Kaidu River Basin, where located within the Xinjiang Uygur Autonomous Region, is one of the four major tributaries of the Tarim River. This study examines the challenges in forecasting nonlinear and nonstationary flood processes, with particular emphasis on high-altitude mountainous watersheds.Based on the observed daily runoff data from 1959 to 2018 at the Dashankou Hydrological Station and the extracted 152 flood events by POT (Peaks Over Threshold) method. A hybrid model combining VMD(Variational Mode Decomposition) and LSTM (Long Short-Term Memory) was developed to flood simulate and forecast without any future referenceinformation, and analyze the role of different sub-modal decomposition features in runoff sequence. After Bayesian optimization, both the standalone LSTM and the hybrid VMD-LSTM models achieved excellent performance in simulating runoff sequences, with R² values of 0.99 and 0.93, respectively. The VMD-LSTM model significantly reduces the complexity of runoff signals by decomposing them into multiple sub-modal components across low- to high-frequency bands, This decomposition enhances the model’s ability to capture the dynamics of flood events. Notably, the VMD-LSTM model achieves an 88.2% same-day peak discharge detection rate, substantially outperforming the standalone LSTM model, which only reaches 2.6%. Through the sub-model characteristics from low to high frequencies, the hybrid model hierarchically resolves the determinism of runoff accumulation and the uncertainty of short-term fluctuations during flood development. Even without access to future observational data, it can still effectively capture the flood evolution process, and At all four forecasting horizons—1, 5, 10, and 15 days—it consistently exhibits lower cumulative errors than the single LSTM model. While the LSTM model performed well in capturing the overall runoff regime in high-altitude mountain basins, its integration with the VMD significantly improved flood event detection and forecasting performance. This hybrid framework provides valuable scientific insights for improving mountain flood early warning and defense strategies. • LSTM model demonstrates outstanding simulation performances for both the long-term runoff sequence and its sub-models • VMD-LSTM excels in flood events detection by leaning the more stable signals from sub-modal sequences. • VMD-LSTM model forecasts floods process well by leveraging its memory of stable sub-sequences, even without any future signal inputs.
Shi et al. (Wed,) studied this question.