Haghiabi et al. (2025) A comparative study between time series and soft computing models for river discharge forecasting

Identification

• Journal: Applied Water Science
• Year: 2025
• Date: 2025-10-28
• Authors: Amir Hamzeh Haghiabi, Zahra Askari, Naseer Mohammad
• DOI: 10.1007/s13201-025-02632-w

Research Groups

• Department of Water Science, Lorestan University, Lorestan, Iran

Short Summary

This study compared six predictive models, including machine learning (SVR, RF, KNN, LSTM) and time series (CARMA, CARMA-GARCH), for monthly river discharge forecasting in the Kashkan River Basin, Iran, under two input scenarios. The Random Forest (RF) model demonstrated the highest accuracy among machine learning methods, while CARMA-GARCH was the best-performing time series model, with time series models generally showing superior performance.

Objective

• To compare the performance of various machine learning (Support Vector Regression, Random Forest, K-Nearest Neighbors, Long Short-Term Memory) and time series (Contemporaneous Autoregressive Moving Average, CARMA with Generalized Autoregressive Conditional Heteroskedasticity) models for monthly river discharge forecasting in the Kashkan River Basin, western Iran.
• To evaluate model performance under two input scenarios: one using lagged discharge data and another incorporating upstream station discharge data.
• To investigate the impact of heteroskedasticity on river discharge simulation.

Study Configuration

• Spatial Scale: Kashkan River Basin (KRB) in western Iran, spanning approximately 9,502.7 square kilometers, with data from eight hydrometric stations. Geographical coordinates range from latitude 33.5° N to 34.0° N and longitude 48.0° E to 49.0° E.
• Temporal Scale: Monthly river discharge data. Two scenarios were evaluated: one using a one-month lag in the data, and the second using upstream hydrometric station data for forecasting.

Methodology and Data

• Models used:
  • Machine Learning: Support Vector Regression (SVR), Random Forest (RF), K-Nearest Neighbors (KNN)
  • Deep Learning: Long Short-Term Memory (LSTM)
  • Time Series: Contemporaneous Autoregressive Moving Average (CARMA), CARMA with Generalized Autoregressive Conditional Heteroskedasticity (CARMA-GARCH)
• Data sources: Monthly river discharge data collected from eight hydrometric stations within the Kashkan River Basin. Data was pre-processed for missing values, trend analysis (Modified Mann–Kendall test), and normalization.

Main Results

• Overall Model Performance: Time series models generally provided better performance than machine learning models.
• Best Performing Models:
  • Among machine learning models, Random Forest (RF) consistently achieved the best results in both scenarios (average RMSE 4.6 m³/s for Scenario 1, 4.1 m³/s for Scenario 2).
  • Among time series models, CARMA-GARCH consistently outperformed CARMA (average RMSE 2.5 m³/s for Scenario 1, 2.51 m³/s for Scenario 2).
• Worst Performing Model: Support Vector Regression (SVR) did not produce satisfactory results in either scenario (average RMSE 18 m³/s for Scenario 1, 36.6 m³/s for Scenario 2).
• Performance Improvement: RF improved results by 36% compared to SVR, 11.75% compared to KNN, and 9.6% compared to LSTM.
• Impact of Heteroskedasticity: The CARMA-GARCH model, by incorporating heteroskedasticity modeling, achieved a remarkable 32% reduction in Root Mean Square Error (RMSE) values compared to the standalone CARMA model in Scenario 1.
• Scenario Comparison: Including upstream river discharge data (Scenario 2) generally improved the precision and efficiency of the CARMA model, with RMSE reductions of 1.2%, 9.2%, and 2.6% for Doab-V, Kashkan-AF, and Poldokhtar stations, respectively.
• Station-Specific Challenges: Doab-V, Poldokhtar, and Kashkan-AF stations consistently showed higher RMSE values across all models, likely due to high standard deviations and hydrological complexities.

Contributions

• Comprehensive comparative analysis of a diverse set of six predictive models (three ML, one DL, two time series) for river discharge forecasting.
• Evaluation of model performance under two distinct and practically relevant input scenarios (lagged data and upstream data).
• Demonstrated the significant benefit of incorporating heteroskedasticity modeling (CARMA-GARCH) for improving forecast accuracy in hydrological time series.
• Provided specific insights into the strengths and weaknesses of each model type for monthly river discharge prediction in a complex, flood-prone basin (Kashkan River Basin, Iran).
• Identified Random Forest and CARMA-GARCH as the most effective models for the study area, offering valuable guidance for water resource management.

Funding

This study was not funded by any organization or institution.

Citation

@article{Haghiabi2025comparative,
  author = {Haghiabi, Amir Hamzeh and Askari, Zahra and Mohammad, Naseer},
  title = {A comparative study between time series and soft computing models for river discharge forecasting},
  journal = {Applied Water Science},
  year = {2025},
  doi = {10.1007/s13201-025-02632-w},
  url = {https://doi.org/10.1007/s13201-025-02632-w}
}