Enhancing Water Level Prediction Using Ensemble Machine Learning Models: A Comparative Analysis

Accurate water level prediction is crucial for effective water resource management, flood forecasting, and prevention. This study assesses the performance of XGBoost, CatBoost, LGBM, and Random Forest models in predicting water levels in the Narmada Basin. The models were evaluated using statistical metrics, including MAE, MSE, RMSE, NRMSE, RMSPE, and R². The training results demonstrate that XGBoost outperforms the other models, yielding the lowest error values and an R² of 0.99. However, on the validation and testing datasets, Random Forest demonstrates the highest robustness and generalizability, achieving the lowest prediction errors (MAE of 0.27, RMSE of 0.56) and the highest R² (0.79). In contrast, XGBoost exhibits overfitting, leading to reduced accuracy on unseen data, while CatBoost and LGBM show strong predictive capabilities but with greater variability in predictions. This analysis underscores the importance of model selection in hydrological forecasting, with Random Forest emerging as the most reliable choice for real-world applications.