Shabbir et al. (2025) Hybrid method for river inflow prediction: an integration of Hampel filter, decomposition techniques, and support vector machine

Identification

• Journal: Applied Water Science
• Year: 2025
• Date: 2025-11-12
• Authors: Maha Shabbir, Sohail Chand, Rana Abdul Wajid
• DOI: 10.1007/s13201-025-02660-6

Research Groups

• Department of Mathematics and Statistical Sciences, Lahore School of Economics, Lahore, Pakistan
• College of Statistical Sciences, University of the Punjab, Lahore, Pakistan

Short Summary

This study develops a novel hybrid model (HEVS: Hampel filter, EEMD, VMD, SVM) for daily river inflow prediction, demonstrating significantly improved accuracy over existing models by effectively addressing outliers, noise, randomness, and multi-scale characteristics in hydrological time series within the Indus River Basin.

Objective

• To develop and validate a novel three-layer hybrid model (HEVS) that effectively predicts daily river inflow by addressing outliers, noise, randomness, and multi-scale characteristics in hydrological time series.

Study Configuration

• Spatial Scale: Four main tributaries of the Indus River Basin (IRB) in Pakistan: Indus River, Chenab River, Kabul River, and Jhelum River.
• Temporal Scale: Daily inflow series from 1 January 2015 to 27 March 2027 (3374 observations).

Methodology and Data

• Models used: Hampel filter (HF), Ensemble Empirical Mode Decomposition (EEMD), Variational Mode Decomposition (VMD), Support Vector Machine (SVM). The proposed hybrid model is HEVS (HF–EEMD–VMD–SVM).
• Data sources: Daily inflow data acquired from the WAPDA website (https://www.wapda.gov.pk/).

Main Results

• The proposed HEVS hybrid model consistently outperformed standalone SVM and other hybrid models (HF-SVM, EEMD-SVM, VMD-SVM, HF-EEMD-SVM, HF-VMD-SVM, EEMD-VMD-SVM) across all four studied rivers.
• For the Chenab River (S1) in the testing phase, HEVS demonstrated a 57.99% lower Mean Squared Error (MSE) compared to the standalone SVM model. Similar significant reductions in MSE were observed against other competing models: 48.91% lower than HF-SVM, 46.07% lower than EEMD-SVM, 38.3% lower than VMD-SVM, 36.27% lower than HF-EEMD-SVM, 24.68% lower than HF-VMD-SVM, and 18.35% lower than EEMD-VMD-SVM.
• The Diebold-Mariano (DM) test confirmed that the HEVS hybrid model exhibited significantly higher prediction accuracy and lower prediction errors than all comparative models at 10%, 5%, and 1% significance levels.
• The average daily inflow rates for the rivers were: Chenab (S1) 884.08 m³/s (31.2 × 1000 ft³/s), Jhelum (S2) 781.53 m³/s (27.6 × 1000 ft³/s), Kabul (S3) 773.08 m³/s (27.3 × 1000 ft³/s), and Indus (S4) 2191.98 m³/s (77.4 × 1000 ft³/s).
• The standard deviations of daily inflow were: Chenab (S1) 812.87 m³/s (28.7 × 1000 ft³/s), Jhelum (S2) 591.82 m³/s (20.9 × 1000 ft³/s), Kabul (S3) 724.91 m³/s (25.6 × 1000 ft³/s), and Indus (S4) 2328.09 m³/s (82.2 × 1000 ft³/s).

Contributions

• Introduction of a novel three-layer hybrid model (HEVS) that uniquely integrates Hampel filter for outlier correction, Ensemble Empirical Mode Decomposition for denoising, and Variational Mode Decomposition for multi-scale decomposition, all feeding into a Support Vector Machine model.
• Addresses a critical gap in hydrological modeling by effectively handling the joint presence of nonlinearity, randomness, outliers, and multi-scale characteristics in river inflow time series.
• Demonstrates superior predictive performance and significant error reduction compared to existing individual and multi-layer hybrid models, providing a more robust tool for hydrological forecasting.
• Offers a practical and efficient approach for river flow management, which can aid in mitigating the impacts of droughts, heat waves, and floods.

Funding

• No funding for this research work.

Citation

@article{Shabbir2025Hybrid,
  author = {Shabbir, Maha and Chand, Sohail and Wajid, Rana Abdul},
  title = {Hybrid method for river inflow prediction: an integration of Hampel filter, decomposition techniques, and support vector machine},
  journal = {Applied Water Science},
  year = {2025},
  doi = {10.1007/s13201-025-02660-6},
  url = {https://doi.org/10.1007/s13201-025-02660-6}
}