MODELLING FLUCTUATIONS OF GROUNDWATER LEVEL USING MACHINE LEARNING ALGORITHMS IN THE SOKOTO BASIN

Abstract

Groundwater level forecasting is essential for the sustainable management of water resources, especially water scarce regions such as the Sokoto Basin. This study investigates the application of machine learning models, specifically Long Short-Term Memory (LSTM), eXtreme Gradient Boosting (XGBoost)and Random Forest (RF) algorithms to predict groundwater levels across six boreholes within the Sokoto Basin. A thorough preprocessing procedure was applied to the daily groundwater data spanning a range of three to four years. This included removing null values, interpolating missing data and downsampling to weekly intervals engineering to improve model performance. Time series decomposition and the creation of lag features were also utilized to capture temporal dependencies effectively. Among the models, the XGBoost algorithm demonstrated the highest performance, providing precise predictions that closely aligned with the actual groundwater levels. Hyperparameters for the XGBoost model were fine-tuned using grid search techniques, resulting in optimal settings that significantly enhanced predictive accuracy with Mean Absolute Error (MAE) ranging from 0.016 – 0.757m and Root Mean Square Error (RMSE) ranging from 0.051 - 2.859m. The LSTM model also showed strong performance, particularly in capturing the peaks and valleys of the groundwater level time series, with MAE and RMSE values ranging from 0.016 – 0.757m and 0.051 – 2.859m, respectively. The RF model exhibited reliable performance across most locations. The research findings offer a practical method for forecasting groundwater levels. The frameworks could be used to manage water resources, particularly in dry years, where water restrictions and drought alerts can also be rapidly issued.