REAL-TIME FLOW FORECASTING IN THE PARANA RIVER: A COMPARISON BETWEEN ARMAX AND ANN MODELS Pujol Reig, Lucas (1); Ortiz, Enrique (1); Cifres, Enrique (2) and García-Bartual, Rafael (2) (1)
HidroGaia S.L. Avda. Juan de la Cierva 27, 46980 - Paterna - Valencia - Spain phone: +34 96136 6072; fax: +34 96136 6073; emails: lucaspujol@yahoo.com.ar; eortiz@hidrogaia.com (2) Universidad Politécnica de Valencia (Valencia-Spain) phone: +34 963879891 ; fax: +34 963877618; e-mail: enrique@cifres.com ; rgarciab@hma.upv.es
ABSTRACT Linear and nonlinear approaches are compared herein, in terms of their performance in 5 and 10 days ahead flow forecasts for a given section of the Paraná River, located 34 km upstream the city of Santa Fe (Argentina). The models used are an autoregressive moving average with exogenous variable (ARMAX) and a multilayer feedforward artificial neural network (ANN). The variables used as inputs to the models are daily flows at the mentioned section and another cross section located 571 km upstream, together with daily rainfall measured in an intermediate rain gauge station, all during the period 1994-1998. Different configurations of the models have been tested, varying the number of inputs and parameters, but in general the resulting forecast quality with either method is similar and quite satisfactory.Nash Sutcliffe coefficient of efficiency is in all cases over 0.94 for the 10days ahead forecasts in the validation sample. Among the compared models, best results are obtained with the non linear approach ANN. Some uncertainty considerations are also pointed out, after the analysis of the empirical error distributions under different modelling strategies. It is shown that the apparently centred normal distribution expected for errors really departs significantly from zero-average when different parts of the hydrograph are taken into account. Keywords: ARMAX, Neural Network, flow forecast, Real-time 1 INTRODUCTION Forecasting flows and corresponding water levels at different sections of the Paraná River (average flow around 18,000 m3/s), is crucial to minimize the effects of floods affecting properties and human activities. On the other hand, low flows can affect navigation conditions through the river. The National Institute of Water and Hydrologic Alert System for the Plata Basin - INA SiyAH – provides forecasts with time leads of 5 and 10 days, at several river harbours located along the middle and lower Paraná. This anticipation is enough to alert population in different agricultural and urban areas and take the appropriate actions to reduce damages and losses. Such predicted flows estimations are obtained with the hydraulic model Ezeiza-V [Jaime and Menéndez, 1997]. The research reported herein is developed under the same practical framework, applying alternative modelling schemes for an improved flow forecasts in the Paraná River, using exactly the same time horizons. The application of artificial neural networks (ANNs) to various aspects of hydrological modelling has yielded to interesting and promising results during recent years. In particular, rainfall-runoff models and real time forecasting models based on ANN’s schemes have received special attention [Kang et al., 1993; Karunanithi et al., 1994; Smith and Eli, 1995; Minns and Hall, 1996; Shamseldin, 1997; Fernando and Jayawardena, 1998; Sajikumar and Thandaveswara, 1999; Tokar and Markus, 2000; Kumar and Minocha, 2001; García-Bartual, 2002; Jain and Prasad, 2003; Abrahart et al., 2004; Cigizoglu, 2005; Sahoo et al., 2006]. Their ability to incorporate in a systematic approach non linear relationships between variables