Improving Seasonal Snowmelt Runoff Forecasts in Manic-5 Catchment using ECMWF SEAS5 Data and Variational Data Assimilation through HBV Hydrological Model

Código

I-EBHE0216

Autores

Gokcen Uysal, Marie-Amelie Boucher, Charles Mathieu, Rodolfo Alvarado-Montero, Ali Arda Sorman

Tema

WG 1.10: Hydrologic Design - Solutions & Communication

Resumo

Hydropower serves as the primary electricity source in Québec. Traditional seasonal planning for hydroelectric production relies on extended streamflow predictions, but this approach is increasingly challenged by non-stationarity due to global climate change. To address this issue, the study aims at improving seasonal snowmelt runoff forecasts through data assimilation in a hydrological modeling framework, enhancing initial state accuracy, and evaluating the performance of seasonal numerical weather forecasts compared to benchmark perfect forecasts. The selected study area is the Manic 5 Reservoir Basin (24,717 km2) within the upper part of the Manicouagan System, where snow and ice persist for over seven months annually. A daily rainfall-runoff relation is developed by semi-distributed conceptual HBV model, partitioning the basin into four elevation zones (346-1178 m). The model is calibrated within the 1980-2007 period via with the Particle Swarm Optimization (PSO) algorithm and validated for the two periods of 2008-2016 and 2017-2022. Within these periods, Nash-Sutcliffe Efficiency (NSE) and Kling Gupta Efficiency (KGE) for runoff range between 0.77-0.80 and 0.76-0.87, respectively. Simulated snow cover area (SCA) and snow water equivalent (SWE) are compared with global remote sensing products. SCA data is derived from the Interactive Multisensor Snow and Ice Mapping System (IMS), ensuring minimal impact from meteorological variables like cloud cover. Observed in-situ SWE measurements are used in data assimilation, while ERA5-Land reanalysis SWE product is employed for model comparison. The simulated SCA against IMS and simulated SWE against ERA5-Land show mean absolute error (MAE) values of 6.9 % and 31.0 mm within the entire validation period of 2008-2022, respectively. The simulated SWE is underestimated against ERA5-Land SWE with an average percent-bias of -27%. Sequential data assimilation methods like Ensemble Kalman Filter are widely used in literature. However, this study analyses Variational Data Assimilation (VarDA) configured with the less common Moving Horizon Estimation (MHE) approach. We introduce disturbances to the model states and model forcings, as well as observations (i.e., runoff, SCA and SWE) in an objective function and it is configured in a 6-month predefined assimilation window. The ECMWF SEAS5 data is composed of a 51-member ensemble forecasts with a forecast horizon of 215 days (~7 months) and a resolution of ~36-km. They are issued sub-daily for the first day of each month since 2017. Hindcasting simulations conducted in a closed-loop environment employ both deterministic perfect forecasts and probabilistic SEAS5 ensemble forecasts for the period 2017 to 2022. According to the analyses, utilization of VarDA against no DA considerably improves the quality of SEAS5 based runoff and SWE forecasts in terms of mean- Continuous Ranked Probability Score with a diminishing performance up to a 3-month lead time. Higher gain in terms of MAE and longer lead time effects are observed for perfect forecasts. Additionally, VarDA enables us to effectively incorporate sparse discrete in-situ SWE data into our modeling framework. Overall, this study assesses the impact of the forecasting scheme on hydropower generation, utilizing advanced modeling and assimilation techniques to enhance seasonal predictions in a changing climate context.