La régionalisation des précipitations : une revue bibliographique des développements récents Regional precipitation frequency analysis : a literature review of recent developments

Résumé

L'estimation de l'intensité de précipitations extrêmes est un sujet de recherche en pleine expansion. Nous présentons ici une synthèse des travaux de recherche sur l'analyse régionale des précipitations. Les principales étapes de l'analyse régionale revues sont les méthodes d'établissement de régions homogènes, la sélection de fonctions de distributions régionales et l'ajustement des paramètres de ces fonctions.

De nombreux travaux sur l'analyse régionale des précipitations s'inspirent de l'approche développée en régionalisation des crues. Les méthodes de types indice de crues ont été utilisées par plusieurs auteurs. Les régions homogènes établies peuvent être contiguës ou non-contiguës. L'analyse multivariée a été utilisée pour déterminer plusieurs régions homogènes au Canada. L'adéquation des sites à l'intérieur d'une région homogène a souvent été validée par une application des L-moments, bien que d'autres tests d'homogénéité aient aussi été utilisés.

La loi générale des valeurs extrêmes (GEV) est celle qui a le plus souvent été utilisée dans l'analyse régionale des précipitations. D'autres travaux ont porté sur la loi des valeurs extrêmes à deux composantes (TCEV), de même que sur des applications des séries partielles.

Peu de travaux ont porté sur les relations intensité durée dans un contexte régional, ni sur les variations saisonnières des paramètres régionaux. Finalement, les recherches ont débuté sur l'application des concepts d'invariance d'échelle et de loi d'échelle. Ces travaux sont jugés prometteurs.

Abstract

Research on the estimation of extreme precipitation events is currently expanding. This field of research is of great importance in hydraulic engineering not only for the design of dams and dikes, but also for municipal engineering designs. In many cases, local data are scarce. In this context, regionalization methods are very useful tools. This paper summarizes the most recent work on the regionalization of precipitation. Steps normally included in any regionalization work are the delineation of homogenous regions, selection a regional probability distribution function and fitting the parameters.

Methods to determine homogenous regions are first reviewed. A great deal of work on precipitation was inspired by methods developed for regional flow analysis, especially the index flood approach. Homogenous regions can be contiguous, but in many cases they are not. The region of influence approach, commonly used in hydrological studies, has not been often applied to precipitation data. Homogenous regions can be established using multivariate statistical approaches such as Principal Component Analysis or Factorial Analysis. These approaches have been used in a number of regions in Canada. Sites within a homogenous region may be tested for their appropriateness by calculating local statistics such as the coefficient of variation, coefficient of skewness and kurtosis, and by comparing these statistics to the regional statistics. Another common approach is the use of L-moments. L-moments are linear combinations of ordered statistics and hence are not as sensitive to outliers as conventional moments. Other homogeneity tests have also been used. They include a Chi-squared test on all regional quantiles associated with a given non-exceedance probability, and a Smirnoff test used to validate the inclusion of a station in the homogenous region.

Secondly, we review the distributions and fitting methods used in regionalization of precipitation. The most popular distribution function used is the General Extreme Value (GEV) distribution. This distribution has been recommended for precipitation frequency analysis in the United Kingdom. For regional analysis, the GEV is preferred to the Gumbel distribution, which is often used for site-specific frequency analysis of precipitation extremes. L-moments are also often used to calculate the parameters of the GEV distribution. Some applications of the Two-Component Extreme Value (TCEV) distribution also exist. The TCEV has mostly been used to alleviate the concerns over some of the theoretical and practical restrictions of the GEV.

Applications of the Partial Duration Series or Peak-Over-Threshold (POT) approach are also described. In the POT approach, events with a magnitude exceeding a certain threshold are considered in the analysis. The occurrence of such exceedances is modelled as a Poisson process. One of the drawbacks of this method is that it is sometimes necessary to select a relatively high threshold in order to comply with the assumption that observations are independent and identically distributed (i.i.d.). The use of a re-parameterised Generalised Pareto distribution has also been suggested by some researchers.

Research on depth-duration relations on a regional scale is also discussed. Empirical approaches used in Canada and elsewhere are described. In most cases, the method consists of establishing a non-linear relationship between a quantile associated with a given duration and its return period to a reference quantile, such as a 1-hour rainfall with a 10-year return period. Depth duration relationships cannot be applied uniformly across Canada for events with durations exceeding two hours. Seasonal variability studies in regionalization are relatively scarce, but are required because of the obvious seasonality of precipitation. In many cases, seasonal regimes may lead to different regionalization approaches for the wet and the dry season. Some research has focused on the use of periodic functions to model regional parameters. Another approach consists of converting the occurrence data of a given event in an angular measurement and developing seasonal indices based on this angular measurement.

Other promising avenues of research include the scaling approach. The debate over the possibility of scale invariance for precipitation is ongoing. Simple scaling was studied on a number of precipitation data, but the fact that intermittence is common in precipitation regimes and the presence of numerous zero values in the series does not readily lead to proper application of this approach. Recent research has shown that multiple scaling is likely a more promising avenue.