Nonlinear complex behavior of pore-water pressure responses to rainfall was modelled using support vector regression (SVR). Pore-water pressure can rise to disturbing levels that may result in slope failure during or after rainfall. Traditionally, monitoring slope pore-water pressure responses to rainfall is tedious and expensive, in that the slope must be instrumented with necessary monitors. Data on rainfall and corresponding responses of pore-water pressure were collected from such a monitoring program at a slope site in Malaysia and used to develop SVR models to predict pore-water pressure fluctuations. Three models, based on their different input configurations, were developed. SVR optimum meta-parameters were obtained using k-fold cross validation and a grid search. Model type 3 was adjudged the best among the models and was used to predict three other points on the slope. For each point, lag intervals of 30 min, 1 h and 2 h were used to make the predictions. The SVR model predictions were compared with predictions made by an artificial neural network model; overall, the SVR model showed slightly better results. Uncertainty quantification analysis was also performed for further model assessment. The uncertainty components were found to be low and tolerable, with d-factor of 0.14 and 74 % of observed data falling within the 95 % confidence bound. The study demonstrated that the SVR model is effective in providing an accurate and quick means of obtaining pore-water pressure response, which may be vital in systems where response information is urgently needed.

Irrigation is essential for achieving food security to the burgeoning global population but unplanned and injudicious expansion of irrigated areas causes waterlogging and salinization problems. Under this backdrop, groundwater resources management is a critical issue for fulfilling the increasing water demand for agricultural, industrial, and domestic uses. Various simulation and optimization approaches were used to solve the groundwater management problems. This paper presents a review of the individual and combined applications of simulation and optimization modeling for the management of groundwater-resource problems associated with irrigated agriculture. The study revealed that the combined use of simulation-optimization modeling is very suitable for achieving an optimal solution for groundwater-resource problems, even with a large number of variables. Independent model tools were used to solve the problems of uncertainty analysis and parameter estimation in groundwater modelling studies. Artificial neural networks were used to minimize the problem of computational complexity. The incorporation of socioeconomic aspects into the groundwater management modeling would be an important development in future studies.

A fundamental aspect in groundwater heat pump (GWHP) plant design is the correct evaluation of the thermally affected zone that develops around the injection well. This is particularly important to avoid interference with previously existing groundwater uses (wells) and underground structures. Temperature anomalies are detected through numerical methods. Computational fluid dynamic (CFD) models are widely used in this field because they offer the opportunity to calculate the time evolution of the thermal plume produced by a heat pump. The use of neural networks is proposed to determine the time evolution of the groundwater temperature downstream of an installation as a function of the possible utilization profiles of the heat pump. The main advantage of neural network modeling is the possibility of evaluating a large number of scenarios in a very short time, which is very useful for the preliminary analysis of future multiple installations. The neural network is trained using the results from a CFD model (FEFLOW) applied to the installation at Politecnico di Torino (Italy) under several operating conditions. The final results appeared to be reliable and the temperature anomalies around the injection well appeared to be well predicted.

Groundwater pollution source identification (GPSI) provides information about the temporal and spatial distribution of pollution sources and helps decision makers design pollution remediation plans to protect the groundwater environment. The Bayesian approach based on the Markov Chain Monte Carlo (MCMC) approach provides an efficient framework for GPSI. However, MCMC sampling entails multiple model calls to converge to the posterior probability distribution of unknown pollution source parameters and entails a massive computational load if the simulation model is directly called. This study aimed to develop an innovative framework in which an improved MCMC approach was linked to a surrogate model. Sensitivity analysis was incorporated into the MH-MCMC approach, named SAMH-MCMC (sensitivity analysis based Metropolis Hastings-Markov Chain Monte Carlo), to speed up the convergence of the posterior distribution in a novel way to control the search step size. Three computationally inexpensive surrogate models for the simulation model were proposed: support vector regression, Kriging (KRG), and multilayer perceptron, and the most accurate model was chosen. The feasibility and advantages of the developed framework were evaluated and validated through two hypothetical numerical cases with homogenous and heterogeneous media. The proposed approach has strong convergence robustness as it considers the sensitivities of the unknown parameters that characterise groundwater pollution sources and can achieve high identification accuracy. Furthermore, the KRG surrogate model has a higher accuracy than other surrogate models, owing to its linear unbiased estimation characteristic. Overall, the framework developed in this study is a promising solution for identifying groundwater pollution source parameters.

Jinci Spring in Shanxi, north China, is a major local water source. It dried up in April 1994 due to groundwater overexploitation. The groundwater system is complex, involving many nonlinear and uncertain factors. Artificial neural network (ANN) models are statistical techniques to study parameter nonlinear relationships of groundwater systems. However, ANN models offer little explanatory insight into the mechanisms of prediction models. Sensitivity analysis can overcome this shortcoming. In this study, a back-propagation neural network model was built based on the relationship between groundwater level and its sensitivity factors in Jinci Spring Basin; these sensitivity factors included precipitation, river seepage, mining drainage, groundwater withdrawals and lateral discharge to the associated Quaternary aquifer. All the sensitivity factors were analyzed with Garson’s algorithm based on the connection weights of the neural network model. The concept of “sensitivity range” was proposed to describe the value range of the input variables to which the output variables are most sensitive. The sensitivity ranges were analyzed by a local sensitivity approach. The results showed that coal mining drainage is the most sensitive anthropogenic factor, having a large effect on groundwater level of the Jinci Spring Basin.

Change in groundwater level is predicted for a special site where transient natural factors affecting the groundwater level are mixed with very irregular anthropogenic influences. When there is not enough hydrogeological information about the area to be analyzed, an artificial neural network (ANN) is a powerful tool for groundwater level forecasting in highly irregular and uncertain groundwater systems. In this study, groundwater levels were predicted by using ANN models with input variables composed of one natural factor and two anthropogenic factors in Yangpyeong riverside area, South Korea. Complex and irregular change of the groundwater level was monitored due to the operation of a groundwater heat pump system and winter intensive pumping for water curtain cultivation (by which greenhouses are warmed). The prediction results showed good performance with root mean square errors of 3–6 cm when the average groundwater level is about 25.59 m, the correlation coefficient is >0.9 and the Nash–Sutcliffe efficiency is >0.75, indicating that the ANN models are well suited for assessing complex groundwater systems. Along with the prediction, an extraction method was devised to calculate contributions and relative impacts of the input variables in the time-series-based ANN models. As a result, it was proved that the river level dominantly affects the groundwater level fluctuation, and the contributions of each influencing factor were obtained reliably according to spatial distribution and temporal variance. This makes the scheme effective for managing and using groundwater resources with consideration of every crucial influencing factor of the groundwater level fluctuation.

An artificial neural network model (ANN) and a geographic information system (GIS) are applied to the mapping of regional groundwater productivity potential (GPP) for the area around Pohang City, Republic of Korea. The model is based on the relationship between groundwater productivity data, including specific capacity (SPC) and its related hydrogeological factors. The related factors, including topography, lineaments, geology, and forest and soil data, are collected and input into a spatial database. In addition, SPC data are collected from 44 well locations. The SPC data are randomly divided into a training set, to analyse the GPP using the ANN, and a test set, to validate the predicted potential map. Each factor’s relative importance and weight are determined by the back-propagation training algorithms and applied to the input factor. The GPP value is then calculated using the weights, and GPP maps are created. The map is validated using area under the curve analysis with the SPC data that have not been used for training the model. The validation shows prediction accuracies between 73.54 and 80.09 %. Such information and the maps generated from it could serve as a scientific basis for groundwater management and exploration.

Several groundwater-level forecasting studies have shown that data-driven models are simpler, faster to develop, and provide more accurate and precise results than physical or numerical-based models. Five data-driven models were examined for the forecasting of groundwater levels as a result of recharge via tailings from an abandoned mine in Quebec, Canada, for lead times of 1 day, 1 week and 1 month. The five models are: a multiple linear regression (MLR); an artificial neural network (ANN); two models that are based on de-noising the model predictors using the wavelet-transform (W-MLR, W-ANN); and a W-ensemble ANN (W-ENN) model. The tailing recharge, total precipitation, and mean air temperature were used as predictors. The ANN models performed better than the MLR models, and both MLR and ANN models performed significantly better after de-noising the predictors using wavelet-transforms. Overall, the W-ENN model performed best for each of the three lead times. These results highlight the ability of wavelet-transforms to decompose non-stationary data into discrete wavelet-components, highlighting cyclic patterns and trends in the time-series at varying temporal scales, rendering the data readily usable in forecasting. The good performance of the W-ENN model highlights the usefulness of ensemble modeling, which ensures model robustness along with improved reliability by reducing variance.

Artificial intelligence (AI) techniques have increasingly become efficient alternative modeling tools in the water resources field, particularly when the modeled process is influenced by complex and interrelated variables. In this study, two AI techniques—artificial neural networks (ANNs) and support vector machine (SVM)—were employed to achieve deeper understanding of the salinization process (represented by chloride concentration) in complex coastal aquifers influenced by various salinity sources. Both models were trained using 11 years of groundwater quality data from 22 municipal wells in Khan Younis Governorate, Gaza, Palestine. Both techniques showed satisfactory prediction performance, where the mean absolute percentage error (MAPE) and correlation coefficient (R) for the test data set were, respectively, about 4.5 and 99.8% for the ANNs model, and 4.6 and 99.7% for SVM model. The performances of the developed models were further noticeably improved through preprocessing the wells data set using a k-means clustering method, then conducting AI techniques separately for each cluster. The developed models with clustered data were associated with higher performance, easiness and simplicity. They can be employed as an analytical tool to investigate the influence of input variables on coastal aquifer salinity, which is of great importance for understanding salinization processes, leading to more effective water-resources-related planning and decision making.