Estimación de datos faltantes en estaciones meteorológicas de Venezuela vía un modelo de redes neuronales

Revista de Climatología 10/2008; 8(1578-8768):51-70.

ABSTRACT En el presente trabajo se propone un método de reconstrucción de series de tiempo de precipitaciones, para ser
aplicado a las estaciones meteorológicas de Venezuela con el propósito de corregir el problema de datos faltantes.
La metodología se fundamenta en dos técnicas: la primera reconstruye la dinámica y el tiempo de retardo del
sistema dinámico de la serie temporal, y la segunda utiliza un modelo de redes neuronales para predecir los datos
faltantes. Los modelos de redes neuronales exploran la dependencia espacio temporal de los atributos meteorológicos
de las series y constituyen una herramienta importante para la propagación de la información relacionada con
el clima, y además proveen soluciones prácticas de incertidumbre asociados con la interpolación y la captura de la
estructura espacio temporal de los datos. Para llevar a cabo estos procedimientos, se ha determinado la dimensión
de inmersión del atractor de las series y el tiempo de retardo, y luego se han usado estas medidas para definir la
arquitectura de la red neuronal. El algoritmo utilizado para estimar los parámetros de la red neuronal ha sido el
de retropropagación, que básicamente actualiza los pesos del modelo en la dirección en que el gradiente decrece
más rápidamente. Para seleccionar la arquitectura de la red, se ha usado el criterio de información de Bayes (BIC),
que consiste en penalizar el error cuadrático medio de los parámetros utilizados en el ajuste del modelo. Los resultados
indican que las series de precipitaciones en Venezuela tienen alguna estructura subyacente no lineal, y
provienen de un sistema caótico de bajas dimensiones. Los modelos de redes neuronales se han revelado útiles
para la reconstrucción de los datos faltantes de las series.

Download full-text


Available from: Saba Infante, Jul 28, 2015
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: We present the first algorithms that allow the estimation of non-negative Lyapunov exponents from an experimental time series. Lyapunov exponents, which provide a qualitative and quantitative characterization of dynamical behavior, are related to the exponentially fast divergence or convergence of nearby orbits in phase space. A system with one or more positive Lyapunov exponents is defined to be chaotic. Our method is rooted conceptually in a previously developed technique that could only be applied to analytically defined model systems: we monitor the long-term growth rate of small volume elements in an attractor. The method is tested on model systems with known Lyapunov spectra, and applied to data for the Belousov-Zhabotinskii reaction and Couette-Taylor flow.
    Physica D Nonlinear Phenomena 07/1985; 16(3-16):285-317. DOI:10.1016/0167-2789(85)90011-9 · 1.83 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: A a-step downscaling method for general circulation model (GCM) outputs is presented. The investigation is based on daily upper air geopotential and humidity fields (Atlantic-European sector) and surface observations (52 German climate stations) from 1966 to 1993. In the first step, significant circulation patterns are identified using cluster analysis which takes advantage of information exclusively from upper air fields. This leads to composite charts which can be readily interpreted synoptically, but which only explain a portion of the Variance of the local weather elements. In the second step, a conditional (weather pattern-dependent) stepwise screening regression analysis is performed for each weather element and 6 German climate regions. A principal finding is that the modelled (downscaled) local climate is in good agreement with observations, particularly for the temperature regime, due to the fact that a major part of the variance is explained after the 2 steps. Including 700 hPa humidity slightly improves the explained variance. An application to a GCM control run is added. It shows that the method is capable of reconstructing interannual variability of local weather elements.
    Climate Research 10/1997; 8(3):195-207. DOI:10.3354/cr008195 · 2.71 Impact Factor
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: Climatological records users, frequently, request time series for geographical locations where there is no observed meteorological attributes. Climatological conditions of the areas or points of interest have to be calculated interpolating observations in the time of neighboring stations and climate proxy. The aim of the present work is the application of reliable and robust procedures for monthly reconstruction of precipitation time series. Time series is a special case of symbolic regression and we can use Artificial Neural Network (ANN) to explore the spatiotemporal dependence of meteorological attributes. The ANN seems to be an important tool for the propagation of the related weather information to provide practical solution of uncertainties associated with interpolation, capturing the spatiotemporal structure of the data. In practice, one determines the embedding dimension of the time series attractor (delay time that determine how data are processed) and uses these numbers to define the network's architecture. Meteorological attributes can be accurately predicted by the ANN model architecture: designing, training, validation and testing; the best generalization of new data is obtained when the mapping represents the systematic aspects of the data, rather capturing the specific details of the particular training set. As illustration one takes monthly total rainfall series recorded in the period 1961?2005 in the Rio Grande do Sul ? Brazil. This reliable and robust reconstruction method has good performance and in particular, they were able to capture the intrinsic dynamic of atmospheric activities. The regional rainfall has been related to high-frequency atmospheric phenomena, such as El Niño and La Niña events, and low frequency phenomena, such as the Pacific Decadal Oscillation.
    Advances in Geosciences 01/2007; 10. DOI:10.5194/adgeo-10-67-2007
Show more