Grid search based parameter tuning of dynamic neural network to forecast daily reservoir inflow

![CDATA[Machine learning algorithms frequently require careful tuning of model parameters. In this paper, a grid search based parameter tuning for neural network was carried out to obtain the best set of parameters for the selected model structure in order to model the inflow of the Kotmale reservoir at Mahaweli river basins. A Non-linear Autoregressive Artificial Neural Network (NAR-ANN) approach is considered which is relatively novel in the related literature to forecast inflow. The original inflow series, first differenced series, In transformed series, first difference of ln transformed series were applied to NAR-ANN through grid search. The NAR-ANN for the original series outperformed the rest. A lookback window consisting with 21 time bands starting from year 2015 backwards until 1995 was analysed in order to find the history of information required for optimum model. Diagnostics of the final model suggested a further improvement to the model that is identified to be done by addressing the heavy noise existing in the inflow series.]]