径流预报是提升水资源管理效率、保障流域水安全的重要手段,以湘江流域为研究对象,对比了新安江模型、SWAT模型、LSTM模型在不同时间尺度下的径流预报精度及适应性。研究结果表明:①各模型的计算效率有LSTM模型>新安江模型>SWAT模型。②在日尺度下,LSTM模型模拟效果整体最佳,新安江模型汛期模拟更优,SWAT模型在日尺度中的表现一般。③月尺度下,SWAT模拟精度明显提升,LSTM受输入条件影响较大。④总而言之,LSTM模型适用性最高,但缺乏物理机制支撑;新安江模型在汛期有较好的应用;SWAT模型能够反映流域内多种水循环要素的时空变化过程,但日尺度径流预报精度相对一般,适用于大范围的流域水资源综合评价。研究成果可为不同目标需求下的流域径流预报提供借鉴。

[Objective] The simplified structures of traditional hydrological models often limit their adaptability under complex hydroclimatic and anthropogenic conditions. Meanwhile, data-driven models such as deep learning frameworks typically lack explicit physical interpretability. To address these challenges, this study compares the performance of three representative runoff prediction methods: two physically based models—the Xinanjiang (XAJ) model and Soil & Water Assessment Tool (SWAT) model—and one data-driven model, the Long Short-Term Memory (LSTM) model. The Xiangjiang River Basin, a major tributary of the Yangtze River in southern China, is selected as the study area. This study aims to evaluate model performance and adaptability across multiple temporal scales. [Methods] The dataset included daily runoff records from 1971 to 2020 at the Xiangtan hydrological station, along with concurrent meteorological observations from 32 meteorological stations. Land use data with a spatial resolution of 1 km × 1 km were obtained from the Chinese Academy of Sciences, and soil property data were derived from the Harmonized World Soil Database (HWSD). The employed XAJ model estimated evaporation from three soil layers, routed surface runoff using the unit hydrograph method, and modeled interflow and groundwater components through the linear reservoir approach. The SWAT model divided the river basin into 13 subbasins and 192 hydrological response units (HRU) using high-resolution spatial inputs. For both models, the SUFI-II algorithm was employed for parameter calibration based on the Nash-Sutcliffe Efficiency (NSE) coefficient and the total water balance error. The LSTM model was trained using climate indices as inputs under different strategies (areal averages and multi-station series). The input indices included precipitation, temperature, and relative humidity. To evaluate the effects of temporal dependencies, lag times from 0 to 5 days were tested. All three models were calibrated using data from 1971 to 2010 (with the first two years as a warm-up period) and were validated over 2011-2020. [Results] (1) For daily-scale prediction, the LSTM model achieved the highest accuracy, with NSE values reaching 0.99 during calibration and 0.87 during validation when using multi-station meteorological inputs. Incorporating spatially distributed meteorological inputs significantly improved LSTM performance compared to using areal averages, highlighting the importance of spatial heterogeneity in data-driven hydrological forecasting. The XAJ model performed robustly (NSE>0.76), especially during flood seasons, but tended to overestimate dry-season flows and underestimate flood peaks. The SWAT model (NSE≈0.6) reproduced the overall hydrograph patterns but showed systematic biases similar to those of the XAJ model. (2) At the monthly scale, the performance of the SWAT model improved significantly (NSE=0.92 during calibration, 0.83 during validation), while LSTM accuracy declined (NSE = 0.86 during validation). The reduced training sample size (480 months) likely caused overfitting in the LSTM model and limited its generalization ability. Incorporating temperature and humidity as input features enhanced the stability of the LSTM model, indicating that variables related to evapotranspiration became more influential at coarser temporal resolutions. (3) Both hydrological models showed distinct seasonal bias patterns: overestimation during dry seasons and underestimation during wet seasons. The main reason was the fixed parameterization that failed to represent temporal variability in infiltration and storage processes. In contrast, the flexibility of the LSTM model enabled better adaptation, though at the cost of physical interpretability. The observed discrepancies also reflected the impact of nine cascade hydropower projects along the mainstream of the Xiangjiang River, which regulated flow seasonality but were not explicitly modeled in the hydrological frameworks used in this study. [Conclusion] In summary, this study systematically evaluates the performance of the XAJ, SWAT, and LSTM models for runoff forecasting in the Xiangjiang River Basin at daily and monthly scales. The results demonstrate that the LSTM model achieves the highest forecasting accuracy and computational efficiency, particularly at the daily scale. In contrast, the XAJ model is more reliable during flood seasons, and the SWAT model is more suitable for representing the long-term spatiotemporal variability of hydrological processes. Although the data-driven LSTM model lacks explicit physical mechanisms, it offers substantial advantages in adaptability and predictive precision, whereas physically based models maintain interpretability and stability under limited data conditions. The innovation of this study lies in bridging process-based and deep learning methods under unified experimental conditions, emphasizing the potential of hybrid modeling frameworks that integrate physical constraints with deep neural architectures to enhance both accuracy and interpretability in future runoff forecasting applications.