Correction of sub-seasonal predictions of summer precipitation in Southwest China based on the Transformer-Seq2Seq-DNN ensemble deep learning model.

Numerical climate models usually cannot meet the operational service needs for sub-seasonal projections in East Asia. Modification of the preliminary predictions with downscaling methods is essential to improve prediction skills. In recent years, the downscaling process using deep learning algorithms has brought unprecedented changes for improving numerical climate forecasts. This study evaluated the sub-seasonal prediction skills of the China Meteorological Administration BCC-CPS-S2Sv2 (BCC S2S) model and the National Centers for Environmental Prediction CFSv2 (NCEP S2S) model for summer precipitation in Chongqing in Southwest China. To improve the prediction skills of the BCC S2S and NCEP S2S models, the ensemble deep learning models were established by integrating the Transformer, the Sequence-to-Sequence neural network (Seq2Seq), and the deep neural network (DNN). The ensemble deep learning models were established with both single-layer single-mode (single-layer model) and multi-layer multi-mode (multi-layer model) structures. The results indicate noticeable spatio-temporal variations in the sub-seasonal predictions of summer precipitation for both the BCC S2S and NCEP S2S models. The prediction skills of these two S2S models decrease with the extension of prediction timeliness and obtain a final skill of two pentads (10 days). Both the single-layer and multi-layer models improve the prediction skills for summer precipitation. However, great variabilities are found in the predictions using the single-layer model across the six pentads. Compared with the single-layer model, the multi-layer model shows higher skills by providing more stable predictions across the entire Chongqing from the first to the sixth pentad. The anomaly correlation coefficient (ACC) values are improved by 0.29, 0.46, 0.37, 0.31, 0.48, and 0.44 for the BCC S2S, and 0.21, 0.41, 0.40, 0.29, 0.47, and 0.40 for the NCEP S2S from the first to the sixth pentad, respectively. Generally, the multi-layer Transformer-Seq2Seq-DNN model shows excellent potential for sub-seasonal predictions of summer precipitation. [ABSTRACT FROM AUTHOR]