Makumbura et al. (2026) QPred: A Lightweight Deep Learning-Based Web Pipeline for Accessible and Scalable Streamflow Forecasting

Identification

• Journal: Computers, materials & continua/Computers, materials & continua (Print)
• Year: 2026
• Date: 2026-01-01
• Authors: Randika K. Makumbura, Hasanthi Wijesundara, Hirushan Sajindra, Upaka Rathnayake, Vikram Kumar, Dineshbabu Duraibabu, Sumit Sen
• DOI: 10.32604/cmc.2026.075539

Research Groups

• Faculty of Engineering and Design, Atlantic Technological University, Sligo, Ireland
• Department of Civil Engineering, Sri Lanka Institute of Information Technology, Malabe, Sri Lanka
• Department of Hydrology, Indian Institute of Technology, Roorkee, India

Short Summary

This study developed QPred, a lightweight, web-delivered application for daily streamflow forecasting using deep learning models, demonstrating that accurate, context-aware hydrological predictions can be made accessible and scalable through low-cost web platforms. The application integrates field data acquisition with deep learning models (LSTM, GRU) to provide on-demand streamflow forecasts for specific watersheds.

Objective

• To develop a regional, lightweight, cost-effective, web-delivered application for daily streamflow forecasting to bridge the gap between advanced data-driven models and their real-world deployment limitations.

Study Configuration

• Spatial Scale: Aglar and Paligaad watersheds.
• Temporal Scale: Daily streamflow forecasting using daily datasets of precipitation, river water level, and discharge.

Methodology and Data

• Models used: Vanilla Long Short-Term Memory (LSTM), Stacked LSTM, Bidirectional LSTM, and Gated Recurrent Unit (GRU).
• Data sources: High-resolution rainfall data recorded with tipping-bucket gauges and loggers; river water depth converted to discharge using site-specific rating curves. This resulted in daily datasets of precipitation, river water level, and discharge.
• Web Pipeline Components: Google Colab (execution environment), Flask (backend inference framework), Google Drive (artefact storage), Ngrok (secure HTTPS tunnelling).

Main Results

• All deep learning models achieved satisfactory to excellent performance during calibration (Coefficient of Determination (R²) > 0.91, Nash-Sutcliffe Efficiency (NSE) > 0.91 for both watersheds).
• For the Aglar watershed, the vanilla LSTM demonstrated the best generalization during validation (R² = 0.88, NSE = 0.82, Root-Mean-Square-Error (RMSE) = 0.12 m³/s).
• For the Paligaad watershed, the GRU achieved the highest validation accuracy (R² = 0.88, NSE = 0.88, RMSE = 0.49 m³/s).
• The QPred web application successfully integrates the highest-performing models, providing a user-friendly interface for on-demand streamflow forecasting.

Contributions

• Development of QPred, a novel, lightweight, and cost-effective web-based pipeline for accessible and scalable daily streamflow forecasting using deep learning.
• Demonstration of an end-to-end workflow from field data acquisition to web-based deployment, addressing the challenges of real-world application of data-driven hydrological models.
• Provision of a reproducible and scalable framework for hydrological applications that can be adapted to other watersheds and practitioners with limited infrastructure.

Funding

• Not specified in the provided text.

Citation

@article{Makumbura2026QPred,
  author = {Makumbura, Randika K. and Wijesundara, Hasanthi and Sajindra, Hirushan and Rathnayake, Upaka and Kumar, Vikram and Duraibabu, Dineshbabu and Sen, Sumit},
  title = {QPred: A Lightweight Deep Learning-Based Web Pipeline for Accessible and Scalable Streamflow Forecasting},
  journal = {Computers, materials & continua/Computers, materials & continua (Print)},
  year = {2026},
  doi = {10.32604/cmc.2026.075539},
  url = {https://doi.org/10.32604/cmc.2026.075539}
}