| Title | Sub-seasonal 9+1 member hindcasts: 1989-2016: NN bias correction |
|---|---|
| Description | The flow-dependent bias-correction experiment consists of sub-seasonal hindcasts performed with IFS cycle 48r1, in which model-error tendencies are estimated online using neural networks. Hindcasts are initialised on the 15th of each month from 1989 to 2016 and integrated forward for 32 days. For each start date, an ensemble of 10 members is generated, comprising one unperturbed control and nine perturbed members to sample initial-condition uncertainty. The model is run at TCo319 horizontal resolution with 137 vertical levels. The bias correction is implemented as an additive tendency applied to temperature and horizontal winds at each model time step. These tendencies are predicted by three neural networks (one each for zonal wind, meridional wind, and temperature) trained to diagnose 6-hourly mean model-error tendencies from the instantaneous model state. The networks follow the architecture of Espinoza et al. (2022) and use inputs including three-dimensional wind and temperature fields, surface pressure, solar geometry, and geographical location. This experiment is designed to assess the impact of a flow-dependent, state-aware representation of model error on sub-seasonal forecast skill. |
| DOI | 10.21957/xtnt-2y33 |
| Experiment ID | ia5q |
| Experiment class | ECMWF Research Department |
| Examples |
Retrieve perturbed mebers from 15 September start hindcasts of T/U/V for years 1989 to 2016. Retrieve perturbed mebers from 15 April start hindcasts of 2m temperature for years 1989 to 2016. |
Sub-seasonal 9+1 member hindcasts: 1989-2016: NN bias correction
To retrieve the data described in this experiment, you will need to use the ECMWF Web API with the example(s) given on this page. Please note that when accessing the data you are bound by the ECMWF terms of use.