Abstract: Facing the challenges in reconstructing historical extreme low runoffs and quantifying current drought defense capabilities, this study constructs a framework integrating the Transformer model with an implicit encoding of current defense conditions strategy. By using self-attention mechanisms, we develop a new model that captures the hydrological memory and embeds the modern land surface and engineering regulation patterns via training on the data of the present time. This model is validated against the extreme drought in the Chongzhen era of the late Ming Dynasty (1637-1643) in the middle Yellow River Basin, which achieves a high accuracy in modern testing (NSE = 0.82) and reproduces the historical drought events effectively. Scenario analysis indicates that a recurrence of such a drought event today, despite the existing current drought defense system, would cause an annual grain yield loss of greater than 40% and an economic loss above 1.5% of the existing GDP. These findings reveal the vulnerability boundaries of the current defense systems under extreme climate conditions, helping enhance the resilience of a river basin.

Key words: historical hydrology, transformer deep learning, extreme drought in Chongzhen era, middle Yellow River Basin