A Hybrid Feature-Engineered XGBoost Framework for Rainfall-Runoff Based Inflow Prediction of Ukai Dam, India

Technical Summary: A Hybrid Feature-Engineered XGBoost Framework for Rainfall-Runoff Based Inflow Prediction of Ukai Dam, India

Problem Statement
Accurate reservoir inflow prediction is critical for water resource management, flood control, hydropower optimization, and irrigation planning, particularly for strategically important systems like the Ukai Dam in Gujarat, India. While traditional physically-based hydrological models offer interpretability, they often require extensive calibration and detailed watershed data. Conversely, purely data-driven machine learning models can capture complex nonlinear relationships but may fail to fully exploit physical hydrological knowledge. This study addresses the need for a robust forecasting system that integrates physical runoff estimation with advanced machine learning to handle the nonlinear hydrological processes and high variability of the Ukai catchment, which experiences distinct peak and low-flow periods driven by monsoon rainfall.

Methodology
The study proposes a hybrid feature-engineered XGBoost framework tailored to the hydrological and meteorological attributes of the Ukai Dam. The methodology follows a structured workflow:

1. Data Collection and Preprocessing:

   • Sources: Data was aggregated from the Ukai Dam station (1985–2025) and 52 upstream meteorological stations, covering hydrological, meteorological, and spatial variables.
   • Variables: Inputs include reservoir attributes (opening capacity, discharge, volume, balance capacity, inflow) and meteorological data (temperature, rainfall, humidity, wind speed, evaporation). Spatial data (elevation, slope, Land Use/Land Cover) were derived from Google Earth/Maps and Bhuvan (NRSC).
   • Processing: Raw data (approx. 200k reservoir and 278k runoff records) was cleaned, temporally aggregated to daily averages, and missing values imputed using mean values. Outliers were capped at the 99.5th percentile.

2. Hybrid Runoff Estimation (Physical Component):

   • An enhanced Soil Conservation Service-Curve Number (SCS-CN) method was employed to estimate runoff.
   • The Curve Number (CN) was assigned based on Land Use/Land Cover (LULC) and geomorphology.
   • A slope-based adjustment factor was introduced to account for watershed variability: slopes >10% increased runoff generation (factor 1.2), while slopes <3% increased infiltration (factor 0.8).
   • This generated an "Adjusted Runoff" variable used as a primary input for the machine learning model.

3. Feature Engineering:

   • Lag Features: Inflow lags of 1, 2, 3, 5, and 7 days were included to capture temporal persistence and autoregressive relationships.
   • Rolling Statistics: 7-day rolling averages for rainfall and runoff were calculated to represent cumulative watershed response and reduce noise.
   • Seasonal Indicators: A "month" feature was extracted to capture seasonal variability.
   • Target Transformation: To stabilize variance and reduce skewness, the target variable (inflow) was transformed using a low-power logarithmic function: $y_t = \log(1 + \text{Inflow}^{0.7})$.

4. Model Training and Optimization:

   • Algorithm: Extreme Gradient Boosting (XGBoost) was used with 800 decision trees, a maximum depth of 5, and a learning rate of 0.025.
   • Peak-Flow Weighting: To improve prediction of extreme events, observations exceeding the 90th percentile of historical inflows were assigned double the training weight ($w_i = 2$) compared to normal flows.
   • Validation: A 5-fold TimeSeriesSplit cross-validation strategy was employed to prevent data leakage and preserve temporal ordering.

5. Comparative Analysis:

   • The proposed framework was benchmarked against Decision Tree (DT), Random Forest (RF), Gradient Boosting Regressor (GBR), and Long Short-Term Memory (LSTM) networks using identical input features and validation protocols.

Key Contributions

• Integration of Physics and ML: The study successfully integrates physically-based runoff estimation (enhanced SCS-CN with slope adjustments) with machine learning, bridging the gap between data-driven and process-based modeling.
• Hybrid Feature Engineering: Development of a comprehensive feature set combining lagged inflow, rolling hydrological statistics, reservoir state variables, and seasonality.
• Peak-Aware Learning: Introduction of a specific weighting mechanism to enhance the model's sensitivity to extreme inflow events, which are critical for flood forecasting.
• Comprehensive Benchmarking: A rigorous comparison demonstrating the superiority of the hybrid approach over classical ML, ensemble methods, and deep learning (LSTM) for this specific reservoir context.

Results
The proposed Hybrid XGBoost model demonstrated exceptional predictive accuracy, significantly outperforming all comparison models:

• Performance Metrics: The model achieved an $R^2$ of 0.992 and a Nash-Sutcliffe Efficiency (NSE) of 0.992. The Root Mean Square Error (RMSE) was 7.415, and the Mean Absolute Error (MAE) was 0.977.
• Comparison:
  • vs. Lag-less Model: Removing lag features caused performance to collapse (RMSE increased to 5409.4, $R^2$ dropped to -17602), highlighting the critical importance of temporal dependencies.
  • vs. Other ML/DL Models: The Hybrid XGBoost model vastly outperformed DT ($R^2=0.18$), RF ($R^2=0.11$), GBR ($R^2=0.08$), and LSTM ($R^2=0.15$).
  • Efficiency: While Random Forest required the highest computational time (67.94 minutes), the proposed model trained in approximately 4.05 minutes, offering a balance of high accuracy and computational efficiency.
• Feature Importance: Analysis revealed that lagged inflow variables were the most influential predictors, followed by runoff estimates and reservoir operation parameters (balance/opening capacity).

Significance and Claims
The authors claim that the proposed framework serves as a reliable tool for flood forecasting, reservoir operation, and sustainable water management. By achieving high accuracy in predicting both low-flow and peak-flow conditions, the model supports critical decision-making for:

• Real-time flood forecasting.
• Reservoir operation optimization.
• Water allocation planning.
• Hydropower scheduling.
• Climate resilience assessment.

The study concludes that hybrid approaches, which combine physical runoff estimation with advanced feature engineering and gradient boosting, significantly outperform conventional machine learning and deep learning models for inflow prediction in complex reservoir systems like Ukai Dam. The authors suggest future work could incorporate radar rainfall products, satellite observations, and transformer-based architectures to further enhance performance.