Publication

Abstract : Growth of digital technologies increased the energy consumptions and associated greenhouse emissions, contributing to rising carbon footprints. Accurate assessment and associated prediction of carbon footprints are critical for formulating effective sustainability strategies. This paper analyzes using random forest, Federated learning and contrastive learning for modeling and prediction. Random forest is modeled for capturing nonlinear relationship between the features, which enable robust baseline predictions. Federated learning address privacy by collaborating model across different distributed dataset without centralizing. Contrastive learning enhances the representation by differencing variations in turn improving the generalization across heterogeneous data. The dataset contained life-cycle attributes, manufacturing attributes, and physical attributes of different devices such as laptops, desktops, monitors, and servers associated with Global Warming Potential (GWP). Each model trained and evaluated with K-Fold cross-validation, using various performance metrics, such as mean absolute error (MAE), Root Mean Squared Error (RMSE), coefficient of determination (R2), and prediction and accuracy. Noted estimations were done by using Random Forest model considered as the strong baseline. Federated Learning came with a competitive accuracy while guaranteeing privacy. The triplet loss for Contrastive Learning got the highest unconditional prediction accuracy plus the smallest error boundaries. The study illustrated how successfully representation learning can be used for true mapping and drawing complex relations among features. This analysis enables data driven decision making process to minimize environmental impacts and promote sustainable technology adoption.