Evaluación de estrategias de modelado para la predicción y reconstrucción de variables meteorológicas en estaciones costeras

Abstract

This monograph addresses the importance of collecting meteorological data using drifting buoys to understand and predict ocean behavior and climatic conditions in coastal areas. It analyzes how, due to the remote location and various technical and environmental factors, periods with missing data occur in the time series. To tackle this challenge, modeling approaches were investigated that first allow for the reconstruction of incomplete series and then for the prediction of the same variables—air temperature, wind speed, and wind direction—up to a two-week horizon. The study developed and implemented a comprehensive strategy based on preprocessing (interpolation of missing data, error elimination, outlier correction, and normalization) and the evaluation of methods such as KNN, ARIMA, ARMA, and recurrent neural networks (RNN). The results indicate that, by enriching RNNs with exogenous information, complex patterns are captured, achieving significantly superior metrics in the reconstruction and prediction of the series, at the expense of a higher computational cost in terms of execution time and memory consumption.