Apak et al. (2026) Multi-resolution adaptive channel fusion transformer encoder LSTM for accurate streamflow prediction

Identification

• Journal: Scientific Reports
• Year: 2026
• Date: 2026-02-21
• Authors: Sina Apak, Huseyin Cagan Kilinc, Adem Yurtsever, Hilal Haznedar, Furkan Özkan
• DOI: 10.1038/s41598-026-40713-1

Research Groups

• Department of Management Information Technology, Istanbul Aydın University, Istanbul, Turkey
• Department of Civil Engineering, Istanbul Aydın University, Istanbul, Turkey
• Department of Environmental Engineering, Istanbul University-Cerrahpaşa, Turkey
• Department of Computer Engineering, Hasan Kalyoncu University, Gaziantep, Turkey
• Department of Computer Engineering, Çukurova University Adana, Turkey

Short Summary

This study proposes a novel hybrid deep learning model, MR-ACF-TE-LSTM, for accurate and interpretable univariate streamflow prediction by effectively capturing multi-scale temporal patterns. The model consistently outperforms baseline and state-of-the-art methods across benchmark datasets, demonstrating significant reductions in prediction error and enhanced generalization capabilities.

Objective

• To develop a hybrid deep learning architecture that simultaneously captures short-term variability and long-term dependencies in univariate streamflow time series, thereby improving predictive accuracy and interpretability for water management and flood mitigation.

Study Configuration

• Spatial Scale: Western Black Sea Basin, Turkey, covering approximately 28,855 square kilometers. The study focused on three Flow Measurement Stations (FMSs): Darıören (D13A049, catchment area 137 km²), Dereevi (D13A032, catchment area 71.50 km²), and Akhasan (D13A022, catchment area 76.50 km²).
• Temporal Scale: 10 years of daily flow data (2001–2011). The models used a 96-day input window to predict the streamflow value 5 days in advance.

Methodology and Data

• Models used:
  • Proposed Model: Multi-Resolution Adaptive Channel Fusion Transformer Encoder LSTM (MR-ACF-TE-LSTM). This hybrid architecture integrates:
    • Multi-Resolution Input Construction (lagged observations, statistical features, seasonal/temporal indicators).
    • Adaptive Channel Fusion (ACF) module with attention-like gating network.
    • Transformer Encoder module for long-range dependencies.
    • Temporal Attention Mechanism for dynamic importance scoring.
    • LSTM-Based Sequential Decoder for local temporal continuity.
  • Baseline Models: Transformer, Transformer-LSTM, Fusion LSTM-GRU, Bayesian Convolutional Neural Network (BCNN).
• Data sources: Daily flow data collected from Darıören, Dereevi, and Akhasan Flow Measurement Stations (FMS) in the Western Black Sea Basin. The dataset for each station was split into 70% for training and 30% for testing.

Main Results

• The MR-ACF-TE-LSTM model consistently outperformed all baseline and state-of-the-art models across the three benchmark streamflow datasets (Dereevi, Darıören, Akhasan), achieving the lowest RMSE and highest R² scores.
• Ablation studies confirmed the critical contributions of each architectural component:
  • Multi-resolution inputs yielded performance enhancements of up to 13% (e.g., Akhasan FMS RMSE decreased from 0.3370 to 0.2935).
  • Overall RMSE reductions varied from 28% to 48% compared to baseline and state-of-the-art models.
  • Specifically, the RMSE at Dereevi FMS improved by 39%, from 1.436 (Transformer) to 0.874 (MR-ACF-TE-LSTM).
• The model demonstrated strong generalization capabilities and robustness in cross-dataset evaluations, particularly when trained on mid-complexity datasets.
• Attention weight visualizations provided physical interpretability, showing the model's ability to selectively focus on recent time steps during high-flow periods and broader temporal spans during low-flow conditions, consistent with hydrological memory effects.
• For Dereevi FMS, MR-ACF-TE-LSTM achieved an RMSE of 0.874, MAE of 0.494, R² of 0.700, KGE of 0.808, and NSE of 0.705.
• For Darıören FMS, MR-ACF-TE-LSTM achieved an RMSE of 2.854, MAE of 2.017, R² of 0.492, KGE of 0.808, and NSE of 0.705.
• For Akhasan FMS, MR-ACF-TE-LSTM achieved an RMSE of 0.293, MAE of 0.156, R² of 0.658, KGE of 0.645, and NSE of 0.663.

Contributions

• Proposes a novel hydrology-sensitive hybrid deep learning architecture, MR-ACF-TE-LSTM, that integrates multi-resolution learning, adaptive channel coupling, dense Transformer encoders, and LSTM units for accurate and interpretable univariate streamflow prediction.
• Addresses the limitations of conventional deep learning models (e.g., LSTMs, CNNs, pure Transformers) in simultaneously capturing multi-scale hydrological dependencies, maintaining temporal continuity, and ensuring physical consistency.
• Introduces an adaptive channel fusion mechanism that dynamically weighs contributions from different temporal resolutions, enhancing feature interactions and regime-sensitive predictive relevance.
• Incorporates a temporal attention mechanism that improves interpretability by highlighting critical time steps contributing most to predictions, revealing internal decision-making rationale.
• Establishes a new benchmark for robust flow prediction by demonstrating superior performance and generalization capabilities across heterogeneous catchments compared to existing state-of-the-art models.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Citation

@article{Apak2026Multiresolution,
  author = {Apak, Sina and Kilinc, Huseyin Cagan and Yurtsever, Adem and Haznedar, Hilal and Özkan, Furkan},
  title = {Multi-resolution adaptive channel fusion transformer encoder LSTM for accurate streamflow prediction},
  journal = {Scientific Reports},
  year = {2026},
  doi = {10.1038/s41598-026-40713-1},
  url = {https://doi.org/10.1038/s41598-026-40713-1}
}