Ensemble Kalman Filter for Hourly Streamflow Forecasting in Huaynamota River, Nayarit, México

Abstract

Hydrological phenomena are characterized by the formation of a non-linear dynamic system, and streamflows are not unrelated to this premise. Data assimilation offers an alternative for flow forecasting using the Ensemble Kalman Filter, given its relative ease of implementation and lower computational effort in comparison with other techniques. The hourly streamflow of the Chapalagana station was forecasted based on that of the Platanitos station in northwestern México. The forecasts were made from one to six steps forward, combined with set sizes of 5, 10, 20, 30, 50, and 100 members. The Nash-Sutcliffe coefficients of the Discrete Kalman filter were 0,99 and 0,85 for steps one and six, respectively, achieving the best fit with a tendency to shift the predicted series, similar to the persistent forecast. The Ensemble Kalman Filter (EnKF) obtained 0,99 and 0,05 in steps one and six. However, it converges on the observed series with the limitation of considerable overestimation in higher steps. All three algorithms have equal statistical adjustment values in step one, and there are progressive differences in further steps, where ARX and DKF remain similar and EnKF is differentiated by the overestimation. EnKF enables capturing non-linearity in sudden streamflow changes but generates overestimation at the peaks.