Sharma et al. (2026) Improving Daily Streamflow Predictions over Large Watersheds: Introducing a Novel Enhanced Long Short-Term Memory (En-LSTM) Model

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-01-01
• Authors: Purshottam Sharma, Vaibhav Tripathi, Mohit Prakash Mohanty
• DOI: 10.1007/s11269-025-04475-1

Research Groups

• Department of Water Resources Development and Management, Indian Institute of Technology Roorkee, Roorkee, 247667, India

Short Summary

This study introduces a novel Enhanced Long Short-Term Memory (En-LSTM) model, integrating Temporal Convolutional Networks, an attention mechanism, a peak-aware hybrid loss function, and multi-scale temporal features, to significantly improve daily streamflow prediction and peak-flow simulation over large, data-scarce watersheds, demonstrating robust transferability to ungauged locations.

Objective

• Develop a unified deep-learning architecture for daily streamflow prediction.
• Compare the performance of the proposed En-LSTM with a conventional LSTM to quantify the benefits of the enhancements.
• Evaluate the generalizability of the basin-wide En-LSTM (En-LSTM(i)) to unseen stations, representing ungauged environments.

Study Configuration

• Spatial Scale: Mahanadi River Basin, India, spanning approximately 142,000 square kilometers, with varied topography (0 to 1311 meters elevation) and 20 hydrological gauging stations for training/testing, plus additional unseen stations for validation.
• Temporal Scale: Daily streamflow prediction. Data period for LSTM models: June 1, 2003 – May 31, 2023 (Training: June 1, 2003 – May 31, 2018; Testing: June 1, 2018 – May 31, 2023). SWAT+ model period: January 1998 – December 2022 (Warm-up: 1998–2001; Calibration: 2002–2016; Validation: 2017–2022).

Methodology and Data

• Models used:
  • Enhanced Long Short-Term Memory (En-LSTM): A novel deep learning framework augmenting conventional LSTM with:
    • Temporal Convolutional Networks (TCN) for multi-scale temporal pattern extraction.
    • An attention mechanism for dynamic predictor weighting.
    • A peak-aware hybrid loss function (combining Peak Error Swish and Mean Absolute Error).
    • Multi-scale lag (1, 3, 7, 14 days) and rolling-window (3, 7, 14 days) features.
  • Conventional Long Short-Term Memory (LSTM): Baseline model for comparison.
  • SWAT+ (Soil and Water Assessment Tool+): A process-based hydrological model used as a benchmark.
• Data sources:
  • Observed Streamflow: Daily gauge discharge data from the Central Water Commission (CWC) for 20 stations.
  • Hydro-meteorological Data: Daily rainfall data from the India Meteorological Department (IMD).
  • Catchment Descriptors (CDs): Derived from Multi-Error-Removed Improved-Terrain (MERIT) DEM (elevation, slope, aspect, drainage networks), land use, soil properties, and geology.

Main Results

• Overall Performance Improvement: The integrated En-LSTM (En-LSTM(i)) consistently outperformed the conventional LSTM (LSTM(i)) and station-specific En-LSTM (En-LSTM(r)) models.
  • En-LSTM(i) vs. LSTM(i) (Training): En-LSTM(i) achieved KGE of 0.84, NSE of 0.83, and r of 0.95, significantly higher than LSTM(i)'s KGE of 0.66, NSE of 0.62, and r of 0.79.
  • En-LSTM(i) vs. LSTM(i) (Testing): En-LSTM(i) achieved KGE of 0.71, NSE of 0.63, and r of 0.85, compared to LSTM(i)'s KGE of 0.56, NSE of 0.50, and r of 0.71.
• Peak-Flow Simulation: En-LSTM(i) demonstrated marked improvements in capturing peak-flow timing and magnitude, addressing a known weakness of many existing hydrological models, including conventional LSTMs.
• Transferability to Unseen Stations: When evaluated on completely unseen stations (ungauged contexts), En-LSTM(i) achieved robust performance with mean NSE of 0.72, KGE of 0.68, and r of 0.87, indicating strong generalization capabilities in data-scarce locations.
• Comparison with SWAT+: At an independent pilot gauging station (Station 24) on a monthly timescale (2018–2023), En-LSTM(i) achieved superior performance (NSE = 0.89, KGE = 0.66) compared to SWAT+ (NSE = 0.64, KGE = 0.72), suggesting data-driven approaches can match or exceed physics-based models under limited calibration resources.

Contributions

• Introduction of a novel, unified Enhanced LSTM (En-LSTM) framework that synergistically integrates Temporal Convolutional Networks, an attention mechanism, a peak-aware hybrid loss function, and multi-scale lag and rolling-window features, a combination not previously explored in hydrological deep learning studies.
• Development of a basin-wide integrated En-LSTM (En-LSTM(i)) model that demonstrates robust transferability and superior performance in ungauged or data-scarce locations, by learning shared hydrological patterns across diverse physiographic and climatic conditions.
• Significant improvement in peak-flow simulation accuracy compared to conventional LSTMs and a process-based SWAT+ model, which is crucial for flood forecasting and water resource management in monsoon-dominated basins.
• Validation of a streamflow normalization strategy (scaling by basin area and mean annual precipitation) that enhances balanced learning across hydrologically contrasting catchments, promoting regional transferability.

Funding

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

Citation

@article{Sharma2026Improving,
  author = {Sharma, Purshottam and Tripathi, Vaibhav and Mohanty, Mohit Prakash},
  title = {Improving Daily Streamflow Predictions over Large Watersheds: Introducing a Novel Enhanced Long Short-Term Memory (En-LSTM) Model},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-025-04475-1},
  url = {https://doi.org/10.1007/s11269-025-04475-1}
}