The Short-Term Ensemble Prediction System (STEPS) is implemented in real-time at the Royal Meteorological Institute (RMI) of Belgium. The main idea behind STEPS is to quantify the forecast uncertainty by adding stochastic perturbations to the deterministic Lagrangian extrapolation of radar images. The stochastic perturbations are designed to account for the unpredictable precipitation growth and decay processes and to reproduce the dynamic scaling of precipitation fields, i.e., the observation that large-scale rainfall structures are more persistent and predictable than small-scale convective cells. This paper presents the development, adaptation and verification of the STEPS system for Belgium (STEPS-BE). STEPS-BE provides in real-time 20-member ensemble precipitation nowcasts at 1 km and 5 min resolutions up to 2 h lead time using a 4 C-band radar composite as input. In the context of the PLURISK project, STEPS forecasts were generated to be used as input in sewer system hydraulic models for nowcasting urban inundations in the cities of Ghent and Leuven. Comprehensive forecast verification was performed in order to detect systematic biases over the given urban areas and to analyze the reliability of probabilistic forecasts for a set of case studies in 2013 and 2014. The forecast biases over the cities of Leuven and Ghent were found to be small, which is encouraging for future integration of STEPS nowcasts into the hydraulic models. Probabilistic forecasts of exceeding 0.5 mm h−1 are reliable up to 60–90 min lead time, while the ones of exceeding 5.0 mm h−1 are only reliable up to 30 min. The STEPS ensembles are slightly under-dispersive and represent only 75–90 % of the forecast errors.
An analogue-based approach for nowcasting the spatio-temporal evolution of orographic rainfall at the southern side of the Swiss Alps, NORA, was developed by Panziera et al. (2011). Analogues were found by retrieving a set of similar mesoscale situations and rainfall fields, and the forecast was given by the evolution of the precipitation observed after the analogues. This strategy avoids the explicit space–time rainfall simulation to obtain ensembles that characterize the forecast uncertainty. In this study the choice of the most similar rainfall fields is further explored by means of principal components analysis. The latter is used to represent the sequences of radar images in a phase space constructed with a low number of principal components. The principal components explain the main patterns which characterize the spatial distribution of rainfall, a feature that was not implemented in the original NORA. The alternative version of the nowcasting tool is described and the forecasts are verified in detail. Due to the ability to represent forecast uncertainty, the ensemble prediction system has superior value for probabilistic short-term forecasting compared to Eulerian persistence, which is more suited for deterministic forecasts. It is also demonstrated that retrieving similar sequences of images instead of single images does not improve forecast skill, which leads to the conclusion that the past trend in rainfall evolution is not a good predictor of its future evolution.
The goal of this research is to study the effect of orography on the spatial distribution of rainfall nowcasting errors. Twenty months of forecasts of hourly rainfall accumulations from the short‐term ensemble prediction system (STEPS) for the weather radar composite over eastern Victoria, Australia, were verified against radar observations. The verification was stratified according to six different flow regimes derived by applying a k‐means clustering algorithm on the archive of motion fields. STEPS biases were found to have a strong flow‐regime dependence, which appears as an underestimation of rainfall on the windward side of terrain features followed by an overestimation on the leeward side. In the regions characterized by strong orographic forcing, the relative contribution of the bias to the root mean square error (RMSE) of the nowcast is locally important. The results also highlighted areas where the errors in the radar rainfall estimates contributed a significant fraction of the observed forecast error. These promising results open ways to integrate flow‐dependent and spatially inhomogeneous growth and decay mechanisms in the stochastic simulation of radar rainfall fields performed by STEPS.