Daily quantitative precipitation forecasts based on the analogue method: Improvements and application to a French large river basin
Introduction
Many water-related stakeholders need quantitative precipitation forecasts as reliable as possible to anticipate discharges in river basins several hours or days ahead. For example, French operational flood forecasting services require precipitation forecasts several days ahead to anticipate flood risks (Lacaze et al., 2008), and hydroelectricity power producers need accurate and reliable precipitation forecasts to anticipate the discharge evolution along the regulated rivers and their tributaries in order to perform their activities and to provide safety control (Bompart et al., 2009).
Nowadays forecast uncertainties related to meteorological predictions and implied by modelling processes are more and more taken into account, leading to probabilistic quantitative precipitation forecasts (PQPFs) that provide ensemble forecasts. In particular, this kind of forecast enables to evaluate the risk of extreme events. At least, two approaches for producing PQPFs are commonly used: (i) regional ensemble weather forecasts based on dynamical approaches (e.g. COSMO-LEPS, Marsigli et al., 2005; PEARP, Thirel et al., 2008; EFAS, Thielen et al., 2009; ECMWF-ENS, Miller et al., 2010), (ii) statistical approaches based on a search for analogues (e.g. Obled et al., 2002, Hamill and Whitaker, 2006, Messner and Mayr, 2010, Marty et al., 2012).
The principle of analogue methods (AMs) is to use deterministic meteorological outputs of a target day D as an input for a statistical search of similar past days in terms of general circulation patterns. Similarity is measured on a set of relevant predictor variables computed by Numerical Weather Prediction (NWP) models. Finally the analogue situations selected are used to produce a sample of observed daily precipitations, which provides a probabilistic forecast for the day D.
The AMs assume a relationship between large-scale variables (predictors) and a local-scale target variable (predictand). Different predictors can be found in the literature. Table 1 presents a short description of a selection of representative models developed to predict precipitation fields under various climates. Note that this table does not include methods using combination of predictors in linear regression to estimate predictands. One can note that the number of predictors may differ from one region to another. Timbal et al. (2008) suggest defining combinations of predictors adapted for each season and each climatic region.
The predictors can be divided into two categories. Most AM developments deal with one or more low flow atmospheric fields, i.e. descriptors of the meteorological state of the atmosphere at synoptic scale and a set of geopotential heights is usually chosen. The number of geopotential heights and their corresponding pressure level varies from one study to another: Altava-Ortiz et al. (2006) used three geopotential heights at 500, 850 and 1000 hPa to estimate precipitation amounts for one major heavy rainfall event in Catalonia (Spain) while Diomede et al. (2008) selected the geopotential at 500 hPa to estimate precipitation in Northern Italy. Following Bontron and Obled (2005), Horton et al. (2012) and Chardon et al. (2014) selected analogous situations based on geopotential heights at 500 and 1000 hPa-level in Switzerland and in France, respectively.
The second category of predictors is made of moisture related variables with among them, the large-scale precipitation outputs PRCP from NWP models (e.g. Diomede et al., 2014, Turco et al., 2011). Themeßl et al. (2011) found PRCP the most important predictors for local precipitation far beyond the other variables related to the state of the atmosphere. However Dayon et al. (2015) have compared different versions of AM and pointed out large possible biases in an application to France, when PRCP is used. In Catalonia, Barrera et al. (2007) added the humidity at 1000 hPa to the set of predictors suggested formerly by Altava-Ortiz et al. (2006). Similarly Bliefernicht and Bárdossy (2007) selected three predictors (two geopotential heights and the moisture flux at 700 hPa defined by the product of the specific humidity at 700 hPa and the westerly wind) to predict rainfall in the Rhine basin in Germany. Bontron (2004) and later Gibergans-Báguena and Llasat (2007) and Diomede et al. (2008) tested different combinations of variables as predictors. Diomede et al. (2008) showed that geopotential at 500 hPa and vertical velocity at 700 hPa are the most relevant set of predictors to estimate precipitation in Northern Italy while Bontron (2004) highlighted that the best moisture related variable to be used in addition to synoptic scale variables to estimate precipitation in South-Eastern France is a combination of relative humidity and total column precipitable water. More exhaustively, Gibergans-Báguena and Llasat (2007) have identified 7 variables among a list of 33 thermodynamic descriptors and the AM combining these 7 variables to geopotential fields as predictors outperforms the other tested AMs. These works show that improved analogues sorting technique can be expected by incorporating new levels of analogy (i.e. incorporating new predictors and related similarity criteria). Note that analogy between meteorological situations can be also used in a post-processing of precipitation forecasts to improve the prediction skill of NWP models (e.g. Wang and Fan, 2009).
We will focus here on the application of AMs, and more specifically on one of the reference PQPF methods already used in France in an operational context and suggested by Bontron and Obled (2005) following the work of Bontron (2004). This former method was developed considering small basins located in South-Eastern France, subject to flash-flood events. The aim of this paper is to extend the scope of this AM, considering large French river basins under oceanic influences. This study presents then two improvements of the reference method and compares the relative performances of the latter with two alternatives, on the large French basin of the Saône river. We consider two contexts: (i) a “perfect prognosis context” where the meteorological fields of the target day D are directly built with observed data or re-analysed by a NWP model; (ii) an “operational forecast” context where the meteorological fields are provided by an operational NWP model, including forecast errors.
Section 2 presents the study area and the datasets used, and Section 3 introduces the scores selected for the evaluation of probabilistic forecasts. Section 4 gives the main principles of the benchmark method for searching analogues, based on a two-step procedure, plus two alternatives adding supplementary explanatory variables, air temperature and vertical air motion. The main results of the comparison of the methods are analysed in the fifth section, considering the “perfect prognosis” and the “operational forecast” contexts. The last section draws general conclusions.
Section snippets
Study area
This application of the AM focuses on the Saône River basin located in eastern France. With a response time of 12 to 24 h, the Saône River is one of the most important tributaries of the Rhône River. This basin is under oceanic influences and westerly fluxes generating rainfall when large fronts pass but it can be also affected by heavy rainfall events extending from the Mediterranean region. For this application, the total drainage area (~ 30,000 km2) is divided into three sub-catchments defined
Numerical scores for performance evaluation
Performance evaluation of any forecasting system relies on numerical scores to assess the accuracy of the forecasts (see Jolliffe and Stephenson (2003) for an extended list of scores). In this study, several scores are used:
- i)
A global performance criterion on the overall forecasts (i.e. considering no rain situations as well as precipitation event). The Continuous Ranked Probability Skill Score was chosen to perform the optimisation process;
- ii)
End-user oriented criteria based on contingency tables
Development of the analogue method
Duband (1970) initiated the development of the AM in France, which was followed by several academic works by Guilbaud (1997), Bontron (2004) and Ben Daoud (2010). This paper relates to the latter work, focussed on the case of large basin under oceanic influences, such as the Saône river basin. We will show how the introduction of additional variables enables to provide a better forecast skill.
Each time a new variable is introduced, a new step of analogues selection is created in the algorithm.
Perfect prognosis forecasts
During the development phase of the algorithm, parameters were optimised in a perfect prognosis mode over the calibration period, and improvements were confirmed over the validation period. In this section, the total length of available record was considered for searching analogues dates, i.e. from 01/08/1970 to 30/08/2002. That means an extension of 7 years, compared to the previous section (5 years of the validation period defined for Section 4, plus the 2 years 1970 and 1971 that were then
Conclusions
The aim of this work was the adaptation of an analogue method, initially developed in the 2000s for small to medium-sized mountain catchments, to large river basins mainly influenced by frontal systems at large scale. In this study, we focused on the Saône basin located in eastern France.
Hence several tests were performed in order to improve the selection of analogues over large river basins, by proposing new predictors that are useful for a weather forecaster. Especially, different predictors
Acknowledgements
The financial support provided by CNR and Cemagref-Irstea for the PhD research of A. Ben Daoud is gratefully acknowledged. The NCEP/NCAR and ECMWF ERA-40 re-analysis data were obtained from the NOAA web site (http://www.cdc.noaa.gov/) and from the ECMWF data server (http://data.ecmwf.int/data/d/era40_daily/), respectively. The authors acknowledge Météo-France for providing the ECMWF forecasts and the Safran data set for research purposes. The authors acknowledge the anonymous reviewers for
References (60)
Nonlinear analog predictor analysis: a coupled neural network/analog model for climate downscaling
Neural Netw.
(2007)- et al.
Improvement of the analog forecasting method by using local thermodynamic data. Application to autumn precipitation in Catalonia
Atmos. Res.
(2007) - et al.
Quantitative precipitation forecasts: a statistical adaptation of model outputs through an analogues sorting approach
Atmos. Res.
(2002) Mapping mean annual river discharges: geostatistical developments for incorporating river network dependencies
J. Hydrol.
(2006)- et al.
Assessment of an analogue downscaling method for modelling climate change impacts on runoff
Journal of Hydrology
(2012) - et al.
Statistical precipitation downscaling in central Sweden with the analogue method
J. Hydrol.
(2005) - et al.
Statistical downscaling of climate forecast system seasonal predictions for the Southeastern Mediterranean
Atmos. Res.
(2012) - et al.
Application of the MM5 and the analogous method to heavy rainfall events. The case of 16–18th October 2003 in Catalonia (NE Spain)
Adv. Geosci.
(2006) - et al.
Heavy rain prediction using deterministic and probabilistic models — the flash flood cases of 11–13 October 2005 in Catalonia (NE Spain)
Adv. Geosci.
(2007) Améliorations et développements d'une méthode de prévision probabiliste des pluies par analogie
Application à la prévision hydrologique sur les grands bassins fluviaux de la Saône et de la Seine (in French)
(2010)