A novel approach to forecast water table rise in arid regions using stacked ensemble machine learning and deep artificial intelligence models

Accurate prediction of water table rise (WTR), or groundwater flooding is crucial for water resources management, flood risk mitigation, infrastructure planning, and environmental protection. This study utilized novel shallow learning (SL) and deep learning (DL) models for WTR forecasting for one-, two-, and three-week steps ahead at two levels of modeling in Muscat Governate, Oman. At Level 1, the predictive outcomes of SL models (CBR, XGB, LGBM) and DL models (LSTM, GRU, TR) were utilized as input for producing stacked SL and DL models at level 2 to improve the overall forecasting. This research used a total of 19,465 datasets with high-resolution (half-hour interval) measurements between Dec 2017 and Jan 2019. The data were split into training and testing sets, with 90% for training (the first 17,976 datasets) and 10% (the remaining 1,489 datasets) for testing. Feature engineering techniques, such as lags, differences, and aggregates of WTR, were used as input data for the individual SL and DL models at level 1 of modeling. The novel rolling forecasting method was applied alongside the models to capture patterns and trends changes over time. The results showed that the stacked SL and DL models at the level 2 strategy outperformed the standalone SL and DL models at level 2, based on NSE and RSR statistical metrics of the testing data. Furthermore, the modeling of one-week step-ahead provided more accurate WTR forecasting compared with the two- and three-week step-ahead forecasts. The improvement percentage of the stacked SL and DL models reached 32.80%, and there was a 41.17% reduction in RSR, respectively. The approach of this research could be effectively applied to several arid regions affected by WTR worldwide.