Towards scalable and generic data-driven solutions for energy flexibility control in buildings

Energy flexibility of an electric utility consumer is the capability of changing its power consumption from the normal consumption pattern. Energy flexibility on the demand side is a major resource for ensuring the grid stability, avoiding activation of carbon-intensive energy resources, avoiding costly investments for the reinforcement of the grid infrastructure and construction of new power plants. In this context, buildings have a great potential due to their thermal inertia (capacity of storing and discharging thermal energy).The objective of this thesis is to study and propose efficient solutions for the estimation and optimal control of energy flexibility in buildings. The diversity of buildings in terms of thermal characteristics, Heating, Ventilation and Air-Conditioning (HVAC) systems, meters, sensors and actuators is considerable. It follows that the flexibility can be generated by different manners, and the exact impact of the different available options is often unknown. Given this complexity, available methods are rarely scalable nor easily implementable.The solutions investigated in this thesis are designed to address real-life implementation issues: availability of data, scalability, genericity, integration to available building management systems, thermal comfort requirements. Consequently, they are based on commonly available data (sensors, meters and actuators) in buildings and do not impact the thermal comfort levels defined by the building managers.A solution that addresses the above objective and implementation constraints is HVAC load control via indoor temperature setpoint adjustments. The first part of this thesis is therefore focused on exploration of methods for day-ahead forecasting of the total HVAC power consumption profile, based on commonly available features: global indoor temperature setpoints, indoor temperature, weather data, time related data. Since indoor temperature reflects the thermal state of the building, similar methods have been investigated for forecasting its dynamics. Cascaded predictions have been tested, which imply using indoor temperature forecasts for predicting the power consumption. Regarding this option, results show that the propagation of the indoor temperature forecasts errors is prohibitive.Based on the first investigations, a predictive model for power consumption forecasting has been implemented, based on supervised learning. To capture the dynamical behavior of power consumption, without forecasting the evolution of indoor temperature, actual and past values of the selected features are used for predicting the actual HVAC load.A second part of the thesis is focused on refining and integrating the predictive model in a demand response framework, consisting in load shedding incentives. Load shedding is achieved using inherent energy storage capability of the building (preheating/precooling). Based on a simplistic demand response program that compensates the shedded load, an optimization-based strategy has been proposed to perform optimized load shedding operations. An estimation of the possible income based on the given program has been carried out.This thesis has been prepared as part of a partnership bewteen GIPSA-lab laboratory, Schneider-Electric and MIAI Grenoble Alpes Institute, supported by Association Nationale de la Recherche et de la Technologie (ANRT).