Are ensemble NWP forecasts now so good that calibration is unnecessary?

For streamflow forecasting, calibration of ensemble numerical weather prediction (NWP) models has long been considered a necessary evil. Necessary, because NWP forecasts are usually too biased to force calibrated hydrological models, they often produce unreliable ensembles and may produce forecasts that are less accurate than simple climatology at longer lead times. Evil, because calibration adds complexity to any forecasting system and the calibration process destroys spatial, temporal and inter-variable correlations in the ensemble, which then must be reconstructed in various and usually unsatisfying ways. As ensemble NWPs improve, the degree to which calibration is ‘necessary’ declines.

Here we investigate recent versions of two ensemble NWP models – the European Centre for Medium-range Weather Forecasts ensemble NWP (ECMWF-ens) and the Bureau of Meteorology’s Australian Community Climate and Earth-System Simulator Global Ensemble (ACCESS-GE) NWP. The models are tested over Tasmania, where CSIRO is working with Hydro Tasmania, Australia’s largest generator of hydropower, to establish new ensemble streamflow forecasting systems. Tasmania is mountainous and temperate and features strong rainfall gradients. We apply an existing calibration method – the Catchment-scale Hydrological Precipitation Processor (CHyPP) – which uses a Bayesian Joint Probability model to calibrate ensemble precipitation forecasts.

We show that CHyPP improves reliability in both the ECMWF-ens and ACCESS-GE ensembles, but these improvements come at the cost of a slight reduction in skill at short lead times. Uncalibrated ACCESS-GE forecasts generally produce more biased and less reliable forecasts than ECMWF-ens, and we conclude that calibration is necessary for the ACCESS-GE model, both to reduce biases and improve reliability. However, the improvements in bias from calibrating the ECMWF-ens are negligible in some catchments, with the main benefit being improved reliability at longer lead times. This brings into question the need for calibration of the ECMWF-ens model with CHyPP. We note that these findings may not hold outside the Tasmanian catchments tested, where high resolution ensemble NWP forecasts generally perform well. We discuss the implications of these findings with respect to streamflow forecasts.