JMEE Home · Issue Contents · Forthcoming Papers

Multi-step Marine-oriented Daily Runoff Forecasting by Integrating Convolutional and Sparse Attention Mechanisms
Yuanxuan Zhu, Xiaohong Chang, Huanhuan Zhang and Chupeng Yang

Accurate runoff marine-influenced forecasting is essential for extending lead time, optimizing nearshore water resource management, and supporting coastal flood control and disaster mitigation. A novel model, called LogConvFormer, has been developed to address the temporal lag in multi-step runoff prediction and improve peak flood forecasting performance. The model integrates three progressive mechanisms: an Inverted Embedding Mechanism Optimization, a Convolutional Attention Enhancement Module, and a Log-Sparse Attention Pattern, which together enhance the macro- and micro-level perception of land-ocean interactive hydrological processes and modeling efficiency. Ablation experiments confirm the contribution of each component. Applied to multistep daily discharge forecasting for rivers in Maine, USA, LogConvFormer achieved superior performance compared to CNN, LSTM, and iTransformer models, with NSE values up to 0.910. At the USGS 01013500 site, LogConvFormer’s NSE only slightly drops to 0.807 (6 steps) and 0.762 (7 steps), remaining consistently strong. By contrast, CNN, LSTM, and iTransformer show steep NSE declines (e.g., CNN plummets from 0.8 to 0.1). Sensitivity analyses further demonstrated its robust predictive capacity across increasing lead times, indicating substantial improvements in multi-step marine-oriented runoff forecasting and providing critical technical support for coastal water management and disaster prevention.

Keywords: Marine-influenced runoff forecasting, multi-step prediction, peak flood forecasting, deep learning, attention mechanism, hydrological modeling, LogConvFormer