INDUCTIVE AND MODEL-TREE-BASED APPROACHES FOR FORECASTING TEMPERATURE

Inductive and model-tree (MT) approach-based models are developed and evaluated for forecasting mean, minimum and maximum monthly temperature in this study. The models are developed and tested using long-term historical temperature time series data derived from U.S. Historical Climatology Network at 22 sites located in the state of Florida. Inductive models developed include conceptually simple naïve models to multiple regression models utilizing lagged temperature values, sea surface temperatures (SSTs), correction factors derived using historical data. A global model using data from all the sites is also developed. The performances of the models were evaluated using observed temperature records and several error and performance measures. A composite measure combining multiple error and performance measures is developed to select the best model. MT approach-based and regression models with SSTs and correction factors along with lagged temperature values are found to be best models for forecasting temperature based on assessments of composite measures and error diagnostics.