Energy Management Based On Charging Behavior Prediction [Report]

NESL Technical Report #: 2013-8-1


Abstract: As mobile applications transition into providing richer content and supporting more use-cases, it has become increasingly apparent that smart-phones are constrained by battery life. Industry has put great effort into energy-aware hardware platforms, but little attention has been given to designing energy efficiency around user charging behavior. If smart-phones were to become capable of accurately predicting the time and duration of charging events of specific users, task scheduling could be more intelligently catered towards the user. Our project examines whether machine learning can be reliably implemented on smart-phones to infer aspects of charging behavior, as well as the appropriate classification schemes required to accurately predict future charging events. By implementing SystemSens, we have developed an alternative model to offload model training to a server and provide user-specific models to the phone. The new application, Tree-Diagram, manages classification on the client-side and keeps track of prediction accuracy. Many server side scripts also do additional processing that produce attribute-relation files that we can analyze externally. Using the machine learning suite Weka[1], we found that the “Random Tree” classifier yielded the most accurate results initially. Later analysis showed that classifier strength varies across different users and over time as the datasets grow. We found that the prediction capability was strongest when using the attributes of time, battery level, charging status, and statistics of the most recent charging session. Our application utilized our classifier and predicted the time until the user’s next charge accurately, and without high-resource use.

Publication Forum: WHI Summer Intern Final Report

Date: 2013-08-01

Publisher: Los Angeles, CA

Public Document?: Yes

NESL Document?: Yes

Document category: Report