A Mixture-of-Experts Deep Learning Architecture for Daily Streamflow Prediction across the Sierra Nevada in California

Reliable daily streamflow prediction is essential for water resources management in California’s Sierra Nevada, where increasing hydroclimatic variability has intensified flood and drought risks. Although regional machine learning models can improve generalization across watersheds, conventional shared architectures may not adequately capture watershed-specific characteristics and distinct flow regimes. To address this limitation, we propose MoE- τ , a mixture-of-experts framework that combines multiple LSTM experts through an attention-based temporal gating network with a learnable temperature mechanism for adaptive expert weighting. We evaluated MoE- τ across 14 heterogeneous Sierra Nevada watersheds against learning-based models and the SAC-SMA process-based model. Under the in-domain temporal prediction benchmark, MoE- τ attained the highest mean NSE (0.873) among the learning-based models, a modest improvement over the strongest recurrent baseline (Bi-LSTM, 0.857) and a larger margin over SAC-SMA (0.762). It also remained robust under an alternative train–test split. Ablation and parameter-scale analyses showed that the performance gains arise from the proposed gating and expert-based design rather than increased model size. Model interpretation analyses revealed expert specialization across flow regimes and a strong relationship between hydrometeorological distribution differences and expert-usage differences (Spearman’s r=0.687). In held-out watershed experiments, MoE- τ outperformed LSTM on nearly all metrics across watersheds with distinct hydrologic conditions. These results demonstrate that adaptive expert selection improves both in-domain prediction and generalization to unseen watersheds, providing a robust and interpretable method for regional streamflow modeling.