To reach the objectives of reducing the energy consumption and increasing the flexibility of buildings energy demand, it is necessary to have load forecast models easy to adapt on site and efficient for the implementation of energy optimization and load shedding strategies. This thesis compares several inverse model architectures ("black box", "grey box"). A 2nd order semi-physical model (R6C2) has been selected to forecast load curves and the average indoor temperature for heating and cooling. It is also able to simulate unknown situations (load shedding), absent from the learning phase. Three energy optimization and load shedding strategies adapted to operational constraints are studied. The first one optimizes the night set-back to reduce consumption and to reach the comfort temperature in the morning. The second strategy optimizes the set-point temperatures during a day in the context of variable energy prices, thus reducing the energy bill. The third strategy allows load curtailment in buildings by limiting load while meeting specified comfort criteria. The R6C2 model and strategies have been faced with a real building (elementary school). The study shows that it is possible to forecast the electrical power and the average temperature of a complex building with a single-zone model; the developed strategies are assessed and the limitations of the model are identified.
