采用机理和机器学习模型的山洪事件集合模拟与校准

Key Points

• The research aims to enhance the precision of flood event simulations in mountainous small watersheds.
• Utilized MIX_HM, DTVGM, and LSTM models in a multi-model framework.
• Employed quantile mapping and LSTM residual correction strategies for secondary calibration.
• Validated the framework using 38 flood events from 2006 to 2021 in the Yangwu River catchment.
• LSTM showed Nash efficiency coefficients of 0.73 during calibration and 0.63 during validation, outperforming MIX_HM and DTVGM.
• Residual correction improved Nash efficiency for MIX_HM, DTVGM, and LSTM during validation by 93.3%, 103.5%, and 21.2%, respectively.
• Ensemble simulation achieved Nash efficiency coefficients of 0.80 and 0.79 during calibration and validation, improving on LSTM's performance.

Abstract