This article was downloaded by: [Mine, Miriam Rita Moro] On: 2 May 2011 Access details: Access

Details: [subscription number 936895525] Publisher Taylor & Francis Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 3741 Mortimer Street, London W1T 3JH, UK

Hydrological Sciences Journal

Publication details, including instructions for authors and subscription information: http://www.informaworld.com/smpp/title~content=t911751996

Monthly rainfall-runoff modelling using artificial neural networks

Fernando Machadoa; Miriam Minea; Eloy Kaviskia; Heinz Filla a Departamento de Hidrulica e Saneamento - Programa de Ps-graduao em Engenharia de Recursos Hdricos e Ambiental, Universidade Federal do Paran, Curitiba PR, Brasil Online publication date: 26 April 2011

To cite this Article Machado, Fernando , Mine, Miriam , Kaviski, Eloy and Fill, Heinz(2011) 'Monthly rainfall-runoff

modelling using artificial neural networks', Hydrological Sciences Journal, 56: 3, 349 361 To link to this Article: DOI: 10.1080/02626667.2011.559949 URL: http://dx.doi.org/10.1080/02626667.2011.559949

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf This article may be used for research, teaching and private study purposes. Any substantial or systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or distribution in any form to anyone is expressly forbidden. The publisher does not give any warranty express or implied or make any representation that the contents will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses should be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or arising out of the use of this material.

Hydrological Sciences Journal Journal des Sciences Hydrologiques, 56(3) 2011

349

Monthly rainfallrunoff modelling using articial neural networks

Fernando Machado, Miriam Mine, Eloy Kaviski & Heinz Fill
Departamento de Hidrulica e Saneamento - Programa de Ps-graduao em Engenharia de Recursos Hdricos e Ambiental, Universidade Federal do Paran, Caixa Postal 19011, CEP 80531-990 Curitiba PR, Brasil mrmine.dhs@ufpr.br
Received 22 August 2008; accepted 22 November 2010; open for discussion until 1 October 2011 Citation Machado, F ., Mine, M., Kaviski, E. & Fill, H. (2011) Monthly rainfallrunoff modelling using articial neural networks. Hydrol. Sci. J. 56(3), 349361.
Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

Abstract Rainfallrunoff models usually present good results, but parameter calibration sometimes is tedious and subjective, and in many cases it depends on additional data surveys in the eld. An alternative to the conceptual models is provided by empirical models, which relate input and output by means of an arbitrary mathematical function that bears no direct relationship to the physical characteristics of the rainfallrunoff process. This category includes the articial neural networks (ANNs), whose implementation is the main focus of this paper. This study evaluated the capacity of ANNs to model with accuracy the monthly rainfallrunoff process. The case study was performed in the Jangada River basin, Paran, Brazil. The results of the three ANNs that produced the best results were compared to those of a conceptual model at monthly time scale, IPHMEN. The ANNs presented the best results with highest correlation coefcients and Nash-Sutcliffe statistics and the smallest difference of volume.
Key words articial neural networks; rainfallrunoff models; empirical models, Brazil

Modlisation pluie-dbit mensuelle laide de rseaux de neurones articiels

Rsum Les modles pluie-dbit donnent en gnral de bons rsultats, mais le calage des paramtres est parfois fastidieux et subjectif, et dpend souvent de donnes additionnelles observes in situ. Une alternative aux modles conceptuels est fournie par les modles empiriques qui relient des donnes dentre et de sortie par lintermdiaire dune fonction mathmatique arbitraire sans relation directe avec les caractristiques physiques du processus de transformation pluiedbit. Cette catgorie inclut les rseaux de neurones articiels (RNAs), dont limplmentation est le sujet de cet article. Cette tude value la capacit de RNAs modliser avec prcision le processus pluie-dbit mensuel. Une tude de cas a t ralise pour le bassin de la Rivire Jangada, Paran, Brsil. Les rsultats des trois RNAs les plus performants ont t compars avec ceux du modle conceptuel pas de temps mensuel, IPHMEN. Les RNAs donnent les meilleurs rsultats, avec les plus forts coefcients de corrlation et statistiques de Nash-Sutcliffe et les plus faibles diffrences de volume.
Mots clefs rseaux de neurones articiels; modles pluiedbit; modles empiriques; Brsil

INTRODUCTION The purpose of this study is to apply an articial neural network (ANN) to monthly rainfallrunoff modelling in the Jangada River basin, a left bank tributary of the Iguau River, in the state of Paran, Brazil, and to validate the results by confronting them with the results of the IPHMEN conceptual model, a monthly rainfallrunoff model developed by the Instituto de Pesquisas Hidrulicas (IPH) at the Federal University of Rio Grande do Sul, Brazil. Transforming rainfall into runoff is a difcult process to formulate, due to the large number of
ISSN 0262-6667 print/ISSN 2150-3435 online 2011 IAHS Press doi: 10.1080/02626667.2011.559949 http://www.informaworld.com

variables that are relevant, which change in both space and time. Evaluating this process with accuracy is what allows rational management of the different water uses, such as supply, irrigation, electric power generation, as well as forecasting of extreme ood events and dry periods. Generally mathematical models known as rainfallrunoff models perform the evaluation of this process. Rainfallrunoff models may be divided into two major groups: conceptual and empirical models. The conceptual models describe mathematically the processes of the hydrological cycle based on physical

350

Fernando Machado et al.

laws governing each of these processes. However, despite the generally good results that are achieved, some aspects of the conceptual models are challenging. Calibration is not easy and, in many cases, depends on eld surveys of data that are scarce. Also, the use of basin averages for relevant parameters together with the nonlinear character of those processes leads to additional difculties. These characteristics often render the implementation of the conceptual model difcult and nancially burdensome. Empirical models are an alternative to the conceptual models. The main characteristic of this type of model consists of establishing a stable relationship between input and output variables without accounting to the physical laws that govern the natural processes when rainfall is transformed into runoff. These models are easy to apply and supposedly cheaper. Examples of these models are multivariable equations with parameters estimated by least squares, and articial neural networks (ANNs). An ANN is a structure of elements formed by nodes or neurons, mathematically interconnected, representing a function. The coefcients and intercepts of the input variables of this function are called weights and biases. There are different types of ANN, and the most common is the ANN multilayer perceptron (MLP), with the neurons distributed in layers, usually three of them (Galvo et al., 1999). Herein, the ANN nomenclature is adopted to designate an ANN multilayer perceptron (MLP). Haykin (1994), Galvo et al. (1999) and Fernandes et al. (1996) mention that a three-layer ANN can approach any function with nonlinear characteristics. In water resources, ANNs have been used to solve several problems: inow forecasting and reservoir operation (Jain et al., 1999), simulating and optimizing reservoir operation (Neelakantan & Pundarikanthan, 2000), tting rating curves (Machado et al., 2005), among others. One major application of ANNs in hydrology is related to streamow or rainfall forecasting. Recent contributions of Cigizoglu (2003a,b, 2005a,b), Jain et al. (1999) and Partal & Cigizoglu (2009) deal with this topic. Another application has been the estimation of sediment transport (Cigizoglu, 2004; Cigizoglu & Alp, 2006; Cigizoglu & Kisi, 2006; Alp & Cigizoglu, 2007). A comprehensive review of ANN applications in hydrology can be found in the publication of the ASCE Task Committee (2000). The modelling of the rainfallrunoff process by ANN has been used extensively for at least 15 years and is the main subject of this paper.

Hsu et al. (1995) used an ANN to model the relationship between rainfall and daily runoff in the Leaf River, Collins, Mississippi, USA. In order to evaluate the ANN performance, they compared the results obtained to the results of the SACRAMENTO model and autoregressive models of the ARMAX type. In this situation, ANNs presented the smallest errors and the smallest differences in volume. Good results were also obtained by Elshorbagy et al. (2000), Anmala et al. (2000) and Tokar & Markus (2000). Elshorbagy et al. (2000) applied ANN to ow forecasting for ood control in the Red River, Canada, and compared the results to linear and nonlinear multiple regression models; Anmala et al. (2000) applied ANN to three basins in Kansas, USA, to model the mean monthly ow, and also conpared the results with empirical models; and Tokar & Markus (2000) applied the ANN to three different USA basins: the Fraser River, Colorado, Raccoon Creek River, Iowa and the Little Patuxent River in Maryland, for monthly and daily forecasting horizons. Modarres (2009) tested multilayer perceptron models in the Zayandehrud watershed in Iran. In all cases, the ANNs presented very good results, superior to the models to which they were compared. Rajurkar et al. (2002) presented an application of ANNs for the Narmada River in India (17 157 km2 ) using sub-area linearly transformed rainfall data as input. The output was daily ows during ood events. Six events were used for calibration and four for verication. The Nash-Sutcliffe statistic for the best case was 83.2%. Another ANN application, also for the daily rainfallrunoff relationship, but for seven different catchments around the world with drainage areas varying from 1207 to 26 200 km2 , was published by Rajurkar et al. (2004). That study used effective rainfall data (rainfall minus evapotranspiration) at current and antecedent time as input. For the larger basins, rainfall in sub-areas was also considered. The results were very reasonable, with Nash-Sutcliffe statistics generally higher than 75%. In previous studies it was observed that, in most cases, the ANNs are applied to shorter-term modelling or forecasting, with hourly or daily scales. There are few studies that have investigated the application of ANNs to longer-term modelling. Lungu (cited in Anmala et al., 2000) states that among the daily, monthly and annual scales, the monthly rainfallrunoff relationship is probably the most difcult since it has to take into account both short-term and long-term hydrological processes.

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

Monthly rainfallrunoff modelling using articial neural networks

351

In order for an ANN to be able to model properly the rainfallrunoff process, it should undergo calibration, or process called training, in which its weights and biases are tted. During the training, paired sets of inputs and outputs are presented to the ANN. Based on a specic set of weights and biases, the ANN outputs are calculated and compared to the observed output; if the deviation exceeds an allowable value, the weights and biases are corrected and new outputs are computed until the deviation is smaller than, or equal to, the allowable value. This process is controlled by special optimization algorithms referred to as back-propagation, which include the descending gradient with momentum and the LevenbergMaquardt method. During the training process, care should be taken with the ANN architecture, the number of iterations, the initialization of weights and the length of series for training. Architecture is the way the neurons are distributed among the layers. It is the architecture that denes the functional form of the ANN. The ANN architecture has been investigated by: Kadowaki & Andrade (1997), Ballini et al. (1997), Campolo et al. (1999), Tokar & Johnson (1999), Thirumalaiah & Deo (2000), Lima & Ferreira Filho (2003) and Ramos & Galvo (2001). Tokar & Johnson (1999) performed an evaluation comparing the length of the data series and the number of ANN inputs, and suggested that any increase in the number of ANN inputs should comply with an increase in the lengths of the data series. Lima & Ferreira Filho (2003), studying the semi-arid region of Cear, Brazil, evaluated different combinations of the number of inputs and the number of neurons in the hidden layers. Twenty-four ANNs were evaluated, each of them trained with three different sets of data. Lima & Ferreira Filho (2003) did not make recommendations about the ideal architecture, but presented a good method to map it. Ramos & Galvo (2001) proposed a methodology to determine the ANN architecture, based on the initialization of weights, and changes in the transfer and training functions with data series of different lengths. The number of iterations is the number of times the ANN is trained, or the number of iterations of the optimization algorithm to determine weights and biases. Anmala et al. (2000) analysed the inuence of the number of iterations, but did not make any special recommendation. It is known that, in the case of an excessive number of iterations, the ANNs memorize the sample data and do not generalize the problem proposed. This process is call overtting.

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

Initialization represents the set of initial values for the weights and biases at the beginning of training. Ramos & Galvo (2001) comment that, as a good practice, the initialization of weights and biases should be repeated. The length of the series represents the size of the data series used during ANN training. Different lengths of series were considered by Lima & Ferreira Filho (2003) and Sajikumar & Thandaveswara (1999). The latter applied temporal ANNs in rainfallrunoff modelling, in the Lee River basin, UK, and Thuthapuzha, in Kerala, India, and compared the results obtained with the results of other empirical models. The ANNs presented the best results. Only a few studies have considered monthly rainfallrunoff modelling using ANNs. The specic contribution of this study, beside modelling and evaluating the monthly rainfallrunoff process in the Jangada River basin, is to check the inuence on the accuracy of results of: the ANN architecture, the number of iterations, initialization and the length of the data series during training.

DESCRIPTION OF THE RIVER BASIN The case study was developed for the Jangada River basin, at the Jangada stream gauging station. The Jangada River is located close to the cities of Unio da Vitria (PR) and Porto Unio (SC), between latitude 26 20 26 40 S and longitude 51 30 51 10 W (Fig. 1). The choice of the Jangada River basin for this study was based on three main factors: (a) the size of the basinthe drainage area at the Jangada stream gauging station is 1055 km2 and, since IPHMEN is a concentrated conceptual model, larger basins could produce distortions of the results (larger river basins are better represented by distributed models); (b) the station densitythe Jangada River basin has ve rainfall gauging stations within the basin, two nearby, and four stream gauging stations, one of which is remotely monitored; and (c) the quality and length of the data seriesdata for the period March 1976 February 1986 were used in the training, and data for March 1976July 1984 were used in the validation of the ANN model. Table 1 presents the relevant statistics of the rainfallrunoff data. Data on precipitation, potential evapotranspiration and mean monthly discharge were selected for the case study. All the data were analysed for consistency:

352

Fernando Machado et al.

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

Fig. 1 Jangada River basin. The rainfall, ow and meteorological stations are referenced by 8-digit code. Table 1 Statistics of rainfallrunoff data for the period March 1976July 1994 (221 months).
Rainfall (basin average) (mm/month) Mean St. dev. Skewness Kurtosis Autocorrelation (lag 1) Cross-correlation Rainfallow Flowevapotranspiration 146.9 77.7 1.03 5.39 0.01 0.76 Flow (m3 /s) 25.9 21.0 2.9 18.07 0.35 Evapotranspiration (mm/month) 82.4 33.1 0.14 1.93 0.74

0.29

(i) in the case of precipitation, the erroneous data were dropped, and the overall behaviour was evaluated using the cumulative double mass curve. One of the seven stations available presents a large number of gaps and was not used in this study. The mean precipitation was computed using the Thiessen method based on six stations. Data of 221 months were used, beginning in March 1976 and ending in July 1994; (ii) evapotranspiration was computed based on the Penman method recommended by Muller (1995) for the State of Paran. For meteorological data on temperature, wind velocity, humidity and hours of sunshine, the Teixeira Soares meteorological station (02550025, Fig. 1) was usedit is located nearby, with characteristics similar to those of the Jangada River basin; and (iii) mean monthly discharge refers to the data from the Jangada stream gauging station (65370000,

Fig. 1) with observations since 1945. To evaluate the overall behaviour, the basin response was assessed qualitatively (visually) on a rainfall event basis. ARTIFICIAL NEURAL NETWORK MODELLING General Basically, an articial neuron is composed of three elements: (a) a set of weights and biases, responsible for ANN learning; (b) summation units computing the linear combination of the inputs; and (c) a transfer function, which determines the response. The most frequently used transfer functions are the sigmoid and linear functions (Haykin, 1994). Usually, the values are normalized, thus transforming the real data into a scale compatible with the characteristics of the transfer functions.

Monthly rainfallrunoff modelling using articial neural networks

353

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

Although other ANN methods are reported in the literature, the ANN employed in this study uses the feed-forward back-propagation method. In this method, the connection of several neurons is distributed across layers. Within the ANN, the data ow in a single direction (feed-forward), i.e. the input data are propagated through the ANN, layer by layer, in the forward direction. The inputs in the input layer are multiplied by the weights of the respective connections. Each neuron in the middle layer receives a linear combination of the input elements. This combination generates a stimulus to the transfer function that emits an output. The responses of the transfer functions are the inputs to the next layer. The input to the output layer is the linear combination of the outputs from the middle layer. The output from the output layer is the ANN response (Haykin, 1994). In mathematical terms, the output from a three-layer ANN is represented by equation (1). yk =
j =1 q p

wjk = (H wjk )1 ( wjk )

(2)

where (H | wjk ) is the Hessian matrix of the objective function; and ( | wjk ) is the gradient of the objective function. The objective function is quadratic in the following form:
n

( n) =
k =1

[ek (n)]2

(3)

where ek (n) is the error between the desired output and the output calculated by the ANN in the output layer, and n is the length of the data series. For this function, Hagan & Menhaj (1994) proposed to compute the corrections of the weights and biases by: wjk = [J T (n)J (n) + I ]1 J T (n)e(n) (4)

wji xi + bj + bk (1)

wkj
i=1

where x are the input elements, w are the weights between the connections, b the biases, p the number of neurons in the input layer, q the number of neurons in the middle layer, is the transfer function, y is the ANN output and i, j and k are neurons of the input, middle and output layers, respectively. Choosing an ANN to solve a problem consists of solving two sub-problems: the choice of the functional form of f (x,w), and estimating the weights vector W (Fernandes et al., 1996). From a statistical perspective f (x,w) is a regression function used to t a vector of inputs, X , to a vector of outputs, D. The elements x are the exogenous variables and w the set of parameters. Thus, the function f (x,w) represents a family of curves, and the statistical problem is to obtain the optimum estimator W , which will minimize the sum of square of the residues. Several specic optimization algorithms for an ANN to t the weights are available in the literature. In this study, the Levenberg-Maquardt method (Hagan & Menhaj, 1994) was used.

where J is the Jacobian matrix containing the rst derivatives of the error function with respect to the weights; > 0 is a parameter, and T denotes the transpose of a matrix. ANN training The problem of training an ANN to solve a rainfall runoff type problem is to t a suitable function to a data sample. The rainfallrunoff process is nonlinear, and the functional form for the t is unknown. In this case, the aim of applying an ANN is not only to t the best weights and biases to the sample of observed data, but also to investigate, by varying the ANN architecture, which is the best functional form for the observed data. The ANNs used are of the three-layer MLP type. The input layer does not have transfer functions. All neurons of the middle layer and the output layer have a transfer function of the sigmoid and linear type, respectively. In order to investigate the best functional form, 24 ANNs were created with variations in the number of inputs (six combinations) and the number of neurons in the middle layers (four alternatives). The output is always runoff. Each combination of inputs was referred to as a model, and Table 2 shows the input and output for each of these models. Input variables are precipitation (P), evapotranspiration (EVT) and runoff (Q), all of them expressed as mean monthly values. The precipitation and evapotranspiration were

Levenberg-Maquardt optimization The Levenberg-Maquart optimization method is an extension of the Newton-Raphson method (Machado, 2005):

354

Fernando Machado et al.

Table 2 Models proposed.

Model 1 2 3 4 5 6 Inputs P(t), EVT(t) P(t), EVT(t), Q(t 1) P(t 1), P(t), EVT(t 1), EVT(t) P(t 1), P(t), EVT(t 1), EVT(t), Q(t 1) P(t 2), P(t 1), P(t), EVT(t 2), EVT(t 1), EVT(t) P(t 2), P(t 1), P(t), EVT(t 2), EVT(t 1), EVT(t), Q(t 2), Q(t 1) Outputs Q(t) Q(t) Q(t) Q(t) Q(t) Q(t)

P: monthly precipitation (mm/month); EVT: potential evapotranspiration (mm/month); Q: mean monthly discharge (m3 /s).

evaluated at the current and previous time steps, and runoff at the previous time step is used as input, because it is believed that this variable gives indication about the moisture content of the basin. This belief is supported by the strong autocorrelation of mean monthly discharge series. The use of this input variable poses no problem to the runoff generation process, because the streamow values are generated sequentially over time. For each model, three, ve, eight and 10 neurons were used in the middle layer, making a total of 24 ANN combinations. Varying both the number of neurons in the middle layer and the number of inputs, allows the evaluation of ANN sensitivity in terms of its architecture. During the training of the 24 ANNs, the following parameters were considered: length of data series; number of iterations (epochs); and initialization of weights. Figure 2 shows a schematic representation of the training process with variation of those

parameters. Before the training, all inputs were normalized between 0.1 and 0.9, according to Sajikumar & Thandaveswara (1999). All the ANNs were trained with three different sets of data. From the 221 months available, sets of 60, 120 and 180 items were used for training, and 161, 101 and 41 items for validation. These lengths of the data series were chosen according to Lima & Ferreira Filho (2003). An excessive number of iterations during training puts the calculated values very close to the observed values and does not generalize the process. Because the Levenberg-Maquardt training algorithm converges rapidly, all 24 pre-established ANNs were trained arbitrarily over 30, 60 and 90 iterations. The weights and biases were initialized at interval [1,1]. During the training process, the values of the weights and biases change in order to reduce the error. When the initial weights and biases are on a point close to a local minimum, the optimization process will inevitably converge the solution to the local minimum. To avoid this problem, the ANNs are initialized six times, at different values. During the training, the combination of the ANN architecture, the input sets, the number of initializations and the number of iterations generated a total of 1296 results for analysis. In order to evaluate the inuence of all elements proposed in the ANN training, an algorithm was created in MATLAB software, that manipulates the data, trains, simulates, computes the statistics of the results, and stores all responses in an output le. The statistics used at this time were the correlation coefcient and the percentage difference of the volumes.

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

RESULTS OF ANN Sensitivity analysis Only 108 of the 1296 results were selected for further analysis. They were selected based on the following criterion: for each initialization, model and input set, the model with the best performance according to the correlation coefcient in the validation was selected. Each model is represented by four ANNs with three, ve, eight and 10 neurons in the middle layer and 30, 60 and 90 inputs during the training. This means there are 12 possible results for each model, and the best of these has been selected for each model. Based on the 108 best results, ve different analyses were performed:

Fig. 2 Training method.

Monthly rainfallrunoff modelling using articial neural networks

355

Table 3 Frequency of the best results in terms of the number of parameters and the number of iterations.
Weights + Bias Number of inputs during training: 30 015 1630 3145 4660 6175 7690 91105 Total 5 26 25 7 2 6 1 72 60 0 5 7 1 1 1 0 15 90 6 4 5 4 2 0 0 21 11 35 37 12 5 7 1 108 Total

Fig. 3 Frequency of the number of iterations.

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

(a) Number of iterations to achieve the best results The frequency of the best result for each number of iteration was calculated. It was observed that 72, about 67%, achieved the best results when they were trained only for 30 iterations; 14 or about 13% for 60 iterations, and 22 or 20% for 90 iterations (see Fig. 3). (b) Relationship between the number of iterations and the length of data series The number of iterations for the best result was also evaluated considering the input data. It was believed that for more input data, the number of iterations during training should be greater (see Fig. 3). However, Fig. 3 shows that the different input sets used in the training did not inuence the number of iterations. For all inputs, the best results were achieved by the models trained only 30 times. (c) Relationship between the number of weights and biases of the ANN and the number of iterations Analysis was done to identify whether the number of weights and biases inuences the number of iterations. In order to determine the total number of weights and biases of an ANN, it is enough to know the number of connections, the number of neurons in the middle layer and the number of neurons in the output layer. The simplest ANN in this study has a total of 13 weights and biases, and the most complex has 101. In order to evaluate the relationship between the number of iterations and the number of weights and biases of the ANNs, they were divided into seven

classes. For each interval, their frequencies were computed among the 108 best results, as shown in Table 3. Independent of the number of iterations, the best results were most frequent for ANNs with weights and biases ranging between 16 and 45. (d) Analysis of the inuence of re-initializing the weights and biases in the training process All 24 ANNs were initialized six times. The rst initialization is represented by the letter A, and the last by F. Initialization A does not refer to a particular set of values, but rather the rst initialization for a particular ANN. Table 4 shows the average, maximum and minimum R2 values and the Nash-Sutcliffe efciency coefcient (NS) for these groups of models, both for training and for validation. In order to evaluate the effect of initialization, the best model was chosen for each pair of initialization and input set. The results are shown in Table 5, together with the respective values of R2 and NashSutcliffe coefcient (NS) for the validation period. It may be observed that initialization inuences the choice of both model and architecture. Also, for the same input the results present a certain consistency. For Input 1, independent of initialization, Model 2 with three or 10 neurons in the middle layer presented the best performance. For Input 2, models 4 and 5 presented the best results, and for Input 3, it was Model 4 with eight neurons in the middle layer. Most of the models listed in Table 5 were trained over 30 iterations with the number of weights and biases varying from 15 to 57, consistent with the results of sub-sections (a) and (c) above. (e) Relationship between the number of ANN inputs and length of data series The relationship

356

Fernando Machado et al.

Table 4 Average, maximum and minimum R2 and Nash-Sutcliffe (NS) values.

Epoch Weight+Bias Training: 30 R2 Average: 015 1630 3145 4660 6175 7690 91105 Maximum: 015 1630 3145 4660 6175 7690 91105 Minimum: 015 1630 3145 4660 6175 7690 91105 0.77 0.88 0.88 0.89 0.92 0.92 0.94 0.81 0.93 0.95 0.94 0.93 0.95 0.94 0.65 0.70 0.75 0.77 0.92 0.88 0.94 NS 0.74 0.83 0.84 0.87 0.87 0.90 0.94 0.81 0.93 0.95 0.93 0.88 0.95 0.94 0.62 0.62 0.75 0.77 0.86 0.81 0.94 60 R2 NS 90 R2 0.66 0.86 0.87 0.87 0.94 NS 0.63 0.82 0.86 0.85 0.94 Validation: 30 R2 0.56 0.72 0.68 0.76 0.74 0.74 0.68 0.67 0.79 0.80 0.79 0.76 0.76 0.48 0.50 0.51 0.74 0.72 0.70 0.68 NS 0.50 0.70 0.71 0.80 0.71 0.74 0.67 0.66 0.76 0.89 0.90 0.74 0.91 0.30 0.34 0.35 0.72 0.65 0.67 0.67 60 R2 NS 90 R2 0.67 0.76 0.76 0.79 0.71 NS 0.66 0.78 0.80 0.79 0.70

0.89 0.91 0.94 0.89 0.94

0.87 0.87 0.94 0.89 0.87

0.71 0.74 0.76 0.70 0.75 . 0.76 0.79 0.76 0.70 0.75

0.77 0.75 0.78 0.70 0.74 0,00 0.86 0.89 0.78 0.70 0.74

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

0.92 0.95 0.94 0.89 9.94

0.92 0.93 0.94 0.89 0.87

0.79 0.93 0.95 0.94 0.96 0.68

0.63 0,90 0.95 0.93 0.95 0.67 0.62 0.71 0.74 0.71 0.93

0.69 0.77 0.82 0.82 0.75

0.68 0.88 0.86 0.92 0.73

0.86 0.88 0.94 0.84 0.94

0.84 0.82 0.94 0.84 0.87

0.63 0.71 0.74 0.71 0.93

0.69 0.71 0.76 0.70 0.75

0.64 0.65 0.78 0.70 0.74

0.57 0.75 0.73 0.73 0.68

0.55 0.74 0.70 0.72 0.68

Table 5 Inuence of initialization on the determination of the best model for each input set.
Init. Input 1 Model A B C D E F 5 2 2 2 2 2 ARC 3 10 8 3 10 3 R2 0.75 0.82 0.82 0.78 0.75 0.76 NS 0.73 0.81 0.79 0.76 0.72 0.78 Input 2 Model 4 5 5 4 5 4 ARC 3 3 3 5 3 3 R2 0.76 0.72 0.77 0.76 0.75 0.79 NS 0.75 0.72 0.76 0.77 0.76 0.78 Input 3 Model 4 4 4 4 4 4 ARC 8 8 5 5 8 8 R2 0.79 0.81 0.80 0.79 0.78 0.80 NS 0.76 0.72 0.77 0.89 0.90 0.92

Init.: initialization; ARC: architecture of the middle layer; NS: Nash-Sutcliffe efciency coefcient.

between the number of inputs and the length of the data series was also analysed. According to Table 5, for Input 1, Model 2 presented the best result in most cases. This model presents only three inputs: P(t), EVT(t) and Q(t 1). For inputs 2 and 3, models 4 and 5 gave the best results, with ve and six inputs, respectively: Model 4: P(t 1), P(t), EVT(t 1), EVT(t) and Q(t 1); and Model 5: P (t 2), P(t 1), P(t), EVT(t 2), EVT(t 1) and EVT(t). Models with a greater number of inputs only achieved good results when the number of data used in the training was increased from 60 to 120 or 180, i.e. the relationship between the input variables

was better identied with more data. The same behaviour was also reported by Tokar & Johnson (1999). Statistical analysis Three models were selected for statistical and graphical analysis of the results, one for each input set. For Input 1, Model 2 was the best model, for all initializations except one. For inputs 2 and 3, models 4 and 5 performed best (see Table 5). The best performance for Input 1 using Model 2, was the one trained on initialization B, shown on Fig. 4.

Monthly rainfallrunoff modelling using articial neural networks

357

From Fig. 4, it can be observed that, during training, Model 2 showed a relatively good representation of low runoff, but some dispersion occurred from medium to high runoff values. The tendency was to minimize these ows. A particular feature of this model is that, despite the ANN being trained using predominantly low runoffs, during validation it did very well in forecasting higher runoffs, although it preserved the tendency to under-predict the latter. The difference of volume in training was 0.03%, and in validation 0.17%. The R2 value and the NashSutcliffe coefcient during validation were 0.82 and 0.81, respectively. During training the values were 0.71 for both statistics. For Input 2, Model 4 or Model 5 should be selected. Model 4, trained in F initialization, presented the best result, as shown in Fig. 5. For this case, Model 4 gave excellent performance during training. Both R2 and the NS were 0.92. The correlation coefcient was 0.96 and the difference in volume 0.2%. During training, both the low and high runoffs were well represented. However, during validation, R2 and NS dropped to 0.79 and 0.78, respectively. The model also presented a strong tendency to reduce the highest runoffs. The best result based on R2 value for Input 3 was obtained with Model 4 and initialization B. The performance for this case is shown in Fig. 6. Among all models, based on R2 values both for training and validation that shown in Fig. 6, was chosen as the model to represent the relationship between rainfall and monthly runoff through the ANNs. The model performed well, for both low and high runoffs. The R2 values in training and validation were respectively equal to 0.94 and 0.81. The difference in volume was insignicant in training and 1.28% in validation. The Nash-Sutcliffe coefcient, NS was 0.88 and 0.72 during training and validation, respectively.

Fig. 4 Best result for Input 1 ANN.

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

Fig. 5 Best result for Input 2 ANN.

THE IPHMEN MODEL General The IPHMEN model (Tucci, 1998) is a conceptual monthly rainfallrunoff model which includes three algorithms to calculate volume distribution, surface runoff and groundwater runoff, and an algorithm for parameter optimization. There are seven parameters: a, b and c from Hortons inltration equation; related to evapotranspiration; Ks, a surface propagation coefcient; Kb, groundwater reservoir variation

Fig. 6 Best result for Input 3 ANN.

358

Fernando Machado et al.

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

time; and Cr, a surface runoff factor. The volume distribution performs the determination of runoff and inltration volume. Inltration is determined by Hortons equation, and the runoff volume by the continuity equation. The inltrated volume less the naturally evaporated and percolated volumes reduces the inltration until the soil is saturated. At that time all precipitation turns into surface runoff. The approach using a monthly scale distorts the inltration equation parameters, since the process occurs at a much smaller time scale. Even so, the model provides a basis to evaluate inltration losses over time. When precipitation amount is less than the inltration capacity, all the water inltrates. However, in this case, considering the spatial and temporal distribution within the month, at a given time and place, precipitation may be greater than the inltration capacity, and surface runoff occurs. In order to simulate this behaviour, the parameter Cr is used to compute the surface runoff from a relationship between precipitation and inltration (equation (7)). The higher the value of Cr, the lower the surface runoff. The inltration is determined by the difference between precipitation and surface runoff: Vs = P (P/I )2 P/I + Cr (7)

Table 6 Fitted parameters of the IPHMEN model.

Input 160 1120 1160 a 317.5 216 313 b 2.7 3.4 3.2 c 0.39 0.18 0.25 Ks 0.4 0.46 0.39 Kb 2.0 2.0 2.0 Cr 0.8 0.65 0.67 a 5.2 4.3 4.4

of IPHMEN itself. In this study calibration was made by trial and error. Three different calibrations were performed, one for each set of inputs of 60, 90 and 120 data items, with the objective to compare the results obtained by both models. The tted parameters are shown in Table 6. Conceptual model results Just as for the ANN, the results for the IPHMEN model are presented and analysed statistically. The result for the rst calibration used 60 items of data. It can be seen that IPHMEN provides a good simulation of low and high ows, both in calibration and in validation. The volume difference increases from 5.6% in calibration to 9.3% in validation. The correlation coefcients in calibration and validation were, respectively, 80.7% and 87.4%. In the result of the calibration with 120 items of data, it can be observed that, even with a signicant change in the parameter values, the results remained representative. During calibration, the model ts well the data for all ows, and the correlation coefcient is 91.1%. The dispersion is acceptable and there is a compensation of volumes. The difference in volumes was equal to 1.3% in calibration. In validation the errors increase, with a correlation coefcient of only 76.4%, but the differences in volume are still satisfactory, reaching 2.7%. The calibration of IPHMEN with 180 items of data shows the worst performance. The correlation coefcients were equal to 88.8% in calibration and 68.7% in validation. The attempt to t a particular point of the hydrograph, or to provide good compensation for the volumes throughout the series, generated a reasonable result for calibration, but during validation it showed a sharp drop in performance. In calibration there is a slight tendency to increase the low runoffs and to reduce high runoffs. The difference in volume was 4.81%. In validation, a few low ows are well represented, but, in general, the result was poor and the difference in volume increases to 5.99%.

where Vs is surface runoff volume, P is precipitation and I is inltration. The second module of the IPHMEN model uses the runoff and inltration volumes to determine the amount of surface and groundwater ow. The third module is a parameter optimization module using the absolute or relative error function. This module is optional.

Model calibration There are two alternatives to calibrate the IPHMEN model: using the parameter optimization module of the model itself, or tting the parameters manually by trial and error. The manual calibration is preferred sometimes because hydrologists with experience prefer to t the parameters from scratch by trial and error. In addition, since the optimization is a mathematical process in some cases, parameter values without any physical meaning may result. Galvez et al. (2001) performed calibration based on the optimization routine

Monthly rainfallrunoff modelling using articial neural networks

359

Table 7 Summary of the results of ANN and IPHMEN for calibration.

Period analysed 160 1120 1180

ANN model: R
2

IPHMEN model: DV (%) 0.03 0.20 0.00 R2 0.65 0.76 0.79 NS 0.62 0.83 0.78 DV (%) 5.60 1.30 4.81

NS 0.71 0.92 0.88

0.71 0.92 0.94

The period analysed refers to months 160, 1120 and 1180. DV: difference of volumes.

Table 8 Summary of statistical results of ANNs and the IPHMEN model.

Period analysed 61221 121221 181221 ANN model: R2 0.82 0.78 0.81 NS 0.81 0.78 0.72 DV (%) 0.17 5.55 1.28 IPHMEN model: R2 0.76 0.58 0.47 NS 0.75 0.52 0.28 DV (%) 9.38 2.78 5.99

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

COMPARING THE RESULTS The summary of results for the three input sets, for ANNs and the IPHMEN model during calibration and in validation is presented in Tables 7 and 8, respectively. The ANN model used in this comparison was always that which performed best for each input (period analysed). Thus: for Input 1, Model 2 with initialization B was used; for Input 2, Model 4 with initialization F; and for Input 3, Model 4 with initialization B. Independent of the input set used, the ANNs always presented a better performance than the IPHMEN model for both calibration and validation. The correlation coefcients were higher at all times and the difference in volumes was always smaller, except for the series with 121 to 221 data. In this case, during validation, the ANNs had a very good approach to low ows, but reduced the high ows. The IPHMEN model does approach the average runoff well, but presents a large dispersion of the computed ows. This different behaviour is reected in the statistics. The ANNs present a good correlation coefcient because, for low ows, the computed values are close to the observed ones; however, the differences in volume are large because the same model reduces the high ows without compensating them. The IPHMEN model presents a small difference in volumes because of compensation throughout the entire range of ows, but the correlation coefcient is low. Three other specic features were observed when comparing the results:

1. Neither model represented the high ows well. This behaviour was recorded for both the ANNs and the IPHMEN model, in training, calibration, as well as validation. 2. Neither model was able to simulate well a particular reach in the hydrograph, between intervals 150 and 180. The analysis of precipitation and streamow data observed for this part of the hydrograph showed inconsistency in the data, accounting for the fact that neither model was able to approach computed and recorded ows. However, when this information was used in training the ANN, training with Input 3, the ANN ts the computed wrong observed ows. This did not happen in the case of the IPHMEN model, which, was not able to approach the computed and the observed hydrograph during calibration. 3. The ANNs and the IPHMEN model using Input 1 presented higher correlation coefcients in validation than in the training or calibration processes. This probably results from a numerical coincidence associated with the future hydrograph, resulting in a better approach between the computed and observed ows for both models.

CONCLUSIONS The main objective of this study was to model the monthly rainfallrunoff process using ANNs. For this purpose an algorithm using the MATLAB software was developed, which considers the architecture, the inuence of the number of iterations, the initialization, and the length of data series while training the ANNs. To attain that objective it was necessary to study the inuence of these elements on the performance of the ANNs. For this purpose 24 four ANNs were trained, generating 1296 results with the 108 best results selected for analysis. In order to compare these results, three ANNthe best among the 108 pre-selectedwere selected, each of them trained with set of different inputs using 60, 120 and 180 items of data. All of these three ANNs were able to identify clearly the relationship between rainfall and monthly discharge in the Jangada River basin, but Model 4, trained with 180 items of data, in initialization B, was chosen as the best model. This model has ve inputs: Q(t) = f (P(t 1), P(t), EVT(t 1), EVT(t), Q(t 1)), and eight neurons in the intermediate layer. It was trained 90 times and has a total number of 57 weights and biases.

360

Fernando Machado et al.

Cigizoglu, H. K. (2005a) Application of the generalized regression neural networks to intermittent ow forecasting and estimation. J. Hydrol. Engng ASCE 10(4), 336341. Cigizoglu, H. K. (2005b) Generalized regression neural network in monthly ow forecasting. Civil Engng Environ. Systems 22(2), 7184. Cigizoglu, H. K. & Alp, M. (2006) Generalized regression neural network in modelling river sediment yield. Adv. Engng Software 37(2), 6368. Cigizoglu, H. K. & Kisi, O. (2006) Methods to improve the neural network performance in suspended sediment estimation. J. Hydrol. 317(3-4), 221238. Elshorbagy, A., Simonovic, S. P. & Panu, U. S. (2000) Performance evaluation of articial neural networks for runoff prediction. J. Hydrol. Engng ASCE 5(4), 424427. Fernandes, L. G. L., Navaux, P. O. A. & Portugual, M. S. (1996) Previso de sries de tempo: redes neurais articiais e modelos estruturais. Pesquisa e Planejamento Econmico, PPE 26(2), 253276. Galvo, C. O., Valena, M. J. S., Vieira, V . P. P. B., Diniz, L. S., Lacerda, E. G. M. de, Carvalho, A. C. P. L. F. & Ludermir, T. B. (1999) Sistemas Inteligentes: aplicaes a recursos hdricos e cincias ambientais. Porto Alegre, Brazil: Universidade/ UFRGS/ABRH. Galvez, J. J. O., Pinto, C. A. L. & Parraga, E. O. (2001) Anlisis hidrometeorologico y aplicacin del modelo de simulacin IPHMEN en la cuenca del Rio Pachitea. Santiago, Chile: III Encuentro de Las Aguas, 3. Hagan, M. T. & Menhaj, M. B. (1994) Training feedforward networks with the Marquardt algorithm. neural networks, IEEE Trans. 5(6), 989993. Haykin, S. (1994) Neural Networks A Comprehensive Foundation. New York: Macmillan College Publishing Co.. Hsu, K. L., Gupta, H. V . & Sorroshian, S. (1995) Articial neural network modelling of the rainfallrunoff process. Water Resour. Res. 31(10), 25172530. Jain, S. K., Das, A. & Srivastava, D. K. (1999) Application of ANN for reservoir inow prediction and operation. J. Water Resour. Plan. Manage. ASCE 125(5), 263271. Kadowaki, M. & Andrade, M. G. (1997) Previso de vazes mensais utilizando redes neurais multicamadas com algoritmo backpropagation. Simpsio Brasileiro de Redes Neurais. Goiana, Brazil. Lima, C. H. R. & Ferreira Filho, W. M. (2003) Anlise de modelos de redes neurais aplicados ao processo chuva-devio no semi-rido. XV Simpsio Brasileiro de Recursos Hdricos. Curitiba. Relao de trabalhos. Curitiba: Associao Brasileira de Recursos Hdricos, ABRH. CD ROM. Machado, F. W. (2005) Modelagem chuva-vazo mensal utilizando redes neurais articiais. MSc Thesis, Universidade Federal do Paran, Curitiba, Paran, Brazil. Machado, F. W., Santos, I., Perreira Filho, D. L. B. & Mine, M. R. M. (2005) Avaliao do ajuste e extrapolao de curvas de descarga atravs de redes neurais. XX Congreso Nacional del Agua CONAGUA & III Simposio de Recursos Hidricos del Cono Sur. Mendonza, Argentina. Modarres R. (2009) Multi-criteria validation of articial neural network rainfallrunoff modeling. Hydrol. Earth System Sci. 13(3), 411421. Muller, I. I. (1995) Mtodos de avaliao da evaporao e evapotranspirao: anlise comparativa para o Estado do Paran. MSc Thesis, Universidade Federal do Paran, Curitiba, Paran, Brazil. Neelakantan, T. R. & Pundarikanthan, N. V . (2000) Neural networkbased simulation-optimization model for reservoir operation. J. Water Resour. Plan. Manage. ASCE 126(2), 5764.

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011

The good result achieved is associated with the methodology used. During training it was observed that the number of inputs is directly related to the length of the data series, and that an increased number of inputs should be followed by an increase in the length of the data series used. This is a qualitative approach and should be further investigated in future research. The number of iterations should be small and is not related either to the total number of weights and biases, or to the length of the data series. This fact was associated with the fast convergence of the Levenberg-Maquardt training algorithm. The best results were obtained for the ANNs trained 30 times with weights and biases in total of between 15 and 57 (see Table 3). The initialization of weights and biases did not inuence the best result, but should be considered, since in some cases the solution may converge to a local minimum. Three sets of parameters have been established to calibrate the IPHMEN model with the objective to simulate a real situation in which the conceptual model is calibrated by trial and error. In this situation, the three values of the parameters are unknown with the main focus being to approach the computed ows to the observed ones. Comparing the results of the ANNs with the IPHMEN model showed that the ANNs were superior in reproducing the observed ows.

REFERENCES
Alp, M. & Cigizoglu, H. K. (2007) Suspended sediment estimation by feed forward back propagation method using hydrometeorological data. Environ. Model. Software 22(1), 213. Anmala, J., Zhang, B. & Govindaraju, R. S. (2000) Comparison of ANNs and empirical approaches for predicting watershed runoff. J. Water Resour. Plan. Manage. 126(3), 156166. ASCE Task Committee (2000) Articial neural networks in hydrology II. J. Hydrol. Engng ASCE 5(2), 124132. Ballini, R., Franca, E., Kadowaki, M., Soares, S. & Andrade, M. (1997) Modelos de redes neurais e Box e Jenkins para previso de vazes mdias mensais. XII Simpsio Brasileiro de Recursos Hdricos. Goiana. Relao de trabalhos. Goiana: Associao Brasileira de Recursos Hdricos, ABRH. CD ROM. Campolo, M., Andreussi, P. & Soldati, A. (1999) River ood forecasting with a neural network model. Water Resour. Res. 35(34), 11911198. Cigizoglu, H. K. (2003a) Incorporation of ARMA models into ow forecasting by articial neural networks, Environmetrics 14(4), 417427. Cigizoglu, H. K. (2003b) Estimation, forecasting and extrapolation of ow data by articial neural networks. Hydrol. Sci. J. 48(3), 349361. Cigizoglu, H. K. (2004) Estimation and forecasting of daily suspended sediment data by multilayer perceptions. Adv. Water Resour. 27, 185195.

Monthly rainfallrunoff modelling using articial neural networks

Partal, T. & Cigizoglu, H. K. (2009) Prediction of daily precipitation using wavelet-neural networks. Hydrol. Sci. J. 54(2), 234246. Rajurkar, M. P., Kothyari, U. C. & Chaube, U. C. (2002) Articial neural network for daily rainfallrunoff modelling. Hydrol. Sci. J. 47(6), 865877. Rajurkar, M. P., Kothyari, U. C. & Chaube, U. C. (2004). Modeling of daily rainfallrunoff relationship with articial neural network. J. Hydrol. 285, 96113. Ramos, A. M. & Galvo, C. O. (2001) Estabilidade de redes neurais em modelagem hidrometeorolgica. XIV Simpsio Brasileiro de Recursos Hdricos, Aracaj. Relao de trabalhos. Aracaj: Associao Brasileira de Recursos Hdricos, ABRH. CD ROM.

361

Sajikumar, N. & Thandaveswara, B. S. (1999) A non-linear rainfall runoff model using articial neural network. J. Hydrol. 216 (1-2), 3255. Thirumalaia, K. & Deo, M. (2000) Hydrological forecasting using neural networks. J. Hydrol. Engng ASCE 5(2), 180189. Tokar, A. S. & Johnson, P. A. (1999) Rainfallrunoff modelling using articial neural networks. J. Hydrol. Engng ASCE 4(3), 232239. Tokar, A. S. & Markus, M. (2000) Precipitation runoff modelling using articial neural networks and conceptual models. J. Hydrol. Engng ASCE 5(2), 156161. Tucci, C. E. M. (1998) Modelos Hidrolgicos. Porto Alegre, Brazil: Universidade/UFRGS/ABRH.

Downloaded By: [Mine, Miriam Rita Moro] At: 15:05 2 May 2011