Atlantic Tropical Cyclone Rapid Intensification Probabilistic Forecasts from an Ensemble of Machine Learning Methods
| Autor: | Alexandria Grimes, Andrew E. Mercer |
|---|---|
| Rok vydání: | 2017 |
| Předmět: |
010504 meteorology & atmospheric sciences
Artificial neural network Computer science business.industry Probabilistic logic Forecast skill Feature selection 02 engineering and technology Rapid intensification Machine learning computer.software_genre Linear discriminant analysis 01 natural sciences Random forest Support vector machine 0202 electrical engineering electronic engineering information engineering General Earth and Planetary Sciences 020201 artificial intelligence & image processing Artificial intelligence Tropical cyclone Predictability business computer 0105 earth and related environmental sciences General Environmental Science |
| Zdroj: | Procedia Computer Science. 114:333-340 |
| ISSN: | 1877-0509 |
| DOI: | 10.1016/j.procs.2017.09.036 |
| Popis: | Atlantic tropical cyclone (TC) rapid intensification (RI) continues to be a major forecasting challenge, with forecast skill scores only about 15% better than climatology. To date, RI forecasts have been completed using linear discriminant analysis (LDA) on predictors optimized for RI forecasts, and no study has directly addressed machine learning’s (hereafter AI) capability in forecasting RI. As such, the objective of this study is to quantify the RI predictability using proxy forecast model data and an ensemble of AI methods to generate probabilistic RI forecasts. Atlantic RI events from 1985 to 2011 were retained for all valid times (over water) for each TC, and these cases were used to train an AI ensemble optimized (through three steps) for RI prediction. First, backwards elimination feature selection was used on a blend of the proxy forecast data, predictors from the currently utilized LDA model, and observed TC track information (such as intensity and position) to optimize the predictor suite. Second, numerous configurations of three AI methods (support vector machines [SVMs], artificial neural networks [ANNs], and random forests [RFs]) were tested using bootstrap-based cross-validation to ascertain the best configurations of each AI method. Finally, the best AI configurations were used to generate probabilistic output for RI, weighted by each ensemble member’s individual cross-validation performance. Resulting probabilistic forecasts were in line with the current LDA method, though the upper skill limit of the ensemble exceeded 30% improvement over climatology, which far exceeds the current LDA scheme. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|