Rainfall‐runoff modelling using artificial neural networks: comparison of network types

Growing interest in the use of artificial neural networks (ANNs) in rainfall‐runoff modelling has suggested certain issues that are still not addressed properly. One such concern is the use of network type, as theoretical studies on a multi‐layer perceptron (MLP) with a sigmoid transfer function enlightens certain limitations for its use. Alternatively, there is a strong belief in the general ANN user community that a radial basis function (RBF) network performs better than an MLP, as the former bases its nonlinearities on the training data set. This argument is not yet substantiated by applications in hydrology. This paper presents a comprehensive evaluation of the performance of MLP‐ and RBF‐type neural network models developed for rainfall‐runoff modelling of two Indian river basins. The performance of both the MLP and RBF network models were comprehensively evaluated in terms of their generalization properties, predicted hydrograph characteristics, and predictive uncertainty. The results of the study indicate that the choice of the network type certainly has an impact on the model prediction accuracy. The study suggests that both the networks have merits and limitations. For instance, the MLP requires a long trial‐and‐error procedure to fix the optimal number of hidden nodes, whereas for an RBF the structure of the network can be fixed using an appropriate training algorithm. However, a judgment on which is superior is not clearly possible from this study. Copyright © 2004 John Wiley & Sons, Ltd.     

Identification of physical processes inherent in artificial neural network rainfall runoff models

The emergence of artificial neural network (ANN) technology has provided many promising results in the field of hydrology and water resources simulation. However, one of the major criticisms of ANN hydrologic models is that they do not consider/explain the underlying physical processes in a watershed, resulting in them being labelled as black‐box models. This paper discusses a research study conducted in order to examine whether or not the physical processes in a watershed are inherent in a trained ANN rainfall‐runoff model. The investigation is based on analysing definite statistical measures of strength of relationship between the disintegrated hidden neuron responses of an ANN model and its input variables, as well as various deterministic components of a conceptual rainfall‐runoff model. The approach is illustrated by presenting a case study for the Kentucky River watershed. The results suggest that the distributed structure of the ANN is able to capture certain physical behaviour of the rainfall‐runoff process. The results demonstrate that the hidden neurons in the ANN rainfall‐runoff model approximate various components of the hydrologic system, such as infiltration, base flow, and delayed and quick surface flow, etc., and represent the rising limb and different portions of the falling limb of a flow hydrograph. Copyright © 2004 John Wiley & Sons, Ltd.     

A stochastic framework to assess the performance of flood warning systems based on rainfall‐runoff modeling

Nowadays, Flood Forecasting and Warning Systems (FFWSs) are known as the most inexpensive and efficient non‐structural measures for flood damage mitigation in the world. Benefit to cost of the FFWSs has been reported to be several times of other flood mitigation measures. Beside these advantages, uncertainty in flood predictions is a subject that may affect FFWS's reliability and the benefits of these systems. Determining the reliability of advanced flood warning systems based on the rainfall–runoff models is a challenge in assessment of the FFWS performance which is the subject of this study. In this paper, a stochastic methodology is proposed to provide the uncertainty band of the rainfall–runoff model and to calculate the probability of acceptable forecasts. The proposed method is based on Monte Carlo simulation and multivariate analysis of the predicted time and discharge error data sets. For this purpose, after the calibration of the rainfall–runoff model, the probability distributions of input calibration parameters and uncertainty band of the model are estimated through the Bayesian inference. Then, data sets of the time and discharge errors are calculated using the Monte Carlo simulation, and the probability of acceptable model forecasts is calculated by multivariate analysis of data using copula functions. The proposed approach was applied for a small watershed in Iran as a case study. The results showed using rainfall–runoff modeling based on real‐time precipitation is not enough to attain high performance for FFWSs in small watersheds, and it seems using weather forecasts as the inputs of rainfall–runoff models is essential to increase lead times and the reliability of FFWSs in small watersheds. Copyright © 2013 John Wiley & Sons, Ltd.     

Uncertainties in real‐time flood forecasting with neural networks

Although artificial neural networks (ANNs) have been applied in rainfall runoff modelling for many years, there are still many important issues unsolved that have prevented this powerful non‐linear tool from wide applications in operational flood forecasting activities. This paper describes three ANN configurations and it is found that a dedicated ANN for each lead‐time step has the best performance and a multiple output form has the worst result. The most popular form with multiple inputs and single output has the average performance. In comparison with a linear transfer function (TF) model, it is found that ANN models are uncompetitive against the TF model in short‐range predictions and should not be used in operational flood forecasting owing to their complicated calibration process. For longer range predictions, ANN models have an improved chance to perform better than the TF model; however, this is highly dependent on the training data arrangement and there are undesirable uncertainties involved, as demonstrated by bootstrap analysis in the study. To tackle the uncertainty issue, two novel approaches are proposed: distance analysis and response analysis. Instead of discarding the training data after the model's calibration, the data should be retained as an integral part of the model during its prediction stage and the uncertainty for each prediction could be judged in real time by measuring the distances against the training data. The response analysis is based on an extension of the traditional unit hydrograph concept and has a very useful potential to reveal the hydrological characteristics of ANN models, hence improving user confidence in using them in real time. Copyright © 2006 John Wiley & Sons, Ltd.     

Comparative study of neural network,fuzzy logic and linear transfer function techniques in daily rainfall‐runoff modelling under different input domains

This paper compares artificial neural network (ANN), fuzzy logic (FL) and linear transfer function (LTF)‐based approaches for daily rainfall‐runoff modelling. This study also investigates the potential of Takagi‐Sugeno (TS) fuzzy model and the impact of antecedent soil moisture conditions in the performance of the daily rainfall‐runoff models. Eleven different input vectors under four classes, i.e. (i) rainfall, (ii) rainfall and antecedent moisture content, (iii) rainfall and runoff and (iv) rainfall, runoff and antecedent moisture content are considered for examining the effects of input data vector on rainfall‐runoff modelling. Using the rainfall‐runoff data of the upper Narmada basin, Central India, a suitable modelling technique with appropriate model input structure is suggested on the basis of various model performance indices. The results show that the fuzzy modelling approach is uniformly outperforming the LTF and also always superior to the ANN‐based models. Copyright © 2010 John Wiley & Sons, Ltd.     

Rainfall‐runoff modelling using artificial neural networks technique: a Blue Nile catchment case study

A rainfall‐runoff model based on an artificial neural network (ANN) is presented for the Blue Nile catchment. The best geometry of the ANN rainfall‐runoff model in terms of number of hidden layers and nodes is identified through a sensitivity analysis. The Blue Nile catchment (about 300 000 km²) in the Nile basin is selected here as a case study. The catchment is classified into seven subcatchments, and the mean areal precipitation over those subcatchments is computed as a main input to the ANN model. The available daily data (1992–99) are divided into two sets for model calibration (1992–96) and for validation (1997–99). The results of the ANN model are compared with one of physical distributed rainfall‐runoff models that apply hydraulic and hydrologic fundamental equations in a grid base. The results over the case study area and the comparative analysis with the physically based distributed model show that the ANN technique has great potential in simulating the rainfall‐runoff process adequately. Because the available record used in the calibration of the ANN model is too short, the ANN model is biased compared with the distributed model, especially for high flows. Copyright © 2005 John Wiley & Sons, Ltd.     

Modelling suspended sediment concentration using artificial neural networks for Gangotri glacier

The dynamics of suspended sediment involves inherent non‐linearity and complexity because of existence of both spatial variability of the basin characteristics and temporal climatic patterns. This complexity, therefore, leads to inaccurate prediction by the conventional sediment rating curve (SRC) and other empirical methods. Over past few decades, artificial neural networks (ANNs) have emerged as one of the advanced modelling techniques capable of addressing inherent non‐linearity in the hydrological processes. In the present study, feed‐forward back propagation (FFBP) algorithm of ANNs is used to model stage–discharge–suspended sediment relationship for ablation season (May–September) for melt runoff released from Gangotri glacier, one of the largest glaciers in Himalaya. The simulations have been carried out on primary data of suspended sediment concentration (SSC) discharge and stage for ablation season of 11‐year period (1999–2009). Combinations of different input vectors (viz. stage, discharge and SSC) for present and previous days are considered for development of the ANN models and examining the effects of input vectors. Further, based on model performance indices for training and testing phase, a suitable modelling approach with appropriate model input structure is suggested. The conventional SRC method is also used for modelling discharge–sediment relationship and performance of developed models is evaluated by statistical indices, namely; root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R²). Statistically, the performance of ANN‐based models is found to be superior as compared to SRC method in terms of the selected performance indices in simulating the daily SSC. The study reveals suitability of ANN approach for simulation and estimation of daily SSC in glacier melt runoff and, therefore, opens new avenues of research for application of hybrid soft computing models in glacier hydrology. Copyright © 2015 John Wiley & Sons, Ltd.     

Gridded climate data products are an alternative to instrumental measurements as inputs to rainfall–runoff models

Rainfall–runoff models are widely used to predict flows using observed (instrumental) time series of air temperature and precipitation as inputs. Poor model performance is often associated with difficulties in estimating catchment‐scale meteorological variables from point observations. Readily available gridded climate products are an underutilized source of temperature and precipitation time series for rainfall–runoff modelling, which may overcome some of the performance issues associated with poor‐quality instrumental data in small headwater monitoring catchments. Here we compare the performance of instrumental measured and E‐OBS gridded temperature and precipitation time series as inputs in the rainfall–runoff models “PERSiST” and “HBV” for flow prediction in six small Swedish catchments. For both models and most catchments, the gridded data produced statistically better simulations than did those obtained using instrumental measurements. Despite the high correspondence between instrumental and gridded temperature, both temperature and precipitation were responsible for the difference. We conclude that (a) gridded climate products such as the E‐OBS dataset could be more widely used as alternative input to rainfall–runoff models, even when instrumental measurements are available, and (b) the processing applied to gridded climate products appears to provide a more realistic approximation of small catchment‐scale temperature and precipitation patterns needed for flow simulations. Further research on this issue is needed and encouraged.     

Error analysis for the evaluation of model performance: rainfall–runoff event time series data

This paper provides a procedure for evaluating model performance where model predictions and observations are given as time series data. The procedure focuses on the analysis of error time series by graphing them, summarizing them, and predicting their variability through available information (recalibration). We analysed two rainfall–runoff events from the R‐5 data set, and evaluated 12 distinct model simulation scenarios for these events, of which 10 were conducted with the quasi‐physically‐based rainfall–runoff model (QPBRRM) and two with the integrated hydrology model (InHM). The QPBRRM simulation scenarios differ in their representation of saturated hydraulic conductivity. Two InHM simulation scenarios differ with respect to the inclusion of the roads at R‐5. The two models, QPBRRM and InHM, differ strongly in the complexity and number of processes included. For all model simulations we found that errors could be predicted fairly well to very well, based on model output, or based on smooth functions of lagged rainfall data. The errors remaining after recalibration are much more alike in terms of variability than those without recalibration. In this paper, recalibration is not meant to fix models, but merely as a diagnostic tool that exhibits the magnitude and direction of model errors and indicates whether these model errors are related to model inputs such as rainfall. Copyright © 2004 John Wiley & Sons, Ltd.     

Green‐Ampt Curve‐Number mixed procedure as an empirical tool for rainfall–runoff modelling in small and ungauged basins

The Soil Conservation Service Curve Number (SCS‐CN) method is a popular rainfall–runoff model that is widely used to estimate direct runoff from small and ungauged basins. The SCS‐CN is a simple and valuable approach to quantify the total streamflow volume generated by storm rainfall, but its use is not appropriate for estimating the sub‐daily incremental rainfall excess. To overcome this drawback, we propose to include the Green‐Ampt (GA) infiltration model into a mixed procedure, which is referred to as Curve Number for Green‐Ampt (CN4GA), aiming to distribute in time the information provided by the SCS‐CN method. For a given storm, the computed SCS‐CN total net rainfall amount is employed to calibrate the soil hydraulic conductivity parameter of the GA model. The proposed procedure is evaluated by analysing 100 rainfall–runoff events that were observed in four small catchments of varying size. CN4GA appears to provide encouraging results for predicting the net rainfall peak and duration values and has shown, at least for the test cases considered in this study, better agreement with the observed hydrographs than the classic SCS‐CN method. Copyright © 2012 John Wiley & Sons, Ltd.     

Prediction and modelling of rainfall–runoff during typhoon events using a physically-based and artificial neural network hybrid model

Abstract

The accurate prediction of hourly runoff discharge in a watershed during heavy rainfall events is of critical importance for flood control and management. This study predicts n-h-ahead runoff discharge in the Sandimen basin in southern Taiwan using a novel hybrid approach which combines a physically-based model (HEC-HMS) with an artificial neural network (ANN) model. Hourly runoff discharge data (1200 datasets) from seven heavy rainfall events were collected for the model calibration (training) and validation. Six statistical indicators (i.e. mean absolute error, root mean square error, coefficient of correlation, error of time to peak discharge, error of peak discharge and coefficient of efficiency) were employed to evaluate the performance. In comparison with the HEC-HMS model, the single ANN model, and the time series forecasting (ARMAX) model, the developed hybrid HEC-HMS–ANN model demonstrates improved accuracy in recursive n-h-ahead runoff discharge prediction, especially for peak flow discharge and time.     

Efficacy of neural network and genetic algorithm techniques in simulating spatio‐temporal fluctuations of groundwater

Groundwater modelling has emerged as a powerful tool to develop a sustainable management plan for efficient groundwater utilization and protection of this vital resource. This study deals with the development of five hybrid artificial neural network (ANN) models and their critical assessment for simulating spatio‐temporal fluctuations of groundwater in an alluvial aquifer system. Unlike past studies, in this study, all the relevant input variables having significant influence on groundwater have been considered, and the hybrid ANN technique [ANN‐cum‐Genetic Algorithm (GA)] has been used to simulate groundwater levels at 17 sites over the study area. The parameters of the ANN models were optimized using a GA optimization technique. The predictive ability of the five hybrid ANN models developed for each of the 17 sites was evaluated using six goodness‐of‐fit criteria and graphical indicators, together with adequate uncertainty analyses. The analysis of the results of this study revealed that the multilayer perceptron Levenberg–Marquardt model is the most efficient in predicting monthly groundwater levels at almost all of the 17 sites, while the radial basis function model is the least efficient. The GA technique was found to be superior to the commonly used trial‐and‐error method for determining optimal ANN architecture and internal parameters. Of the goodness‐of‐fit statistics used in this study, only root‐mean‐squared error, r² and Nash–Sutcliffe efficiency were found to be more powerful and useful in assessing the performance of the ANN models. It can be concluded that the hybrid ANN modelling approach can be effectively used for predicting spatio‐temporal fluctuations of groundwater at basin or subbasin scales. Copyright © 2014 John Wiley & Sons, Ltd.     

Copula entropy coupled with artificial neural network for rainfall–runoff simulation

The rainfall–runoff relationship is not only nonlinear and complex but also difficult to model. Artificial neural network (ANN), as a data-driven technique, has gained significant attention in recent years and has been shown to be an efficient alternative to traditional methods for hydrological modeling. However, for different input combinations, ANN models can yield different results. Therefore, input variables and ANN types need to be carefully considered, when using an ANN model for stream flow forecasting. This study proposes the copula-entropy (CE) theory to identify the inputs of an ANN model. The CE theory permits to calculate mutual information (MI) and partial MI directly which avoids calculating the marginal and joint probability distributions. Three different ANN models, namely multi-layer feed (MLF) forward networks, radial basis function networks and general regression neural network, were applied to predict stream flow of Jinsha River, China. Results showed that the inputs selected by the CE method were better than those by the traditional linear correlation analysis, and the MLF ANN model with the inputs selected by CE method obtained the best predicted results for the Jinsha River at Pingshan gauging station.     

Predicting catchment flow in a semi‐arid region via an artificial neural network technique

A model of rainfall–runoff relationships is an essential tool in the process of evaluation of water resources projects. In this paper, we applied an artificial neural network (ANN) based model for flow prediction using the data for a catchment in a semi‐arid region in Morocco. Use of this method for non‐linear modelling has been demonstrated in several scientific fields such as biology, geology, chemistry and physics. The performance of the developed neural network‐based model was compared against multiple linear regression‐based model using the same observed data. It was found that the neural network model consistently gives superior predictions. Based on the results of this study, artificial neural network modelling appears to be a promising technique for the prediction of flow for catchments in semi‐arid regions. Accordingly, the neural network method can be applied to various hydrological systems where other models may be inappropriate. Copyright © 2004 John Wiley & Sons, Ltd.     

Hybrid denoising-jittering data pre-processing approach to enhance multi-step-ahead rainfall–runoff modeling

Successful modeling of stochastic hydro-environmental processes widely relies on quantity and quality of accessible data and noisy data might effect on the functioning of the modeling. On the other hand in training phase of any Artificial Intelligence based model, each training data set is usually a limited sample of possible patterns of the process and hence, might not show the behavior of whole population. Accordingly in the present article first, wavelet-based denoising method was used in order to smooth hydrological time series and then small normally distributed noises with the mean of zero and various standard deviations were generated and added to the smoothed time series to form different denoised-jittered training data sets, for Artificial Neural Network (ANN) and Adaptive Neuro-Fuzzy Inference System (ANFIS) modeling of daily and multi-step-ahead rainfall–runoff process of the Milledgeville station of the Oconee River and the Pole Saheb station of the Jighatu River watersheds, respectively located in USA and Iran. The proposed hybrid data pre-processing approach in the present study is used for the first time in modeling of time series and especially in modeling of hydrological processes. Furthermore, the impacts of denoising (smoothing) and noise injection (jittering) have been simultaneously investigated neither in hydrology nor in any other engineering fields. To evaluate the modeling performance, the outcomes were compared with the results of multi linear regression and Auto Regressive Integrated Moving Average models. Comparing the achieved results via the trained ANN and ANFIS models using denoised-jittered data showed that the proposed data pre-processing approach which serves both denoising and jittering techniques could improve performance of the ANN and ANFIS based single-step-ahead rainfall–runoff modeling of the Milledgeville station up to 14 and 12% and of the Pole Saheb station up to 22 and 16% in the verification phase. Also the results of multi-step-ahead modeling using the proposed data pre-processing approach showed improvement of modeling for both watersheds.     

A new magnitude category disaggregation approach for temporal high‐resolution rainfall intensities

Simulation of quick runoff components such as surface runoff and associated soil erosion requires temporal high‐resolution rainfall intensities. However, these data are often not available because such measurements are costly and time consuming. Current rainfall disaggregation methods have shortcomings, especially in generating the distribution of storm events. The objectives of this study were to improve point rainfall disaggregation using a new magnitude category rainfall disaggregation approach. The procedure is introduced using a coupled disaggregation approach (Hyetos and cascade) for multisite rainfall disaggregation. The new procedure was tested with ten long‐term precipitation data sets of central Germany using summer and winter precipitation to determine seasonal variability. Results showed that dividing the rainfall amount into four daily rainfall magnitude categories (1–10, 11–25, 26–50, >50 mm) improves the simulation of high rainfall intensity (convective rainfall). The Hyetos model category approach (Hyetos_Cat) with seasonal variation performs representative to observed hourly rainfall compared with without categories on each month. The mean absolute percentage accuracy of standard deviation for hourly rainfall is 89.7% in winter and 95.6% in summer. The proposed magnitude category method applied with the coupled Hyetos_Cat–cascade approach reproduces successfully the statistical behaviour of local 10‐min rainfall intensities in terms of intermittency as well as variability. The root mean square error performance statistics for disaggregated 10‐min rainfall depth ranges from 0.20 to 2.38 mm for summer and from 0.12 to 2.82 mm for the winter season in all categories. The coupled stochastic approach preserves the statistical self‐similarity and intermittency at each magnitude category with a relatively low computational burden. Copyright © 2014 John Wiley & Sons, Ltd.     

Using recurrent neural networks to reconstruct rainfall‐runoff processes

Many novel techniques for reconstructing rainfall‐runoff processes require hydrometeorologic and geomorphologic information for modelling. However, certain information is not always measurable. In this paper, we employ a special recurrent neural network to reconstruct the rainfall‐runoff process by using collected rainfall data. In addition, we propose an indirect system identification to overcome the drawback of a traditional, time‐consuming trial‐and‐error search. The indirect system identification is an efficient method to recognize the structure of a recurrent neural network. The unit hydrograph can be derived directly from the weights of the network due to a state‐space form embedded in the recurrent neural network. This improves the link between the weights of the network and the physical concepts that most neural networks fail to connect. The case studies of 41 events from 1966 to 1997 have been implemented in Taiwan's Wu‐Tu watershed, where the runoff path‐lines are short and steep. Two recurrent neural networks and one state‐space model are utilized to simulate the rainfall‐runoff processes for comparison. The results are validated by four criteria: coefficient of efficiency; peak discharge error; time to peak arrival error; total discharge volume error. The resulting data from the recurrent neural network reveal that the neural network proposed herein is appropriate for hydrological systems. Copyright © 2005 John Wiley & Sons, Ltd.     

A self‐organization algorithm for real‐time flood forecast

The group method of data handling (GMDH) algorithm presented by A. C. Ivakhnenko and colleagues is an heuristic self‐organization method. It establishes the input–output relationship of a complex system using a multilayered perception‐type structure that is similar to a feed‐forward multilayer neural network. This study provides a step towards understanding and evaluating a role for GMDH in the investigation of the complex rainfall–runoff processes in a heterogeneous watershed in Taiwan. Two versions of the revised GMDH model are implemented: a stepwise regression procedure and a recursive formula. Eleven typhoon events in the Shen‐cei Creek watershed, Taiwan, are used to build the model and verify its usefulness. The prediction results of the revised GMDH models and the instantaneous unit hydrograph (IUH) model are compared. Based on the criteria of forecasting precision and the rate and time of peak error, a much better performance is obtained with the revised GMDH models. Copyright © 1999 John Wiley & Sons, Ltd.     

Wavelet–copula‐based mutual information for rainfall forecasting applications

Under a climate change, the physical factors that influence the rainfall regime are diverse and difficult to predict. The selection of skilful inputs for rainfall forecasting models is, therefore, more challenging. This paper combines wavelet transform and Frank copula function in a mutual information‐based input variable selection (IVS) for non‐linear rainfall forecasting models. The marginal probability density functions (PDFs) of a set of potential rainfall predictors and the rainfall series (predictand) were computed using a wavelet density estimator. The Frank copula function was applied to compute the joint PDF of the predictors and the predictand from their marginal PDFs. The relationship between the rainfall series and the potential predictors was assessed based on the mutual information computed from their marginal and joint PDFs. Finally, the minimum redundancy maximum relevance was used as an IVS stopping criterion to determine the number of skilful input variables. The proposed approach was applied to four stations of the Nigerien Sahel with rainfall series spanning the period 1950–2016 by considering 24 climate indices as potential predictors. Adaptive neuro‐fuzzy inference system, artificial neural networks, and random forest‐based forecast models were used to assess the skill of the proposed IVS method. The three forecasting models yielded satisfactory results, exhibiting a coefficient of determination between 0.52 and 0.69 and a mean absolute percentage error varying from 13.6% to 21%. The adaptive neuro‐fuzzy inference system performed better than the other models at all the stations. A comparison made with KDE‐based mutual information showed the advantage of the proposed wavelet–copula approach.     

An approach for evaluating the hydrological effects of urbanization and its application

This study focuses mainly on observing urban development in Taiwan's Wu‐Tu watershed from the perspective of urban hydrological theory. An approach is proposed for developing a method for incorporating available meteorological data to define the degree of change in a runoff hydrograph for urbanizing basins. The mean rainfall was estimated using the Kriging method. For calibration, two methods of calculating the effective rainfall (the Φ‐index method and the non‐linear‐programming (NLP) method) were used as model inputs, and the optimal global parameters of the linear reservoir model were then obtained from the shuffled complex evolution (SCE) algorithm. Twenty‐six (1966–1991) and eight (1994–1997) rainfall–runoff events were used for calibration and verification, respectively. The NLP method yielded better results than the Φ‐index method, especially for multipeak rainfall–runoff events. The regression equation determined the relationship between the parameters of the model and impervious areas. A comparison based on the results of the instantaneous unit hydrograph of the study area revealed that three decades of urbanization had increased the peak flow by 27%, and the time to peak was decreased by 4 h. The study simply describes the results of the impact of imperviousness on hydrological modelling. Copyright © 2002 John Wiley & Sons, Ltd.