Tuning Stacked Auto-encoders for Energy Consumption Prediction: A Case Study

Abstract

Energy is a requirement for electronic devices. A processor is a substantial part of computer components in terms of energy consumption. A great concern has risen over recent years about computers with regard to the energy consumption. Taking accurate information about energy consumption of a processor allows us to predict energy flow features. However, using traditional classifiers may not enhance the accuracy of the prediction of energy consumption. Deep learning shows great promise for predicting energy consumption of a processor. Stacked auto-encoders has emerged a robust type of deep learning. This work investigates the effects of tuning stacked auto-encoder in computer processor with regard to the energy consumption. To search parameter space, a grid search based training method is adopted. To prepare data to prediction, a data preprocessing algorithm is also proposed. According to the obtained results, on average, the method provides 0.2% accuracy improvement along with a remarkable success in reducing parameter tuning error. Further, in receiver operating curve analysis, tuned stacked auto-encoder was able to increase value of are under the curve up to 0.5.