Anomalies detection using Ensemble Learning in carbon report data

Details

Supervisors

Deville, Yves

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

In 2022, according to Eurostat, more than 2500 millions of tons of CO2e were emitted by economics activities (industries and households). There is a clear need for carbon accounting/reduction and this is done in carbon reports following specific protocols. However, a tiny error in the data collected or encoded can lead to a huge misunderstanding of the emission sources and the reduction strategy to adopt. Hence, carbon experts are spending a lot of time and efforts on data verification of the carbon reports created by companies. In this work, we propose a tool to detect automatically various pre-determined type of anomalies contained in a carbon report: the Anomalies Detection ToolKit (ADTK). Such data are highly versatile and no clear rules can be established to determine whether a data is an anomaly or not. Therefore, we train machine learning models using Ensemble Learning to build a tool that is able to classify each new record as correct or as anomaly. We were able to establish that for such data, the chosen supervised learning approach outperforms the traditional unsupervised learning that often uses clustering to detect anomalies. Tests computed on real carbon reports filled by Tapio’s clients and exempt from any modifications from carbon experts showed that our tool manages to detect anomalous records with high precision.

Citations

• APA

Mounzer, A. (2023). Anomalies detection using Ensemble Learning in carbon report data. https://dial-mem.test.bib.ucl.ac.be/handle/2078.2/34044