Effect of input variables selection on energy demand prediction based on intelligent hybrid neural networks

Numerous techniques have been applied by the researchers to predict the future electrical energy demand, which can be broadly categorized as parametric (statistical) and non-parametric (intelligent) techniques. The non-parametric or intelligent methods which are based on artificial intelligence are gaining a lot of attention during the recent past years. As compared to the other intelligent techniques, the Artificial Neural Network (ANN) has a tendency to map and memorize the non-linear relations between inputs and output variables. Because of this ability, they are extensively implemented in modern predictive model development. The efficacy of these models depends upon many factors such as, neural network architecture, type of training algorithm, input training and testing data set and initial values of synaptic weights. Among the others, the selection of most influential input variables has a critical effect on the forecast results. In this paper, the important issues related with the best input variable selection for a hybrid model is addressed. A hybrid approach that combines ANN and an evolutionary optimization technique, genetic algorithm (GA) is used for the development of a short term load forecast (STLF) model. GA and correlation analysis are used for the selection of the most influential input variables for the training and testing of the hybrid model. Multiple cases are developed using different optimally selected input variable vectors to train and test the back propagation neural network (BP-NN) and the hybrid model. The results show that hybrid forecast model provide better performance when it is trained and tested with optimally selected input variable vector (IV), containing historical load and meteorological variables. The proposed input variable selection approach not only improves that forecast accuracy but also reduces the computational efforts and training time of forecasting models. © 2006-2015 Asian Research Publishing Network (ARPN).