par Levasseur, Guillaume 
Président du jury Birattari, Mauro
Promoteur Bersini, Hugues
Publication Non publié, 2023-11-13
Président du jury Birattari, Mauro
Promoteur Bersini, Hugues
Publication Non publié, 2023-11-13
Thèse de doctorat
| Résumé : | In the Anthropocene, droughts are caused by human-made water scarcity. Water scarcity arises when the demand for freshwater exceeds the renewable supply and constrains the freshwater available to households. On the other hand, human activities and practices at home have a direct impact on residential water demand and on product footprint. Water disaggregation models allow the end-use classification of water meter data. It offers the opportunity to quantify the water consumption of each activity at scale and understand their impact. In this thesis, my objective is the development of classification algorithms to disaggregate human activities in smart water meter data alone. In the process, I make interdisciplinary contributions. First, I investigate the best heuristic to find the optimal window size when applying deep neural networks on real-world time series. Then, I compare self-reporting tools for the annotation of human activities, while collecting a new dataset of residential water meter records, with my colleagues at Procter & Gamble. My results indicate that Internet-connected buttons are better tools than voice diaries to record human activities in daily life settings. Using these data, I show that existing water disaggregation algorithms do not transferto the disaggregation of human activities. Therefore, I characterize the problem and list the open challenges for future research. Finally, I conclude with perspectives to improve the generalization and performance of disaggregation algorithms. |
