An Effective Optimizer based on Global and Local Searched Experiences for Short-term Electricity Consumption Forecasting

Abstract

A precise forecasting for the future short-term electricity consumption will be quite useful for making a good plan for the power demand management. Since deep neural network (DNN) provides an effective way for the short-term load forecasting, one of the focuses of this research is thus to use it to construct the prediction model. The gradient-based optimizers (GBOs) have been widely used in DNN algorithms in recent years; however, they make it easy for DNN to trap in poor local regions during the training process. In this research, we propose a new optimizer that is based on the searched experiences to solve this problem to enhance the performance of GBOs. More precisely, the proposed optimizer integrates the best position searched, Lévy flight, and gradient descent to preserve not only the diversification but also the intensification of search during the training process. To evaluate the performance of the proposed optimizer, we compare it with several state-of-the-art GBOs for DNN; namely, Adagrad, RMSprop, and Adam, for training a forecasting model for the short-term electricity consumption forecasting problem. The simulation results show that the proposed optimizer outperforms all the other GBOs in terms of the mean absolute percentage error.