Towards Reliable Stochastic Data-Driven Models Applied to the Energy Saving in Buildings
We aim at the elaboration of Information Systems able to optimize energy consumption in buildings while preserving human comfort. Our focus is in the use of state-based stochastic modeling applied to temporal signals acquired from heterogeneous sources such as distributed sensors, weather web services, calendar information and user triggered events. Our general scientic objectives are: (1) global instead of local optimization of building automation sub-systems (heating, ventilation, cooling, solar shadings, electric lightings), (2) generalization to unseen building conguration or usage through self-learning data-driven algorithms and (3) inclusion of stochastic state-based modeling to better cope with seasonal and building activity patterns. We leverage on state-based models such as Hidden Markov Models (HMMs) to be able to capture the spatial (states) and temporal (sequence of states) characteristics of the signals. We envision several application layers as per the intrinsic nature of the signals to be modeled. We also envision room-level systems able to leverage on a set of distributed sensors (temperature, presence, electricity consumption, etc.). A typical example of room-level system is to infer room occupancy information or activities done in the rooms as a function of time. Finally, building-level systems can be composed to infer global usage and to propose optimization strategies for the building as a whole. In our approach, each layer may be fed by the output of the previous layers. More specically in this paper, we report on the design, conception and validation of several machine learning applications. We present three different applications of state-based modeling. In the rst case we report on the identication of consumer appliances through an analysis of their electric loads. In the second case we perform the activity recognition task, representing human activities through state-based models. The third case concerns the season prediction using building data, building characteristic parameters and meteorological data.