Day-ahead surface runoff forecasting using a parallel VMD-CNN-BiLSTM-attention framework

Accurate daily surface runoff forecasting is essential for watershed management and flood-risk assessment. This study proposes a one-day-ahead forecasting framework, VCBLSTM-PMWA, integrating Variational Mode Decomposition, 1-D Convolutional Neural Networks, Bi-directional Long Short-Term Memory, and an attention mechanism. Hydro-meteorological data from 1990–2019 at Jingcun station in the Jing River Basin were chronologically divided into training, validation, and independent testing sets. On the testing set, the model achieved R2 = 0.9725, NSE = 0.9601, RMSE = 7.3931 m3/s, and PBIAS = −4.0002%, indicating high predictive accuracy with slight underestimation. The framework demonstrates potential for short-term watershed runoff forecasting and may support downstream urban river management, reservoir or gate scheduling, and flood preparedness. However, multi-basin validation, real-time meteorological inputs, multi-step forecasting, and event-based peak-flow assessment are still needed before operational application.