Dataset Information

Deep learning for twelve hour precipitation forecasts.

ABSTRACT: Existing weather forecasting models are based on physics and use supercomputers to evolve the atmosphere into the future. Better physics-based forecasts require improved atmospheric models, which can be difficult to discover and develop, or increasing the resolution underlying the simulation, which can be computationally prohibitive. An emerging class of weather models based on neural networks overcome these limitations by learning the required transformations from data instead of relying on hand-coded physics and by running efficiently in parallel. Here we present a neural network capable of predicting precipitation at a high resolution up to 12 h ahead. The model predicts raw precipitation targets and outperforms for up to 12 h of lead time state-of-the-art physics-based models currently operating in the Continental United States. The results represent a substantial step towards validating the new class of neural weather models.

SUBMITTER: Espeholt L

PROVIDER: S-EPMC9436943 | biostudies-literature | 2022 Sep

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Deep learning for twelve hour precipitation forecasts.

Espeholt Lasse L Agrawal Shreya S Sønderby Casper C Kumar Manoj M Heek Jonathan J Bromberg Carla C Gazen Cenk C Carver Rob R Andrychowicz Marcin M Hickey Jason J Bell Aaron A Kalchbrenner Nal N

Nature communications 20220901 1

Existing weather forecasting models are based on physics and use supercomputers to evolve the atmosphere into the future. Better physics-based forecasts require improved atmospheric models, which can be difficult to discover and develop, or increasing the resolution underlying the simulation, which can be computationally prohibitive. An emerging class of weather models based on neural networks overcome these limitations by learning the required transformations from data instead of relying on han ...[more]

PMID: 36050311

Similar Datasets

Project description:Abstract Seasonal precipitation forecasting is highly challenging for the northwest fringes of Europe due to complex dynamical drivers. Hybrid dynamical–statistical approaches offer potential to improve forecast skill. Here, hindcasts of mean sea level pressure (MSLP) from two dynamical systems (GloSea5 and SEAS5) are used to derive two distinct sets of indices for forecasting winter (DJF) and summer (JJA) precipitation over lead‐times of 1–4 months. These indices provide predictors of seasonal precipitation via a multiple linear regression model (MLR) and an artificial neural network (ANN) applied to four Irish rainfall regions and the Island of Ireland. Forecast skill for each model, lead time, and region was evaluated using the correlation coefficient (r) and mean absolute error (MAE), benchmarked against (a) climatology, (b) bias corrected precipitation hindcasts from both GloSea5 and SEAS5, and (c) a zero‐order forecast based on rainfall persistence. The MLR and ANN models produced skilful precipitation forecasts with leads of up to 4 months. In all tests, our hybrid method based on MSLP indices outperformed the three benchmarks (i.e., climatology, bias corrected, and persistence). With correlation coefficients ranging between 0.38 and 0.81 in winter, and between 0.24 and 0.78 in summer, the ANN model outperformed MLR in both seasons in most regions and lead‐times. Forecast skill for summer was comparable to that in winter and for some regions/lead times even superior. Our results also show that climatology and persistence performed better than direct use of bias corrected dynamical outputs in most regions and lead‐times in terms of MAE. We conclude that the hybrid dynamical–statistical approach developed here—by leveraging useful information about MSLP from dynamical systems—enables more skilful seasonal precipitation forecasts for Ireland, and possibly other locations in western Europe, in both winter and summer. In this study, hindcasts of mean sea level pressure (MSLP) from dynamical systems (GloSea5 and SEAS5) are used to derive two distinct sets of indices to forecast winter and summer precipitation. These indices provide predictors of seasonal precipitation via a multiple linear regression model (MLR) and an artificial neural network (ANN) applied to four Irish rainfall regions and the Island of Ireland. Both MLR and ANN models produced skilful precipitation forecasts of up to 4 months lead‐times with our hybrid method based on MSLP indices outperforming direct use of dynamical forecasts, climatology and persistence in every case.

Dataset Information

Deep learning for twelve hour precipitation forecasts.

Publications

Deep learning for twelve hour precipitation forecasts.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets