Physical Oceanography, Vol. 16, No. 3, 2006
ANALYSIS OF OBSERVATIONS AND METHODS FOR CALCULATING
HYDROPHYSICAL FIELDS IN THE OCEAN
APPLICATION OF THE METHOD OF ARTIFICIAL NEURAL NETWORKS
TO THE DOWNSCALING OF PRECIPITATION FORECASTS IN THE
COASTAL REGION OF THE BLACK SEA
V. V. Efimov and V. L. Pososhkov
We correct the diurnal data on precipitation obtained from the output of the global system of re-
analysis as applied to the observed daily amounts of precipitation at certain geographic points of
the coastal region of the Black Sea. The estimations of the actual amounts of precipitation are
taken from the ECAD (European Climate Assessment and Dataset) database. We analyze the
amounts of precipitation for three winter months. As a working tool for the investigation of cor-
relations between the regular meteorological variables (predictors) taken from the reanalysis and
the local amounts of precipitation at certain geographic points (predictants), we use the method
of artificial neural networks (ANN). A numerical criterion of adequacy of the estimates of the
daily amounts of precipitation performed according to reanalysis and the ANN method is pro-
posed. By using this criterion, we show that the efficiency of the ANN method in simulating
precipitation is higher as compared with the procedure of reanalysis.
Precipitation is one to the least predictable parameters in contemporary numerical weather forecasts due to,
first of all, its high space and time variability. The space resolution even of the most advanced numerical models
appears to be insufficient for the adequate description of the processes of precipitation. Therefore, the dynamic
statistical methods of downscaling of the forecasts of meteorological parameters are now extensively developed
with an aim, e.g., to improve the output estimates of numerical models for specific geographic points. For the
most part, the methods of downscaling are developed in the following two directions: First, in the direction of
application of local numerical models with high space resolution based on the use of large-scale calculated fields
as boundary conditions with regard for the regional characteristics. The second approach is connected with the
so-called statistical downscaling . This method becomes more and more popular due to its relative simplicity
and much lower costs as compared with the application of local models. The basic idea of statistical downscal-
ing is to use the observed correlations between the large-scale circulation and local climate in order to establish
statistical models relating the anomalies of large-scale fields (predictors) to the anomalies of local climatic varia-
bles (predictants), e.g., to the anomalies of atmospheric precipitation.
In the present work, we use an empirical downscaling approach based on the application of the method of
artificial neural networks as a diagnostic tool. Artificial neural networks appeared in the mid-1950s as mathem-
atical models aimed at the approximate description of functioning of the nervous system of human beings. The
first one-layer perceptron created in 1959  was used for the solution of some simple problems of classifica-
Marine Hydrophysical Institute, Ukrainian Academy of Sciences, Sevastopol.
Translated from Morskoi Gidrofizicheskii Zhurnal, No.
23–35, May–June, 2006. Original article submitted January 20, 2005;
revision submitted February 10, 2005.
0928-5105/06/1603–0141 © 2006 Springer Science+Business Media, Inc. 141