Flash Flood Forecasting Using Support Vector Regression Model in a Small Mountainous Catchment

Jian Wu, Haixing Liu, Guozhen Wei, Tianyu Song, Chi Zhang, Huicheng ZhouWater 2019 · DOI

TL;DR

This paper presents a Support Vector Regression (SVR) model for flash flood forecasting in a small mountainous catchment, achieving satisfactory predictions up to three hours ahead. The study highlights that while lagged runoff data is crucial for predicting peak flow magnitude, rainfall data plays a key role in mitigating the timing error (phase lag) of these predictions.

What Problem Does This Paper Address?

Flash floods in mountainous regions pose a significant threat due to their rapid onset and destructive potential. Accurately forecasting these events is challenging for hydrologists and managers because of the inherent nonlinearity, instantaneity, and complex dependencies within hydrological processes. Traditional hydrological models often struggle with the limited data availability and accuracy required for effective forecasting in such dynamic environments. Data-driven models, particularly those leveraging machine learning, have emerged as a promising alternative due to their ability to capture complex, nonlinear relationships. This study specifically addresses the need for reliable, multi-hour ahead flash flood forecasting in mountainous catchments, a critical requirement for effective flood warning and evacuation.

Methods

The core of this research is the application of a Support Vector Regression (SVR) model. SVR is a powerful machine learning technique that maps input data into a high-dimensional feature space to model nonlinear relationships. The authors employed a nonlinear SVR to capture the complex hydrological processes.

Key methodological aspects include:

  • Input Variable Selection: Lagged average rainfall and runoff data were chosen as input variables. The time lags for these variables were determined using the hydrological concept of the "time of response," which is the time it takes for a catchment to respond to rainfall.
  • Parameter Calibration: A two-step grid search method was used to optimize the SVR hyperparameters (C, ε, and γ). This method is favored for its simplicity and efficiency in finding optimal parameters.
  • Kernel Function: The Radial Basis Function (RBF) kernel was selected for the SVR model, as it has shown strong performance in previous hydrological applications.
  • Sensitivity Analysis: A Leave-One-Out Method (LOOM) was adapted to assess the contribution of different input variables (rainfall vs. runoff) to the prediction accuracy.
  • Data Preprocessing: Input variables were normalized using the MinMaxScaler method to scale them to a range of [0, 1], ensuring that variables with different magnitudes did not disproportionately influence the model.

Datasets

The study utilized historical rainfall and discharge data from the Anhe (AH) catchment in Jiangxi province, China.

  • Data Source: Hourly rainfall and discharge data were collected from local hydrological bureaus.
  • Flood Events: A total of 69 flash flood events spanning from 1984 to 2012 were used.
  • Data Split: The data was chronologically split into 80% for training (55 events) and 20% for testing (14 events).
  • Rainfall Data: Average rainfall was calculated using data from eight rainfall stations via the Thiessen polygon method.

Key Insights and Ideas

  • Runoff Dominance for Magnitude, Rainfall for Timing: A crucial finding is the distinct roles of rainfall and runoff data. Lagged runoff variables are more influential in predicting the magnitude of peak flows, while rainfall variables are more effective in reducing the timing error (phase lag) of these predictions. This suggests a synergistic approach is best for comprehensive forecasting.
  • Multi-Hour Ahead Forecasting is Feasible but Challenging: The SVR model demonstrates satisfactory performance for forecasting up to three hours ahead, which is a significant advancement for operational flood warnings. However, prediction accuracy naturally declines with increasing lead time.
  • Lagged Runoff is Key: The study emphasizes that historical runoff data is a strong predictor of future runoff, often more so than rainfall alone, especially for predicting the peak flow magnitude.
  • Phase Lag is a Persistent Issue: For longer lead times (2-3 hours), a noticeable phase lag (difference between observed and predicted peak times) exists. This highlights the inherent difficulty in precisely predicting the timing of rapid hydrological events.

Key Results

  • Satisfactory Predictive Performance: The SVR models achieved satisfactory predictive performances for one to three-hour ahead flash flood forecasting.
  • Lead Time Impact: The three-hour ahead predictions were generally worse than one- and two-hour ahead forecasts, indicating a trade-off between lead time and accuracy.
  • Runoff Lagged Variables' Significance: Lagged runoff variables had a more significant effect on predicting peak flow magnitude compared to rainfall variables.
  • Phase Lag Mitigation: While phase lag was present in two- and three-hour ahead forecasts, the inclusion of rainfall information helped to mitigate this timing error.
  • Quantitative Metrics:
    • Qualified Rate (QR): For lead times of 1, 2, and 3 hours, the testing QR values were 1.000, 1.000, and 0.786, respectively, indicating high accuracy for shorter lead times.
  • Root Mean Square Error (RMSE): Average RMSE values for testing were 4.1 m³/s (1h), 10.1 m³/s (2h), and 12.2 m³/s (3h), showing increasing error with longer lead times.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Catchment Specificity: The study was conducted on a single, small mountainous catchment. Future work should investigate the SVR model's performance in catchments with different spatial scales, topography, and climatic conditions (e.g., arid, semi-arid, semi-humid).
  • Climate and Land Use Change: The hydrological data used spans from 1984 to 2012. Given the increasing impact of climate and land use changes on hydrological processes, future studies should incorporate more recent data to assess these evolving dynamics.
  • Model Variations: The study focused on a specific SVR configuration. Exploring different kernel functions and comparing SVR with other machine learning techniques could provide further insights.
  • Input Variable Exploration: While rainfall and runoff were analyzed, further investigation into other potential input variables (e.g., antecedent soil moisture, temperature) could enhance model performance.

Summary updated

Abstract

Flash floods in mountainous catchments are often caused by the rainstorm, which may result in more severe consequences than plain area floods due to less timescale and a fast-flowing front of water and debris. Flash flood forecasting is a huge challenge for hydrologists and managers due to its instantaneity, nonlinearity, and dependency. Among different methods of flood forecasting, data-driven models have become increasingly popular in recent years due to their strong ability to simulate nonlinear hydrological processes. This study proposed a Support Vector Regression (SVR) model, which is a powerful artificial intelligence-based model originated from statistical learning theory, to forecast flash floods at different lead times in a small mountainous catchment. The lagged average rainfall and runoff are identified as model input variables, and the time lags associated with the model input variables are determined by the hydrological concept of the time of response. There are 69 flash flood events collected from 1984 to 2012 in a mountainous catchment in China and then used for the model training and testing. The contribution of the runoff variables to the predictions and the phase lag of model outputs are analyzed. The results show that: (i) the SVR model has satisfactory predictive performances for one to three-hours ahead forecasting; (ii) the lagged runoff variables have a more significant effect on the predictions than the rainfall variables; and (iii) the phase lag (time difference) of prediction results significantly exists in both two- and three-hours-ahead forecasting models, however, the input rainfall information can assist in mitigating the phase lag of peak flow.
BibTeX
@inproceedings{wu2019flash,
  author = {Jian Wu and Haixing Liu and Guozhen Wei and Tianyu Song and Chi Zhang and Huicheng Zhou},
  title = {{Flash Flood Forecasting Using Support Vector Regression Model in a Small Mountainous Catchment}},
  booktitle = {Water},
  year = {2019},
  doi = {10.3390/w11071327},
}
Figure