MODIFIED NARX NETWORK FOR LOW VOLTAGE CONSUMER LOAD PREDICTION

This paper is a continuation of our previous work on Nonlinear autoregressive witheXogenous input (NARX) for load prediction. Application of NARX network inreal-time might be difficult since the tapped delay was selected by trial and error,leading to nine hidden neurons which makes the network complex. This NARXnetwork was modified based on Genetic Algorithm (GA) and Differential Evolution(DE) resulting in a new model coded NARX-DE-GA. GA and DE search for thenumber of hidden neurons and tapped delay automatically. The NARX-DE-GA wasused to predict the consumer load using one month energy data with 8928 datapoints. The results show that NARX- DE -GA outperformed the NARX network.The training mean square error (MSE) value for NARX- DE -GA is 0.0253 whilevalidation is 0.0612. These values are slightly higher when compared with NARXnetwork in previous study which are 0.0225 and 0.0533 respectively. However, thenetwork structure which is one input and output tapped delay, and one hidden neuronis simple and applicable in real time.