Data-driven one-day probabilistic forecasts of precipitation occurrence and amount for northern tropical Africa

Despite the overall improvement of numerical weather prediction (NWP) models, precipitation forecasts in the tropics remain a great challenge. Several studies have shown that NWP models have difficulties in outperforming climatological forecasts, which motivates the usage of alternative approaches from statistics and machine learning. Vogel et al. (2021) for instance implement a fairly simple, purely data-driven logistic regression model for 24-hour precipitation occurrence, which outperforms climatological and NWP forecasts for the summer monsoon season in West Africa.

The present work extends this approach in several ways. First, for validation we replaced the satellite-based rainfall estimate TRMM by the newer IMERG product. Furthermore, in order to find relevant predictors for statistical and machine learning models, a wide range of weather variables from ERA5, the latest re-analysis product of the European Centre for Medium-Range Weather Forecasts (ECMWF), are assessed with respect to their relation to rainfall. Moreover, the analysis is extended to five different seasons covering the entire year. In addition, the more difficult problem of producing probabilistic forecasts for accumulated precipitation is investigated, for which more involved statistical methods are required.

In this contribution we will give a detailed overview of all considered forecasting approaches and a comprehensive comparison between the performance of each model covering state-of-the-art methods from NWP (ECMWF, raw and post-processed), statistics, and machine learning. We use a logistic regression model to predict probability of precipitation occurrence and methods based on the recently introduced isotonic distributional regression (IDR) technique to produce probabilistic forecasts for accumulated precipitation. For both tasks, using ERA5 variables as additional predictors improves forecast performance. Overall the performance of the statistical models are similar to the post-processed operational ECMWF ensemble or slightly worse in the case of rainfall amount. Using a more sophisticated machine learning model, namely a convolutional neural network, which maintains the two dimensional gridded data structure and thus exploits spatial relations clearly outperforms all considered approaches.

The results demonstrate that the employed methods have potential to improve probabilistic forecasts of rainfall in the tropics in operational procedures.

Vogel, P., Knippertz, P., Gneiting, T., Fink, A. H., Klar, M., and Schlueter, A., 2021: Statistical forecasts for the occurrence of precipitation outperform global models over northern Tropical Africa. Res. Lett., 47, e2020GL091022, doi:10.1029/2020GL091022.