A dual branch deep learning framework with trend gating for nonstationary monthly runoff prediction

Study region: The Source Region of the Yangtze River in the Three-River Headwaters of the Tibetan Plateau, with three representative stations: Tuotuo River, Zhimenda, and Xinzhai. Study focus: This study addresses monthly runoff prediction in alpine basins with pronounced nonstationarity and multi-scale variability. A variational mode decomposition (VMD)-based trend guided dual branch gated collaborative network (VCLT-Net) is proposed. VMD decomposes runoff into a slowly varying trend component statistically dominated by baseflow-like variability and a high-frequency fluctuation component containing rapid response signals, with no unique mapping to physical processes. An LSTM simulates the trend, multi-scale convolutional Transformers extract high-frequency features, and a trend-driven gated interaction network (TDGN) adaptively fuses both branches in a data-driven manner. Hyperparameters are optimized by a heuristic optimization algorithm (HOA); SHAP quantifies driver attribution. VCLT-Net is benchmarked against multiple models using seven metrics. New hydrological insights: VCLT-Net robustly outperforms benchmarks at all three stations. At Zhimenda, it increases Nash–Sutcliffe efficiency (NSE) by 16.21%, reduces root mean square error (RMSE) by 152.47 m 3 /s, improves Kling–Gupta efficiency (KGE) by 13.37%, and lowers mean absolute percentage error (MAPE) by 6% over standalone LSTM; both NSE and KGE exceed 0.95. The consistent outperformance demonstrates the model’s ability to capture complex nonstationary runoff dynamics, suggesting that integrating decomposition, dual-branch feature extraction, and trend-driven interaction improves the representation of long-term and short-term variability in monthly runoff.