The traditional Muskingum model has difficulty representing complex hydraulic behaviors under high-flow conditions because it relies on simplified assumptions and fixed parameters. To address the pronounced nonlinearity and non-stationarity of flood routing in the arid Tarim River Basin, a hybrid forecasting framework was developed by coupling the Muskingum method with multiple machine learning algorithms (Ridge, LASSO, RF, and LSTM) to predict and correct Muskingum residuals. Global Muskingum parameters were identified using the L-BFGS-B algorithm to represent basin-scale routing characteristics. For rolling forecast, a multidimensional feature space was constructed by integrating routing gradients and hydraulic interaction terms. The results indicated that all hybrid models outperformed the traditional Muskingum method across lead times. The Ridge-based hybrid model achieved the best performance at short lead times, with the Nash–Sutcliffe efficiency (NSE) at a 4 h lead time increasing from 0.56 for the physical baseline to 0.977. For longer lead times (12–24 h), the LASSO-based hybrid model demonstrated higher robustness, which was attributed to L1-regularization-based feature selection. The key scientific contribution of this work lies in proposing a lead-time-dependent adaptive modeling strategy, revealing the structural characteristics of the residuals of the Muskingum model, and demonstrating that, in the study basin, simple linear models outperform complex models in multi-step correction. Overall, the proposed framework alleviates systematic underestimation during high-flow periods and provides a predictive scheme for arid-region rivers that preserves physical interpretability while improving forecasting accuracy.

The traditional Muskingum model has difficulty representing complex hydraulic behaviors under high-flow conditions because it relies on simplified assumptions and fixed parameters. To address the pronounced nonlinearity and non-stationarity of flood routing in the arid Tarim River Basin, a hybrid forecasting framework was developed by coupling the Muskingum method with multiple machine-learning algorithms (Ridge, LASSO, RF, and LSTM) to predict and correct Muskingum residuals. Global Muskingum parameters were identified using the L-BFGS-B algorithm to represent basin-scale routing characteristics. For rolling forecast, a multidimensional feature space was constructed by integrating routing gradients and hydraulic interaction terms. The results indicated that all hybrid models outperformed the traditional Muskingum method across lead times. The Ridge-based hybrid model achieved the best performance at short lead times, with the Nash–Sutcliffe efficiency (NSE) at a 4-h lead time increasing from 0.56 for the physical baseline to 0.977. For longer lead times (12–24 h), the LASSO-based hybrid model demonstrated higher robustness, which was attributed to L1-regularization-based feature selection. Overall, the proposed framework alleviates systematic underestimation during high-flow periods and provides a predictive scheme for arid-region rivers that preserves physical interpretability while improving forecasting accuracy.

In the past decades，some progress has been made in the theory and practice of water and sediment regulation of the Yellow River．However，under the current engineering conditions, the feasibility and model of the whole river water and sediment regulation combined with the cascade reservoirs of Long ( Longyangxia) , Liu ( Liujiaxia) , Wan ( Wanjiazhai) , San ( Sanmenxia) and Xiao ( Xiaolangdi) are rarely discussed. Based on a comprehensive analysis of the adjustable water volume of the Long-Liu cascade reservoirs, interval inflow and diversion water volume of the whole river under the multi-year average condition, this paper demonstrated the feasibility of the whole river water and sediment regulation under the current engineering conditions. Two modes were put forward including the conventional mode and the unconventional mode, and the effects of two modes were compared and analyzed. The results show that the sediment discharge of Xiaolangdi Reservoir under the two modes is 145.1 million tons and 191.3 million tons respectively and the sediment deposition of the downstream channel is 11.3 million tons and 22.1 million tons respectively. No matter which mode of water and sediment regulation is adopted, Xiaolangdi Reservoir can achieve effective erosion in reservoir during the period of discharging before flood, and at the same time, it has little influence to the siltation of river channe. Therefore, under the premise of strengthening the fine operation of the reservoir and strict interval diversion management, the whole river water and sediment regulation mode under the current engineering conditions is completely feasible.