The advancement in electrical load forecasting techniques with new algorithms offers reliable solutions to operators for operational cost reduction, optimum use of available resources, effective power management, and a reliable planning process. The focus is to develop a comprehensive understanding regarding the forecast accuracy generated by employing a state of the art optimal autoregressive neural network (NARX) for multiple, nonlinear, dynamic, and exogenous time varying input vectors. Other classical computational methods such as a bagged regression tree (BRT), an autoregressive and moving average with external inputs (ARMAX), and a conventional feedforward artificial neural network are implemented for comparative error assessment. The training of the applied method is realized in a closed loop by feeding back the predicted results obtained from the open loop model, which made the implemented model more robust when compared with conventional forecasting approaches. The recurrent nature of the applied model reduces its dependency on the external data and a produced mean absolute percentage error (MAPE) below 1%. Subsequently, more precision in handling daily grid operations with an average improvement of 16%–20% in comparison with existing computational techniques is achieved. The network is further improved by proposing a lightning search algorithm (LSA) for optimized NARX network parameters and an exponential weight decay (EWD) technique to control the input error weights.