Weather Regime Prediction Using Statistical Learning

Abstract

Two novel statistical methods are applied to the prediction of transitions between weather regimes. The methods are tested using a long, 6000-day simulation of a three-layer, quasigeostrophic (QG3) model on the sphere at T21 resolution.The two methods are the k nearest neighbor classifier and the random forest method. Both methods are widely used in statistical classification and machine learning; they are applied here to forecast the break of a regime and subsequent onset of another one. The QG3 model has been previously shown to possess realistic weather regimes in its northern hemisphere and preferred transitions between these have been determined. The two methods are applied to the three more robust transitions; they both demonstrate a skill of 35%-40% better than random and are thus encouraging for use on real data. Moreover, the random forest method allows one, while keeping the overall skill unchanged, to efficiently adjust the ratio of correctly predicted transitions to false alarms.A long-standing conjecture has associated regime breaks and preferred transitions with distinct directions in the reduced model phase space spanned by a few leading empirical orthogonal functions of its variability. Sensitivity studies for several predictors confirm the crucial influence of the exit angle on a preferred transition path. The present results thus support the paradigm of multiple weather regimes and their association with unstable fixed points of atmospheric dynamics.

A mechanistic model of mid-latitude decadal climate variability

“…The existence of such a nongaussian behavior in the observational data sets is still a subject of an ongoing debate (Hsu and Zwiers 2001;Stephenson et al 2004;Deloncle et al 2007). Berner and Branstator (2007) identified significant deviations from gaussianity in a four-dimensional phase space of an atmospheric general circulation model and showed that the corresponding phase-space structure can be described in terms of two off-centered Gaussian distributions.…”

Application of machine learning for fuel consumption modelling of trucks

“…[24], [25]). A forest is a combination of tree predictors such that each tree depends on a vector of independently and randomly sampled values, or features, with the same distribution for all trees in the forest.…”

Energy-based predictions in Lorenz system by a unified formalism and neural network modelling

“…A first necessary step in view of a possible empirical application of the energetic argument to real systems, particularly within low-frequency variability of atmospheric patterns, in the same spirit of Kondrashov et al (2004) where a Lorenz model is also used as a test case -and Deloncle et al (2007), is to check whether a given statistical tool working solely on the solutions of the model (i.e., ignoring dynamical laws) is able to capture the relation between predictors and predictands in the system itself.…”