为融合过程驱动与数据驱动模型的优势，本研究提出了一种基于深度学习模型融入气象水文预报信息的短中期径流预报方法，并以沅水流域的三板溪、白市、托口、五强溪4座水库为研究对象，对比评估了该方法与仅采用过程驱动模型或数据驱动模型的预报表现。具体而言，首先评估并校正了欧洲中期天气预报中心发布的降水与气温回算预报产品，并将其作为过程驱动模型GR4J与新安江模型的驱动；然后，筛选与径流相关性较强的过程变量，并将其与过程驱动模型的径流预报结果共同作为长短期记忆网络（LSTM）模型的额外输入，构建了两种混合预报模型。结果表明：相较于单一模型，混合模型可明显改善径流预报精度。其中，相关系数提高了0.1~0.4；平均绝对百分比误差降低了10%~40%；3/5/10/15日累积水量误差最多可降低15%，降至20%~35%；洪水事件中的洪峰命中率最大提高45%，峰值流量误差最大减小20%。以上结果表明，将过程驱动模型的径流预报和关键过程变量同时作为深度学习模型的额外输入所构建的混合模型可有效提高梯级水库短中期径流预报的精度，具有潜在的应用价值。

[Objectives] To synergize the strengths of process-driven and data-driven hydrological models, this study develops a short- to medium-term runoff forecasting method that integrates meteorological and hydrological forecasting information via a deep learning architecture. The proposed methodology is implemented and validated in the Yuanjiang River basin, focusing on four large‑scale cascade reservoirs—Sanbanxi, Baishi, Tuokou, and Wuqiangxi. [Methods] The retrospective precipitation and temperature forecasts issued by the European Centre for Medium‑Range Weather Forecasts (ECMWF) are systematically evaluated and bias‑corrected using statistical post‑processing techniques, so as to enhance the reliability of meteorological forcing. These corrected products are then used as inputs to drive two widely adopted conceptual rainfall‑runoff models, namely GR4J and the Xinanjiang model, which were calibrated using observed meteorological and hydrological datasets. Two rainfall‑runoff models represent distinct hydrological processes with different structural complexities and the number of free parameters. Concurrently, a set of intermediate process‑related variables that exhibit strong correlations with observed runoff—such as soil moisture storage and evapotranspiration deficits—are extracted from the model simulations. Both the direct runoff outputs from the process‑driven models and these selected process variables are fed as additional input features into a Long Short‑Term Memory (LSTM) network, which is well suited for learning long‑term temporal dependencies in sequential data. This yields two hybrid variants that couple physically based simulations with data‑driven corrections. The performance of the hybrid models is comprehensively compared against that of the standalone process‑driven and data‑driven benchmarks, using multiple error and correlation metrics. [Results] The results demonstrate that the hybrid approach consistently outperforms all single‑model alternatives across almost evaluation metrics and forecast horizons. Specifically, the correlation coefficient between forecasted and observed runoff increases by 0.1 to 0.4 across four reservoirs and various lead times, indicating a notably improved agreement with observations. The mean absolute percentage error drops substantially, with reductions ranging from 10% to 40%, reflecting enhanced predictive stability. Moreover, the cumulative runoff volume errors accumulated over 3‑, 5‑, 10‑, and 15‑day windows are reduced by up to 15%, eventually falling within a range of 20% to 35%. In addition, the hybrid model effectively curtails both the frequency and magnitude of overestimated streamflow peaks, confining the upper envelope of forecasted discharges to a physically reasonable interval, which is particularly beneficial for reservoir operation and flood control. Beyond general flow simulation, the hybrid model also demonstrates marked superiority in flood event forecasting. It achieves a substantially higher flood peak hit rate, with improvements of up to 45% relative to the LSTM model, and reduces the relative error in peak flow magnitude by 20% across most events, highlighting the model's capability in capturing extreme hydrological responses. Most notably, the hybrid model corrects the extreme overestimation of peak flows at the Tuokou Reservoir—a persistent issue in the LSTM model, where the peak error once reached 95%—and brings this error down to within 10%, thereby offering a much more reliable warning signal for extreme flood events. Collectively, these results confirm that the hybrid architecture not only raises overall forecast accuracy but also enhances the model's robustness in capturing high-flow extremes, which is critical for real-time reservoir management and flood mitigation. [Conclusions] Overall, these findings indicate that enriching a deep learning model with both process‑based runoff simulations and key intermediate variables can effectively mitigate the deficiencies of either modelling paradigm alone, leading to more accurate and reliable short‑ to medium‑term forecasts for cascade reservoir systems. The proposed framework is computationally feasible and holds considerable promise for operational implementation in real‑time reservoir scheduling and early warning systems, especially in regions where high‑quality observed data are available for training and bias correction.