Ensemble forecast for monthly reservoir inflow : a dynamic neural network approach

Mahaweli cascaded reservoir system is built contiguous to the Mahaweli river, enhancing the water storage and transferring ability to reinforce the needs of water in accordance with the climatic changes. Therefore, modeling the changes of inflow to the upmost reservoir of the system is substantial to effective water management in the whole reservoir system. Modeling the inflow on a monthly basis using the dynamic neural network approach, non-linear autoregressive with exogenous input (NARX-ANN) is presented. The fitted model exhibits satisfactory results while opening paths for further improvement of the model. Further improvements of the can be investigated through fitting the model to a deseasonalised series, changing the training algorithms to choose the best or by adding an MA term as an input (NARMAX). A rainfall measurement that represents the variability of the whole catchment can be obtained to be used as a modified input to the model. Adding a categorical variable to the NARX, which represent the fluctuations in the variable inflow can also be scrutinized.