Statistical post‐processing of ensemble forecasts of temperature in Santiago de Chile

Abstract

Modelling forecast uncertainty is a difficult task in any forecasting problem. In weather forecasting a possible solution is the use of forecast ensembles, which are obtained from multiple runs of numerical weather prediction models with various initial conditions and model parametrizations to provide information about the expected uncertainty. Currently all major meteorological centres issue forecasts using their operational ensemble prediction systems. However, it is a general problem that the spread of the ensemble is too small compared to observations at specific sites resulting in under‐dispersive forecasts, leading to a lack of calibration. In order to correct this problem, various statistical calibration techniques have been developed in the last two decades. In the present work different post‐processing techniques were tested for calibrating nine member ensemble forecasts of temperature for Santiago de Chile, obtained by the Weather Research and Forecasting model using different planetary boundary layer and land surface model parametrizations. In particular, the ensemble model output statistics and Bayesian model averaging techniques were implemented and, since the observations are characterized by large altitude differences, the estimation of model parameters was adapted to the actual conditions at hand. Compared to the raw ensemble, all tested post‐processing approaches significantly improve the calibration of probabilistic forecasts and the accuracy of point forecasts. The ensemble model output statistics method using parameter estimation based on expert clustering of stations (according to their altitudes) shows the best forecast skill.

Operational statistical postprocessing of temperature ensemble forecasts with station‐specific predictors

“…Many studies have shown that a good choice for the length of the sliding‐window training period ranges from 20 to 40 days (e.g. Feldmann et al ., 2015; Möller et al ., 2015; Díaz et al ., 2019). To test the sensitivity of coefficient calibration to the length of the training period, we compared 15, 30 and 90 day sliding windows when estimating the calibration coefficients.…”

“…Other spatially adaptive approaches have been studied for temperature and wind speed forecasts. A similar idea as that here is used in cluster‐based approaches, where the calibrated stations are grouped by their altitude (Díaz et al ., 2019) or climatological characteristics (Lerch and Baran, 2017). Instead of grouping the stations, we try to include the station‐specific characteristics linearly to the model.…”

Intercomparison of multiple statistical methods in post‐processing ensemble precipitation and temperature forecasts

“…Diaz et al . (2019) pointed out that the distance measure should include the station climatology and ensemble forecast errors. Therefore, when using the K ‐means method for classifying these stations, the similarity between two stations is defined by the forecast performance (including calibration metric Δ, forecast error metric MAE and forecast skill metric CRPSS) and climate division.…”

“…For example, Lerch and Baran (2017) found using the distribution of forecast errors as the similarity measure augments the training data, which helps to improve the predictive performance of the post-processing methods. Diaz et al (2019) pointed out that the distance measure should include the station climatology and ensemble forecast errors. Therefore, when using the K-means method for classifying these stations, the similarity between two stations is defined by the forecast performance (including calibration metric Δ, forecast error metric MAE and forecast skill metric CRPSS) and climate division.…”

Comparison of the BMA and EMOS statistical methods for probabilistic quantitative precipitation forecasting

“…Díaz et al . (2019) used the BMA and EMOS post‐processing methods for 2 min temperature probabilistic forecasts. They showed that these post‐processing methods result in a significant decrease in the continuous ranked probability scores (CRPS) of probabilistic forecasts.…”