Abstract
Accurate forecasting of river discharge and water levels is essential for effective water resource management, flood mitigation, and public safety. This study compares correlation-based and PCA-based feature selection methods for LSTM forecasting models in the study area at Uba River basin, within Shemonaiha city in the East Kazakhstan region. The dataset spans from 1995 to 2021, with 1995 to 2019 used for training and validation and 2020 to 2021 for testing. Both feature selection methods reduced the original predictor set to 13 features while generally maintaining predictive accuracy. An ensemble of 10 LSTM models was trained using 60-day input sequences to forecast discharge and water levels over a 10-day horizon, reducing variance from random initialization and stabilizing predictions. Performance was evaluated using the Nash-Sutcliffe Efficiency. Results showed that correlation-based selection performed comparably to the full-feature baseline in 2020 test set, suggesting that removing highly correlated predictors did not decrease short-term forecasts capacity of the model. The model with PCA-based selected features, while slightly lagging at longer lead times in 2020, exhibited advantages in most lead times with 2021 forecasts. However, overall predictive performance declined in 2021 compared to 2020, indicating that the hydrological conditions deviate more from the historical training record, and suggesting the need for model updates with relevant historical training data. Both feature selection methods successfully reduced dimensionality, while preserving performance capacity, though neither was universally superior across all forecast lead times. These results emphasize the value of systematic feature selection in hydrological modeling and highlight the importance of model adaptability to evolving environmental conditions.