Energy Management Based on Charging Behavior Prediction [Report]

NESL Technical Report #: 2013-8-2


Abstract: As mobile device applications continue to include richer content, the need for maximizing power efficiency on the devices increases. While a great deal of effort is put into designing more energy efficient hardware platforms, not much attention has been given to individual users’ charging behaviors. Assuming that every user will charge his/her device differently, we can potentially increase battery life by dynamically scheduling tasks to take place when their impact on the battery is minimal (such as during an overnight charge). In order to allow this optimization, we made use of machine learning algorithms to predict charging behavior. We gathered data using SystemSens, an Android program developed by Dr. Hossein Falaki. Through testing and cross-validation in the machine learning suite Weka, we found that the “RandomTree” classifier yielded the most accurate results in tandem with our feature selection. We used the raw reported values of battery status, battery level, and time, as well as extracted features for the last charging time, last charging level, and duration of last charge. These proved to be the most important features for predicting the time before the next expected charge, giving us approximately 90% accuracy on our training sets. After finalizing our feature and classifier selection, we developed an architecture in which the phone locally polls data, uploads it to a server for training a model, and then receives that model and runs it on the phone, providing a platform for accurately predicting individual charging behavior.

Publication Forum: WHI Summer Intern Final Report

Date: 2013-08-01

Place: Los Angeles, CA

Public Document?: Yes

NESL Document?: Yes

Document category: Report