Self-Learning Statistical Short-Term Climate Predictive Model for Europe

A multivariate, self-learning fuzzy-neural model is developed to describe predictive relationships between evolving large-scale patterns in Northern Hemisphere surface atmospheric pressure and air temperature fields (predictors) and subsequent patterns in the Europe surface temperature and precipitation (predictands). A lead interval of varying length (from 1 to 6 months) is placed between a series of consecutive predictor periods and a single predictand period. Objective evaluation of strength of such relationships is a primary aim of this study. Statistical analyses provide empirical knowledge that can lead to more skilful forecasts in the absence of explicit physical understanding. Additionally, acquired information may provide guidance towards identification of the physical process, contributing to or limiting the predictability. The choice to use an empirical approach reflect the fact that both simple and complex general circulation models (GCMs), either with prescribed boundary conditions or with actual oceanic coupling, currently do not adequately reproduce the processes of the real atmosphere at the lead times an averaging periods of concern here. We hope and assume that eventually, with advances in physical understanding, dynamic prediction approaches will outperform statistical ones. Prediction of time-averaged surface climate has received considerable attention over last two decades. First, the potentially predictable portion of the total variability of a given predictand has been empirically estimated using ratios of predictand variability at different frequencies (Trenberth, 1984). Second, direct attempts at forecasting and verification have been made using analog approaches (Barnston and Livezey, 1987) and linear statistical approaches with either several pre-selected predictor elements or whole predictor fields (Barneft, 1981). Neural network (NN) is a powerful nonlinear scheme based on 'black box' statistics, where one can tune the model parameters to arrive at a good prediction, but can see neither the phase relation between the predicand and predictors, nor the origin of skills. Therefore, we assume that the predictability of seasonal climate is connected with forcing fields such as the SST or others. The key to a truly successful application of NN model lies in the understanding of the underlying physical mechanism for the relation between predictor and predictand fields (Pokrovsky, 2000). The decision to use the principal fuzzy patterns (PFPS) of the surface atmospheric pressure and temperature as predictor fields is based on findings of other studies on the field tele connections (Namias, 1982; Lanzante, 1984). Comparison of EOFs (empirical orthogonal function) and PFPs computed for surface atmospheric pressure and air temperature fields has been carried out and its results are analyzed. PFP advantages for field anomaly performance is demonstrated (Pokrovsky et al, 2002). Following tele connection spatial areas (Barnston and Livezey, 1987) were selected to derive PFPs for Europe model: North-Atlantic Oscillation (NAO), East Atlantic (EA), EA Jet (EAJ), East-Atiantic/West Russia (EAWR), Scandinavia (S), Polar/Eurasian (PE). It is known (Barnston, 1994; Wang, 2001) that the effectiveness of statistical models depends crucially on whether the relevant components (with respect to space and time scales) to be used as predictors are suitable incorporated into prediction model, and whether the relationships between predictor and predicands (which may not necessarily be linear, particularly, when mid and high latitude are concerned) are properly established. In this respect implementation of optimal design technique (Pokrovsky and Roujean, 2003) permits to determine an optimal set of predictors (PFPS) representing key low-oscillation patterns, which are most informative with respect to the predictand field for a prescribed lead interval. Five layers neural network utilizes fuzzy classification input and out layers and radial basis functions (RBF) for PFPs as the activated units. In order to reduce the p roblem of artificial skill produced from over-fitting and thus receive a more representative estimate of real skill we used cross-validation method, in which forecast model is developed using only part of available data set and then applied to the independent data. Monthly time series (reanalysis NCEP) for 1948-1998 was split in two parts: the learning and verification samples. In contrast to the GCMs our self-learning model accumulates all past observing data in so way that after 35 years of learning process it could provide very competitive prediction results for surface temperature fields. It captured both positive and negative phases of above climate indexes as well as transition periods in their relationships with predicands fields. It is demonstrated a series of monthly observed and presented.(Author)e experience gainover Europe for last decade of twentieth century. Forecasted fields reproduce main features of analysis grids. Deviation and other proper statistics are discussed. In particularly, achieved level for explained variance of predicted fields is much higher than those accessible for linear regression (see Blender et al, 2003). Partition of Europe at the set of the homogeneous climate ranges by fuzzy regioning is discussed. Forecasted and observed temperature and precipitation time series for several climate areas are considered. Skilful magnitudes are analysed as well. (Author)