Water table rise forecasting using machine and deep learning models in arid regions, Oman

Accurately forecasting water table rise (WTR) is essential for effective water resource management, infrastructure development, flood risk mitigation, and environmental conservation. This research employed multiple machine learning (ML) models, namely Ridge Linear Regression (RLR), Radial Basis Function Support Vector Machine (RBF-SVM), Linear SVM (LSVM), Random Forest (RF), and a hybrid deep learning Transformer (TR) with Bi-Long Short-Term Memory (BiLSTM), to forecast WTR one and two weeks ahead in the Muscat Governorate, Oman. A total of 19,465 high-resolution datasets, measured at half-hour intervals between December 2017 and January 2019, were utilized. The data were divided into training and testing sets, with 90% (17,976 datasets) used for training and the remaining 10% (1,489 datasets) reserved for testing. A two-way time series analysis was employed to analyze dynamic interactions between two time-dependent behaviors over time. Additionally, the rolling forecasting method was used alongside the models to capture patterns and provide updated predictions based on the most recent data trends. The results demonstrated that RLR outperformed both the individual ML models and the hybrid deep learning TR-BiLSTM models, as indicated by the NSE and RSR statistical metrics applied to the testing data. Furthermore, the one-week step-ahead forecasting achieved greater accuracy in predicting WTR compared to the two-week step-ahead forecast. However, the average computational time of the hybrid deep learning TR-BiLSTM models was notably higher compared to the standalone models. Linear models such as RLR and LSVM demonstrated accurate forecasting results due to their ability to prevent overfitting in correlated features and effectively capture the simplicity of the relationship between the data. The approach presented in this research can be effectively useful to various arid regions worldwide that are influenced by WTR.