Comparative Study Of Machine Learning Models In Predicting Water Table Fluctuations In Azarshahr Plain, Iran
Auteurs : Molavi Ahad . Ahmadzadeh Kaleybar Fariborz . Fatemi Seyyed Majid .
Résumé
The management of drought and water scarcity concerning underground water resources requires modeling methodologies with a simple yet effective framework. Due to temporal and financial constraints, machine learning models (MLMs) play a crucial role in this context. This study aims to predict daily underground water levels (UWL) in Azarshar plain, Tabriz, Iran, using three MLMs (SVM, GEP, MLP). Covering 126 annual datasets for 34 wells from 2018 to 2021, various combinations were tested with different UWLs and lag-times. Performance evaluation metrics including RMSE, MAE, R2, and DDR were employed. Results show satisfactory accuracy for all three MLMs, with SVM, GEP, and MLP being more accurate in 53%, 26%, and 20% of cases respectively among the 34 wells. The input configuration with a lag-time of two days (M2) emerged as the most optimal, yielding the most accurate simulations. Average values of RMSE, MAE, R2, and DDR for M2 during the testing period were calculated as 0.2457, 0.2077, 0.9482, and 31.53 respectively. In conclusion, these MLMs offer viable alternatives to numerical models for managing and predicting UWL, facilitating better water resource management in areas prone to drought and water scarcity.
Mots clés
Simulation ; Aquifer ; Prediction ; Performance Assessment
