Evaluating End-User Device Energy Models in Sustainability Reporting of Browser-Based Web Services

TL;DR

This work interrogates the accuracy of prevalent end-user energy models used in digital sustainability reporting (Digst, DIMPACT) by empirically comparing measured energy during realistic browser interactions to model estimates. Using four laptops and eight representative flows across four non-streaming categories (shopping, booking, navigation, news), it tests the common $E_{model} = P_{avg} \times t$ assumption and contrasts fixed constants of $P_{Digst} = 22~\text{W}$ and $P_{DIMPACT} = 15~\text{W}$ against actual device behavior. The study finds systematic underestimation by both models, with errors that grow with task duration, and shows that incorporating context in the form of hardware and category factors dramatically improves estimation accuracy ($R^2$ approaching $0.994$). The results argue for category-aware and device-reflective power parameters to enable fair, transparent, and reproducible sustainability reporting, while maintaining simplicity for cross-service comparisons. The authors provide open data and code to support reproducibility and outline future work to broaden hardware coverage and include additional service classes such as streaming and AI-assisted interactions.

Abstract

Sustainability reporting in web-based services increasingly relies on simplified energy and carbon models such as the Danish Agency of Digital Government's Digst framework and the United Kingdom-based DIMPACT model. Although these models are widely adopted, their accuracy and precision remain underexplored. This paper presents an empirical study evaluating how well such models reflect actual energy consumption during realistic user interactions with common website categories. Energy use was measured across shopping, booking, navigation, and news services using predefined user flows executed on four laptop platforms. The results show that the commonly applied constant-power approximation (P * t) can diverge substantially from measured energy, depending on website category, device type, and task characteristics. The findings demonstrate that model deviations are systematic rather than random and highlight the need for category-aware and device-reflective power parameters in reproducible sustainability reporting frameworks.

Figures (9)

• Figure 1: The setup for the experiment with 4 Laptops, two Siglents and two data-collecting computers with user-controlled synchronization software.
• Figure 2: Measured versus modeled energy consumption for the DIMPACT framework. The red dashed line indicates perfect prediction ($y = x$).
• Figure 3: Measured versus modeled energy consumption for the Digst framework. The stronger deviation from the reference line reflects systematic underestimation.
• Figure 4: Relationship between task duration and absolute percentage error in DIMPACT energy estimates ($\beta = 5.0\,\text{J s}^{-1}, R^{2} = 0.86$).
• Figure 5: Relationship between task duration and absolute percentage error in Digst energy estimates ($\beta = 11.8\,\text{J s}^{-1}, R^{2} = 0.96$).