10 Janvier – Thesis defense - Huynh Phan

14 h Amphi 2 - Building A9 (University of Bordeaux | Talence campus)

Modelling household activities in instrumented dwellings.

Simulation tools support engineers to size HVAC systems and optimize the building design. It is proved that occupant behaviors are important to explain the discrepancy between simulated and actual energy consumption. Better information on the occupant activities might reduce this gap. Recently, many scientists have focused on modeling occupants' activities to improve the performance of the simulation. Though, an activity is affected by the context in houses, which gives the information to fully understand this activity. To address this question, stochastic approaches were proposed with statistical data to model activity profiles based on household characteristics (building type, income, etc.). Nevertheless, their outputs only provide average information based on different contexts (locations, appliances, etc.). Thus, they are not specific and are unverifiable for particular houses. The agent-based approach is another way to model, in a contextualized manner, the occupants' behaviors in houses. It has been developed to satisfy occupant needs. However, the agent-based approach is challenging to implement in households with many members due to their numerous interactions. Data-based models with measurements have also been proposed to estimate occupant activities. They aim to determine the systematic relationships between variables using machine learning techniques. Due to the measurements, the context is considered. However, most of these models are not analytical and they do not provide occupant-understandable explanations. The problem tackled by this thesis is to use data from a given context, which comes from sensors and questionnaires, to estimate contextualized occupant activities (cooking, sleeping, taking a shower, etc.) and their related energy impacts. The necessary features for the estimation should be determined to reduce data noise and save the instrumentation budget. Activity estimations should consider the context to have more precise results. These approaches must be verifiable and provide understandable information for inhabitants to help them enhance their behaviors in order to reduce energy consumption. In this thesis, a mobile application has been developed to collect labels of activities from the occupants. These labels have then been combined with the data from sensors and the context information of the house to detect and simulate occupants' activities and their related energy impacts. The proposed method focused on three steps:
-estimating occupant activities with sensors, the context, and the activity labels. It starts by determining the most relevant features for the activity estimation. Then, a consequences-based Bayesian Network is built using context factors and the knowledge coming from observations and questionnaires which are given to occupants to ask about their activities. Two houses have been used for testing. Cross-validation and F1-score are used to evaluate the model quality. The results show that it could estimate some activities (cooking, washing dishes, washing clothes, personal care, etc.) with the F1-score from 0.7 to 0.9.
-evaluating the energy impacts of occupant activities with a set of sensors, knowledge, and activity labels. The thesis focuses on the effects of activities on electricity consumption. A Bayesian Network is built for each activity to evaluate its impacts on electricity consumption. The model has been tested with several activities in the studied houses. To validate, the model compares statistically the electricity consumption between activities and between the weekdays.
-simulating occupant activities and their energy impacts, which is particularly useful in the building renovation. Results from previous models are re-used and combined with Monte-Carlo simulation to simulate sequences of occupants' activities and their associated electricity consumption. Comparisons between simulated and actual electricity consumption of appliances are made to verify the model quality.

Event localization