Comparison of four updating models for real-time river flow forecasting

Abstract

Four different error-forecast updating models are investigated in terms of their capability of providing real-time river flow forecast accuracy superior to that of rainfall-runoff models applied in the simulation (nonupdating) mode. The first and most widely used is the single autoregressive (AR) model, the second being an elaboration of that model, namely the autoregressive-threshold (AR-TS) updating model. A fuzzy autoregressive-threshold (FU-AR-TS) updating model is proposed as the third form of model, the fourth and final error-forecast updating model applied being the artificial neural network (ANN) model. In the application of these four updating models, the lumped soil moisture accounting and routing (SMAR) conceptual model has been selected to simulate the observed discharge series on 11 selected test catchments. As expected, it is found that all of these four updating models are very successful in improving the flow forecast accuracy, when operating in real-time forecasting mode. A less expected, but nonetheless welcome, result is that the three updating models having the most parameters, i.e. AR-TS, FU-AR-TS, and ANN, do not show any considerable advantages in improving the real-time flow forecast efficiency over that of the simple standard AR model. Thus it is recommended that, in the context of real-time river flow forecasting based on error-forecast updating, modellers should continue to use the AR model.     

Univariate streamflow forecasting using commonly used data-driven models: literature review and case study

Eight data-driven models and five data pre-processing methods were summarized; the multiple linear regression (MLR), artificial neural network (ANN) and wavelet decomposition (WD) models were then used in short-term streamflow forecasting at four stations in the East River basin, China. The wavelet–artificial neural network (W-ANN) method was used to predict 1-month-ahead monthly streamflow at Longchuan station (LS). The results indicate better performance of MLR and wavelet–multiple linear regression (W-MLR) in analysing the stationary trained dataset. Four models showed similar performance in 1-day-ahead streamflow forecasting, while W-MLR and W-ANN performed better in 5-day-ahead forecasting. Three reservoirs were shown to have more influence on downstream than upstream streamflow and models had the worst performance at Boluo station. Furthermore, the W-ANN model performed well for 1-month-ahead streamflow forecasting at LS with consideration of a deterministic component.     

Monthly streamflow forecasting using neuro-wavelet techniques and input analysis

ABSTRACT

Combinations of low-frequency components (also known as approximations) resulting from the wavelet decomposition are tested as inputs to an artificial neural network (ANN) in a hybrid approach, and compared to classical ANN models for flow forecasting for 1, 3, 6 and 12 months ahead. In addition, the inputs are rewritten in terms of the flow, revealing what type of information was being provided to the network, in order to understand the effect of the approximations on the forecasting performance. The results show that the hybrid approach improved the accuracy of all tested models, especially for 1, 3 and 6 months ahead. The input analyses show that high-frequency components are more important for shorter forecast horizons, while for longer horizons, they may worsen the model accuracy.     

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

ABSTRACT

An appropriate streamflow forecasting method is a prerequisite for implementation of efficient water resources management in the water-limited, arid regions that occupy much of Iran. In the current research, monthly streamflow forecasting was combined with three data-driven methods based on large input datasets involving 11 precipitation stations, a natural streamflow, and four climate indices through a long period. The major challenges of rainfall–runoff modelling are generally attributed to complex interacting processes, the large number of variables, and strong nonlinearity. The sensitivity of data-driven methods to the dimension of input/output datasets would be another challenge, so large datasets should be compressed into independently standardized principal components. In this study, three pre-processing techniques were applied: singular value decomposition (SVD) provided more efficient forecasts in comparison to principal component analysis (PCA) and average values of inputs in all networks. Among the data-driven methods, the multi-layer perceptron (MLP) with 1-month lag-time outperformed radial basis and fuzzy-based networks. In general, an increase in monthly lag-time of streamflow forecasting resulted in a decline in forecasting accuracy. The results reveal that SVD was highly effective in pre-processing of data-driven evaluations.     

Wavelet regression model as an alternative to neural networks for monthly streamflow forecasting

The accuracy of the wavelet regression (WR) model in monthly streamflow forecasting is investigated in the study. The WR model is improved combining the two methods—the discrete wavelet transform (DWT) model and the linear regression (LR) model—for 1‐month‐ahead streamflow forecasting. In the first part of the study, the results of the WR model are compared with those of the single LR model. Monthly flow data from two stations, Gerdelli Station on Canakdere River and Isakoy Station on Goksudere River, in Eastern Black Sea region of Turkey are used in the study. The comparison results reveal that the WR model could increase the forecast accuracy of the LR model. In the second part of the study, the accuracy of the WR model is compared with those of the artificial neural networks (ANN) and auto‐regressive (AR) models. On the basis of the results, the WR is found to be better than the ANN and AR models in monthly streamflow forecasting. Copyright © 2009 John Wiley & Sons, Ltd.     

Practical aspects of calibrating and selecting unit hydrograph-based models for continuous river flow simulation

Abstract

A continuous simulation rainfall-streamflow modelling approach that identifies unit hydrographs for total streamflow has been applied to an 11-year record from a national hydrometric monitoring network catchment in the UK. The model is of the parametrically parsimonious conceptual model (PPCM) type that can make efficient use of rainfall, streamflow and air temperature data readily available from established national and regional monitoring networks. A multiple split-sample model calibration and simulation analysis is presented that reveals some guiding principles for calibrating and applying PPCMs generally. The inadequacy of a one-dimensional objective function for calibrating best PPCMs is demonstrated. A two-dimensional objective function approach is superior but is shown to be unreliable in some cases, confirming the need for additional critical inspection of other model performance statistics, model parameters and time series plots as an integral part of the model calibration process. A strong tendency evident from the multiple split-sample analysis is that, for the catchment studied, models that fit relatively well in calibration mode perform relatively poorly in simulation mode.     

Improving the performance of data-driven techniques through data pre-processing for modelling daily reservoir inflow

Abstract

The effect of data pre-processing while developing artificial intelligence (AI) -based data-driven techniques, such as artificial neural networks (ANN), model trees (MT) and linear genetic programming (LGP), is studied for Pawana Reservoir in Maharashtra, India. The daily one-step-ahead inflow forecasts are compared with flows generated from a univariate autoregressive integrated moving average (ARIMA) model. For the full-year data series, a large error is found mainly due to the occurrence of zero values, since the reservoir is located in an intermittent river. Hence, all the techniques are evaluated using two data series: 18 years of daily full-year inflow data (from 1 January to 31 December); and 18 years of daily monsoon season inflow data (from 1 June to 31 October) to take into account the intermittent nature of the data. The relevant range of inputs for each category is selected based on autocorrelation and partial autocorrelation analyses of the inflow series. Conventional pre-processing methods, such as transformation and/or normalization of data, do not perform well because of the large variation in magnitudes, as well as the many zero values (65% of the full-year data series). Therefore, the input data are pre-processed into un-weighted moving average (MA) series of 3 days, 5 days and 7 days. The 3-day MA series performs better, maintaining the peak inflow pattern as in the actual data series, while the coarser-scale (5-day and 7-day) MA series reduce the peak inflow pattern, leading to more errors in peak inflow prediction. The results indicate that AI methods are powerful tools for modelling the daily flow time series with appropriate data pre-processing, in spite of the presence of many zero values. The time-lagged recurrent network (TLRN) ANN modelling technique applied in this study maps the inflow forecasting in a better way than the standard multilayer perceptron (MLP) neural networks, especially in the case of the seasonal data series. The MT technique performs equally well for low and medium inflows, but fails to predict the peak inflows. However, LGP outperforms the other AI models, and also the ARIMA model, for all inflow magnitudes. In the LGP model, the daily full-year data series with more zero inflow values performs better than the daily seasonal models.

Citation Jothiprakash, V. & Kote, A. S. (2011) Improving the performance of data-driven techniques through data pre-processing for modelling daily reservoir inflow. Hydrol. Sci. J. 56(1), 168–186.     

Stream flow forecasting using neuro‐wavelet technique

This paper proposes the application of a neuro‐wavelet technique for modelling monthly stream flows. The neuro‐wavelet model is improved by combining two methods, discrete wavelet transform and multi‐layer perceptron, for one‐month‐ahead stream flow forecasting and results are compared with those of the single multi‐layer perceptron (MLP), multi‐linear regression (MLR) and auto‐regressive (AR) models. Monthly flow data from two stations, Gerdelli Station on Canakdere River and Isakoy Station on Goksudere River, in the Eastern Black Sea region of Turkey are used in the study. The comparison results revealed that the suggested model could increase the forecast accuracy and perform better than the MLP, MLR and AR models. Copyright © 2008 John Wiley & Sons, Ltd.     

Stream temperature modelling using artificial neural networks: application on Catamaran Brook,New Brunswick,Canada

Fish habitat and aquatic life in rivers are highly dependent on water temperature. Therefore, it is important to understand andto be able to predict river water temperatures using models. Such models can increase our knowledge of river thermal regimes as well as provide tools for environmental impact assessments. In this study, artificial neural networks (ANNs) will be used to develop models for predicting both the mean and maximum daily water temperature. The study was conducted within Catamaran Brook, a small drainage basin tributary to the Miramichi River (New Brunswick, Canada). In total, eight ANN models were investigated using a variety of input parameters. Of these models, four predicted mean daily water temperature and four predicted maximum daily water temperature. The best model for mean daily temperature had eight input parameters: minimum, maximum and mean air temperatures of the current day and those of the preceding day, the day of year and the water level. This model had an overall root‐mean‐square error (RMSE) of 0·96 °C, a bias of 0·26 °C and a coefficient of determination R² = 0·971. The model that best predicted maximum daily water temperature was similar to the first model but excluded mean daily air temperature. Good results were obtained for maximum water temperatures with an overall RMSE of 1·18 °C, a bias of 0·15 °C and R² = 0·961. The results of ANN models were similar to and/or better than those observed from the literature. The advantages of artificial neural networks models in modelling river water temperature lie in their simplicity of use, their low data requirement and their good performance, as well as their flexibility in allowing many input and output parameters. Copyright © 2008 Crown in the right of Canada and John Wiley & Sons, Ltd.     

Genetic programming model for forecast of short and noisy data

Though forecasting of river flow has received a great deal of attention from engineers and researchers throughout the world, this still continues to be a challenging task owing to the complexity of the process. In the last decade or so, artificial neural networks (ANNs) have been widely applied, and their ability to model complex phenomena has been clearly demonstrated. However, the success of ANNs depends very crucially on having representative records of sufficient length. Further, the forecast accuracy decreases rapidly with an increase in the forecast horizon. In this study, the use of the Darwinian theory‐based recent evolutionary technique of genetic programming (GP) is suggested to forecast fortnightly flow up to 4‐lead. It is demonstrated that short lead predictions can be significantly improved from a short and noisy time series if the stochastic (noise) component is appropriately filtered out. The deterministic component can then be easily modelled. Further, only the immediate antecedent exogenous and/or non‐exogenous inputs can be assumed to control the process. With an increase in the forecast horizon, the stochastic components also play an important role in the forecast, besides the inherent difficulty in ascertaining the appropriate input variables which can be assumed to govern the underlying process. GP is found to be an efficient tool to identify the most appropriate input variables to achieve reasonable prediction accuracy for higher lead‐period forecasts. A comparison with ANNs suggests that though there is no significant difference in the prediction accuracy, GP does offer some unique advantages. Copyright © 2006 John Wiley & Sons, Ltd.     

Data-based modelling approach for variable density flow and solute transport simulation in a coastal aquifer

Data-based models, namely artificial neural network (ANN), support vector machine (SVM), genetic programming (GP) and extreme learning machine (ELM), were developed to approximate three-dimensional, density-dependent flow and transport processes in a coastal aquifer. A simulation model, SEAWAT, was used to generate data required for the training and testing of the data-based models. Statistical analysis of the simulation results obtained by the four models show that the data-based models could simulate the complex salt water intrusion process successfully. The selected models were also compared based on their computational ability, and the results show that the ELM is the fastest technique, taking just 0.5 s to simulate the dataset; however, the SVM is the most accurate, with a Nash-Sutcliffe efficiency (NSE) ≥ 0.95 and correlation coefficient R ≥ 0.92 for all the wells. The root mean square error (RMSE) for the SVM is also significantly less, ranging from 12.28 to 77.61 mg/L.     

Real‐time recurrent learning neural network for stream‐flow forecasting

Various types of neural networks have been proposed in previous papers for applications in hydrological events. However, most of these applied neural networks are classified as static neural networks, which are based on batch processes that update action only after the whole training data set has been presented. The time variate characteristics in hydrological processes have not been modelled well. In this paper, we present an alternative approach using an artificial neural network, termed real‐time recurrent learning (RTRL) for stream‐flow forecasting. To define the properties of the RTRL algorithm, we first compare the predictive ability of RTRL with least‐square estimated autoregressive integrated moving average models on several synthetic time‐series. Our results demonstrate that the RTRL network has a learning capacity with high efficiency and is an adequate model for time‐series prediction. We also investigated the RTRL network by using the rainfall–runoff data of the Da‐Chia River in Taiwan. The results show that RTRL can be applied with high accuracy to the study of real‐time stream‐flow forecasting networks. Copyright © 2002 John Wiley & Sons, Ltd.     

Hybrid modelling approach to prairie hydrology: fusing data-driven and process-based hydrological models

Abstract

Much of the prairie region in North America is characterized by relatively flat terrain with many depressions on the landscape. The hydrological response (runoff) is a combination of the conventional runoff from the contributing areas and the occasional overflow from the non-contributing areas (depressions). In this study, we promote the use of a hybrid modelling structure to predict runoff generation from prairie landscapes. More specifically, the Soil and Water Assessment Tool (SWAT) is fused with artificial neural networks (ANNs), so that SWAT and the ANN module deal with the contributing and non-contributing areas, respectively. A detailed experimental study is performed to select the best set of inputs, training algorithms and hidden neurons. The results obtained in this study suggest that the fusion of process-based and data-driven models can provide improved modelling capabilities for representing the highly nonlinear nature of the hydrological processes in prairie landscapes.

Editor D. Koutsoyiannis; Associate editor L. See     

Comparing the performance of empirical black‐box models for river flow forecasting in the Heihe River Basin,Northwestern China

For many practical reasons, the empirical black‐box models have become an increasingly popular modelling tool for river flow forecasting, especially in mountainous areas where very few meteorological observatories exist. In this article, precipitation data are used as the only input to estimate river flow. Using five empirical black‐box models—the simple linear model, the linear perturbation model, the linearly varying gain factor model, the constrained nonlinear system model and the nonlinear perturbation model–antecedent precipitation index—modelling results are compared with actual results in three catchments within the Heihe River Basin. The linearly varying gain factor model and the nonlinear perturbation model yielded excellent predictions. For better simulation accuracy, a commonly used multilayer feed‐forward neural network model (NNM) was applied to incorporate the outputs of the individual models. Comparing the performance of these models, it was found that the best results were obtained from the NNM model. The results also suggest that more reliable and precise predictions of river flow can be obtained by using the NNM model while also incorporating the combined outputs of different empirical black‐box models. Copyright © 2012 John Wiley & Sons, Ltd.     

Diary of forthcoming events

Abstract

Currently, environmental modelling is frequently conducted with the aid of artificial neural networks (ANNs) in an effort to achieve greater accuracy in simulation and forecasting beyond that typically obtained when using solely linear models. For the design of an ANN, modellers must contend with two key issues: (a) the selection of model input and (b) the determination of the number of hidden neurons. A novel approach is introduced to address the optimal design of ANNs based on a multi-objective strategy that enables the user to find a set of feasible ANNs, determined as optimal trade-off solutions between model simplicity and accuracy. This is achieved in a multi-objective fashion by simultaneously minimizing three different cost functions: the model input dimension, the hidden neuron number and the generalization error computed on a validation set of data. The multi-objective approach is based on the Pareto dominance criterion and an evolutionary strategy has been employed to solve the combinatorial optimization problem. From a theoretical perspective, the choice of a multi-objective approach marks an attempt to account for, and overcome, the “curse of dimensionality” and to circumvent the drawbacks of “overfitting” that are inherent in ANNs. Moreover, it is demonstrated that the strategy renders the choice of the ANN more robust, as is evident by “unseen data” in the testing stage, since structure determination is not merely based on the statistical evaluation of the generalization performance. The methodology is tested and the results are reported in a case study relating groundwater level predictions to total monthly rainfall.     

A two‐step‐ahead recurrent neural network for stream‐flow forecasting

In many engineering problems, such as flood warning systems, accurate multistep‐ahead prediction is critically important. The main purpose of this study was to derive an algorithm for two‐step‐ahead forecasting based on a real‐time recurrent learning (RTRL) neural network that has been demonstrated as best suited for real‐time application in various problems. To evaluate the properties of the developed two‐step‐ahead RTRL algorithm, we first compared its predictive ability with least‐square estimated autoregressive moving average with exogenous inputs (ARMAX) models on several synthetic time‐series. Our results demonstrate that the developed two‐step‐ahead RTRL network has efficient ability to learn and has comparable accuracy for time‐series prediction as the refitted ARMAX models. We then investigated the two‐step‐ahead RTRL network by using the rainfall–runoff data of the Da‐Chia River in Taiwan. The results show that the developed algorithm can be successfully applied with high accuracy for two‐step‐ahead real‐time stream‐flow forecasting. Copyright © 2003 John Wiley & Sons, Ltd.     

Improving extreme hydrologic events forecasting using a new criterion for artificial neural network selection

The issue of selecting appropriate model input parameters is addressed using a peak and low flow criterion (PLC). The optimal artificial neural network (ANN) models selected using the PLC significantly outperform those identified with the classical root‐mean‐square error (RMSE) or the conventional Nash–Sutcliffe coefficient (NSC) statistics. The comparative forecast results indicate that the PLC can help to design an appropriate ANN model to improve extreme hydrologic events (peak and low flow) forecast accuracy. Copyright © 2001 John Wiley & Sons, Ltd.     

Prediction of hourly actual evapotranspiration using neural networks,genetic programming,and statistical models

The complexity of the evapotranspiration process and its variability in time and space have imposed some limitations on previously developed evapotranspiration models. In this study, two data‐driven models: genetic programming (GP) and artificial neural networks (ANNs), and statistical regression models were developed and compared for estimating the hourly eddy covariance (EC)‐measured actual evapotranspiration (AET) using meteorological variables. The utility of the investigated data‐driven models was also compared with that of HYDRUS‐1D model, which makes use of conventional Penman–Monteith (PM) model for the prediction of AET. The latent heat (LE), which is measured using the EC method, is modelled as a function of five climatic variables: net radiation, ground temperature, air temperature, relative humidity, and wind speed in a reconstructed landscape located in Northern Alberta, Canada. Several ANN models were evaluated using two training algorithms of Levenberg–Marquardt and Bayesian regularization. The GP technique was used to generate mathematical equations correlating AET to the five climatic variables. Furthermore, the climatic variables, as well as their two‐factor interactions, were statistically analysed to obtain a regression equation and to indicate the climatic factors having significant effect on the evapotranspiration process. HYDRUS‐1D model as an available physically based model was examined for estimating AET using climatic variables, leaf area index (LAI), and soil moisture information. The results indicated that all three proposed data‐driven models were able to approximate the AET reasonably well; however, GP and regression models had better generalization ability than the ANN model. The results of HYDRUS‐1D model exhibited that a physically based model, such as HYDRUS‐1D, might be comparable or even inferior to the data‐driven models in terms of the overall prediction accuracy. Based on the developed GP and regression models, net radiation and ground temperature had larger contribution to the AET process than other variables. Copyright © 2010 John Wiley & Sons, Ltd.     

Artificial neural networks as an alternative approach to groundwater numerical modelling and environmental design

Classical optimization methodologies based on mathematical theories have been developed for the solution of various constrained environmental design problems. Numerical models have been widely used to represent an environmental system accurately. The use of methodologies such as artificial neural networks (ANNs), which approximate the complicated behaviour and response of physical systems, allows the optimization of a large number of case scenarios with different set of constraints within a short period of time, whereas the corresponding simulation time using a numerical model would be prohibitive. In this paper, a combination of an ANN with a differential evolution algorithm is proposed to replace the classical finite‐element numerical model in water resources management problems. The objective of the optimization problem is to determine the optimal operational strategy for the productive pumping wells located in the northern part of Rhodes Island in Greece, to cover the water demand and maintain the water table at certain levels. The conclusions of this study show that the use of ANN as an approximation model could (a) significantly reduce the computational burden associated with the accurate simulation of complex physical systems and (b) provide solutions very close to the optimal ones for various constrained environmental design problems. Copyright © 2008 John Wiley & Sons, Ltd.     

An evaluation of a traditional and a neural net modelling approach to flood forecasting for an upland catchment

This study evaluates two (of the many) modelling approaches to flood forecasting for an upland catchment (the River South Tyne at Haydon Bridge, England). The first modelling approach utilizes ‘traditional’ hydrological models. It consists of a rainfall–runoff model (the probability distributed model, or PDM) for flow simulation in the upper catchment. Those flows are then routed to the lower catchment using two kinematic wave (KW) routing models. When run in forecast‐mode, the PDM and KW models utilize model updating procedures. The second modelling approach uses neural network models, which use a ‘pattern‐matching’ process to produce model forecasts.Following calibration, the models are evaluated in terms of their fit to continuous stage data and flood event magnitudes and timings within a validation period. Forecast times of 1 h, 2 h and 4 h are selected (the catchment has a response time of approximately 4 h). The ‘traditional’ models generally perform adequately at all three forecast times. The neural networks produce reasonable forecasts of small‐ to medium‐sized flood events but have difficulty in forecasting the magnitude of the larger flood events in the validation period. Possible modifications to the latter approach are discussed. © Crown copyright 2002. Reproduced with the permission of Her Majesty's stationery office. Published by John Wiley & Sons, Ltd.