Bagheri et al. (2026) RetroSight and ForeSight ensemble model (ReForM) for improved time series prediction: A case study on river temperature prediction

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-03-06
• Authors: Faezeh Bagheri, Jon M. Hathaway, Nathan Michael Barber, Colleen Rice Montgomery, Anahita Khojandi
• DOI: 10.1016/j.jhydrol.2026.135244

Research Groups

• Department of Industrial and Systems Engineering, University of Tennessee, Knoxville, United States
• Department of Civil and Environmental Engineering, University of Tennessee, Knoxville, United States
• River Management, Tennessee Valley Authority, Knoxville, United States

Short Summary

This study introduces ReForM, a novel data-driven and physics-informed ensemble model that leverages both historical data and future physics-based simulations to significantly improve time series predictions. Applied to river temperature forecasting, ReForM demonstrates superior accuracy, especially for long-term predictions, outperforming state-of-the-art machine learning benchmarks.

Objective

• To introduce and evaluate the RetroSight and ForeSight Ensemble Model (ReForM), a new class of data-driven and physics-informed ensemble prediction models, for improved time series prediction by leveraging both historical data and physics-based simulations of future conditions.
• To apply ReForM to a case study on river temperature prediction in the Buffalo River, Tennessee, USA, to facilitate proactive river management and assess its performance across various lead times, particularly for long-term forecasts.

Study Configuration

• Spatial Scale: Buffalo River, Tennessee, USA.
• Temporal Scale: Predictions evaluated across various lead times, with particular effectiveness noted for long-term predictions (beyond 24 hours).

Methodology and Data

• Models used: RetroSight and ForeSight Ensemble Model (ReForM). Comparative statistical analyses were performed against Autoregressive, Wavelet Artificial Neural Network, Random Forest, XGBoost, and SVM models.
• Data sources: Historical data and physics-based simulations of future conditions.

Main Results

• ReForM achieved a root mean square error (RMSE) of 1.456 °C across all experiments for river temperature prediction.
• This represents an average improvement of approximately 38% (or approximately 1 °C) in RMSE compared to state-of-the-art machine learning benchmarks that rely solely on historical data.
• ReForM demonstrated significant improvement in river temperature prediction, particularly for long-term predictions (beyond 24 hours), and can achieve good performance with a minimal feature set.

Contributions

• Introduction of ReForM, a novel data- and physics-informed ensemble prediction model that uniquely combines historical learning with insights from physics-based simulations of future conditions.
• Demonstrated significant improvement in prediction accuracy (approximately 38% reduction in RMSE) for river temperature forecasting compared to existing state-of-the-art machine learning models.
• Showcased the model's particular effectiveness in long-term predictions (beyond 24 hours) and its ability to achieve good performance with a minimal feature set, enhancing practical applicability for river management.

Funding

• Not specified in the provided text.

Citation

@article{Bagheri2026RetroSight,
  author = {Bagheri, Faezeh and Hathaway, Jon M. and Barber, Nathan Michael and Montgomery, Colleen Rice and Khojandi, Anahita},
  title = {RetroSight and ForeSight ensemble model (ReForM) for improved time series prediction: A case study on river temperature prediction},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135244},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135244}
}