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

Abstract

Precise inflow prediction is very important for water resource management, flood control and hydropower optimization, especially for strategically important reservoir like Ukai Dam in Gujarat, India. This study proposes a novel hybrid feature-engineered XGBoost model tailored to the hydrological and meteorological attributes of the Ukai catchment. The model integrates an enhanced Soil Conservation Curve Number (SCS-CN) method, incorporating with LULC, geomorphology, slope and elevation data, alongside a slope-based runoff adjustment factor to better represent watershed variability. Advanced feature engineering techniques, including lag features, rolling statistics, seasonal indicators, and reservoir state variables, are employed to capture temporal dynamics. Variance stabilization is achieved through low-power transformation, while peak-flow-aware weighting enhances the prediction of extreme inflow events. The structured modelling workflow ensures reproducibility and scalability. Evaluation of model is done by matrices like RMSE, MAE, R² and NSE. The findings of proposed model is compared against traditional models like Decision Tree, Gradient Boosting, Random Forest and Long-Short-Term Memory. The proposed hybrid framework achieved excellent predictive accuracy (R²= 0.992, NSE=0.992) with low prediction error(RMSE=7.415, MAE=0.977) outperforming Decision Tree, Gradient Boosting, Random Forest and LSTM models and demonstrated robust capability for reservoir inflow forecasting. The findings make the proposed model as reliable tool for flood forecasting, reservoir operation and sustainable water management.