Seamless Hydrological prediction of east Indian summer monsoon and Variance Analysis of its meteorological and hydrological uncertainty (SHIVA)
Final Report Abstract
We have analyzed predictions of streamflow for the Mahanadi river in North-East India. The predictions are based on daily predicted large-scale atmospheric fields ("reforecasts") from a lead time of one day to 6 months (1d to 180d), as regularly conducted by the ECMWF. The fields are "downscaled" to station-level representative of the Mahanadi basin and corresponding time series of temperature and precipitation fed in a hydrologic model. Such streamflow predictions are naturally uncertain, caused by imperfections of the participating models as well as errors in estimating the atmospheric initial state, which grow dynamically with increasing lead time. The meteorological and hydrological uncertainties were addressed each in its own way. While model and initial-state errors are already reflected in the predicted atmospheric fields, which come in the form of an ensemble of 25 members, hydrological uncertainty had to be developed in SHIVA itself. Hydrologic uncertainty was confined to parametric, using an ensemble of 11 different calibration periods for parameter estimation. This setting provided for each forecast case an ensemble of 25x11 single predictions coupling meteorological and hydrological realization, with the number of cases being 180x12x12 (lead time, issue month, year). To keep the number of combinations and analyses feasible, we have focused on the times of the Indian monsoon, from June to September, and also mostly on the months when predictions are most needed, that is, from May to July. During monsoon season, the uncertainty in the predictions are dominated by the meteorological uncertainty. Implementing state updating data assimilation in the hydrological model reduced the forecast errors, in particular at the short-term scale. However, for longer range forecasts, and in particular in during the monsoon season, the meteorological forecasts are the main factor determining the predicted flow. Uncertainty decomposition was systematically investigated using an ANOVA approach. Overall, the decomposition yielded a very large component of unresolvable variance ( 90%). That is, that part of the Q variations cannot be traced back to a single factor alone (MET and HYD) but can only be explained by interactions between MET and HYD. For the remaining variance, it turned out that in the crucial May predictions both factors contribute equally until about July, after which time the MET becomes dominant. When predicted in the middle of the monsoon (July 1), the first week is characterized by HYD initialization errors, despite the application of DA; starting late July MET errors again dominate.
Publications
- (2019). Dynamical prediction of Indian monsoon: Past and present skill. International Journal of Climatology, 39(8), 3574–358
Köhn-Reich, L., & Bürger, G.
(See online at https://doi.org/10.1002/joc.6039) - (2020). Intraseasonal Oscillation Indices from Complex EOFs. Journal of Climate, 34(1), 107–122
Bürger, G.
(See online at https://doi.org/10.1175/jcli-d-20-0427.1)
