A comparison of performance of several artificial intelligence methods for forecasting monthly discharge time series

Developing a hydrological forecasting model based on past records is crucial to effective hydropower reservoir management and scheduling. Traditionally, time series analysis and modeling is used for building mathematical models to generate hydrologic records in hydrology and water resources. Artificial intelligence (AI), as a branch of computer science, is capable of analyzing long-series and large-scale hydrological data. In recent years, it is one of front issues to apply AI technology to the hydrological forecasting modeling. In this paper, autoregressive moving-average (ARMA) models, artificial neural networks (ANNs) approaches, adaptive neural-based fuzzy inference system (ANFIS) techniques, genetic programming (GP) models and support vector machine (SVM) method are examined using the long-term observations of monthly river flow discharges. The four quantitative standard statistical performance evaluation measures, the coefficient of correlation (R), Nash–Sutcliffe efficiency coefficient (E), root mean squared error (RMSE), mean absolute percentage error (MAPE), are employed to evaluate the performances of various models developed. Two case study river sites are also provided to illustrate their respective performances. The results indicate that the best performance can be obtained by ANFIS, GP and SVM, in terms of different evaluation criteria during the training and validation phases.     

Comparative analysis of fuzzy inference systems for water consumption time series prediction

Two types of fuzzy inference systems (FIS) are used for predicting municipal water consumption time series. The FISs used include an adaptive neuro-fuzzy inference system (ANFIS) and a Mamdani fuzzy inference systems (MFIS). The prediction models are constructed based on the combination of the antecedent values of water consumptions. The performance of ANFIS and MFIS models in training and testing phases are compared with the observations and the best fit model is identified according to the selected performance criteria. The results demonstrated that the ANFIS model is superior to MFIS models and can be successfully applied for prediction of water consumption time series.     

A Model-Averaging Method for Assessing Groundwater Conceptual Model Uncertainty

This study evaluates alternative groundwater models with different recharge and geologic components at the northern Yucca Flat area of the Death Valley Regional Flow System (DVRFS), USA. Recharge over the DVRFS has been estimated using five methods, and five geological interpretations are available at the northern Yucca Flat area. Combining the recharge and geological components together with additional modeling components that represent other hydrogeological conditions yields a total of 25 groundwater flow models. As all the models are plausible given available data and information, evaluating model uncertainty becomes inevitable. On the other hand, hydraulic parameters (e.g., hydraulic conductivity) are uncertain in each model, giving rise to parametric uncertainty. Propagation of the uncertainty in the models and model parameters through groundwater modeling causes predictive uncertainty in model predictions (e.g., hydraulic head and flow). Parametric uncertainty within each model is assessed using Monte Carlo simulation, and model uncertainty is evaluated using the model averaging method. Two model-averaging techniques (on the basis of information criteria and GLUE) are discussed. This study shows that contribution of model uncertainty to predictive uncertainty is significantly larger than that of parametric uncertainty. For the recharge and geological components, uncertainty in the geological interpretations has more significant effect on model predictions than uncertainty in the recharge estimates. In addition, weighted residuals vary more for the different geological models than for different recharge models. Most of the calibrated observations are not important for discriminating between the alternative models, because their weighted residuals vary only slightly from one model to another.     

Monthly total sediment forecasting using adaptive neuro fuzzy inference system

Accurate forecasting of sediment is an important issue for reservoir design and water pollution control in rivers and reservoirs. In this study, an adaptive neuro-fuzzy inference system (ANFIS) approach is used to construct monthly sediment forecasting system. To illustrate the applicability of ANFIS method the Great Menderes basin is chosen as the study area. The models with various input structures are constructed for the purpose of identification of the best structure. The performance of the ANFIS models in training and testing sets are compared with the observed data. To get more accurate evaluation of the results ANFIS models, the best fit model structures are also tested by artificial neural networks (ANN) and multiple linear regression (MLR) methods. The results of three methods are compared, and it is observed that the ANFIS is preferable and can be applied successfully because it provides high accuracy and reliability for forecasting of monthly total sediment.     

Testing alternative ground water models using cross-validation and other methods

Many methods can be used to test alternative ground water models. Of concern in this work are methods able to (1) rank alternative models (also called model discrimination) and (2) identify observations important to parameter estimates and predictions (equivalent to the purpose served by some types of sensitivity analysis). Some of the measures investigated are computationally efficient; others are computationally demanding. The latter are generally needed to account for model nonlinearity. The efficient model discrimination methods investigated include the information criteria: the corrected Akaike information criterion, Bayesian information criterion, and generalized cross-validation. The efficient sensitivity analysis measures used are dimensionless scaled sensitivity (DSS), composite scaled sensitivity, and parameter correlation coefficient (PCC); the other statistics are DFBETAS, Cook's D, and observation-prediction statistic. Acronyms are explained in the introduction. Cross-validation (CV) is a computationally intensive nonlinear method that is used for both model discrimination and sensitivity analysis. The methods are tested using up to five alternative parsimoniously constructed models of the ground water system of the Maggia Valley in southern Switzerland. The alternative models differ in their representation of hydraulic conductivity. A new method for graphically representing CV and sensitivity analysis results for complex models is presented and used to evaluate the utility of the efficient statistics. The results indicate that for model selection, the information criteria produce similar results at much smaller computational cost than CV. For identifying important observations, the only obviously inferior linear measure is DSS; the poor performance was expected because DSS does not include the effects of parameter correlation and PCC reveals large parameter correlations.     

Uncertainty-based multi-criteria calibration of rainfall-runoff models: a comparative study

This study compares formal Bayesian inference to the informal generalized likelihood uncertainty estimation (GLUE) approach for uncertainty-based calibration of rainfall-runoff models in a multi-criteria context. Bayesian inference is accomplished through Markov Chain Monte Carlo (MCMC) sampling based on an auto-regressive multi-criteria likelihood formulation. Non-converged MCMC sampling is also considered as an alternative method. These methods are compared along multiple comparative measures calculated over the calibration and validation periods of two case studies. Results demonstrate that there can be considerable differences in hydrograph prediction intervals generated by formal and informal strategies for uncertainty-based multi-criteria calibration. Also, the formal approach generates definitely preferable validation period results compared to GLUE (i.e., tighter prediction intervals that show higher reliability) considering identical computational budgets. Moreover, non-converged MCMC (based on the standard Gelman–Rubin metric) performance is reasonably consistent with those given by a formal and fully-converged Bayesian approach even though fully-converged results requires significantly larger number of samples (model evaluations) for the two case studies. Therefore, research to define alternative and more practical convergence criteria for MCMC applications to computationally intensive hydrologic models may be warranted.     

Reconciling theory with observations: elements of a diagnostic approach to model evaluation

This paper discusses the need for a well‐considered approach to reconciling environmental theory with observations that has clear and compelling diagnostic power. This need is well recognized by the scientific community in the context of the ‘Predictions in Ungaged Basins’ initiative and the National Science Foundation sponsored ‘Environmental Observatories’ initiative, among others. It is suggested that many current strategies for confronting environmental process models with observational data are inadequate in the face of the highly complex and high order models becoming central to modern environmental science, and steps are proposed towards the development of a robust and powerful ‘Theory of Evaluation’. This paper presents the concept of a diagnostic evaluation approach rooted in information theory and employing the notion of signature indices that measure theoretically relevant system process behaviours. The signature‐based approach addresses the issue of degree of system complexity resolvable by a model. Further, it can be placed in the context of Bayesian inference to facilitate uncertainty analysis, and can be readily applied to the problem of process evaluation leading to improved predictions in ungaged basins. Copyright © 2008 John Wiley & Sons, Ltd.     

Analyzing spatial ecological data using linear regression and wavelet analysis

Spatial (two-dimensional) distributions in ecology are often influenced by spatial autocorrelation. In standard regression models, however, observations are assumed to be statistically independent. In this paper we present an alternative to other methods that allow for autocorrelation. We show that the theory of wavelets provides an efficient method to remove autocorrelations in regression models using data sampled on a regular grid. Wavelets are particularly suitable for data analysis without any prior knowledge of the underlying correlation structure. We illustrate our new method, called wavelet-revised model, by applying it to multiple regression for both normal linear models and logistic regression. Results are presented for computationally simulated data and real ecological data (distribution of species richness and distribution of the plant species Dianthus carthusianorum throughout Germany). These results are compared to those of generalized linear models and models based on generalized estimating equations. We recommend wavelet-revised models, in particular, as a method for logistic regression using large datasets.     

基于VC++的梁式桥震害评估软件

震后桥梁结构震害快速评估能够为地震应急救援提供重要参考信息,对减灾工作具有重要价值。为了快速准确地评估遭受地震影响梁式桥的破坏状态,使梁式桥震害评估方法在地震应急中发挥更大作用,基于统计回归模型、神经网络模型和推断模型等三种梁式桥震害评估模型,采用VC++6.0及Access2003数据库软件开发了有关评估软件,并以桥梁震害资料为算例验证了软件的可靠性。     

A Bayesian central equivalent dose model for optically stimulated luminescence dating

In this study, we propose and implement a Bayesian model to estimate a central equivalent dose from a set of luminescence measurements. This model is based on assumptions similar to the ones used in the standard statistical pipeline (typically implemented in the Analyst software followed by a subsequent central equivalent dose analysis) but tackles some of its main limitations. More specifically, it consists of a three-stage hierarchical model that has two main advantages over the standard approach: first, it avoids the introduction of auxiliary variables (typically mean and variance), at each step of the inference process, which are likely to fail to characterise the distributions of interest; second, it ensures a homogeneous and consistent inference with respect to the overall model and data. As a Bayesian model, our model requires the specification of prior distributions; we discuss such informative and non-informative distributions and check the relevance of our choices on synthetic data. Then, we use data derived from Single Aliquot and Regenerative (SAR) dose measurements performed on single grains from laboratory-bleached and dosed samples. The results show that our Bayesian approach offers a promising alternative to the standard one. Finally, we conclude by stressing that, relying on a Bayesian hierarchical model, our approach could be modified to incorporate additional information (e.g. stratigraphic constraints) that is difficult to formalise properly with the existing approaches.     

A partially non-ergodic ground-motion prediction equation for Europe and the Middle East

A partially non-ergodic ground-motion prediction equation is estimated for Europe and the Middle East. Therefore, a hierarchical model is presented that accounts for regional differences. For this purpose, the scaling of ground-motion intensity measures is assumed to be similar, but not identical in different regions. This is achieved by assuming a hierarchical model, where some coefficients are treated as random variables which are sampled from an underlying global distribution. The coefficients are estimated by Bayesian inference. This allows one to estimate the epistemic uncertainty in the coefficients, and consequently in model predictions, in a rigorous way. The model is estimated based on peak ground acceleration data from nine different European/Middle Eastern regions. There are large differences in the amount of earthquakes and records in the different regions. However, due to the hierarchical nature of the model, regions with only few data points borrow strength from other regions with more data. This makes it possible to estimate a separate set of coefficients for all regions. Different regionalized models are compared, for which different coefficients are assumed to be regionally dependent. Results show that regionalizing the coefficients for magnitude and distance scaling leads to better performance of the models. The models for all regions are physically sound, even if only very few earthquakes comprise one region.     

Development of stochastic dynamic Nash game model for reservoir operation II. The value of players’ information availability and cooperative behaviors

Stochastic dynamic game models can be applied to derive optimal reservoir operation policies by considering interactions among water users and reservoir operator, their preferences, their levels of information availability and cooperative behaviors. The stochastic dynamic game model with perfect information (PSDNG) has been developed by [Ganji A, Khalili D, Karamouz M. Development of stochastic dynamic Nash game model for reservoir operation. I. The symmetric stochastic model with perfect information. Adv Water Resour, this issue]. This paper develops four additional versions of stochastic dynamic game model of water users interactions based on the cooperative behavior and hydrologic information availability of beneficiary sectors of reservoir systems. It is shown that the proposed models are quite capable of providing appropriate reservoir operating policies when compared with alternative operating models, as indicated by several reservoir performance characteristics. Among the proposed models, the selected model by considering cooperative behavior and additional hydrologic information (about the randomness nature of reservoir operation parameters), as exercised by reservoir operator, provides the highest attained level of performance and efficiency. Furthermore, the selected model is more realistic since it also considers actual behavior of water users and reservoir operator in the analysis.     

Inference and uncertainty of snow depth spatial distribution at the kilometre scale in the Colorado Rocky Mountains: the effects of sample size,random sampling,predictor quality,and validation procedures

Historically, observing snow depth over large areas has been difficult. When snow depth observations are sparse, regression models can be used to infer the snow depth over a given area. Data sparsity has also left many important questions about such inference unexamined. Improved inference, or estimation, of snow depth and its spatial distribution from a given set of observations can benefit a wide range of applications from water resource management, to ecological studies, to validation of satellite estimates of snow pack. The development of Light Detection and Ranging (LiDAR) technology has provided non‐sparse snow depth measurements, which we use in this study, to address fundamental questions about snow depth inference using both sparse and non‐sparse observations. For example, when are more data needed and when are data redundant? Results apply to both traditional and manual snow depth measurements and to LiDAR observations. Through sampling experiments on high‐resolution LiDAR snow depth observations at six separate 1.17‐km² sites in the Colorado Rocky Mountains, we provide novel perspectives on a variety of issues affecting the regression estimation of snow depth from sparse observations. We measure the effects of observation count, random selection of observations, quality of predictor variables, and cross‐validation procedures using three skill metrics: percent error in total snow volume, root mean squared error (RMSE), and R². Extremes of predictor quality are used to understand the range of its effect; how do predictors downloaded from internet perform against more accurate predictors measured by LiDAR? Whereas cross validation remains the only option for validating inference from sparse observations, in our experiments, the full set of LiDAR‐measured snow depths can be considered the ‘true’ spatial distribution and used to understand cross‐validation bias at the spatial scale of inference. We model at the 30‐m resolution of readily available predictors, which is a popular spatial resolution in the literature. Three regression models are also compared, and we briefly examine how sampling design affects model skill. Results quantify the primary dependence of each skill metric on observation count that ranges over three orders of magnitude, doubling at each step from 25 up to 3200. Whereas uncertainty (resulting from random selection of observations) in percent error of true total snow volume is typically well constrained by 100–200 observations, there is considerable uncertainty in the inferred spatial distribution (R²) even at medium observation counts (200–800). We show that percent error in total snow volume is not sensitive to predictor quality, although RMSE and R² (measures of spatial distribution) often depend critically on it. Inaccuracies of downloaded predictors (most often the vegetation predictors) can easily require a quadrupling of observation count to match RMSE and R² scores obtained by LiDAR‐measured predictors. Under cross validation, the RMSE and R² skill measures are consistently biased towards poorer results than their true validations. This is primarily a result of greater variance at the spatial scales of point observations used for cross validation than at the 30‐m resolution of the model. The magnitude of this bias depends on individual site characteristics, observation count (for our experimental design), and sampling design. Sampling designs that maximize independent information maximize cross‐validation bias but also maximize true R². The bagging tree model is found to generally outperform the other regression models in the study on several criteria. Finally, we discuss and recommend use of LiDAR in conjunction with regression modelling to advance understanding of snow depth spatial distribution at spatial scales of thousands of square kilometres. Copyright © 2012 John Wiley & Sons, Ltd.     

A Bayesian hierarchical modeling approach for analyzing observational data from marine ecological studies

We introduce the Bayesian hierarchical modeling approach for analyzing observational data from marine ecological studies using a data set intended for inference on the effects of bottom-water hypoxia on macrobenthic communities in the northern Gulf of Mexico off the coast of Louisiana, USA. We illustrate (1) the process of developing a model, (2) the use of the hierarchical model results for statistical inference through innovative graphical presentation, and (3) a comparison to the conventional linear modeling approach (ANOVA). Our results indicate that the Bayesian hierarchical approach is better able to detect a “treatment” effect than classical ANOVA while avoiding several arbitrary assumptions necessary for linear models, and is also more easily interpreted when presented graphically. These results suggest that the hierarchical modeling approach is a better alternative than conventional linear models and should be considered for the analysis of observational field data from marine systems.     

A comparison of non-linear least square and GLUE for model calibration and uncertainty estimation for pesticide transport in soils

The problems of calibrating soil hydraulic and transport parameters are well documented, particularly when data are limited. Programs such as CXTFIT, UUCODE and PEST, based on well established principles of statistical inference, will often provide good fits to limited observations giving the impression that a useful model of a particular soil system has been obtained. This may be the case, but such an approach may grossly underestimate the uncertainties associated with future predictions of the system and resulting dependent variables. In this paper, this is illustrated by an application of CXTFIT within the generalised likelihood uncertainty estimation (GLUE) approach to model calibration which is based on a quite different philosophy. CXTFIT gives very good fits to the observed breakthrough curves for several different model formulations, resulting in very small parameter uncertainty estimates. The application of GLUE, however, shows that much wider ranges of parameter values can provide acceptable fits to the data. The wider range of potential outcomes should be more robust in model prediction, especially when used to constrain field scale models.     

Assessing effects of model complexity and structure on predictions of hydrological responses using serial and parallel model design

By utilizing functional relationships based on observations at plot or field scales, water quality models first compute surface runoff and then use it as the primary governing variable to estimate sediment and nutrient transport. When these models are applied at watershed scales, this serial model structure, coupling a surface runoff sub-model with a water quality sub-model, may be inappropriate because dominant hydrological processes differ among scales. A parallel modeling approach is proposed to evaluate how best to combine dominant hydrological processes for predicting water quality at watershed scales. In the parallel scheme, dominant variables of water quality models are identified based entirely on their statistical significance using time series analysis. Four surface runoff models of different model complexity were assessed using both the serial and parallel approaches to quantify the uncertainty on forcing variables used to predict water quality. The eight alternative model structures were tested against a 25-year high-resolution data set of streamflow, suspended sediment discharge, and phosphorous discharge at weekly time steps. Models using the parallel approach consistently performed better than serial-based models, by having less error in predictions of watershed scale streamflow, sediment and phosphorus, which suggests model structures of water quantity and quality models at watershed scales should be reformulated by incorporating the dominant variables. The implication is that hydrological models should be constructed in a way that avoids stacking one sub-model with one set of scale assumptions onto the front end of another sub-model with a different set of scale assumptions.     

Variational assimilation of meteorological observations in the lower atmosphere: A tutorial on how it works

Data assimilation combines atmospheric measurements with knowledge of atmospheric behavior as codified in computer models, thus producing a “best” estimate of current conditions that is consistent with both information sources. The four major challenges in data assimilation are: (1) to generate an initial state for a computer forecast that has the same mass-wind balance as the assimilating model, (2) to deal with the common problem of highly non-uniform distribution of observations, (3) to exploit the value of proxy observations (of parameters that are not carried explicitly in the model), and (4) to determine the statistical error properties of observing systems and numerical model alike so as to give each information source the proper weight. Variational data assimilation is practiced at major meteorological centers around the world. It is based upon multivariate linear regression, dating back to Gauss, and variational calculus. At the heart of the method is the minimization of a cost function, which guarantees that the analyzed fields will closely resemble both the background field (a short forecast containing a priori information about the atmospheric state) and current observations. The size of the errors in the background and the observations (the latter, arising from measurement and non-representativeness) determine how close the analysis is to each basic source of information. Three-dimensional variational (3DVAR) assimilation provides a logical framework for incorporating the error information (in the form of variances and spatial covariances) and deals directly with the problem of proxy observations. 4DVAR assimilation is an extension of 3DVAR assimilation that includes the time dimension; it attempts to find an evolution of model states that most closely matches observations taken over a time interval measured in hours. Both 3DVAR and, especially, 4DVAR assimilation require very large computing resources. Researchers are trying to find more efficient numerical solutions to these problems. Variational assimilation is applicable in the upper atmosphere, but practical implementation demands accurate modeling of the physical processes that occur at high altitudes and multiple sources of observations.     

Using atmospheric tracers to reduce uncertainty in groundwater recharge areas

A Monte Carlo-based approach to assess uncertainty in recharge areas shows that incorporation of atmospheric tracer observations (in this case, tritium concentration) and prior information on model parameters leads to more precise predictions of recharge areas. Variance-covariance matrices, from model calibration and calculation of sensitivities, were used to generate parameter sets that account for parameter correlation and uncertainty. Constraining parameter sets to those that met acceptance criteria, which included a standard error criterion, did not appear to bias model results. Although the addition of atmospheric tracer observations and prior information produced similar changes in the extent of predicted recharge areas, prior information had the effect of increasing probabilities within the recharge area to a greater extent than atmospheric tracer observations. Uncertainty in the recharge area propagates into predictions that directly affect water quality, such as land cover in the recharge area associated with a well and the residence time associated with the well. Assessments of well vulnerability that depend on these factors should include an assessment of model parameter uncertainty. A formal simulation of parameter uncertainty can be used to delineate probabilistic recharge areas, and the results can be expressed in ways that can be useful to water-resource managers. Although no one model is the correct model, the results of multiple models can be evaluated in terms of the decision being made and the probability of a given outcome from each model.     

To be or not to be multi-Gaussian? A reflection on stochastic hydrogeology

The multivariate Gaussian random function model is commonly used in stochastic hydrogeology to model spatial variability of log-conductivity. The multi-Gaussian model is attractive because it is fully characterized by an expected value and a covariance function or matrix, hence its mathematical simplicity and easy inference. Field data may support a Gaussian univariate distribution for log hydraulic conductivity, but, in general, there are not enough field data to support a multi-Gaussian distribution. A univariate Gaussian distribution does not imply a multi-Gaussian model. In fact, many multivariate models can share the same Gaussian histogram and covariance function, yet differ by their patterns of spatial continuity at different threshold values. Hence the decision to use a multi-Gaussian model to represent the uncertainty associated with the spatial heterogeneity of log-conductivity is not databased. Of greatest concern is the fact that a multi-Gaussian model implies the minimal spatial correlation of extreme values, a feature critical for mass transport and a feature that may be in contradiction with some geological settings, e.g. channeling. The possibility for high conductivity values to be spatially correlated should not be discarded by adopting a congenial model just because data shortage prevents refuting it. In this study, three alternatives to a multi-Gaussian model, all sharing the same Gaussian histogram and the same covariance function, but with different continuity patterns for extreme values, were considered to model the spatial variability of log-conductivity. The three alternative models, plus the traditional multi-Gaussian model, are used to perform Monte Carlo analyses of groundwater travel times from a hypothetical nuclear repository to the ground surface through a synthetic formation similar to the Finnsjön site in Sweden. The results show that the groundwater travel times predicted by the multi-Gaussian model could be ten times slower than those predicted by the other models. The probabilities of very short travel times could be severely underestimated using the multi-Gaussian model. Consequently, if field measured data are not sufficient to determine the higher-order moments necessary to validate the multi-Gaussian model — which is the usual situation in practice — other alternative models to the multi-Gaussian one ought to be considered.     

The strategy of building a flood forecast model by neuro‐fuzzy network

A methodology is proposed for constructing a flood forecast model using the adaptive neuro‐fuzzy inference system (ANFIS). This is based on a self‐organizing rule‐base generator, a feedforward network, and fuzzy control arithmetic. Given the rainfall‐runoff patterns, ANFIS could systematically and effectively construct flood forecast models. The precipitation and flow data sets of the Choshui River in central Taiwan are analysed to identify the useful input variables and then the forecasting model can be self‐constructed through ANFIS. The analysis results suggest that the persistent effect and upstream flow information are the key effects for modelling the flood forecast, and the watershed's average rainfall provides further information and enhances the accuracy of the model performance. For the purpose of comparison, the commonly used back‐propagation neural network (BPNN) is also examined. The forecast results demonstrate that ANFIS is superior to the BPNN, and ANFIS can effectively and reliably construct an accurate flood forecast model. Copyright © 2005 John Wiley & Sons, Ltd.