<p>Accurate streamflow prediction plays a vital role in water management and flood mitigation. However, conventional deep learning models often fail to simultaneously capture short-term variability and long-term dependencies, particularly in univariate time series common in hydrological settings. To address these challenges, we propose the Multi-Resolution Adaptive Channel Fusion Transformer Encoder LSTM (MR-ACF-TE-LSTM)—hybrid architecture designed to improve predictive accuracy and interpretability by modeling temporal patterns at multiple scales. The model constructs pseudo-multivariate inputs from lagged observations, statistical summaries, and seasonal indicators, which are dynamically fused through an adaptive attention-based mechanism. Before making a prediction with LSTM, this fused representation are encoded by a transformer and then refined over time. Extensive experiments on three benchmark streamflow datasets demonstrate that MR-ACF-TE-LSTM consistently outperforms baseline models, including Transformer, Transformer-LSTM, and Bayesian CNN, achieving lower RMSE and higher R² scores. Ablation and comparative analyses demonstrate that MR-ACF-TE-LSTM attains the minimal RMSE across all FMSs, with multi-resolution yielding enhancements of up to 13% and overall RMSE reductions varying from 28% to 48% in comparison to baseline and state-of-the-art models. In comparison to baseline models, the RMSE at Dereevi FMS improved by 39%, from 1.436 to 0.874. Cross-dataset evaluations further highlight the model’s robustness and generalization capabilities across heterogeneous catchments. Ablation studies confirm the critical contributions of multi-resolution inputs and adaptive fusion, while attention weight visualizations reveal the model’s ability to selectively focus on temporally inputs. These findings establish MR-ACF-TE-LSTM as a powerful and interpretable framework for univariate hydrological forecasting.</p>