Abstract
An input delayed neural network (IDNN) for synthetic inflow generation is presented to establish monthly inflow scenarios for the reservoir that supplies water to the city of Matsuyama, Japan. IDNNs are dynamic networks capable of accounting for nonlinearities and representing temporal information of input sequences. In this study, the IDNN model relates the two previousreservoir inflows in order to estimate the current inflow. The inflow scenarios will be used as input to optimization models in order to construct reservoir operation policies. Twenty years of historical inflows were used for calibrating the IDNN and a new 20-year synthetic series was generated. Besides the comparison with the IDNN-generated inflows, the statistics of historical series were also compared with those of synthetic series generated by a second-order autoregressive (AR-2) model. The IDNN modelproved to be capable of preserving the main statistical characteristics of the historical series.
