As global climate change intensifies, hydrological cycles are exhibiting increasingly pronounced non-stationary, multi-scale, and highly nonlinear behaviors. Traditional mechanistic models face significant challenges in fully capturing the meteorological–runoff response mechanisms, while deep learning approaches still struggle with long-term dependency modeling, spatial correlations, and sequence decoupling. To address these issues, we propose a hybrid forecasting model, JMD-Gformer, which combines Jump plus AM-FM Mode Decomposition (JMD), sparse directed graph networks, and Transformer-based architectures. In this framework, JMD decomposes the non-stationary runoff time series into jump components (representing abrupt runoff events) and periodic components (capturing seasonal and inter-annual fluctuations), effectively mitigating noise and mode aliasing. Additionally, we construct a sparse directed graph based on meteorology-hydrology nodes to represent the upstream-downstream interactions. The model further incorporates a multi-head self-attention mechanism for long-range temporal dependencies. Experiments conducted on multi-time-scale runoff predictions in the Daliao River Basin demonstrate that, compared to the second-best benchmark model, JMD-Gformer reduces MAPE and RMSE by 36.4% and 41.2%, respectively, and shows strong robustness and predictive accuracy under complex hydrological conditions. This model provides a promising new approach for smart watershed management.