Abstract
Accurate water-level prediction is essential for effective water resource management in Kazakhstan. Reservoirs play an important role in irrigation, hydropower generation, and water supply. Traditional forecasting methods often struggle to capture complex interactions between natural processes and human activities, especially when data are limited or conditions change rapidly. This study applies three baseline machine learning models, Decision Tree, Random Forest, and K-Nearest Neighbors, as well as a hybrid model based on Random Forest Regressor and Multilayer Perceptron, to predict daily water levels using historical Kazhydromet data. The data were preprocessed by handling missing values and applying feature scaling. The experimental results showed that the hybrid RF+MLP model achieved the best performance among the baseline and proposed models. Additional TimeSeriesSplit validation and comparison with advanced models confirmed that RF-based models provide relatively stable results under limited data conditions. The findings indicate that machine learning methods, especially hybrid approaches, can support daily water-level forecasting and improve water resource management in Kazakhstan.