<p>With the increasing penetration of renewable energy under the market-oriented background, effective coordination between day-ahead planning and intra-day adjustment has become essential for power systems. This study proposes a coordinated two-stage scheduling framework for wind-photovoltaic (PV)-hydro hybrid systems, integrating short-term forecasting and operational optimization. To enhance forecasting reliability, a Multi-Head Attention (MHA) mechanism is incorporated into four deep learning models, and a Stacking strategy is adopted to reduce prediction uncertainty. Then, a two-stage scheduling model is established: In the day-ahead stage, generation plans are formulated by maximizing generation benefits and smoothing remaining load fluctuations. In the intra-day stage, hydropower flexibility is reallocated to mitigate deviations between planned and realized renewable outputs, thereby reducing operational losses. The proposed framework is validated using data from a large-scale clean energy base in China. Results show that the MHA-enhanced models outperform baseline models, while the Stacking approach further improves forecasting accuracy. Compared with historical dispatch results, the coordinated two-stage strategy significantly reduces intra-day losses. Sensitivity analyses further show that improved forecasting inputs generally lead to lower operational losses and that the framework maintains stable effectiveness under different representative wind-PV operating scenarios. Moreover, a clear trade-off is observed between generation benefit and deviation-related loss, whereas smoother remaining load profiles contribute to improved operational performance. These findings offer decision-makers valuable guidance for balancing overall economic efficiency and operational losses in multi-energy systems.</p>