A Comparative Study of Data-driven Models for Groundwater Level Forecasting

摘要

Abstract Irregular rainfall patterns and limited freshwater availability have driven humans to increase their dependence on groundwater resources. An essential aspect of effective water resources management is forecasting groundwater levels to ensure that sufficient quantities are available for future generations. Prediction models have been widely used to forecast groundwater levels at the regional scale. This study compares the accuracy of five commonly used data-driven models–Holt–Winters’ Exponential Smoothing, Seasonal Autoregressive Integrated Moving Average, Multi-Layer Perceptron, Extreme Learning Machine, and Neural Network Autoregression for simulating the declining groundwater levels of three monitoring wells in the National Capital Territory of Delhi in India. The performance of the selected models was compared using coefficient of determination (R2), Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). Results indicate that Multi-Layer Perceptron had high R2 while fitting the training data and least RMSE and MAE during testing, thus proving to be more accurate in forecasting than the other models. Multi-Layer Perceptron was used to forecast the groundwater level in the study wells for 2025. The results showed that the groundwater level will decline further if the current situation continues. Such studies help determine the appropriate model to be used for regions with limited available data. Additionally, predictions made for the future will help policymakers understand which areas need immediate attention in terms of groundwater management.