Scale-dependent blending of ensemble rainfall nowcasts with NWP in the open-source pySTEPS library

Radar rainfall nowcasting, an observation-based rainfall forecasting technique that statistically extrapolates current observations into the future, is increasingly used for short-term forecasting (<6 hours ahead). These first hours ahead are a key time scale for e.g. (flash) flood warnings and they are generally not sufficiently well captured by the rainfall forecasts of numerical weather prediction (NWP) models.

A recent development in nowcasting is the transition to more community-driven, open-source models. The Python library pySTEPS is an example of this. One of its main features is an efficient Python implementation of the probabilistic nowcasting scheme STEPS. pySTEPS generates an ensemble of rainfall forecasts by perturbing a deterministic extrapolation nowcast with spatially and temporally correlated stochastic noise. It considers the dynamical scaling of the rainfall predictability by decomposing the rainfall fields into a multiplicative cascade and applies different stochastic perturbations for each scale. This results in large-scale features that evolve more slowly than the small-scale features.

Despite pySTEPS' representation of the uncertainty associated with growth and decay of rainfall in the first 1-2 hours of the nowcast, it quickly loses skill after 2 hours, or even less for convective rainfall events or small radar domains. To extend the skillful lead time to the desired time scale of 6 hours or more, a blending with NWP rainfall forecasts is necessary. We have implemented an adaptive scale-dependent blending in pySTEPS based on earlier work in the STEPS scheme. In this blending implementation, the blending of the extrapolation nowcast, NWP and noise components is performed level-by-level, which means that the blending weights vary per cascade level. These scale-dependent blending weights are computed from the recent skill of the forecast components, and converge to a climatological value, which is computed from a 1-month rolling window and can be adjusted to the (operational) needs of the user. To constrain the (dis)appearance of rain in the ensemble members to regions around the rainy areas, we have developed a Lagrangian blended probability matching scheme and incremental masking strategy.

We present a validation of the blending approach in a hydrometeorological testbed using Belgian radar and NWP data for the Belgian and Dutch catchments Dommel, Geul and Vesdre. We compare the resulting ensemble rainfall and discharge forecasts of the blending implementation with ensemble nowcasts from pySTEPS, ALARO (NWP) forecasts and a linear blending strategy.