Research paper
Modeling response of runoff and evapotranspiration for predicting water table depth in arid region using dynamic recurrent neural network

https://doi.org/10.1016/j.gsd.2018.01.007Get rights and content

Highlights

  • We focus on water sustainability in arid region by developing the models.

  • The most relevant techniques are used for computing water table depth prediction.

  • The advances gained from a modeling point of view are considered and discussed.

  • The models are developed using RNN technique.

  • The study is different from previous studies with the inclusion of three transfer functions.

Abstract

The investigation of the dynamic response of aquifer to rainfall with variation of the atmospheric conditions is a key issue for groundwater resource management. A data-driven dynamic recurrent neural network approach, based on multi-objective optimization is used here to forecast groundwater levels as a function of rainfall, temperature, humidity, runoff and evapotranspiration data. Recurrent Neural Network (RNN) with variable transfer functions like tangential sigmoidal, purelin and logarithmic sigmoidal are used to compute the relative performance of the model. The results demonstrate the transition behavior due to changing evapotranspiration and runoff. Statistical and graphical indicators are used to compare the results. The statistical indicators used in this work are Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R2). From the results, evapotranspiration loss and run off are the influencing parameters which affect the depth to water table in the ground water reservoir. It is observed that calculated losses due to evapotranspiration are comparatively less during high precipitation in the Rengali province. Results of R2 suggest that inclusion of evapotranspiration and runoff in different scenario improves the model efficiency in predicting water table depth.

Graphical abstract

Normalized actual versus predicted water table depth for scenario 1–5 in testing phase.

fx1
  1. Download : Download high-res image (86KB)
  2. Download : Download full-size image

Introduction

Aquifers are modeled according to their conceptualization and calibration of model parameters (Anderson et al., 2015). The procedure is based on the physics of the system by calibrating the parameters of the model (Ljung, 1999). The behavior of natural systems is usually characterized by different variables, exaggerated by inputs, which interact to produce observable variations of output (Sun, 2013). The study of the dynamics of evaluation of groundwater level is a relevant topic for groundwater management, especially in arid and semiarid regions, where rainfall amount and distribution vary consistently across the years (Rushton, 2003). The difficulty of modeling is the dynamics of hydro geological systems with their inherent complexity, and nonlinear processes controlling their responses to external inputs (Kresic and Mikszewski, 2012). Water levels are predicted through physically based or data-driven models. Developing a physically based model for assessing water level fluctuations require proper characterization and synthesis of the aquifer parameters, to describe spatial variability of the subsurface hydrogeology and soil characteristics (Taormina et al., 2012, Dadaser-Celik and Cengiz, 2013). Ghose et al. (2010) has predicted water table depth using BPNN and RBFN techniques.

Empirical approaches have been widely used for water table depth modeling (Box and Jenkins, 1976, Tankersley and Graham, 1994, Hipel and McLeod, 1994, Van Geer and Zuur, 1997, Knotters and Van Walsum, 1997). The empirical time series models are described for a longer time series of water table depth. Artificial Neural Networks (ANNs) have been successfully used for water level modeling (Coulibaly et al., 1999, Coulibaly et al., 2001). In hydrological modeling neural networks is found to be a suitable tool (Hsu et al., 1995, Tokar and Marcus, 2000, Coulibaly et al., 2000a). When there is insufficient knowledge of the hydro geologic characteristics of the system, and accuracy in prediction is more important for understanding the physical processes, then black-box-type models are viable options (Nourani and Mano, 2007). Artificial neural networks (ANNs) are among the black-box-type models that can be applied to capture nonlinear behavior of complex systems. ANN models have been used in rainfall–runoff processes (Kumar et al., 2005, RezaeianZadeh et al., 2010). From literature review of previous studies we have focused to predict groundwater levels on monthly basis (Nayak et al., 2006; Jalalkamali et al., 2011; Shirmohammadi et al., 2013).

In fact, aquifer dynamics is related to (1) rainfall intensity (2) infiltration and percolation processes, and (3) aquifer structural characteristics. Thus, estimation along with impact of evapotranspiration and runoff is essential for predicting groundwater levels to rainfall potential in the aquifer. In this context, the choice of a model is important by data availability, mechanism of calibration and validation. The above literature studies point out that ANN model is an acceptable tool for performing the water table prediction, especially in areas where the scarcity of rainfall occurs or where the real-time simulation is needed.

In the present study evapotranspiration and surface runoff is calculated by empirical equation as suggested by Khosla's formulae to Indian conditions on monthly climate data. Then the parameters evapotranspiration and surface runoff are considered for developing models including precipitation, temperature as input and water table depth as output. The purpose of this paper is to identify the complex dynamics of water table depth fluctuation due to the impact of evapotranspiration and surface runoff. The rest of the paper is organized as follows; Section 2 provides a brief description of the study area and the historical data. 3 Artificial neural networks, 4 Results and discussion introduce the architecture and learning algorithm of RNN. Section 5 describes the result and discussion. Section 6 describes the concluding remarks of the study.

Section snippets

Study area and data

Rengali block is located in the western region of Odisha having area of 713sq. km. Its geographical coordinates are 21° 38' 0" North, 84° 3' 0" East. Major soil type of Rengali is red sandy and red loamy soils. Major water bearing formations are granite gneiss, quartzites, sandstone and arenaceous shales. The region is a part of Indian peninsula and is characterized by scanty rainfall. The area is categorized by rippling plains with inaccessible high peaks. The climate is rooted with dry

Artificial neural networks

Simple neuron model receives a vectorial input with components xi. These input components are then multiplied by the appropriate weights wi and accumulated as wixi. This term is called weighted sum. Then the nonlinear mapping function defines the scalar output y = wixi. After this transition, a neuron model and the weights are to be adjusted. Data processing of a neuron is presented in Fig. 2. Data input includes precipitation, maximum temperature, minimum temperature, humidity, runoff, and

Evaluation of evapotranspiration loss and Runoff

Monthly surface runoff is calculated using Khosla's formulae in Indian condition. Monthly average evapotranspiration losses are evaluated using Khosla's equation, an empirical one because of non-availability of data. The relationship for monthly runoff is, Rm = P m – L m. Where Lm = 0.48 Tm for Tm > 4.5 °C. Pm = Monthly precipitation in mm. Lm= Monthly evapotranspiration losses in mm. Tm = Mean monthly temperature of station in °C. Then using Khosla's formulae on Indian infirmity the following

Conclusion

In this study, three different transfer functions of dynamic recurrent neural networks are used for monthly average water level prediction in arid region with five scenarios. A correlation analysis is used to find the associations among the inputs and the target water table depth. One of the advantages of dynamic RNN is that it converges first to produce the target as compared to other neural network techniques. This is related to its capacity of reproducing the physical phenomenon of the

References (24)

  • K.L. Hsu et al.

    Artificial neural network modeling of the rainfall-rainoff process

    Water Resour. Res.

    (1995)
  • A. Jalalkamali et al.

    Monthly groundwater level prediction using ANN and neuro-fuzzy models: a case study on Kerman plain, Iran

    J. Hydroinf.

    (2011)
  • Cited by (34)

    • Rainfall-runoff modeling using long short-term memory based step-sequence framework

      2022, Journal of Hydrology
      Citation Excerpt :

      To account for the sequential information of the time series and improve the performance, the recurrent neural network was especially proposed, which can remember previous information and capture temporal dynamics. Thus, compared with the traditional artificial neural networks, the recurrent neural networks are more suitable for modeling complex hydrological and meteorological time series (Pan et al., 2007; BBhattacharjee and Tollnerattacharjee and Tollner, 2016; Ghose et al., 2018). The traditional recurrent neural networks, however, have the exploding and vanishing gradient problems when they are trained with long-term lags.

    • Rainfall-runoff modeling using LSTM-based multi-state-vector sequence-to-sequence model

      2021, Journal of Hydrology
      Citation Excerpt :

      These two sorts have their own characteristics, where physically-based approaches can provide insights into physical processes but are subject to many simplifying assumptions, and data-driven ones based on such as recurrent neural network (RNN) and long short-term memory (LSTM) can mimic highly nonlinear and complex systems well from the data. The traditional RNNs account for the sequential information of the time series and thus can remember previous information and capture temporal dynamics (Pan et al., 2007; Bhattacharjee and Tollner, 2016; Ghose et al., 2018). However, they have the exploding and vanishing gradient problems when we train them with long-term lags.

    View all citing articles on Scopus
    View full text