| A new methodology for generating synthetic streamflow series was investigated; it was sought that the statistical properties—specially the drought statistics—of the historical series could be better reproduced than by using the traditional linear models. As a result, a stochastic multivariate non-linear model based on multilayer perception artificial neural networks was built, which fitted the proposed objective. Once the model formulation was developed, it was used to analyze four case studies, in which ARMA and disaggregation models were also applied. The case studies consisted of generating a number of synthetic monthly streamflow series by the neural network model; then, identical calculations were carried out using ARMA models and disaggregation models, in order to compare their statistic preservation capabilities. Afterwards, the probabilistic simulation of the management of several water resources systems was performed, considering states of drought. The results yielded by the neural network model show that it outperforms the traditional linear models, and has a longer hydrologic memory than these latter. Therefore, the neural network model is a high performance new alternative into the field of time series synthetic generation. |
