, global climate models (GCMs) are found to be one of the major sources of uncertainty at the local level and it is important to identify for robust water resource planning and management. Therefore, this study demonstrates the separate and multi-model ensemble GCMs uncertainty for the surface runoff projections for near, mid, and far future under representative concentration pathway (RCP) 4.5 (present condition) and RCP 8.5 (worst condition) at medium level river sub-basins scale in the Western Ghats region of India. The results indicate that considered GCMs are not appropriate for use to prediction of peak surface runoff in the wet season. In addition, uncertainty from ensemble GCMs is closer to actual data than individual GCM because of closely associated with ensemble rainfall data which is maximum influencing the peak surface runoff for the near mid, and far future. Furthermore, findings also suggest that the selection of appropriate GCMs for the study of peak flow analysis at the local level is important for the projection of future surface runoff. Therefore, it is also important to make attention to rainfall data while projecting surface runoff for future time periods in the humid tropic regions.

How to cite: Sinha, R. K. and Eldho, T. I.: Estimation of Uncertainty Contribution of Multiple Sources of GCMs in Hydrological Prediction., EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-12896, https://doi.org/10.5194/egusphere-egu21-12896, 2021.

Oliver Bent, Julian Kuehnert, Sekou Remy, Anne Jones, and Blair Edwards
The increase in extreme weather associated with acute climate change is leading to more frequent and severe flood events.  In the window of months and years, climate change adaption is critical to mitigate risk on socio-economic systems. Mathematical and computational models have become widely used tools to quantify the impact of catastrophic flooding and to predict future flood risks. For decision makers to plan ahead and to select informed policies and interventions, it is vital that the uncertainties of these models are well estimated. Besides the inherent uncertainty of the mathematical model, uncertainties arise from parameter calibration and the driving observational climate data.
Here we focus on the uncertainty of seasonal flood risk prediction for which we treat uncertainty propagation as a two step process. Firstly through calibration of model parameter distributions based on observational data. In order to propagate parameter uncertainties, the posed calibration framework is required to infer model parameter posterior distributions, as opposed to a single best-fit estimate. While secondly uncertainty is propagated by the seasonal weather forecasts driving the flood risk prediction models, such model drivers have their own inherent uncertainty as predictions. Through handling both sources of uncertainty and its propagation we investigate the impacts of combined uncertainty quantification methods for flooding predictions. The first step focussing on the flooding models own characterisation of uncertainty and the second characterising how uncertain model drivers impact our future predictions.
In order to achieve the above features of a calibration framework for flood models we leverage concepts from machine learning. At the core we assume a minimisation of a loss function by the methods based on the supervised learning task in order to achieve calibration of the flood model. Uncertainty quantification is equally a growing field in machine learning or AI with regards the interpretability of parametric models. For this purpose we have adopted a Bayesian framework which contains natural descriptions of model expectation and variance. Through combining uncertainty quantification with the steps of supervised learning for parameter calibrations we propose a novel approach for seasonal flood risk prediction.

How to cite: Bent, O., Kuehnert, J., Remy, S., Jones, A., and Edwards, B.: Machine learning approaches to parameter calibration and uncertainty propagation for seasonal flood risk prediction , EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-13262, https://doi.org/10.5194/egusphere-egu21-13262, 2021.

Farshid Rahmani, Kathryn Lawson, Samantha Oliver, Alison Appling, and Chaopeng Shen

Stream water temperature (Ts) is a variable that plays a pivotal role in managing water resources. We used the long short-term memory (LSTM) deep learning architecture to develop a basin centric single Ts model based on general meteorological data and basin meteo-geological attributes. We created a strong tool for long-term Ts projection and subsequently, improved the Ts model using novel approaches. We investigated the impact of both observed and simulated streamflow data on improving the model accuracy. At a national scale, we obtained a median root-mean-square error (RMSE) of 0.69 oC, and Nash-Sutcliffe model efficiency coefficient (NSE) of 0.985, which are marked improvements over previous values reported in previous studies. In order to test the performance of the model on basins ranging from basins with extensive data to unmonitored basins, we used more than 400 basins with different data-availability groups (DAG) across the continent of the United States to explore how to assemble the training dataset for both monitored and unmonitored basins. Best root-mean-square error (RMSE) for sites with extensive (99%), intermediate (60%), scarce (10%) and absent (0%) data for training were 0.75, 0.837, 0.889, and 1.595 oC, respectively. We observed the negative effect of the presence of reservoirs in Ts modeling. Our results illustrated that the most suitable training set should be different in modeling basins with different availability of observed data. for predicting Ts in a monitored basin, including basins that have at least equal DAG with that particular basin will result in most accurate predictions, however, for Ts prediction in ungauged basin, including all basins in training section will generate the best model, showing a more diverse training set. Furthermore, to decrease overfitting produced by attributes for PUB application, we could improve the accuracy of the model using input-selection ensemble method. We got median correlation higher than 0.90 for PUB after seasonality was removed which is still high. While many Ts prediction models showed better performance in summer, our model was on the opposite side. We found a strong relationship between general available daily meteorological variables and catchment attributes with the presented Ts model. However, our results indicate that combining physics-based criteria to the model can improve the prediction of temperature in river networks.


How to cite: Rahmani, F., Lawson, K., Oliver, S., Appling, A., and Shen, C.: Deep Learning Stream Temperature Model: Recommendations for Modeling Gauged and Ungauged Basins, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-13960, https://doi.org/10.5194/egusphere-egu21-13960, 2021.

Rodrigo Valdés-Pineda, Juan B. Valdés, Sungwook Wi, Aleix Serrat-Capdevila, Roy Tirthankar, Eleonora M.C. Demaria, and Matej Durcik

The operational implementation of a Hydrologic Forecasting System (HFS) is limited in many catchments of the world by the lack of historical in-situ hydrologic data, i.e., long temporal records of rainfall or streamflow. By combining high-resolution Satellite Precipitation Products (SPPs), or Regional Climatological Models (RCMs), with Hydrologic Models, baselines can be established for the quantification and reduction of total hydrologic uncertainty in ungauged basins. We have studied how Variational Ensemble Forecasting (VEF) can be combined with Machine Learning (ML) techniques to improve a hydrologic system representation – i.e., raw data processing, model training, model evaluation, model selection, forecasts post-processing, etc. The VEF-ML method is applied and assessed with three general Hydrologic Processing Hypotheses (HPH): (1) Hydrologic Pre-processing (HPR), (2) Hydrologic Processing (HP), and (3) Hydrologic Post-processing (HPP). The operational implementation of VEF-ML was evaluated in the Upper Zambezi River Basin (UZRB) and its sub-basins, by using multiple precipitation products, multiple hydrologic models, and multiple optimal parameter sets. This extended VEF configuration and its coupling with ML techniques (VEF-ML) allows increasing the number of hydrologic ensembles available for the generation of operational streamflow forecasts products. The performance of VEF-ML is evaluated by comparing two hydrologic learning strategies (HLS) i.e. inference- and pattern-based approaches, which are used to improve hydrologic post-processing hypotheses (i.e. reduce total hydrologic uncertainty) in the poorly gauged UZRB.

How to cite: Valdés-Pineda, R., Valdés, J. B., Wi, S., Serrat-Capdevila, A., Tirthankar, R., Demaria, E. M. C., and Durcik, M.: Operational Daily Streamflow Forecasts by coupling Variational Ensemble Forecasting and Machine Learning (VEF-ML) approaches, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-14087, https://doi.org/10.5194/egusphere-egu21-14087, 2021.

Maureen Wanzala, Andrea Ficchi, Hannah Cloke, and Elizabeth Stephens

Information about monitoring of hydrological extremes, agricultural yields and irrigation may be informed by early warning, forecasts and flood management advice through appropriate modelling skills. However hydrological modelling is a challenging task in poorly gauged catchments, especially in developing countries like Kenya. Open access global precipitation and temperature reanalysis datasets with different spatial and temporal resolutions provide alternative sources in data-scarce regions but, individual reanalysis precipitation datasets have significant uncertainties. Inspired by data scarcity issues, significant spatial and temporal gaps in gauge observations, and poor performance of individual reanalysis in hydrological models, this study assess the performance of five new-era reanalysis datasets (ERA5, ERA-Interim, Modern Era Retrospective Analysis for Research and Applications version 2 (MERRA2), Climate Forecast System Reanalysis (CFSR) and Japanese 55-year Reanalysis Project(JRA55)) to simulate daily streamflow using the GR4J model across the 20 catchments in Kenya. Deviating from the modelling normality of calculating the model performance statistics for the calibration and validation periods to investigate whether a model serves as satisfactory representations of the natural hydrologic phenomenon, we couple with sensitivity analysis (SA) to unveil model structural uncertainty and suitability when forced with the different reanalysis products. In this study we use the reanalysis precipitation, maximum (T max) and minimum (T min) temperatures against the observations from the Climate Hazards group Precipitation (CHIRPS) for 1981–2016 to calculate performance statistics, streamflow simulations and sensitivity analysis. In addition, we develop model suitability index (MSI) by coupling the performance statistics with the sensitivity results across the different reanalysis products for our study catchments. Our results show that ERA5 performs better than other reanalysis products in terms of performance statistics and streamflow simulations at catchments scale. MSI results were suitable with ERA5 and lower in JRA55 across most of the Kenyan catchments, with 0.8 and 0.4 MSI respectively. MSI developed in this study is a quantitative measure that can be used for the comparison of reanalysis products for different catchments, thus useful for application to modelling to assess the suitability of both the modelling tools and catchment response to alternative forcings for early warning and inform early action.

How to cite: Wanzala, M., Ficchi, A., Cloke, H., and Stephens, E.: Assessment of Suitability of Global Reanalysis for Hydrological Applications by Coupling Performance Statistics and Sensitivity Analysis in Kenya, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-14311, https://doi.org/10.5194/egusphere-egu21-14311, 2021.

Jens Grundmann, Achim Six, and Andy Philipp

Reliable warnings and forecasts of extreme precipitation and resulting floods are an important prerequisite for disaster response. Especially for small catchments, warning and forecasting systems are challenging due to the short response time of the catchments and the uncertainties of the meteorological forecasts. Thus, ensemble forecasts of precipitation are an option to portray these inherent uncertainties. By this contribution, we present our operational web-based demonstration platform for ensemble hydrological forecasting in small catchments of Saxony, Germany. We use the ICON/COSMO-D2-EPS product of the German Weather Service, which provides an ensemble of 20 members each three hours, for lead times up to 27 hours. Each member is evaluated regarding specific extreme precipitation thresholds for predefined hydrological warning regions. If these thresholds are exceeded in a specific region, rainfall-runoff models for the associated catchments are started to propagate the meteorological uncertainty into the resulting runoff, followed by statistical post processing and visualization. Different options for the visualization of the uncertainty information were discussed and evaluated by a series of (virtual) workshops with locally responsible civil protection forces and water authorities. This leads to the current design of the web-based demonstration platform in an iterative process, which is still ongoing. The web-based demonstration platform is established for three pilot regions with different hydrological settings in Saxony, Germany. Besides layout and technical issues, first experiences with the demonstration platform are presented as well as first results regarding forecast performance in the small pilot regions.

How to cite: Grundmann, J., Six, A., and Philipp, A.: A web based demonstration platform for flood warning in small catchments using ensemble hydrological forecasting, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-15111, https://doi.org/10.5194/egusphere-egu21-15111, 2021.

Urmin Vegad and Vimal Mishra

Ensemble Streamflow Prediction (ESP) is a widely used method in forecasting streamflow, particularly for extremely low or high flows. However, the incorporation of reservoir operations in using ensemble streamflow prediction has not been investigated till yet. We calibrated Variable Infiltration Capacity (VIC) model for daily streamflow for Narmada river basin at four stations (Sandia, Handia, Mandleshwar and Garudeshwar) considering the effect of four reservoirs (Bargi, Tawa, Indira Sagar and Sardar Sarovar). The model is well-calibrated for the selected river basin (R2>0.55) at all locations. Further, routing of streamflow is done considering the reservoir storage dynamics and operating rules. Input data for ensemble prediction is taken from all 16 members of the Extended Range Forecast System (ERFS) developed by Indian Institute of Tropical Meteorology (IITM) and implemented by India Meteorological Department (IMD). Post-processing of the results gave us probabilities of uncertainties associated with streamflow prediction using ERFS members. This study provides key information in predictions of streamflow by incorporating the reservoirs based on the ERFS ensemble members, which can be used to effectively mitigate life and property losses associated with extreme flows in rivers.

How to cite: Vegad, U. and Mishra, V.: Probabilistic Streamflow forecast for Narmada River Basin, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-15745, https://doi.org/10.5194/egusphere-egu21-15745, 2021.

Ervin Zsoter, Christel Prudhomme, Elisabeth Stephens, and Hannah Cloke

Global flood forecasting systems rely on definition of flood thresholds for identifying upcoming flood events. Existing methods for flood threshold definition can often be based on reanalysis datasets and single thresholds, used for all forecast lead times, but this leads to inconsistencies between how the extreme flood events are represented in the flood thresholds and the ensemble forecasts.

This paper explores the potential benefits of using river flow ensemble reforecasts to generate flood thresholds that can deliver improved reliability and skill. Using the Copernicus Emergency Management Service’s Global Flood Awareness System, the impact of the dataset and the method used to sample the annual maxima to define flood thresholds, are analysed in terms of threshold magnitude, forecast reliability and skill for different flood severity levels and lead times.

It was found that the variability of the threshold magnitudes, when estimated from the different annual maxima samples, can be extremely large, as can the subsequent impact on forecast skill. It was also found that reanalysis-based thresholds should only be used for the first few days, after which ensemble-reforecast-based thresholds, that vary with forecast lead time and can account for the forecast bias trends, provide more reliable and skilful flood forecasts.



How to cite: Zsoter, E., Prudhomme, C., Stephens, E., and Cloke, H.: Using ensemble reforecasts to generate flood thresholds for improved global flood forecasting, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-16303, https://doi.org/10.5194/egusphere-egu21-16303, 2021.