Research paper
Groundwater level simulation and forecasting using interior search algorithm-least square support vector regression (ISA-LSSVR)

https://doi.org/10.1016/j.gsd.2020.100447Get rights and content

Highlights

  • The capabilities of four data-driven methods to simulate and forecast groundwater level were compared.

  • Optimizing SVR parameters with Interior Search meta-heuristic optimization algorithm improves the accuracy.

  • Data-driven methods employed to understand the importance of external variables affecting groundwater flow.

Abstract

Least square support vector regression (LSSVR) is a powerful data-driven method for simulation and forecasting, with two parameters to tune. In this study, these parameters were automatically tuned using the interior search algorithm (ISA) and genetic algorithm (GA). The main purpose is in situ simulation and forecast of monthly groundwater level in Karaj plain, Iran, using historical groundwater level, precipitation, and evaporation data. The results of the interior search algorithm-least support vector regression (ISA-LSSVR) and genetic algorithm-least support vector regression (GA-LSSVR) compared with genetic programming (GP) and adaptive neural fuzzy inference system (ANFIS). Based on average Nash-Sutcliffe criterion, the results revealed that the ISA-LSSVR improves the simulation and forecasting accuracy compared to other methods. Also, the results of the different model structure selection indicate that including precipitation and evaporation does not necessarily improve simulation and forecasting accuracy, but it would increase uncertainty. This increase suggests that groundwater level in the case study is affected by groundwater flow, recharge from leaky urban water infrastructure, and reduced recharge from precipitation due to impervious surfaces in urban areas rather than being solely governed by precipitation and evaporation. Finally, a sensitivity analysis was performed to assess the impacts of optimization algorithm parameters on the simulation and forecasting accuracy. The results indicate high and low sensitivity associated with GA and ISA, respectively. In conclusion, ISA-LSSVR was suggested as the best model due to computational efficiency, low sensitivity to its parameters, and high accuracy compared to other methods.

Introduction

With the increase of computational capability of computers, the application of data-driven methods in groundwater level simulation and forecasting has been increased. Artificial neural network (ANN), adaptive neural fuzzy inference system (ANFIS), genetic programming (GP) and support vector regression (SVR) can be mentioned as some of these methods. Several studies reported the use of ANN (Coppola et al., 2003; Maier and Dandy, 2000; Wunsch et al., 2018), ANFIS (Gong et al., 2016; Moosavi et al., 2013; Shiri and Kişi, 2011), and GP (Fallah-Mehdipour et al., 2013, 2014; Shiri et al., 2013) for groundwater level simulation and forecasting. Although these methods provide some degree of flexibility and accuracy in groundwater modelling, some demerits have been reported. High uncertainty associated with ANN was reported by Yoon et al. (2011) for groundwater level forecasting in a coastal aquifer. Over-fitting of ANFIS was reported by Fallah-Mehdipour et al. (2013) in Karaj plain, and high computational demand to find solutions was stated as a major disadvantage of GP by Meier et al. (2013).

The SVR method, first presented by Vapnik (1995), is a relatively new data-driven method in hydrology. The SVR has been applied to forecast groundwater level. Liu et al. (2009) used a least square support vector regression (LSSVR) based on chaos dynamic and a radial basis function (RBF) as the kernel to forecast groundwater level. Behzad et al. (2010) modeled groundwater level under different pumping and weather conditions, by ANN and SVR methods. They used the RBF kernel and identified SVR parameters by grid search (GS-SVR). The results indicated that the SVR, especially when the data is insufficient, is more accurate and generalized in comparison to ANN. Yoon et al. (2011) compared ANN and SVR methods in groundwater level simulation in a coastal aquifer. The results illustrated that the average error in SVR method was lower than those in ANN method. Also, uncertainty analysis showed lower uncertainty for SVR method. Shiri et al. (2013) made a comparison between the ANN, ANFIS, GP, SVR and autoregressive moving average (ARMA) methods for groundwater level forecasting. The results showed that the GP has the highest capability to forecast groundwater level up to 7 days beyond the recorded data. They chose RBF kernel and the SVR parameters were tuned by trial and error. None of these studies dealt with the urban aquifers. The recharge of unconfined aquifers in urban areas is complex due to the impact of impervious surface as a result of urban development (Renouf et al., 2019) and urbanization in general (Salvadore et al., 2015), for example leakage from urban water infrastructure. In this case, the link between groundwater recharge and meteorological variables (e.g. precipitation, temperature, etc.) might not be straightforward and linear. Data-driven methods has potential to be used in this context.

Previous studies showed that the SVR results' accuracy largely depends on tuning its parameters. The SVR theory doesn't provide any specific methodology for parameter selection (Su et al., 2014). Trial and error have been performed in previous studies (Behzad et al., 2010; Bhagwat and Maity, 2013; Yoon et al., 2011). Trial and error approach is a time-consuming process and greatly affects the SVR results' accuracy. Achieving a high accuracy with trial and error approach is computationally expensive because the model needs to be evaluated at various points within the grid for each parameter. Meta-heuristic optimization algorithms can be used to overcome this limitation (Raghavendra and Deka, 2015). The Meta-heuristic optimization algorithms were successfully applied in many water resources problems (Bozorg-Haddad et al., 2015; Fayaz et al., 2020; Hosseini-Moghari et al., 2015, 2017) and their strength and weaknesses have been reviewed (Maier et al., 2014; Moravej, 2017, 2018; Sörensen, 2015).

Some meta-heuristic optimization algorithms were employed to find SVR parameters. Genetic algorithm (GA) was employed to choose the SVR parameters (GA-SVR) to forecast monthly reservoir storage (Su et al., 2014), monthly streamflow (Kalteh, 2015), water quality parameters in rivers (Bozorg-Haddad et al., 2017; Soleimani et al., 2016), daily reference evapotranspiration (Yin et al., 2017), and water temperature (Quan et al., 2020). Bat algorithm, Particle Swarm Optimization (PSO), and Artificial Bee Colony were coupled with SVR for streamflow forecasting (Xing et al., 2016) river-stage forecasting (Seo et al., 2016). The results indicate that PSO-SVR and ABC-SVR provided better results compared to GS-SVR and GA-SVR. Similar studies in groundwater level simulation and forecasting is missing. In addition, all meta-heuristics algorithms used for tuning the SVR parameters' suffer from having two or more parameters to tune themselves. The sensitivity of the meta-heuristic algorithms was reported in previous studies (Moravej, 2017, 2018). Interior search algorithm (ISA) is a new evolutionary algorithm which overcomes this limitation.

ISA is proposed by (Gandomi, 2014) showing its superiority over other optimization algorithms in a wide range of problems. Moravej and Hosseini-Moghari (2016) tuned the ISA parameter automatically using a linear equation as a function of optimization iteration. Therefore, the ISA algorithm potentially provides more reliability compared to other meta-heuristic algorithms. They employed the ISA to solve three reservoir operation problems (i.e. Karun 4, four-reservoir and ten-reservoir operation problems). They compared the results of the ISA with other meta-heuristic algorithms such as honey bee mating optimization (HBMO), water cycle algorithm (WCA), biogeography-based optimization (BBO), imperial competition algorithm (ICA), cuckoo optimization algorithm (COA), ant colony optimization (ACO) and bat algorithm from previous studies. The results indicate that the ISA outperform those algorithms in accuracy, reliability, and fast convergence (i.e., efficiency).

Groundwater is a vital source for municipal, industrial and agricultural usages in Karaj city, Iran; supporting socio-economic activities of more than 2 million people. Therefore, developing a reliable tool for simulating and forecasting water level is of interest in this region. In the current study, the ISA was employed to automatically tune the LSSVR (ISA-LSSVR) parameters for groundwater level simulation and forecasting in Karaj Plain. The aim is to evaluate the capability of the proposed method in groundwater level simulation and forecast. To tackle this, 12 different model structures were considered and the result of ISA-LSSVR in two separate categories of simulation and forecast were compared with GA coupled with LSSVR (GA-LSSVR), GP, and ANFIS.

Section snippets

The LSSVR

The LSSVR method was developed by Suykens et al. (2002). It is a modified version of the original method which was introduced by Vapnik (1995). The LSSVR provides a computational advantage over the regular SVR via converting a quadratic programming problem into a set of linear equations (Kalteh, 2015). The main advantage of this conversion is mentioned as decrease in the computational complexity and cost (Adnan et al., 2017; Raghavendra; Deka, 2015). Therefore, LSSVR is used in this study over

ISA-LSSVR and GA-LSSVR results

The values of the LSSSVR parameters were obtained using the ISA and the GA. The σ values range from 9 (well 1 model 1) to 925 (well 3 model 2) and the values of γ range from 20 (well 1 model 1) to 823 (well 3 model 4) optimized by the ISA. The measures of accuracy of training datasets are presented in Table 2. Also, results obtained by ANFIS and GP, which are presented by Fallah-Mehdipour et al. (2013), are mentioned in the table for comparison purposes. As Fallah-Mehdipour et al. (2013)

Conclusion

In the current study, the ISA was introduced to optimize the LSSVR parameters forming a new method for groundwater level simulation and forecast. The result is compared with other data-driven methods (i.e. GA-LSSVR, GP, and ANFIS). The comparison showed that the ISA-LSSVR improves the simulation results of GA-LSSVR by 71.8, 27.2 and 12.4 percent in terms of NS, RMSE and R2. Similar improvements for forecast were reported to be 62.7, 24.8 and 12.3 percent for NS, RMSE and R2. The results also

Declaration of competing interest

The authors declare no conflict of interest.

Acknowlegement

No financial support is needed to be acknowledged.

References (61)

  • E. Salvadore et al.

    Hydrological modelling of urbanized catchments: a review and future directions

    J. Hydrol.

    (2015)
  • J. Shiri et al.

    Comparison of genetic programming with neuro-fuzzy systems for predicting short-term water table depth fluctuations

    Comput. Geosci.

    (2011)
  • J. Shiri et al.

    Predicting groundwater level fluctuations with meteorological effect implications—a comparative study among soft computing techniques

    Comput. Geosci.

    (2013)
  • J.A. Suykens et al.

    Weighted least squares support vector machines: robustness and sparse approximation

    Neurocomputing

    (2002)
  • I. Tubau et al.

    Quantification of groundwater recharge in urban environments

    Sci. Total Environ.

    (2017)
  • V. Wagh et al.

    Neural network modelling for nitrate concentration in groundwater of Kadava River basin, Nashik, Maharashtra, India

    Groundwater for Sustainable Development

    (2018)
  • A. Wunsch et al.

    Forecasting groundwater levels using nonlinear autoregressive networks with exogenous input (NARX)

    J. Hydrol.

    (2018)
  • H. Yoon et al.

    A comparative study of artificial neural networks and support vector machines for predicting groundwater levels in a coastal aquifer

    J. Hydrol.

    (2011)
  • R.M. Adnan et al.

    Improving accuracy of river flow forecasting using LSSVR with gravitational search algorithm

    Advances in Meteorology

    (2017)
  • A. Alipour et al.

    Spatio-temporal analysis of groundwater level in an arid area

    Int. J. Water

    (2018)
  • S. Araghinejad

    Data-driven Modeling: Using MATLAB® in Water Resources and Environmental Engineering

    (2013)
  • S. Araghinejad et al.

    Development of a hybrid data driven model for hydrological estimation

    Water Resour. Manag.

    (2018)
  • M. Behzad et al.

    Comparative study of SVMs and ANNs in aquifer water level prediction

    J. Comput. Civ. Eng.

    (2010)
  • K. Beven et al.

    A guide to good practice in modeling semantics for authors and referees

    Water Resour. Res.

    (2013)
  • P.P. Bhagwat et al.

    Hydroclimatic streamflow prediction using least square-support vector regression

    ISH Journal of Hydraulic Engineering

    (2013)
  • O. Bozorg-Haddad et al.

    Application of the water cycle algorithm to the optimal operation of reservoir systems

    J. Irrigat. Drain. Eng.

    (2015)
  • O. Bozorg-Haddad et al.

    Modeling water-quality parameters using genetic algorithm–least squares support vector regression and genetic programming

    J. Environ. Eng.

    (2017)
  • N. Chitsaz et al.

    Pre-processing of data-driven river flow forecasting models by singular value decomposition (SVD) technique

    Hydrol. Sci. J.

    (2016)
  • M. Chitsazan et al.

    Analysis of temporal and spatial variations in groundwater nitrate and development of its pollution plume: a case study in Karaj aquifer

    Environmental Earth Sciences

    (2017)
  • E. Coppola et al.

    Application of artificial neural networks to complex groundwater management problems

    Nat. Resour. Res.

    (2003)
  • Cited by (36)

    • FCDT-IWBOA-LSSVR: An innovative hybrid machine learning approach for efficient prediction of short-to-mid-term photovoltaic generation

      2023, Journal of Cleaner Production
      Citation Excerpt :

      In another study, a hybrid prediction model combining wavelet transform, particle swarm optimization, and support vector machines was used for day-ahead generation forecasting of real microgrid PV systems, which had a daily average MAPE of 4.2% and an NMAE of 0.4% (Eseye et al., 2018b). It can be known that the prediction accuracy of the hybrid method is usually higher than that of the single method (Malvoni et al., 2017b), (Moravej et al., 2020), (Wang et al., 2022), (Tran et al., 2020), and that the regression results of the LSSVR model are closely related to the parameters of the model itself. Therefore, these researchers optimize the parameter values of the LSSVR model by means of optimization algorithms using hybrid methods to improve the prediction accuracy of the model.

    • An effective tremor-filtering model in teleoperation: Three-domain Wavelet Least Square Support Vector Machine[Formula presented]

      2022, Applied Soft Computing
      Citation Excerpt :

      Moreover, the generalization performance of the LSSVM is related to the kernel function parameter selection. Moravej et al. [36] used the genetic algorithm and interior search algorithm to optimize the selection of the kernel function parameters, and the result showed that the predictive precision of the optimized model is higher than that of the LSSVM. Some researchers proposed chaotic particle swarm optimization [37], improved bird swarm algorithm [38], and grid search [39] to prove the effectiveness of the kernel parameter optimization.

    • Combining autoregressive integrated moving average with Long Short-Term Memory neural network and optimisation algorithms for predicting ground water level

      2022, Journal of Cleaner Production
      Citation Excerpt :

      Therefore, GWL prediction is an important topic in watershed management (Emamgholizadeh et al., 2014; Yang et al., 2016; Guzman et al., 2017; Yousefi et al., 2021). Based on accurate forecasting of the groundwater level, decision-makers can optimally allocate the groundwater resources for irrigation and industrial demands without exacerbating the water crisis (Moravej et al., 2020; Choubin and Malekian, 2017; Banadkooki et al., 2020; Alsubih et al., 2021). Data-driven models are highly capable of extracting the complex relationships between inputs and outputs.

    • The effect of climate change on groundwater level and its prediction using modern meta-heuristic model

      2022, Groundwater for Sustainable Development
      Citation Excerpt :

      Many researchers have confirmed the potential utility of AI techniques for groundwater level simulation (Coppola et al., 2005; Dash et al., 2010; Adamowski et al., 2011; Kisi et al., 2012; Kakahaji et al., 2013; Moosavi et al., 2013; Sahoo et al., 2013; Shiri et al., 2013a,b; Chang et al., 2016; Suryanarayana et al., 2014a,b; Nourani et al., 2015; Supreetha et al., 2015; Soleymani et al., 2016; Vaheddoost et al., 2016; Wen et al., 2016; Yoon et al., 2016; Ebrahimi et al., 2017; Ghorbani et al., 2017b, 2019; Kaya et al., 2018; Guzman et al., 2019; Jeihouni et al., 2019a, 2019b; Malekzadeh et al., 2019b; Miraki et al., 2019; Mirarabi et al., 2019). Moravej et al. (2020) used hybrid algorithms in conjunction with SVR model to simulate the groundwater level of Karaj plain located in Iran. They employed such parameters as groundwater level, precipitation, and evaporation using Internal Search Algorithm-Minimum Support for Vector Regression (ISA-LSSVR) and Genetic Algorithm-Minimal Vector Regression (GA-LSSVR) in comparison with Genetic Programming (GP) and Adaptive Neural Fuzzy Inference System (ANFIS).

    View all citing articles on Scopus
    View full text