TL;DR
This paper introduces a novel Long Short-Term Memory (LSTM) network-based model (LSTM-FF) for flash flood forecasting in mountainous catchments. The model leverages multivariate time-series data of rainfall and discharge to predict flood events with lead times up to 10 hours, demonstrating strong performance, especially for large flood events.
What Problem Does This Paper Address?
Flash floods pose a significant threat, particularly in mountainous regions, due to their rapid onset and complex contributing factors like terrain and rainfall patterns. Effectively forecasting these events with sufficient lead time is a persistent challenge. Traditional hydrological models often struggle with computational inefficiency and the need for high-resolution input data, limiting their applicability. While machine learning offers promise, accurately capturing the temporal dynamics of rainfall-runoff processes for short-term flood prediction remains an area for improvement.
Methods
The core of this research is the development of a Long Short-Term Memory (LSTM) Flood Forecasting (LSTM-FF) model. This model is built upon T multivariate single-step LSTM networks, where each network is dedicated to forecasting discharge at a specific lead time (from 1 to 10 hours). The model takes into account both observed historical data and short-term rainfall forecasts.
Key aspects of the methodology include:
- Multivariate Input: The model utilizes spatial and temporal dynamics of observed rainfall (from multiple stations) and historical discharge data, along with forecasted rainfall.
- Single-Step Forecasting: For each lead time, a separate LSTM network is trained to predict the discharge at that specific future time step.
- Data Normalization: A crucial preprocessing step involves normalizing rainfall and discharge data using a min-max scaling approach to handle their different physical scales.
- Model Training: A standard machine learning approach of splitting data into training, validation, and test sets (6:2:2 ratio) is employed. Hyperparameters are tuned through trial and error, with Mean Absolute Error (MAE) chosen as the loss function to mitigate issues with low-flow fitting that can arise with Mean Squared Error (MSE).
- Evaluation Metrics: The performance is assessed using Relative Peak Error (RPE), Peak Time Error (PTE), Nash-Sutcliffe coefficient of efficiency (NSE), and a Qualified Rate (QR) which combines these metrics based on predefined thresholds.
Datasets
The study utilizes hydrometeorological data from the Anhe catchment in Jiangxi, China.
- Data Period: 1984 to 2012.
- Data Resolution: Hourly time series for rainfall and discharge.
- Flood Event Selection:
- 19 flood events with peak discharge exceeding a threshold of 176 m³/s (estimated 2-year return period flow).
- 75 flood events with peak discharge exceeding the 1-year return period discharge.
- Training/Validation/Test Split: The dataset was divided into training (45 flood events), validation (15 flood events), and test (15 flood events) sets. To ensure representativeness, some events were reallocated.
- Sample Size: For a 1-hour lead time, the training set comprised 6595 samples, validation 2334, and test 2109. These sizes vary slightly with lead time.
Key Insights and Ideas
- Small Floods Aid Large Flood Prediction: Surprisingly, the study found that including smaller flood events in the training data significantly improved the LSTM-FF model's ability to capture the rainfall-runoff relationship, even when the primary focus is on forecasting large flood events. This suggests that a comprehensive dataset, not just extreme events, is beneficial.
- Discharge Feature Dominance at Short Lead Times: The analysis of LSTM network weights revealed that the observed discharge plays a more critical role in the 1-hour lead-time forecast. This influence diminishes as the lead time increases, which is intuitive as future discharge is less directly tied to immediate past observations.
- Consistent Relationships Across Adjacent Lead Times: The weight matrix analysis also indicated that LSTM networks explore similar input-output relationships for adjacent lead times (e.g., 4-8 hours, 9-10 hours). This suggests the model learns robust temporal patterns rather than arbitrary correlations.
- Strong Extrapolation Ability: The LSTM-FF model demonstrated good generalization capabilities, performing well on unseen test data, which is crucial for real-world forecasting applications.
Key Results
- High Qualified Rates (QR): The LSTM-FF model achieved high QR values across various lead times. For large flood events, the QR for peak discharge was above 94.7% at 1-5 hour lead times and ranged from 84.2% to 89.5% at 6-10 hour lead times. Overall QR values for peak discharge, peak time, and flood process were consistently above 80% for the test period, often exceeding 90%.
- Superior Performance for Large Floods: The model showed more stable and better statistical performance in simulating large flood events, which is critical for early warning systems.
- Practical Tool for Forecasting: The model's ability to fit observed hydrograph trends and provide reasonable forecasts across different lead times positions it as a practical tool for flash flood prevention.
- Benchmark Model Outperformed: The LSTM-FF model generally outperformed a benchmark model (likely a simpler hydrological or statistical model, though not explicitly detailed) in the test period.
Limitations and Future Work
- Underestimation at Longer Lead Times: The study noted a tendency for the LSTM-FF model to underestimate peak discharge for lead times of 4-10 hours.
- Influence of Rainfall Forecast Uncertainty: The authors suggest that future research should investigate the impact of uncertainty in rainfall forecasts on the accuracy of flash flood predictions.
- Integration of Hydrological Knowledge: Further work could focus on integrating hydrological domain knowledge into the LSTM model structure and interpreting the evolution of LSTM cell states.
- Multivariate Multi-Step Time-Series: The paper highlights that LSTM networks are well-suited for more complex multivariate multi-step time-series problems, suggesting their potential for broader hydrological applications.
Summary updated