This study focuses on snowmelt-dominated basins in the Northern Hemisphere, including the Yenisei, Ob, Amur, and Yukon River basins, which are characterized by strong seasonality with snow accumulation in winter and rapid melt runoff during spring. This study proposes a triple-attention evolutionary AI model for monthly runoff forecasting in snowmelt-dominated basins. Multiple climate phenomenon indices are selected through maximal information coefficient analysis, and K -means clustering identifies distinct hydro-climatic patterns. For each cluster, an enhanced LSTM with multi-vector, predictor, and temporal attention mechanisms captures nonlinear climate-runoff relationships, with a cooperation search algorithm used to optimize clustering and model parameters. The proposed model demonstrates superior forecasting accuracy across the four snowmelt-dominated basins. For the Yenisei River with clustering number K = 5, it achieves testing NSE of 0.822, outperforming SVM (0.750) and RNN (0.718). The model reveals that antecedent runoff consistently exerts the strongest influence on monthly runoff, reflecting hydrological persistence, while climate indices show varying and compensatory effects across intra- and inter-annual timescales. Specifically, indices such as the Asia Polar Vortex Intensity and Pacific Subtropical High Ridge Position exhibit inverse relationships with runoff, indicating their compensatory roles in hydrological variability. The findings highlight the complex interactions between large-scale climate patterns and monthly runoff in high-latitude basins, supporting reliable runoff prediction and adaptive water management under climate change. • A triple-attention evolutionary model is developed for monthly runoff forecasting. • Dynamic impacts of climate indices on runoff are analyzed at multiple scales. • High forecasting results are achieved through MIC-based predictor selection and optimized model parameters. • The compensatory role of climate indices in hydrological variability is identified.
Liu et al. (2026) studied this question.