<p>Accurate runoff prediction is critically important for water resource management and flood risk prevention. In recent years, deep learning models have been increasingly applied to hydrological forecasting and have achieved remarkable performance. However, such models still face challenges in interpretability, representing complex hydrological processes, and performing reliable multi-step forecasting. To address these gaps, this study proposes a hydrological-adapted Mixture of Experts (HydMoE) model integrating Time2Vec temporal embedding and Local-Global Hybrid Attention (LGHA). The MoE structure adaptively handles diverse hydrological patterns and river spatiotemporal characteristics. Time2Vec encodes implicit temporal signatures to enhance temporal dynamics understanding. LGHA simultaneously captures short-range critical dependencies and long-range correlations, balancing efficiency and dependency capture. The MoE architecture enables pattern-specific expert specialization and enhances physical interpretability by establishing a correspondence between expert activation behaviors and hydrological scenarios. Trained and evaluated on the CAMELS dataset, HydMoE achieves superior performance over existing baseline models across lead times of 1 to 7 days. It attains an NSE of 0.762 for 1‑day ahead prediction and an average NSE of 0.393 for lead times of 2 to 7 days, outperforming the baseline models by 5.8% and 8.6%, respectively.</p>