Neural network based daily precipitation generator (NNGEN-P)

JEAN-PHILIPPE BOULANGER Æ FERNANDO MARTINEZ Æ OLGA PENALBA Æ ENRIQUE CARLOS SEGURA

CLIMATE DYNAMICS

SPRINGER

Año: 2007 p. 307 - 324

0930-7575

Daily weather generators are used in manyapplications and risk analyses. The present paper exploresthe potential of neural network architectures todesign daily weather generator models. Focusing thisfirst paper on precipitation, we design a collection ofneural networks (multi-layer perceptrons in the presentcase), which are trained so as to approximate theempirical cumulative distribution (CDF) function forthe occurrence of wet and dry spells and for the precipitationamounts. This approach contributes to correctsome of the biases of the usual two-step weathergenerator models. As compared to a rainfall occurrenceMarkov model, NNGEN-P represents fairly wellthe mean and standard deviation of the number of wetdays per month, and it significantly improves the simulationof the longest dry and wet periods. Then, wecompared NNGEN-P to three parametric distributionfunctions usually applied to fit rainfall cumulative distributionfunctions (Gamma, Weibull and doubleexponential).A data set of 19 Argentine stations wasused. Also, data corresponding to stations in the UnitedStates, in Europe and in the Tropics were includedto confirm the results. One of the advantages ofNNGEN-P is that it is non-parametric. Unlike otherparametric function, which adapt to certain types ofclimate regimes, NNGEN-P is fully adaptive to theobserved cumulative distribution functions, which, onsome occasions, may present complex shapes. Ongoingworks will soon produce an extended version ofNNGEN to temperature and radiation.