Battery life is one of the main concerns for today’s mobile users. Due to an imbalance in demand and supply of energy in mobile devices, the burden to make battery last longer upon each charge has shifted towards application developers, who, in turn attempt to create energy efficient applications. However, mobile developers lack the tools to detect energy consumption hot-spots in their code. We aim to provide developers with a technique that helps them to precisely locate energy hot-spots at the method-level. In this paper we present MLEE, a novel approach for estimating energy consumption of methods. MLEE uses machine learning models to predict the energy consumption at method-level using software metrics as features. We use the Snapdragon power profiler to measure the energy consumption of applications using the shortest time interval to develop a method-level energy dataset for training machine learning prediction models. We demonstrate that several structural metrics of methods are highly co-related with energy consumption. Thereafter we use these features to predict the energy consumption of methods using linear regression, random forest and decision tree with an average mean-absolute-error of $2.6e^{-2}$ J.

电池寿命是当今移动用户的主要关注点之一。由于移动设备的能源需求和供应不平衡，每次充电后延长电池续航时间的负担已转移到应用程序开发人员身上，他们反过来又试图创建节能应用程序。然而，移动开发人员缺乏检测代码中能源消耗热点的工具。我们旨在为开发人员提供一种技术，帮助他们在方法级别精确定位能量热点。在本文中，我们介绍了 MLEE，这是一种估算方法能耗的新方法。MLEE 使用机器学习模型，使用软件指标作为特征来预测方法级别的能源消耗。我们使用 Snapdragon 功率分析器使用最短的时间间隔来测量应用程序的能耗，以开发用于训练机器学习预测模型的方法级能量数据集。我们证明了方法的几个结构性指标与能源消耗高度相关。此后，我们使用这些特征来预测使用线性回归、随机森林和决策树的方法的能量消耗，平均绝对误差为$2.6{\mathrm{电子}}^{-2}$ J。