Partager

Inférence causale probabiliste basée sur des données pour la modélisation du comportement des occupants.

Data-driven probabilistic causal inference for occupant behavior modeling.

Numéro : 3291

Auteurs : KO J., LEE S.

Résumé

Occupant behaviors and decision-making have significant impacts on indoor environmental quality, energy consumption, and greenhouse gas emissions in buildings. Taking account of occupant behavior in building solutions using data-driven methods is important to reduce carbon emissions and simultaneously satisfy occupants’ needs in buildings. Especially, identifying potential causal factors of occupant decision-making is especially imperative to (i) properly intervene in occupant behavior and (ii) improve robustness of building energy solutions. However, investigating underlying causal mechanisms of occupant behavior is challenging. This is because of (i) the difficulty in conducting controlled experiments with real occupants and (ii) the limitation of conventional statistical methods to investigate causality under observational environments. To address such difficulties, this study proposes a
probabilistic causal discovery approach based on a Bayesian model comparison and a Monte Carlo method. With the open dataset provided by a thermostat company, the proposed causal discovery method identified four potential causal variables of the household’s thermostat decision-making. With the inferred causal knowledge, two models, (i) a causal
model including direct causal variables and (ii) a non-causal model involving all available variables, were developed. The prediction performances of the two models were evaluated with two test datasets with and without data shift, where the joint variable distribution of the test dataset is not identical to that of the training dataset. The prediction performances of the two models were similar over the test data without data shift. On the other hand, over the test dataset under data shift, the mean absolute errors of the causal model and the non-causal model were 2.26 ℃ and 3.44 ℃. This showed more robust predictions from the causal model compared with those from the non-causal model, under the data shift.

Documents disponibles

Format PDF

Pages : 10 p.

Disponible

Gratuit

Télécharger

Détails

  • Titre original : Data-driven probabilistic causal inference for occupant behavior modeling.
  • Identifiant de la fiche : 30032942
  • Langues : Anglais
  • Source : 2024 Purdue Conferences. 8th International High Performance Buildings Conference at Purdue.
  • Date d'édition : 15/07/2024

Liens


Voir d'autres communications du même compte rendu (63)
Voir le compte rendu de la conférence

Indexation