Nine different ANN models are developed for the Kangsabati reservoir, India with a various sets of input and target variables. A three-layer fully connected feed forward multilayer network is used for all models. As for the input and target dataset required for training of each model, optimal solutions obtained from the deterministic DP model with generated inflow sequences are used. Back propagation algorithm is used during training. Regression parameters obtained from training under each model, are used to identify the better model. For this purpose, a new parameter $\phi$ is defined to select the best model performed. Simulation is done using the observed inflow record of 32 years, resulting nine monthly release policies which are compared with the release policy obtained from conventional DP and through error analysis in terms of statistical parameters. Finally, the best model is selected. The performance of the selected model is analysed by Monthly surplus/deficit plot in Kharif \& Rabi season. Overall, the performance of the selected ANN Model is satisfactory and acceptable as an optimising model for reservoir operation.