Comparison of statistical ensemble methods for probabilistic nowcasting of precipitation

• Main Authors: Wei Tao, Aitor Atencia, Yang Li, Xuexing Qiu, Zhiming Kang, Yong Wang
• Format: Article
• Language: English
• Published: Borntraeger 2020-10-01
• Series: Meteorologische Zeitschrift
• Subjects: probabilistic precipitation forecast; neighborhood method; time-lagged method
• Online Access: http://dx.doi.org/10.1127/metz/2020/1020

In the nowcasting of precipitation, predictability often varies with lead times and rain patterns. Ensemble techniques have been developed to deal with forecast uncertainty. Both “time-shifting” and “time-lagged” methods are pragmatic ensemble approaches to derive probabilistic precipitation from deterministic forecasts. This study synthesizes these two statistical probabilistic methods with customized parameters to serve different lead times and precipitation intensities. Thereafter, an integrated method is proposed by combining the time-shifting and the time-lagged methods. The performance of the three methods under various precipitation intensities and lead times is evaluated with the Brier score and Receiver Operating Characteristic curve (ROC) area. The time-shifting method is found to be superior in the forecasting of light rain. For medium precipitation, the time-shifting method gives a better Brier score while the time-lagged and integrated methods yields a higher ROC area. The advantages of the integrated method are pronounced when heavy precipitation occurs. Based on the performance of the above three methods, this study introduces a weighted index to optimize the integrated method, which gives a higher weight to time-shifting related forecasts for light and medium precipitation and keeps an equal weight for heavy rainfall. The weighting allows the integrated method to maintain or exceed the skill of the time-shifting and time-lagged methods. Afterwards, calibration is conducted based on the reliability diagram. The calibrated results outperform the raw products. Finally, a test with independent samples supports the suggestion that the presented probabilistic approach, including combination, weighting, and calibration, could be applied with confidence in an operational system.