@article{fdi:010053898,
  title = {{S}tochastic downscaling of precipitation with neural network conditional mixture models},
  author = {{C}arreau, {J}ulie and {V}rac, {M}.},
  editor = {},
  language = {{ENG}},
  abstract = {{W}e present a new class of stochastic downscaling models, the conditional mixture models ({CMM}s), which builds on neural network models. {CMM}s are mixture models whose parameters are functions of predictor variables. {T}hese functions are implemented with a one-layer feed-forward neural network. {B}y combining the approximation capabilities of mixtures and neural networks, {CMM}s can, in principle, represent arbitrary conditional distributions. {W}e evaluate the {CMM}s at downscaling precipitation data at three stations in the {F}rench {M}editerranean region. {A} discrete ({D}irac) component is included in the mixture to handle the "no-rain" events. {P}ositive rainfall is modeled with a mixture of continuous densities, which can be either {G}aussian, log-normal, or hybrid {P}areto (an extension of the generalized {P}areto). {CMM}s are stochastic weather generators in the sense that they provide a model for the conditional density of local variables given large-scale information. {I}n this study, we did not look for the most appropriate set of predictors, and we settled for a decent set as the basis to compare the downscaling models. {T}he set of predictors includes the {N}ational {C}enters for {E}nvironmental {P}rediction/{N}ational {C}enter for {A}tmospheric {R}esearch ({NCEP}/{NCAR}) reanalyses sea level pressure fields on a 6 x 6 grid cell region surrounding the stations plus three date variables. {W}e compare the three distribution families of {CMM}s with a simpler benchmark model, which is more common in the downscaling community. {T}he difference between the benchmark model and {CMM}s is that positive rainfall is modeled with a single {G}amma distribution. {T}he results show that {CMM} with hybrid {P}areto components outperforms both the {CMM} with {G}aussian components and the benchmark model in terms of log-likelihood. {H}owever, there is no significant difference with the log-normal {CMM}. {I}n general, the additional flexibility of mixture models, as opposed to using a single distribution, allows us to better represent the distribution of rainfall, both in the central part and in the upper tail.},
  keywords = {},
  booktitle = {},
  journal = {{W}ater {R}esources {R}esearch},
  volume = {47},
  numero = {},
  pages = {{W}10502},
  ISSN = {0043-1397},
  year = {2011},
  DOI = {10.1029/2010wr010128},
  URL = {https://www.documentation.ird.fr/hor/fdi:010053898},
}