Abstract: Against the backdrop of global climate change and intensified human activities, extreme drought events have become increasingly frequent, exacerbating the contradiction between water supply and demand. Runoff forecasting, as the cornerstone of refined water resource allocation, plays a pivotal role in dry-season water management. During low-flow periods, runoff volumes were relatively small while hydrological conditions exhibit greater complexity and variability. Mechanistic models offer strong inter-pretability but suffer from complexity, numerous parameters, and challenges in parameter calibration. Data-driven models excel in non-linear processing capabilities but lack explicit descriptions of runoff generation mechanisms. This study integrated the physical processes of the Xin′anjiang model with the Long Short-Term Memory (LSTM) neural network to develop a mechanism-data hybrid model tailored for low-flow runoff forecasting. The proposed model was applied to the upper reaches of the Wuyang River Basin, where daily-scale runoff simulations and forecasting were conducted from 2020 to 2024. Results indicated that mechanistic-data hybrid model performed well in runoff forecasting during different dry seasons and offered robust support for reservoir operation and aquatic ecosystem protection during drought periods.