Elsevier

Journal of Hydrology

Volume 579, December 2019, 124226
Journal of Hydrology

Research papers
Estimation of prediction interval in ANN-based multi-GCMs downscaling of hydro-climatologic parameters

https://doi.org/10.1016/j.jhydrol.2019.124226Get rights and content

Highlights

  • Lower Upper Bound Estimation method used for constructing prediction intervals.

  • General circulation models applied for downscaling precipitation and temperature.

  • Results compared with the results of classic Bootstrap method.

  • Performance of methods was assessed by the criteria of coverage and width.

  • Prediction intervals for temperature were narrower than precipitation.

Abstract

In this paper, point prediction and prediction intervals (PIs) of artificial neural network (ANN) based downscaling for mean monthly precipitation and temperature of two stations (Tabriz and Ardabil in North West of Iran) were evaluated using general circulation models (GCMs). PIs were constructed by novel Upper Lower Bound Estimation (LUBE) method in which an ANN with two outputs was constructed for estimating the prediction bounds. Also, Bootstrap method as a classic technique for assessing uncertainty of ANN was used to further examine the proposed LUBE method. In this way, the accuracy of PIs was quantified by coverage and width criteria. Three GCMs, Can-ESM2, BNU-ESM, INM-CM4 and ensemble-GCM (ensemble of mentioned models) were used in four grid points around each of station for evaluating ANN-based downscaling of precipitation and temperature parameters. Comparison between the results of two methods indicated that LUBE method could lead to more reliable results than the Bootstrap method. PIs width and coverage probability were 10–40% lower and 2–10% higher than the Bootstrap method for different GCMs, respectively. Ensemble-GCM led to more accurate results so that computed PIs width and coverage probability were 10–60% lower and 2–20% higher than those for the single GCMs.

Introduction

The general circulation models (GCMs) as climate models are employed to forecast the atmospheric conditions, understand the weather and forecast climate change. There are some studies focus on the impact of the climate change on the atmospheric variables over the world (e.g. see McSweeney et al., 2015, Campozano et al., 2016, Osman and Abdellatif, 2016, Ranjbar et al., 2018, Liu et al., 2019). In addition, there are some studies assess the impact of climate change on the temperature and precipitation in Middle East countries. In Alotaibi et al., (2018) various GCMs and data-driven models were applied to study changes in temperature and precipitation of Qassim Region in Saudi Arabia. Results in this study showed that the temperature may be increased in the region and rainfall will differ over various spans of the future. In Okkan (2015) the impact of climate change on monthly precipitation of Tahtali watershed in Turkey was investigated. It was concluded that precipitation is decreased around 12.4% during the period between 2020 and 2039, 10.0% during the period between 2040 and 2059, and 18.3% during the period between 2080 and 2099 and statistically significant changes during the periods 2001–2019 and 2060–2079. Okkan and Kirdemir (2016) downscaled the monthly precipitation for Gediz basin in Turkey and used several GCMs and the obtained results showed a decrease trend for precipitation with different intensities in different scenarios.

Despite the GCMs trustworthy provision of atmospheric data, the coarse spatial resolution of GCMs causes weak application in local scale modeling. Therefore, an appropriate downscaling approach should be used to access local scale weather data from large scale GCMs (Wilby and Wigley, 1997). In order to project local climate change, it is essential to interpret the coarse-scale GCMs and reanalysis output to finer spatial scale by dynamical or statistical downscaling methods. In dynamical downscaling, a regional climate model is attained by GCM boundary conditions and solving equations of motion and thermodynamics (Danandeh Mehr and Kahya, 2016). Although these models have the ability to develop climate variables, the process consists of uncertainty in parameterization of sub‐grid‐scale processes with higher computational costs. Whereas statistical downscaling discovers relationships between local climate variable and GCMs, independent of any information about the physics of the process (Sailor and Li, 1999, Olsson et al., 2004, Harpham and Wilby, 2005, Chen and Adams, 2006, Sousa et al., 2007, Spak et al., 2007, Beecham et al., 2014). However, there are assumptions for statistical downscaling, e.g. as statistical downscaling serves black box modeling tools, so the calibrated model can’t be similarly used in other regions. Also there should be data from a regional station and for downscaling. In addition, developing a statistical downscaling model for finer time scales (such as daily or hourly) often becomes a challenging task due to high memory requirements and slow convergence associated with modeling large data sets (Şen, 2010).

Artificial neural network (ANN) is a black box model, with the ability to extract complicated relations between predictor and predictand variables and simulate the nonlinear and time-varying features of the atmospheric variables at various scales (e.g. monthly, daily, hourly) could lead to prevalent use of ANN in GCM data downscaling (e.g. see, Wilby and Wigley, 1997, Dibike and Coulibaly, 2006, Fistikoglu and Okkan, 2010, Okkan and Fistikoglu, 2014, Okkan and Inan, 2015, Okkan and Kirdemir, 2016).

Although there are intensive applications of ANN for downscaling of GCMs data, point estimation/ prediction of targets by ANN conveys no information about the sampling error, prediction accuracy and uncertainty of the model. With this regard, some studies already examined the uncertainty and reliability of ANN-based GCM downscaling procedure focusing on the confidence intervals (CIs) (e.g., Khan et al., 2006, Samadi et al., 2013).

Prediction intervals (PIs) and CIs are two prevalent measures for assessment of prediction/forecast uncertainty. CIs correspond to the accuracy of the prediction of the regression i.e., the mean of the target probability distribution whereas PIs assimilate the accuracy of the predicted values versus the measured values (Heskes, 1997, Khosravi et al., 2011b). Confidence Interval (CI) corresponds to the modeling uncertainty, which can be classified into different forms based on the variability arises from model inputs, parameters and structures which combine together contributes in producing the prediction uncertainty. PIs correspond to the accuracy of the prediction of targets instead of estimating the true model, therefore PIs are included more sources of the uncertainty and are wider than CIs and as so are more beneficial from practical aspect since the constructed upper and lower bounds of PI are capable of capturing exact observed values (Shrivastava et al., 2015). Decision makers should provide design plan to manage critical issues due to climate change. With this regard in prediction intervals by considering uncertainties, a bound is obtained that is essential in decision making, management and design. Optimistic and pessimistic decision making can be considered by prediction intervals. If we have the bounds in the case that the problem is critical so pessimistic decision making is applied and upper bound will be the bases of planning but it may not be economical; however, in the case that optimistic decision making is considered, the lower bound will be the bases of planning but with a bit high risk.

On the other hand, for the climate change, used GCMs contain various sources of uncertainty which are associated with GCMs data, scenarios and physical rules of GCMs and these uncertainties affect the predictions for future. In climate change studies while using GCMs by applying point prediction for future, designing can be uneconomical or may be critical conditions won’t be considered. In this study these uncertainties have been handled using multi GCMs and also estimating the prediction intervals instead of point prediction which can represent such uncertainties. Therefore, in climate studies, when applying GCMs, while having PI there will be more alternatives for designing.

Some classic techniques are usually used to construct the PIs for ANN methods such as delta, Bayesian, mean-variance estimation, and Bootstrap techniques. The delta technique employs the standard asymptotic theory for constructing the PIs (Chryssolouris et al., 1996, Hwang and Ding, 1997). In this technique, the intervals are constructed under the assumption that the noise is homogeneous and normally distributed which is the main limitation of this technique. The Bayesian technique is on the basis of the Bayesian statistics to express the uncertainty of the network weights in terms of probability distributions and integrates them to obtain the probability distribution of the target conditions. The Bayesian approach requires time-consuming Monte-Carlo integration overweight space in multi-dimensional real-world applications (Mackay, 1992). The mean-variance estimation-based method also assumes that errors are normally distributed around the true mean of targets. Therefore, PIs can easily be constructed if the parameters of this distribution (mean and variance) are known. The dependence of the target variance on the set of inputs is the basic assumption and of course the main limitation of this method (Nix and Weigend, 1994). On the other hand, the Bootstrap technique neither requires the complex computations of derivatives and Hessian-matrix inversion involved in the delta-method nor the Montre Carlo solutions of the integrals involved in the Bayesian approach (Dybowski and Roberts, 2001). Therefore, this simple method has been widely used in several hydro-climatologic studies (e.g., Dibike et al., 2008, Samadi et al., 2013).

In this study, Lower Upper Bound Estimation (LUBE) method which was firstly introduced by Khosravi et al. (2010) and is a novel PIs construction method in hydro-climatologic issues, is applied to construct PIs for ANN-based downscaling of monthly precipitation and temperature parameters at Tabriz and Ardabil stations (located in northwest of Iran). In the LUBE method, an ANN with two output neurons (representing upper and lower bounds of the estimation) is trained. The coverage probability of LUBE is better than other methods and it can effectively establish trade-off between the correctness and informativeness of PIs (Khosravi et al., 2011a). Moreover, the required time for PIs construction is remarkably smaller than other methods (Quan et al., 2014).

In this way, three GCMs and their ensemble are used and the obtained results by LUBE are compared with those obtained by the Bootstrap method. Although in some studies different methods have been already used to construct PIs of ANN methods in hydrological modeling (e.g., Srivastav et al., 2007, Kasiviswanathan et al., 2013, Kasiviswanathan and Sudheer, 2016, Kumar et al., 2015), to the best of authors’ knowledge, as a novel application of LUBE method in the hydro-climatic issues, this is the first study to model PIs of ANN-based GCM data downscaling.

Section snippets

Proposed methodology

In the proposed modeling, dominant atmospheric parameters related to the precipitation and temperature of both stations were distinguished using the MI measure. Thereafter point predictions by ANN method and then PIs of ANN-based downscaling were constructed and evaluated by both LUBE and Bootstrap methods using available data. Finally, point and interval predictions for future (2020–2055) values of precipitation and temperature were estimated for both stations under RCP8.5 climate change

Results and discussion

For evaluating the capability of the LUBE method to construct PIs of downscaled precipitation and temperature values, the dominant inputs from GCMs were selected by MI measure and then classic ANN with one output was trained to find point predictions. Then an ANN with two output neurons (considering ±5% of observed values as the initial guess for upper and lower bounds) was trained to initialize the weights of the network. The perturbed values of initialized weights were used to create networks

Conclusions

Although ANN-based downscaling GCM has been broadly used in prediction of climate variables, point predictions by ANN convey no information about prediction uncertainty. In this content, PI is an essential indicator for quantifying the reliability of ANN-based downscaling of climate parameters using GCMs data.

In this paper, ANN-based downscaling of temperature and precipitation variables was performed using three different GCMs and ensemble-GCMs using the data from1950 to 2012, and from 1975 to

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgement

The resarch was conducted under a finincial grant form University of Tabriz, Research Affairs.

References (73)

  • H. Quan et al.

    Particle swarm optimization for construction of neural network-based prediction intervals

    Neurocomputing

    (2014)
  • S.I. Sousa et al.

    Multiple linear regression and artificial neural networks based on principal components to predict ozone concentrations

    Environ. Modell. Software

    (2007)
  • H.H. Yang et al.

    Relevance of time–frequency features for phonetic and speaker-channel classification

    Speech Commun.

    (2000)
  • A. Ahmadalipour et al.

    Multi-criteria evaluation of CMIP5 GCMs for climate change impact analysis

    Theor. Appl. Climatol.

    (2017)
  • A. Ahmadi et al.

    Identification of dominant sources of sea level pressure for precipitation forecasting over Wales

    J. Hydroinf.

    (2013)
  • A. Ahmadi et al.

    Input selection for long-lead precipitation prediction using large-scale climate variables: a case study

    J. Hydroinf.

    (2015)
  • K. Alotaibi et al.

    Future predictions of rainfall and temperature using GCM and ANN for arid regions: a case study for the Qassim Region, Saudi Arabia

    Water

    (2018)
  • S. Beecham et al.

    Statistical downscaling of multi-site daily rainfall in a south Australian catchment using a generalized linear model

    Int. J. Climatol.

    (2014)
  • L. Campozano et al.

    Comparison of Statistical Downscaling Methods for Monthly Total Precipitation: Case Study for the Paute River Basin in Southern Ecuador

    Adv. Meteorol.

    (2016)
  • J.R. Christy et al.

    What do observational datasets say about modeled tropospheric temperature trends since 1979?

    Remote Sens.

    (2010)
  • G. Chryssolouris et al.

    Confidence interval prediction for neural network models

    IEEE Trans. Neural Networks Learn. Syst.

    (1996)
  • A. Danandeh Mehr et al.

    Grid-based performance evaluation of GCM-RCM combinations for rainfall reproduction

    Theor. Appl. Climatol.

    (2016)
  • M. Devak et al.

    Downscaling of precipitation in Mahanadi Basin, India using support vector machine, K-nearest neighbour and hybrid of support vector machine with K-nearest neighbour

  • Y. Dhanesh et al.

    Predictionsa in ungauged basins: can we use artificial neural networks

    (2011)
  • Y. Dibike et al.

    Uncertainty analysis of statistically downscaled temperature and precipitation regimes in Northern Canada

    Theor. Appl. Climatol.

    (2008)
  • Q. Duan et al.

    Bayesian estimation of local signal and noise in multimodel simulations of climate change

    J. Geophys. Res.

    (2010)
  • R. Dybowski et al.

    Confidence intervals and prediction intervals for feed-forward neural networks

    Clin. Appl. Artif. Neural Networks

    (2001)
  • B. Efron et al.

    An Introduction to the Bootstrasp

    (1993)
  • J. Feng et al.

    Comparison of four ensemble methods combining regional climate simulations over Asia

    Meteorol. Atmos. Phys.

    (2010)
  • O. Fistikoglu et al.

    Statistical downscaling of monthly precipitation using NCEP/NCAR reanalysis data for Tahtali River Basin in Turkey

    J. Hydrol. Eng.

    (2010)
  • J.C. Fyfe et al.

    Overestimated global warming over the past 20 years

    Nat. Clim. Change

    (2013)
  • E.I. Grant et al.

    Statistical Quality and Control, Table A

    (1972)
  • J. Guo et al.

    Prediction of variability of precipitation in the Yangtze River Basin under the climate change conditions based on automated statistical downscaling

    Stoch. Env. Res. Risk Assess.

    (2012)
  • S. Haykin

    Neural Networks: A Comprehensive Foundation

    (1994)
  • T. Heskes

    Practical confidence and prediction intervals

    Adv. Neural Inf. Proces. Syst.

    (1997)
  • A. Hosseini Baghanam et al.

    Conjunction of wavelet-entropy and SOM clustering for multi-GCM statistical downscaling

    Hydrol. Res.

    (2018)
  • Cited by (45)

    • Clustering-based interval prediction of electric load using multi-objective pathfinder algorithm and Elman neural network

      2022, Applied Soft Computing
      Citation Excerpt :

      Hu et al. [33] proposed the double output Least squares support vector machine optimized memetic algorithm for interval load forecasting. Nourani et al. [34] applied a method of ANN-based multi-GCMs downscaling to predict interval. Li and Jin [35] utilized the double output Least squares support vector machine optimized by multi-objective ant lion algorithm to estimate the upper and lower bounds.

    • Interval forecasting for urban water demand using PSO optimized KDE distribution and LSTM neural networks

      2022, Applied Soft Computing
      Citation Excerpt :

      The other interval prediction category involves estimating PI based on the probability density of point forecasting results. The key to obtain the prediction interval by this kind of method depends on the appropriate probability density function (PDF) curve of point prediction errors [25,26]. Karamaziotis et al. [27] built a water demand forecasting system to provide the PI of water demand by using the new time series obtained by the bootstrap method.

    View all citing articles on Scopus
    View full text