Forecasting electrical conductivity of a coastal karstic aquifer with artificial intelligence methods

Groundwater resources are inevitably considered as the primary source of high-quality water for the Mediterranean region. In critical cases, groundwater is essential to complement inadequate, uncertain, expensive, or even lacking surface-water sources. Especially in coastal areas of the Mediterranean, which show increasing development and population growth, karst aquifers represent vital freshwater sources. Karst aquifers are rather complex water systems, and therefore, though to manage, predict, and protect. Groundwater modeling has proved to be a very effective tool for groundwater management. Physically-based modeling is usually applicable to porous aquifers; numerical modeling application to karst aquifers is very challenging because of their complexity, which combines discontinuity, conduit, and porous medium domains. The need to forecast groundwater quantity and quality in karst aquifer systems is high, with groundwater salinity being very critical. Artificial Intelligence (AI) algorithms have been proved to be an effective alternative in simulating groundwater quality and quantity variables. This study aims to develop and test the performance of 6 AI algorithms to forecast groundwater electrical conductivity (EC) in the highly complex, coastal karst aquifer system of Salento (Puglia, Southern Italy). The AI algorithms applied were: 1) Multilayer Perceptron (MLP), 2) Long short-term memory (LSTM), 3) Bidirectional LSTM (BiLSTM), 4) Convolutional Neural Network (CNN), 5) Recurrent Neural Networks (RNN), and 6) Support Vector Machine (SVR). Except for SVR, which is considered a machine learning (ML) algorithm, all the other approaches are deep learning (DL) neural network architectures. Models’ development was based on 3-year groundwater EC daily data from 7 sensors. Other variables used for EC modeling were groundwater level and temperature, precipitation, and air temperature. The above variables were combined in 11 input variable experiments. In addition, various realizations of training times windows were developed under five scenarios. The total number of trained EC models was 2184. The results show AI models can efficiently provide a 30-day groundwater EC forecast for a wide range of EC values varying from slightly saline (0.7-2 mS/cm) to very saline (25-45 mS/cm). BiLSTM proved to be the most effective algorithm, while the least but still effective algorithm was SVR, thus showing the superior performance of DL algorithms compared to legacy ML approaches. Experimental results showed that increasing the number of input variables did not improve the performance of models. In contrast, including 2-time windows for training (one short-term and one long-term) increased it.