Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Hybrid Machine Learning Framework for Improved Short-Term Peak-Flow Forecasting

Gabriele Bertoli, Kai Schroeter, Rossella Arcucci, Enrica Caporali

TL;DR

This work presents a lightweight hybrid forecasting framework that merges Extreme Gradient Boosting (XGBoost) for general streamflow with Random Forest (RF) specialized in peak-flow predictions, tested across 857 LamaH-CE catchments with 6-hour resolution data. The approach tokenizes temporal dynamics via lag features and rolling statistics, and defines peaks as exceedances of the $99.9^{\text{th}}$ percentile, integrating the RF peak insights by a calibrated 0.95 scaling into the XGBoost forecast. Across comparisons with the EFAS system, the XGBoost-RF method achieves high skill, notably improving peak-magnitude accuracy and reducing false alarms, while remaining computationally lightweight and scalable to large watershed domains. The study demonstrates good operational performance, offers detailed evaluation via NSE, KGE, RPE, and WMO-guided metrics, and discusses practical considerations for energy use, data requirements, and future enhancements such as extending forecasting horizons and incorporating additional hydrometeorological variables.

Abstract

Reliable river flow forecasting is an essential component of flood risk management and early warning systems. It enables improved emergency response coordination and is critical for protecting infrastructure, communities, and ecosystems from extreme hydrological events. Process-based hydrological models and purely data-driven approaches often underperform during extreme events, particularly in forecasting peak flows. To address this limitation, this study introduces a hybrid forecasting framework that couples Extreme Gradient Boosting (XGBoost) and Random Forest (RF). XGBoost is employed for continuous streamflow forecasting, while RF is specifically trained for peak-flow prediction, and the two outputs are combined into an enhanced forecast. The approach is implemented across 857 catchments of the LamaH-CE dataset, using rainfall and discharge observations at 6-hour resolution. Results demonstrate consistently high skill, with 71% of catchments achieving a Kling-Gupta Efficiency (KGE) greater than 0.90. Peak-flow detection reaches 87%, with a false-alarm rate of 13%. Compared to the European Flood Awareness System (EFAS), the framework achieves lower peak-magnitude errors, fewer false alarms, and improved streamflow and peak-flow forecasting accuracy. The proposed framework is computationally lightweight, scalable, and easily transferable across watersheds, with training times of only seconds on standard CPUs. These findings highlight the potential of integrating hydrological understanding with efficient machine learning to improve the accuracy and reliability of operational flood forecasting, and outline future directions for hybrid hydrological-machine learning model development.

A Hybrid Machine Learning Framework for Improved Short-Term Peak-Flow Forecasting

TL;DR

This work presents a lightweight hybrid forecasting framework that merges Extreme Gradient Boosting (XGBoost) for general streamflow with Random Forest (RF) specialized in peak-flow predictions, tested across 857 LamaH-CE catchments with 6-hour resolution data. The approach tokenizes temporal dynamics via lag features and rolling statistics, and defines peaks as exceedances of the percentile, integrating the RF peak insights by a calibrated 0.95 scaling into the XGBoost forecast. Across comparisons with the EFAS system, the XGBoost-RF method achieves high skill, notably improving peak-magnitude accuracy and reducing false alarms, while remaining computationally lightweight and scalable to large watershed domains. The study demonstrates good operational performance, offers detailed evaluation via NSE, KGE, RPE, and WMO-guided metrics, and discusses practical considerations for energy use, data requirements, and future enhancements such as extending forecasting horizons and incorporating additional hydrometeorological variables.

Abstract

Reliable river flow forecasting is an essential component of flood risk management and early warning systems. It enables improved emergency response coordination and is critical for protecting infrastructure, communities, and ecosystems from extreme hydrological events. Process-based hydrological models and purely data-driven approaches often underperform during extreme events, particularly in forecasting peak flows. To address this limitation, this study introduces a hybrid forecasting framework that couples Extreme Gradient Boosting (XGBoost) and Random Forest (RF). XGBoost is employed for continuous streamflow forecasting, while RF is specifically trained for peak-flow prediction, and the two outputs are combined into an enhanced forecast. The approach is implemented across 857 catchments of the LamaH-CE dataset, using rainfall and discharge observations at 6-hour resolution. Results demonstrate consistently high skill, with 71% of catchments achieving a Kling-Gupta Efficiency (KGE) greater than 0.90. Peak-flow detection reaches 87%, with a false-alarm rate of 13%. Compared to the European Flood Awareness System (EFAS), the framework achieves lower peak-magnitude errors, fewer false alarms, and improved streamflow and peak-flow forecasting accuracy. The proposed framework is computationally lightweight, scalable, and easily transferable across watersheds, with training times of only seconds on standard CPUs. These findings highlight the potential of integrating hydrological understanding with efficient machine learning to improve the accuracy and reliability of operational flood forecasting, and outline future directions for hybrid hydrological-machine learning model development.
Paper Structure (29 sections, 10 equations, 6 figures, 4 tables)

This paper contains 29 sections, 10 equations, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Paper Structure
  3. Related works and broader context
  4. Data Acquisition and Preprocessing
  5. River Discharge Data
  6. Rainfall Data
  7. Data preprocessing
  8. Methods
  9. XGBoost-RF river flow forecasting framework
  10. Extreme Gradient Boosting (XGBoost)
  11. Implementation of temporal dependencies in XGBoost: lag features and rolling statistics
  12. XGBoost hyperparameters tuning and custom performance metric setup
  13. Random Forest (RF) for Peak forecasting
  14. Peak flow definition
  15. Random Forest hyperparameters tuning and implementation
  16. ...and 14 more sections

Figures (6)

  • Figure 1: Location of the 859 LamaH– CE gauged catchments (following the LamaH-CE's delineation B, light blue polygons) and their gauges (blue triangles). The spatial domain spans approximately $8.8^{\circ}$-$17.6^{\circ}$E and $46.2^{\circ}$-$50.4^{\circ}$N, covering about $170{,}000~\text{km}^2$ mainly across Germany, Switzerland, Austria, and the Czech Republic. The lower right box shows the location of the domain within Central Europe. Background: OpenStreetMap.
  • Figure 2: Flowchart of combining XGBoost streamflow and RF peak flow forecasting using rainfall and river flow observations to provide a refined river flow forecasting framework
  • Figure 3: Spatial distribution of the Kling-Gupta Efficiency (KGE) obtained with the XGBoost-RF river-flow forecasting model for 857 LamaH-CE catchments during the testing period. Colour coding: red: KGE < 0.5 (poor); yellow: 0.7 $\leq$ KGE < 0.8 (good); light-green: 0.8 $\leq$ KGE < 0.9 (very good); green: KGE $\ge$ 0.9 (excellent) performance. Map background: OpenStreetMap
  • Figure 4: Forecasting performance of the XGBoost-RF model (light blue) versus the EFAS benchmark (orange) across all LamaH-CE catchments. The left panel shows box plots of NSE, KGE, KGE_mod, Abs_RPE, and Signed_RPE; the right panel displays peak-timing errors (6-h steps). XGBoost-RF consistently outperforms EFAS in magnitude metrics and exhibits lower absolute peak-error variability, while its peak-timing error is slightly delayed compared with EFAS.
  • Figure 5: Comparative analysis of hydrological forecasting performance between the XGBoost-RF model and the EFAS benchmark. The left panel displays continuous error metrics (RMSE, MAE, ME, MSE Mean Bias Squared, and MSE Second-order Bias), while the right panel features binary verification metrics (POD, POFD, SR, FAR, FB, FC, CSI, ETS, and PSS) in a radar plot.
  • ...and 1 more figures