Técnicas de aprendizaje automatizado para el pronóstico de temperaturas minímas en el Centro Meteorológico de Villa Clara, Santa Clara
Archivos
Fecha
2014-07-08
Autores
Thapelo, Swaabow Tsaone
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Central “Marta Abreu” de Las Villas
Resumen
Los pronósticos se mantienen como una de las tareas más desafiantes y fascinantes actualmente. Esto ha atraído la atención de los especialistas e investigadores de distintos campos, especialmente las ciencias computacionales. Estudios de la literatura muestran que las técnicas de aprendizaje automatizado mantienen un mejor rendimiento que los métodos convencionales. En este trabajo se proponen cuatro técnicas de aprendizaje automatizado: máquinas de vectores soporte, las redes neuronales artificiales, los procesos gaussianos y el aprendizaje basado en instancias. Estos modelos se han aplicado para predecir las temperaturas mínimas en cinco estaciones meteorológicas localizadas en la provincia de Villa Clara. Se desarrolló una aplicación SAGI V 1.0 en Java que emplea las técnicas propuestas en este trabajo.
La capacidad predictiva de estos modelos es evaluada en términos de: error medio absoluto y la desviación estándar con el objetivo de escoger el mejor modelo basado en reducción de errores durante la predicción de las temperaturas para una determinada estación meteorológica. Se realizaron diferentes configuraciones para cada modelo y se escogen las mejores configuraciones según su capacidad predictiva. Los resultados finales fueron comparados con aquellos obtenidos por el Sistema empleado actualmente en el Centro Meteorológico de Villa Clara, el MOS (Model Output Statistics).
De esta comparación se destaca que las técnicas propuestas en este trabajo se pueden utilizar para conseguir una mejor predicción para este tipo de datos en los que existe una relación compleja. En cuanto al porciento de los casos positivos obtenidos por las nuevas técnicas, la capacidad predictiva del modelo SAGI resultó mejor con un 90.3% comparado con el 76.1% alcanzado por el modelo actual para el periodo Marzo 2011.
Prediction remains one of the fascinating yet challenging tasks at the moment. This has drawn a lot of interest especially for researchers from different fields and a lot has been done for the past years. Literature studies have shown that machine learning techniques achieved better performance than conventional methods. In this work four Machine Learning techniques: Support Vector Machine, Artificial Neuronal Networks, Gaussian Processes and Instance Based Learning (IBk) models are applied to make minimum temperature predictions for five weather stations located in the Villa Clara province. An application using these four techniques was developed using the Java language. The new application will be referee to as SAGI V 1.0 for simplicity. Their predictive capability is evaluated in terms of Mean Absolute Error and Standard Deviation with the objective of choosing the better model based on error reductions during temperature predictions for a given weather station. Different configurations for each model were realized and the best ones were selected with regards to their predictive capability. The final results were compared to those obtained by the current system being deployed in the Villa Clara Meteorological Centre (CMPVC). It was concluded that the proposed machine learning techniques can be deployed in the Villa Clara province with the aim of improving weather forecasting for this kind of data with complex relations. As to the positive case results obtained SAGI obtained 90% of success as compared to 76% obtained by the current system (MOS) being used in the CMPVC. The models were evaluated for the March 2011 winter season due to the fact that the relationship of the environmental variables is very complex during this period
Prediction remains one of the fascinating yet challenging tasks at the moment. This has drawn a lot of interest especially for researchers from different fields and a lot has been done for the past years. Literature studies have shown that machine learning techniques achieved better performance than conventional methods. In this work four Machine Learning techniques: Support Vector Machine, Artificial Neuronal Networks, Gaussian Processes and Instance Based Learning (IBk) models are applied to make minimum temperature predictions for five weather stations located in the Villa Clara province. An application using these four techniques was developed using the Java language. The new application will be referee to as SAGI V 1.0 for simplicity. Their predictive capability is evaluated in terms of Mean Absolute Error and Standard Deviation with the objective of choosing the better model based on error reductions during temperature predictions for a given weather station. Different configurations for each model were realized and the best ones were selected with regards to their predictive capability. The final results were compared to those obtained by the current system being deployed in the Villa Clara Meteorological Centre (CMPVC). It was concluded that the proposed machine learning techniques can be deployed in the Villa Clara province with the aim of improving weather forecasting for this kind of data with complex relations. As to the positive case results obtained SAGI obtained 90% of success as compared to 76% obtained by the current system (MOS) being used in the CMPVC. The models were evaluated for the March 2011 winter season due to the fact that the relationship of the environmental variables is very complex during this period
Descripción
Palabras clave
Temperaturas Mínimas, Modelo Estadístico MOS, Máquinas de Vectores Soporte, Redes Neuronales Artificiales, IBk, Procesos Gaussianos, Minimum Temperatures, Statistical Method MOS, Support Vector Machine, Artificial Neuronal Networks, IBk, Gaussian Processes