An adversarial learning approach to forecasted wind field correction with an application to oil spill drift prediction

Reanalysis wind fields are obtained by correcting the numerically forecasted wind fields based on observation data (i.e., either remote sensing or in-situ observations, or both). Although they are more accurate than forecasted wind fields, reanalysis wind fields tend to have time latencies because they can only be released after the observations are obtained. In order to produce accurate estimates of wind fields in a more timely manner, we develop an adversarial learning approach to correcting forecasted wind fields to be close to reanalysis wind fields. The adversarial learning approach is conducted by an adversarial ConvLSTM network (ACLN) framework that consists of a corrector and a discriminator. The corrector aims at comprehensively capturing both spatial and temporal characteristics of a sequence of forecasted wind fields and producing a corrected forecast wind field for the final field in the sequence. The discriminator tries to distinguish corrected forecast wind field from its corresponding reanalysis wind field. The training of ACLN is alternate between the corrector and the discriminator in an adversarial fashion. The adversarial training mechanism enhances the corrector’s representational power. Additionally, the corrector exploits a residual learning architecture that effectively learns the differences between forecasted wind fields and the corresponding reanalysis wind fields. In this scenario, the well trained corrector requires neither reanalysis wind fields nor observations such that it can correct forecasted wind fields in a timely manner. Furthermore, corrected forecast wind fields are employed for oil spill drift prediction. Extensive experiments validate the effectiveness of the proposed ACLN framework in forecasted wind field correction along with oil spill drift prediction. Compared with ECMWF numerical forecasts, the ACLN achieves an average reduction of 6.2%, 6.9%, and 10.6% in RMSE, MAE, and MAPE, respectively. Compared with a basic drift prediction method, the ACLN based prediction method reduces the error by about 5000 m in the Sanchi oil spill accident. The source codes are available at https://github.com/liyongqingupc/ACLN-WindFieldCorrection, providing a baseline for correcting forecasted wind fields.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.