Â鶹´«Ã½AV

Artificial intelligence (AI) is increasingly recognized as a key tool in combating climate change, offering innovative solutions across various sectors. However, the environmental cost of machine learning (ML) models, particularly their carbon footprint, is often overlooked. This study addresses this gap by quantifying the carbon footprint of an AI-powered recommendation system designed to enhance energy efficiency through electricity load shifting. Using a systematic literature review, we identify tools for measuring ML-related COâ‚‚ emissions and apply four of them to an activity-based recommendation system. Our results indicate a net positive environmental impact, defined as the difference between COâ‚‚-equivalent (COâ‚‚eq) emissions generated during computation and the emissions saved through system implementation. Specifically, assuming a 1% adoption rate of the recommended actions over one year, the system yields a net emissions reduction of at least 2,648.63 gCOâ‚‚eq. Notably, implementing just ten recommendations is sufficient to offset the system's own COâ‚‚eq emissions over 365 days. To further evaluate efficiency, we compare the recommendation system using three prediction models - logistic regression, random forest, and a neural network - based on the average energy consumption of a German household in 2021. Our analysis reveals that after 68 implemented recommendations, the neural network-based system surpasses the environmental impact of the random forest-based system. However, for the neural network system to match the impact of the logistic regression model, 132 recommendations must be implemented. These findings suggest that while neural networks require more recommendations to offset their carbon footprint, they still prove to be the most environmentally effective option at higher adoption rates.