Advancing Long-Horizon Hydrological Forecasting: A Mamba-based Approach with Explainable AI for Generalized Streamflow Prediction

Accurate long-horizon streamflow forecasting is crucial for water resource management, but existing models often face efficiency and interpretability challenges. This study comprehensively evaluates the Mamba architecture, which utilizes State Space Models for efficient sequence processing, for 120-hour hourly generalized streamflow prediction across 125 diverse Iowa watersheds using 72-hour historical inputs. Performance was benchmarked against Persistence, LSTM, GRU, Seq2Seq, and Transformer models employing NSE, KGE, Pearson's r, and NRMSE. Results demonstrate that Mamba architecture achieves predictive accuracy comparable to, and in several aspects marginally exceeding, the robust Transformer baseline, with both models significantly outperforming other established methods. Critically, Explainable AI (XAI) using SHAP values provided insights into tested models’ decision-making, revealing distinct feature utilization patterns and enhancing model transparency. This research highlights Mamba's potential as an efficient, accurate, and interpretable alternative for advancing operational long-range hydrological forecasting.