Comparative Analysis of HEC-HMS and Temporal Fusion Transformer for Streamflow Prediction under Climate Change Scenarios in the Swat River Basin

Abstract

Climate change is expected to significantly alter hydrological regimes in mountainous river basins, increasing uncertainty in future streamflow and flood risk. This study assesses historical and future streamflow dynamics of the Swat River Basin, Pakistan. A comparative framework was adopted using a physically based semi-distributed model (HEC-HMS) and a data-driven deep learning model, the Temporal Fusion Transformer (TFT). Eleven downscaled CMIP6 Global Climate Models (GCMs) were first evaluated using Recursive Feature Elimination coupled with a Random Forest algorithm, resulting in the selection of three top-performing models: INM-CM4-8, MRI-ESM-2-0, and GFDL-ESM4. Historical simulations showed that while HEC-HMS performed satisfactorily in reproducing seasonal flows and runoff volumes, it underestimated extreme flood peaks. In contrast, the TFT model demonstrated very high predictive accuracy and effectively captured extreme events, including the catastrophic 2010 flood. Future projections (2020–2100) under SSP2-4.5 and SSP5-8.5 scenarios indicate a substantial increase in peak flows, particularly under high-emission pathways, with several extreme events projected to occur in the near future. Overall, the results highlight the limitations of conventional hydrological models in simulating non-linear extremes and demonstrate the strong potential of advanced machine-learning models for climate change impact assessment and flood risk management in data-scarce mountainous basins.

Introduction

The text addresses the growing challenge of predicting river streamflow under climate change, emphasizing its importance for sustainable water management, flood risk reduction, and climate adaptation. Climate change has significantly altered hydrological processes worldwide by increasing temperatures, modifying precipitation patterns, and intensifying extreme events such as floods and droughts. These impacts are particularly pronounced in snow- and glacier-fed river basins, where rising temperatures affect melt dynamics, seasonal runoff, and flow extremes.

Pakistan is identified as one of the most climate-vulnerable countries due to its dependence on climate-sensitive water resources and the dominance of snow and glacier melt in the Indus River Basin. The Swat River Basin, a critical sub-basin of the Upper Indus Basin, is highlighted as a representative and high-risk region because of its hydrological sensitivity, socio-economic importance, and history of extreme floods, notably the 2010 event. Climate projections suggest increased flow variability, higher flood risks, and potential reductions in dry-season water availability in the basin.

The study focuses on comparing two streamflow prediction approaches under historical and future climate scenarios: the physics-based Hydrologic Engineering Center–Hydrologic Modeling System (HEC-HMS) and the data-driven Temporal Fusion Transformer (TFT) deep learning model. HEC-HMS is widely used for rainfall–runoff simulation but may face limitations under non-stationary climate conditions. In contrast, TFT can capture complex, non-linear relationships in hydrological data and offers improved forecasting accuracy and interpretability through attention mechanisms.

Using the Swat River Basin as a case study, the research integrates topographic, land use, soil, and climate data. Bias-corrected and downscaled climate projections from selected CMIP6 Global Climate Models under SSP2–4.5 and SSP5–8.5 scenarios were employed. Both models were calibrated and validated using historical streamflow data and evaluated using standard performance metrics.

The study aims to assess the strengths and limitations of HEC-HMS and TFT in simulating historical flows and projecting future streamflow, seasonal variability, and extremes. By directly comparing physical and deep learning models, the research seeks to enhance understanding of their applicability under climate change and to support climate-resilient water resources management and flood mitigation strategies in the Swat River Basin and similar glacier-fed systems.

Conclusion

This study investigated the impacts of climate change on streamflow in the Swat River Basin, focusing on the proposed intake of the Mingora Gravity Water Supply Scheme. Two modeling approaches were employed: the physically-based, semi-distributed HEC-HMS model and the data-driven Temporal Fusion Transformer (TFT) model. Eleven downscaled CMIP6 GCMs were evaluated using Recursive Feature Elimination with Random Forest to select the top three models (INM-CM4-8, MRI-ESM-2-0, and GFDL-ESM4) for driving future streamflow projections under moderate (SSP2-4.5) and high-emissions (SSP5-8.5) scenarios. The HEC-HMS model performed satisfactorily in both calibration and validation, reliably simulating seasonal flows and overall water balance, but it significantly underestimated extreme peak events, such as the catastrophic 2010 flood. In contrast, the TFT model demonstrated very high predictive accuracy, capturing both low and high flows effectively, including extreme flood peaks. This highlights the TFT model’s superior capability to learn complex, non-linear hydrological responses from historical data and its robustness for future streamflow forecasting. Future projections indicate substantial increases in peak flows under high-emissions scenarios, particularly for the INM-CM4-8 and TFT simulations, with some extreme events projected to occur in the near future (2020–2040). Comparisons between the two models show that while HEC-HMS provides a reliable baseline, TFT projections capture higher-magnitude floods and greater sensitivity to climate forcing, emphasizing the importance of using advanced machine learning approaches for risk assessment. Overall, this research demonstrates that climate change is likely to intensify streamflow variability and extreme events in the Swat River Basin. The findings underscore the need for adaptive water resource planning and flood risk management, particularly under high-emission pathways. The integration of machine learning models like TFT with physically-based hydrological models provides a robust framework for improving the accuracy of streamflow projections and informing sustainable water management strategies in the region.

References

[1] D. Scott, C. M. Hall, B. Rushton, and S. Gössling, “A review of the IPCC Sixth Assessment and implications for tourism development and sectoral climate action,” J. Sustain. Tour., vol. 32, no. 9, pp. 1725–1742, Sept. 2024, doi: 10.1080/09669582.2023.2195597. [2] P. C. D. Milly et al., “Stationarity Is Dead: Whither Water Management?,” Science, vol. 319, no. 5863, pp. 573–574, Feb. 2008, doi: 10.1126/science.1151915. [3] Z. W. Kundzewicz et al., “Flood risk and climate change: global and regional perspectives,” Hydrol. Sci. J., vol. 59, no. 1, pp. 1–28, Jan. 2014, doi: 10.1080/02626667.2013.857411. [4] M. Zebisch et al., “The vulnerability sourcebook and climate impact chains – a standardised framework for a climate vulnerability and risk assessment,” Int. J. Clim. Change Strateg. Manag., vol. 13, no. 1, pp. 35–59, Mar. 2021, doi: 10.1108/IJCCSM-07-2019-0042. [5] A. F. Lutz, W. W. Immerzeel, P. D. A. Kraaijenbrink, A. B. Shrestha, and M. F. P. Bierkens, “Climate Change Impacts on the Upper Indus Hydrology: Sources, Shifts and Extremes,” PLOS ONE, vol. 11, no. 11, p. e0165630, Nov. 2016, doi: 10.1371/journal.pone.0165630. [6] European Investment Bank., Who’s most at risk?: a global index of climate risk for countries. LU: Publications Office, 2025. Accessed: Dec. 29, 2025. [Online]. Available: https://data.europa.eu/doi/10.2867/5317974 [7] S. Ali, B.-H. Kim, T. Akhtar, and J. Kam, “Past and future changes toward earlier timing of streamflow over Pakistan from bias-corrected regional climate projections (1962–2099),” J. Hydrol., vol. 617, p. 128959, Feb. 2023, doi: 10.1016/j.jhydrol.2022.128959. [8] M. Ismail et al., “Evaluating the Impact of Climate Change on the Stream Flow in Soan River Basin (Pakistan),” Water, vol. 14, no. 22, p. 3695, Nov. 2022, doi: 10.3390/w14223695. [9] M. I. Shah, A. Khan, T. A. Akbar, Q. K. Hassan, A. J. Khan, and A. Dewan, “Predicting hydrologic responses to climate changes in highly glacierized and mountainous region Upper Indus Basin,” R. Soc. Open Sci., vol. 7, no. 8, p. 191957, Aug. 2020, doi: 10.1098/rsos.191957. [10] Z. U. Rahman et al., “GIS-based flood susceptibility mapping using bivariate statistical model in Swat River Basin, Eastern Hindukush region, Pakistan,” Front. Environ. Sci., vol. 11, p. 1178540, July 2023, doi: 10.3389/fenvs.2023.1178540. [11] M. Zhang, X. Fu, S. Liu, and C. Zhang, “Integrating Remote Sensing and Machine Learning for Actionable Flood Risk Assessment: Multi-Scenario Projection in the Ili River Basin in China Under Climate Change,” Remote Sens., vol. 17, no. 7, p. 1189, Mar. 2025, doi: 10.3390/rs17071189. [12] ?. B. Peker, S. Gülbaz, V. Demir, O. Orhan, and N. Beden, “Integration of HEC-RAS and HEC-HMS with GIS in Flood Modeling and Flood Hazard Mapping,” Sustainability, vol. 16, no. 3, p. 1226, Feb. 2024, doi: 10.3390/su16031226. [13] X. Yu and J. Zhang, “The Application and Applicability of HEC-HMS Model in Flood Simulation under the Condition of River Basin Urbanization,” Water, vol. 15, no. 12, p. 2249, June 2023, doi: 10.3390/w15122249. [14] D. Feng and E. Beighley, “Identifying uncertainties in hydrologic fluxes and seasonality from hydrologic model components for climate change impact assessments,” Hydrol. Earth Syst. Sci., vol. 24, no. 5, pp. 2253–2267, May 2020, doi: 10.5194/hess-24-2253-2020. [15] V. D. Nguyen, S. Vorogushyn, K. Nissen, L. Brunner, and B. Merz, “A non-stationary climate-informed weather generator for assessing future flood risks,” Adv. Stat. Climatol. Meteorol. Oceanogr., vol. 10, no. 2, pp. 195–216, Nov. 2024, doi: 10.5194/ascmo-10-195-2024. [16] C. Shen et al., “HESS Opinions: Deep learning as a promising avenue toward knowledge discovery in water sciences,” Hydrol. Earth Syst. Sci. Discuss., vol. 2018, pp. 1–21, 2018 [17] B. Lim, S. Ö. Ar?k, N. Loeff, and T. Pfister, “Temporal fusion transformers for interpretable multi-horizon time series forecasting,” Int. J. Forecast., vol. 37, no. 4, pp. 1748–1764, 2021. [18] X. Jin, Y.-X. Wang, and X. Yan, “Inter-series attention model for COVID-19 forecasting,” presented at the Proceedings of the 2021 SIAM International Conference on Data Mining (SDM), SIAM, 2021, pp. 495–503. [19] J. Fang, L. Yang, X. Wen, W. Li, H. Yu, and T. Zhou, “A deep learning-based hybrid approach for multi-time-ahead streamflow prediction in an arid region of Northwest China,” Hydrol. Res., vol. 55, no. 2, pp. 180–204, 2024. [20] F. Kratzert, M. Gauch, G. Nearing, and D. Klotz, “NeuralHydrology---A Python library for Deep Learning research in hydrology,” J. Open Source Softw., vol. 7, no. 71, p. 4050, 2022. [21] F. Kratzert, M. Gauch, G. Nearing, and D. Klotz, “NeuralHydrology---A Python library for Deep Learning research in hydrology,” J. Open Source Softw., vol. 7, no. 71, p. 4050, 2022. [22] L. Breiman, “Random forests,” Mach. Learn., vol. 45, no. 1, pp. 5–32, 2001. [23] R. Kilgore, A. T. Atayee, D. Curtis, J. Harris, G. R. Herrmann, and D. Thompson, “Highway Hydrology: Evolving Methods, Tools, and Data,” United States. Federal Highway Administration. Office of Bridges and Structures, 2023. [24] B. Lim, S. Ö. Ar?k, N. Loeff, and T. Pfister, “Temporal fusion transformers for interpretable multi-horizon time series forecasting,” Int. J. Forecast., vol. 37, no. 4, pp. 1748–1764, 2021. [25] C. Shen, “A transdisciplinary review of deep learning research and its relevance for water resources scientists,” Water Resour. Res., vol. 54, no. 11, pp. 8558–8593, 2018.

Copyright

Copyright © 2026 Imad khan, Fayaz Ahmad Khan, Hamid Hussain, Afed Ullah Khan. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.