Improving Daily Rainfall Downscaling Over Taiwan's Complex Terrain With Surface‐Wind‐Informed Deep Learning Model

ABSTRACT This study evaluates whether an encoder‐decoder deep neural network with multi‐head attention can bias‐correct and downscale ERA5 daily rainfall to Taiwan's 5‐km TCCIP gridded observations. The Encoder–Decoder with Attention (EDA) model ingests ERA5 rainfall, 10‐m winds, coarsened TCCIP rainfall, and 5‐km topography, is trained with consecutive‐year splits and a weighted MSE loss, and is benchmarked against bias correction spatial disaggregation (BCSD) and a rainfall‐only variant. Across Taiwan's five seasonal rainfall regimes, EDA improves the placement of orographic rainfall, reduces the low‐intensity wet bias in ERA5, and better reproduces RX1day (annual maximum 1‐day precipitation), RX10mm (number of days with daily precipitation ≥ 10 mm), CDD (maximum number of consecutive dry days), and interannual variability than BCSD. The largest gains occur in seasons dominated by monsoon and typhoon forcing, indicating that auxiliary wind information helps the network learn the statistical relationships that better reproduce rainfall patterns associated with synoptic flow interacting with complex topography. The results show that attention‐based statistical downscaling can improve the reproduction of high‐resolution gridded rainfall fields for regions with steep terrain.