Power Consumption Patterns in Android Devices : A Data-Driven Approach

Abstract

Minimizing power consumption in mobile and embedded systems remains a critical challenge for technology companies, as battery limitations directly affect user experience and device lifespan. While laboratory studies on power consumption provide benchmarks, they fail to capture the diversity of real-world usage. Current research often focuses on single devices or controlled scenarios, leaving a limited understanding of how power consumption patterns vary across heterogeneous Android devices, operating systems, and environments.

This thesis addresses this gap in knowledge through large-scale analysis of power data from Google Pixel and Fairphone devices across multiple Android versions, including stock and /e/OS builds. The study pursued three goals: (1) quantifying median power consumption across devices and versions, (2) identifying behavioral archetypes through clustering, and (3) developing predictive models to compare efficiency across datasets and device generations. The methodology combined descriptive statistics, clustering, and machine learning. Battery discharge logs and system state features were standardized into efficiency and consumption metrics. Principal Component Analysis (PCA) and K-means clustering revealed behavioral patterns, validated by hierarchical methods. Predictive models were built using leakage-safe LightGBM regressions on two datasets: one battery-centric, the other enriched with system-level features.

Results showed strong contrasts: Pixel devices consumed more power with modest efficiency, while Fairphone devices displayed wider variability, with /e/OS variants achieving higher and more consistent efficiency. Clustering revealed stable archetypes ranging from Heavy Active Use – High Drain to Efficient Long-Life – Low Use. Regression models performed strongly on battery-level data (R² ≈ 0.94) but less well on system-enriched data, which reduced stability while improving interpretability. Cross-dataset tests confirmed that battery-centric models generalized well, whereas system-level models did not.

The findings show that power behavior is influenced not only by hardware and Android versions, but also by factors such as stability and usage intensity. The analysis framework reflects the logic of Profilence’s Quality Analytics Suites: whereas those tools isolate the causes of software defects, this framework identifies behavioral archetypes and efficiency drivers, making performance anomalies traceable to their underlying conditions. By integrating unsupervised clustering with supervised prediction, it delivers both diagnostic and predictive insights, enabling targeted optimization, improved device reliability, and extended service life cycles in operational settings.