Towards resilient water management in deltaic Bangladesh: benchmarking an exogenous regressor-enhanced prophet framework for multi-Lag river stage forecasting
Open Access
- 1 August 2026
- journal article
- Published by IWA Publishing in Water Policy
- Vol. 28 (8) , 1193-1233
Abstract
Flowchart showing integrated methodological framework for multi-lag river stage forecasting, progressing from data collection through preprocessing, feature engineering with four lag windows, two input regimes (stage-only and exogenous), five forecasting architectures, and multi-metric evaluation. Accurate river stage forecasting is critical for flood and drought management in deltaic regions experiencing extreme hydrological variability. This study benchmarked an exogenous regressor-enhanced Prophet framework against machine learning (Random Forest, Light Gradient Boosting) and deep learning (BiLSTM, GRU) models for the Old Brahmaputra River, Bangladesh, using 15 years of daily observations (2007–2021). Models were assessed across four lag configurations (3, 7, 15, 30 days) under stage-only autoregressive and exogenous hydro-meteorological predictor regimes using nine metrics, PCA composite scoring, and flow-regime-stratified Pareto analysis. Under stage-only conditions, Prophet achieved R2 of 0.9984–0.9985 and RMSE of 0.0732–0.0743 m with NSE exceeding 0.998. These near-perfect metrics, however, primarily reflect autoregressive persistence rather than causal forecasting skill; the exogenous evaluation directly quantifies genuine skill once this advantage is removed. PCA scores of 1.000 confirmed Prophet's dominance in most configurations. Ensemble methods exhibited drought-stage overestimation with MAE_low values 4–5 times larger than Prophet's and prediction intervals 3.1–3.4 times wider. Under exogenous inputs, Prophet retained dominant performance (R2 = 0.91–0.93) with best-calibrated uncertainty bounds. Prophet's dual-regime robustness, computational efficiency, and interpretable decomposition establish it as the optimal framework for operational deployment in resource-constrained deltaic environments.Keywords
Funding Information
- Ministry of Science and Technology, Government of the People’s Republic of Bangladesh (SRG-242251)
This publication has 43 references indexed in Scilit:
- River discharge prediction based multivariate climatological variables using hybridized long short-term memory with nature inspired algorithmJournal of Hydrology, 2024
- Hydro-informer: a deep learning model for accurate water level and flood predictionsNatural Hazards, 2024
- Enhancing streamflow simulations with gated recurrent units deep learning models in the flood prone region with low-convergence streamflow dataPhysics and Chemistry of the Earth, Parts A/B/C, 2024
- A critical review of RNN and LSTM variants in hydrological time series predictionsMethodsX, 2024
- Streamflow forecasting with deep learning models: A side-by-side comparison in Northwest SpainEarth Science Informatics, 2024
- A hybrid SARIMA-Prophet model for predicting historical streamflow time-series of the Sobat River in South SudanDiscover Applied Sciences, 2024
- Model-based prediction of water levels for the Great Lakes: a comparative analysisEarth Science Informatics, 2024
- Applications of XGBoost in water resources engineering: A systematic literature review (Dec 2018–May 2023)Environmental Modelling & Software, 2024
- Climate Change Impacts on Rainwater Tank’s Potential Water Savings, Efficiency and Reliability Presenting Relationship Between ‘Seasonality Index’ and Water Savings EfficiencyWater Resources Management, 2023
- Principal component analysisWIREs Computational Statistics, 2010