Postprocessing of 3-h precipitation over the South of France from November 2019 to October 2023

Context The task is the postprocessing of Météo-France’s ensemble prediction system AROME-EPS. The current dataset focuses on 3-h accumulated precipitation over the South of France (43.1–45.875° N, 1.1–5.875° E) at a forecast lead time of 24-h initialized at 1500 UTC daily. AROME-EPS produces ensembles with one control member and 16 perturbed members for forecasts up to 51 hours on four different initialization times. It produces a gridded ensemble with a horizontal resolution of 0.025°. Three-hour accumulated precipitation data from the gauge-adjusted radar product ANTILOPE are used as observations. Our period of interest spans 4 years from November 2019 to October 2023. The period from November 2019 to October 2022 is used as a training/validation dataset using 7-fold cross-validation to tune hyperparameters of the models. The folds are composed of all Mondays, all Tuesdays, and so on. The period from November 2022 to October 2023 is used as a hold-out test set. More information is available in the associated article (arXiv:2407.02125). Code to run postprocessing models is available on GitHub. Content The dataset is separated into the following files: X_trainval.npy, Y_trainval.npy: predictors and observations for the training/validation set. X_test.npy, Y_test.npy: predictors and observations for the test set. X_trainval_log.npy, X_test_log.npy: alternative predictors for the training/validation and test sets, where the raw precipitation is transformed using a log transformation ( log(x+eps) - log(eps) ) before applying mean/min/max/sd. Use for the DRU models. trainval_dow.npy: day of the week for the training/validation set. X_constant.npy: constant fields (e.g., orography) used for both the training/validation and test sets. X_raw_trainval.npy, X_raw_test.npy: raw ensemble forecasts for the training/validation and test sets. trainval_dow.npy: day of the week for each sample in the training/validation set. dates.txt: comma-separated list of the dates corresponding to the data of the training/validation and test sets. File name Shape X_trainval.npy / X_trainval_log.npy (n_trainval, H, W, n_pred) X_test.npy / X_test_log.npy (n_test, H, W, n_pred) Y_trainval.npy (n_trainval, H, W) Y_test.npy (n_test, H, W) trainval_dow.npy (n_trainval, ) X_constant.npy (H, W, n_constant) X_raw_trainval.npy (n_trainval, H, W, n_member) X_raw_test.npy (n_test, H, W, n_member) where: n_trainval = 1091 is the number of samples in the training/validation set, n_test = 365 is the number of samples in the test set, W = 192 and H = 112 are the width and height of the grid considered, n_pred = 24 is the number of predictors (without the constant fields), n_constant = 7 is the number of constant fields, n_member = 17 is the number of members in the raw ensemble. The predictors are provided in the following order : PRECIP_min, PRECIP_mean, PRECIP_max, PRECIP_sd , CAPE_min, CAPE_mean, CAPE_max, CAPE_sd, MAXREF_min, MAXREF_mean, MAXREF_max, MAXREF_sd, PAPT_min, PAPT_mean, PAPT_max, PAPT_sd, RH_min, RH_mean, RH_max, RH_sd, CVIDX_min, CVIDX_mean, CVIDX_max, CVIDX_sd. With the following notations for variables: PRECIP: total precipitation (including rain, snow, and soft hail) or its log-transformed version, CAPE: convective available potential energy, MAXREF: maximal reflectivity, PAPT: pseudo-adiabatic potential temperature, RH: relative humidity, CVIDX: AROME convection index. The following notations are for summary statistics of the ensemble forecast: min is the minimum, mean is the mean, max is the maximum, and sd is the standard deviation. The constant fields are available in the following order: altitude, land-sea mask (i.e., a binary field indicating the presence of land), the first four components of a principal component analysis (PCA) decomposition called AURHELY, and the distance to sea. For each grid point, AURHELY consists of applying PCA to the altitude vector formed by the 7x7 square patch centered around the grid point of interest. The predictors and constant fields are described in more detail in the article associated with the dataset (arXiv:2407.02125). Additional information The current dataset is obtained from a larger dataset available upon request through the Météo-France open data platform.