Adaptive rolling runoff forecasting model: Combining multi-source correlated sequences and extreme value encoding

The research subject of this study is the control watershed at the inlet cross-section of the Linjiacun Reservoir in the Baoji Gorge Irrigation Area, China. This study proposes MEN, a neural network integrating LSTM and CNN architectures to model multi-source runoff sequences and address extreme value challenges. By synergizing dynamic sequence refinement, Kruskal-Wallis sampling for extreme data imbalance, and gating-controlled extreme value encoding, MEN enhances both general runoff prediction and extreme event accuracy. The framework effectively captures long-term hydrological dependencies while mitigating uncertainty in complex forecasting scenarios. This study applies the MEN model to real-time runoff forecasting for the Linjiacun Reservoir inflow section in the Baoji Gorge Irrigation District, using historical reservoir runoff data and upstream rainfall data for model training. Compared to SARIMAX and LSTM benchmarks, MEN achieves the lowest average relative error and maintains R² > 0.8 across extended lead times, demonstrating robustness. By synergizing multi-source data learning and extreme value encoding, the framework offers enhanced technical support for real-time predictions in complex watersheds. • Integrate various neural network frameworks for deep learning, capturing hidden information and refining representations. • Encode extreme values separately, handling imbalanced data using Kruskal-Wallis sampling and gate-controlled processing. • Efficiently integrate multi-source data to assist in prediction and enhance accuracy.