Use of Neural Networks for Statistical Downscaling of Long-term Air Temperature Forecasts for the Czech Republic

Dynamical forecasts use full three-dimensional climate models to simulate potential changes in the atmosphere and ocean over the next few months based on current conditions.  The ensembles of simulations provide probabilistic weather scenarios that indicate the likelihood of a given period being wetter, drier, warmer, or colder compared to the seasonal average. The added value of various downscaling approaches for seasonal forecasts is a topic of frequent debate. This work focuses on statistical downscaling, which is based on empirical relationships derived between a local observed predictand of interest (summer temperature in this case) and one or several suitable model predictors from global seasonal forecasting systems.

Unlike tropical regions, seasonal predictability in Europe remains limited. This study analyses the seasonal forecast systems available in the Copernicus Climate Change Service (C3S) archive, which provide near-surface air temperature data at 1° to 1° spatial resolution. This study examines the statistical downscaling of summer air temperature forecasts for Central Europe from two weather forecast systems: the European Centre for Medium-Range Weather Forecasting (ECMWF) SEAS5.1 system and the Météo-France 8 (MF) system.

The study analyses the period 1993-2016, which is the longest hindcast period common to all systems, and the domain of the Czech Republic in Central Europe (47-52°N, 11-20°E). To perform statistical downscaling using the neural network method in STATISTICA software, we tested air temperature and sea surface pressure from global forecast models as predictors. The reference data used in this study are gridded E-OBS observational air temperature data sets. The best neural networks found are tested on forecasts for the period 2020–2023.