Abstract: To address the insufficient capability of existing unit hydrographs in representing the runoff routing process, this paper adopts a Long Short-Term Memory (LSTM) network to capture the influence of antecedent and current net rainfall on the runoff routing process, and incorporates a Softmax function to generate a time-varying unit hydrograph that is free from the constraints of empirical formulas, termed LSTM-TSUH. Taking seven small and medium-sized catchments in Fujian Province as study areas, hourly flood simulation experiments were conducted, and the evaluation was carried out from the perspectives of unit hydrograph characteristics and model forecast accuracy. The results show that the LSTM-TSUH model, which considers the influence of net rainfall sequences, ranks best in both Kling-Gupta efficiency and Top Peak Error, second only to LSTM-Q in Nash-Sutcliffe efficiency, and the Softmax-based unit hydrograph outperforms the empirical unit hydrograph in both the overall hydrological process and peak fitting. Moreover, the LSTM-TSUH model successfully captures the synergistic driving mechanism of antecedent and current net rainfall on the runoff routing process, further verifying that data-driven models can explore and reveal unknown hydrological laws. Keywords: time-varying unit hydrograph; net rainfall sequence influence; long short-term memory network; Softmax function; flood simulation