Ensemble Kalman Filter for Data Assimilation in Hydrological Modeling: Case Study of the Capivari Reservoir, Paraná - Brazil

Código

I-EBHE0219

Autores

Danieli Mara Ferreira, Maria Fernanda Dames dos Santos Lima, José Eduardo Gonçalves, Cássia Silmara Aver Paranhos, RAFAEL SCHINOFF MERCIO PEREIRA

Tema

WG 1.12: Development & application of river basin simulators

Resumo

Precise forecasting of reservoir inflows holds significant importance in water resources planning and management. It impacts decisions and strategies for reservoir operation, including flood mitigation, drought response, water allocation, and hydropower production. Conceptual hydrologic models often face limitations in their ability to represent environmental changes across time due to fixed parameterizations. In this context, a promising approach is to consider parameters as potentially time-varying quantities that can evolve based on hydrologic observations. The ensemble Kalman filter (EnKF) is a commonly employed technique for data assimilation. It dynamically updates model error covariance information through sequential generation of model predictions, each perturbed by assumed model errors. Challenges with EnKF include the need to spatially localize covariances to mitigate false correlations and its assumptions of Gaussian distributions for both system states and observations. Furthermore, EnKF often requries substantial computational resources. This study explores the application of EnKF for data assimilation to identify temporal patterns in model parameters using streamflow data. The case study is the simulation of inflow to the Capivari reservoir, in Paraná, Brazil. The system is used for hydroelectric power generation by the Companhia Paranaense de Energia (COPEL) at the Governador Parigot de Souza Hydroelectric Power Plant. The hydrological model is the GR5J, a daily lumped conceptual approach to represent rainfall?runoff processes in a watershed. The model accounts a production store level, that tracks soil moisture evolution based on rainfall input and actual evapotranspiration. The routing function is based on a nonlinear routing store and a symmetric unit hydrograph. Exchanges with the basin boundary is simulated using a groundwater exchange function. The model relies on five parameters: maximum capacities of production (X1 [mm]) and routing stores (X3 [mm]), groundwater exchange coefficient (X2 [mm/d]), time base of the unit hydrograph (X4 [d]), and a limit of groundwater exchange (X5 [-]). The streamflow forecasting system for the Capivari reservoir was implemented using the R package airGR, that continuously updates the levels of the production and routing stores. The procedure begins with the calibration of the lumped model, conducted for the period from March 16, 2021, to October 23, 2023, yielding a Nash-Sutcliffe coefficient of 0.8. Based on the calibrated parameters, the data assimilation procedure updates two model states and produces new results. The performance of this process was comparable to the previous phase, achieving a Nash-Sutcliffe coefficient of 0.8 for the ensemble mean. Despite the technique not yielding significant benefits compared to model calibration, it shows potential for operational-level applications, emphasizing the importance of further advancements.