Application of neural approaches to one-step daily flow forecasting in Portuguese watersheds

Abstract

Listen

Translate article

Summary Since a few years ago, the computational neural networks (CNNs) had become one of the most promising tools for stream flow forecasting. However, most of the works presented in the literature is focused in watersheds that, besides stream flow records, generally utilizes as inputs records of other hydro-meteorological variables. In this study the performance of feed forward CNNs to forecast one-day ahead daily flows at large Portuguese watersheds considering that only flows in previous days are available for the calibration of the models were analyzed. For that purpose several CNN approaches were implemented and compared. Besides the CNN having as inputs the flows in previous days or those flows plus differenced flow data also in previous days, auxiliary inputs were used to apply intervention series to the CNN predictor model; a convolution process of the input variables was carried out; and a hybrid methodology combining CNN and ARIMA models was applied. The CNN having inputs of flows in the three previous days combined with a convolution process of the input sequence proved to be able to provide very accurate estimates of the daily flows. A preliminary analysis of the capability of that approach to forecast daily flows at a watershed different from the one considered in the calibration of the parameters of the model was also carried out. The results showed that it is also possible to get daily flows forecasts at watersheds with insufficient flow data.