Impact of the length of observed records on the performance of ANN and of conceptual parsimonious rainfall-runoff forecasting models

Although attractive to hydrologists, artificial neural network modeling still lacks norms that would help modelers to create and train efficient rainfall-runoff models in a systematic way. This study focuses on the impact of the length of observed records on the performance of multiplelayer perceptrons (MLP), and compare their results with those of a parsimonious conceptual model equipped with an updating scheme. Both models were calibrated (trained) for one-day-ahead stream flow predictions. 92 different model realizations were obtained for 1-, 3-, 5-, 9-, and 15-year time sub-series created from a 24-year training set, shifting by a one-year sliding window. All the model realizations were verified against the same 7-year test set. The results revealed that MLP stream flow mapping was efficient as long as wet weather data were available for the training; the longer series implicitly guarantee that the data contain valuable information of the hydrological behavior; the results were consistent with those reported for conceptual rainfall-runoff models. The physical knowledge in the conceptual models allowed them to make much better use of 1-year training sets than the MLPs. However, longer training sets were more beneficial to the MLPs than to the conceptual model. Both types shared best performance about evenly for 3and 5-year training sets, but MLPs did better whenever the training set was dominated by wet weather. The MLPs continued to improve for input vectors of 9 years and more, which was not the case of the conceptual model.