A parallel attention-based framework for multi-step multivariate runoff forecasting in mountainous watersheds: Wuyuan case study

Wuyuan County, Jiangxi Province, China To address the challenges of multivariate modeling and error accumulation in multi-step runoff forecasting, we propose a Parallel Attention Multivariate Multi-step LSTM Network (PAM 2 -LSTM). This framework integrates parallel subnetworks to capture heterogeneous relationships between hydrological variables and combines direct multi-step forecasting with attention mechanisms to mitigate error propagation. The model was evaluated on high-resolution observational data from Wuyuan. The PAM 2 -LSTM significantly outperforms conventional LSTM and other baseline models, achieving a 70.5% improvement in MAE and a 61.2% reduction in RMSE for 6-hour ahead forecasts. The model maintains robust accuracy across extended prediction horizons, with Nash–Sutcliffe Efficiency (NSE) values exceeding 0.97 for 6 h. Visualization of results confirms precise temporal alignment with observed hydrographs and reliable peak flow estimation, critical for flash flood warning systems. The parallel architecture’s superior performance highlights its capability to leverage multivariate data while minimizing cumulative errors, offering a reliable tool for flood risk management in complex mountainous watersheds. • Proposed PAM 2 -LSTM for complex runoff forecasting task. • Parallel structure extracts variable relationships effectively in hydrological data. • Attention-based encoder–decoder framework reduces error accumulation in multi-step forecasting. • Model shows high accuracy in multi-step forecasts across Mountainous Watersheds.