Groundwater Level Modeling Using Multiobjective Optimization with Hybrid Artificial Intelligence Methods

Abstract

Estimating groundwater level (GWL) uctuations is essential for integrated water resource management in arid and semi-arid regions. This study promotes the multi-layer perceptron (MLP) learning process using hybrid evolutionary algorithms. This hybrid metaheuristic algorithm was applied to overcome MLP di culties in the learning process, including its low conversions and local minimum. Also, the hybrid model bene ts from the advantages of two objective function procedures in nding MLP parameters that result in a robust model regardless of over and under-estimating problems. These algorithms include none dominated sorting genetic algorithm (NSGA II) and multi-objective particle swarm optimisation (MOPSO) in different patterns, including MLP-NSGA-II, MLP-MOPSO, MLP-MOPSO-NSGA-II, and MLP-2NSGA-II-MOPSO. Temperature, precipitation and GWL datasets were used in various combinations and delays as model input candidates. Finally, the best model inputs were selected using the correlation coe cient (R 2 ). Input parameters include temperature and precipitation delays of 3, 6, and 9 months and GWL delays of 1 to 12 months. In the next step, the performance of the different combinations of MLP and hybrid evolutionary algorithms was evaluated using The root mean square error (RMSE), correlation coe cient (R), and mean absolute error (MAE) indices. The outcomes of these evaluations revealed that the MLP-2NSGA-II-MOPSO model, with an RMSE=0.073, R=0.98, and MAE=0.059, outperforms other models in estimating GWL uctuations. The selected model bene ts from the advantages of both MOPSO and NSGA-II regarding accuracy and speed. The results also indicated the superiority of multi-objective optimization algorithms in promoting MLP performance. quanti cation for better management along with environmental protection (Li et al., 2015;Noori et al., 2021;Torabi Haghighi et al., 2020). Furthermore, precise GWL modeling can optimize the monitoring of GWL, which has time and resource limitations (Kazemi et al., 2021) and needs signi cant in-situ measurements (Lee et al., 2019) and interpretation of data quality.The GWL uctuations can be due to (i) climate change and variability and (ii) anthropogenic activity, i.e., GW exploitation (Malekinezhad & Banadkooki, 2018). Therefore, GWL modelling needs deep analysis, extensive and reliable data sources, and an understanding of interactions between different environmental parameters (Liang & Zhang, 2015). In recent decades many attempts have been made to develop accurate Page 3/31 and robust models to predict GWL uctuations properly. These attempts resulted in a variety of GWL models that can be categorized into four groups, (i) the univariate time series models (Khorasani et al.,