Global temperature predictability through univariate empirical modeling

Temperature prediction in atmospheric science research has emphasized dynamical prediction models which use primitive equations and computational methods. For the extended-range (more than two-week) predictions and long-term predictions, the performance of dynamical models developed up to now in general is inferior to that of empirical models which use a historical data base to extract the hidden information. Univariate empirical prediction models are developed in this research for possible operational purposes.;In non-linear prediction models, analog methods are adopted to investigate the possibility of the extended-range prediction using daily mean brightness temperature anomalies. The statistical average skill score by correlation coefficients of the prediction with respect to the idealized observational data is about 0.3 for the 200-day prediction. The relatively short-term prediction skills show lower correlation but small mean square errors.;For the theoretical establishment of the predictability, the ideas from dynamical systems (Lorenz model) are applied to the real MSU observational data. After the proposition on the predictability for the given data is established, possible operational models are developed for the given prediction ranges within a certain error bound.;In two-dimensional models, tropospheric and stratospheric temperature anomalies are analyzed and forecast using daily MSU data. In two-dimensional predictions, two models are developed.;The brightness temperature anomalies from the Microwave Sounding Units (MSUs) on board the National Oceanic and Atmospheric Administration (NOAA) satellites are selected as the data base for global temperature prediction modeling. The emphasis is on data analysis and possible operational extended-range forecasts by one-dimensional time-series (global mean temperature anomalies) prediction models, two-dimensional (zonal mean temperature anomalies) prediction models and three-dimensional (global spherical variations of temperature anomalies) prediction models, respectively. In the one-dimensional linear prediction model, tropospheric temperature anomalies, which show more complicated phenomena than in other layers, are forecast. A multi-regression model is developed using monthly mean temperature anomalies. In the multi-regression model, Fourier spectrum analysis is used to investigate the possible physical relation on each atmospheric variable. The univariate Auto Regressive Integrated Moving Average (ARIMA) method is followed with the same data. Three-month lead global mean temperature anomaly prediction by the multi-regression model has a correlation of 0.72 with actual observations during 20-month test periods. The ARIMA model has an average correlation of 0.61 during twelve-month forecasts.;In three-dimensional (monthly, gridpoint) predictions, lower-tropospheric, mid-tropospheric and lower-stratospheric temperature anomalies are forecast. Structural variations are emphasized in the analysis. Similar to the two-dimensional model, one-dimensional techniques are selected in the prediction model with additional mathematical manipulations.;The univariate prediction model performs its forecasting economically. The results of each univariate empirical prediction model applied on the case studies, even though generalization is not possible now, show encouraging results for operational use. (Abstract shortened by UMI.).