Exploring the Application of Machine Learning Models for Water Level Forecasting: A Case Study of the Arayat Station in the Pampanga River Basin, Philippines

Water level forecasting plays a critical role in flood prediction and early warning systems, particularly for dam operations, reservoir management, and river basin monitoring. This study developed and evaluated a machine learning-based framework for water level forecasting in the Pampanga River Basin, on the Arayat station with a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2 4}$</tex>-hour lead time. Ensemble machine learning models: Random Forest Multi-Output Regressor (RForest) and Extreme Gradient Boosting MultiOutput Regressor (XGBoost) were tested alongside a neural network-based model, the Long Short-Term Memory (LSTM). Using historical observed water level and rainfall data from the telemetry stations across the river basin spanning from 2009-2021, were split to 80 % of the dataset that was used for training and 20 % for testing. Following the time series splitting, two experimental configurations were explored: (1) a univariate model using only the target feature for training and testing, and (2) a multivariate model incorporating additional predictors such as rainfall and discharge in the target location and water level from multiple stations. Model performance was assessed using the Root-Mean Squared Error, Coefficient of Determination (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{R}^{\mathbf{2}}$</tex>), and Nash-Sutcliffe Efficiency (NSE). Results showed that the tree-based multi-output regressors performed better when multiple features were included, while the LSTM model achieved superior predictive accuracy under a single feature input. Overall, the XGBoost and RForest outperformed the LSTM on both multivariate and univariate experiment. The RForest and XGBoost model captured the peaks while LSTM underperformed. These findings highlight the potential of machine learning algorithms for real-time operational water level forecasting operational water level forecasting, even with limited availability of data. Future work should consider additional model architectures and systematic hyperparameter optimization to further increase the lead-time and the forecasting accuracy of the models.

Ҳали таржима қилинмаган