Occupancy forecasting for building energy MPC

Energy consumption of buildings contributes significantly to the global energy usage. Conventional Rule Based Control (RBC) ensures maintaining comfort constraints for occupants even during unoccupied times. The energy usage could, however, be decreased by relaxing the comfort constraints for unoccupied periods. A common approach to solve this problem is to employ Model Predictive Control (MPC) with a preliminary fixed occupancy schedule to impose time-dependent temperature constraints. This is, however, ineffective as prolonged unoccupied periods during working hours are not effectively exploited and comfort constraints are not ensured for occupants working outside working hours. It is, therefore, key to develop a data-driven method for occupancy forecasting for a widespread application of MPCs. We compare feedforward Artificial Neural Networks (ANNs) and Long Short-Term Memory Neural Networks (LSTMs) with a direct multi-step prediction approach and discuss the effects of different engineered features, considering domain knowledge. Results show that engineered features have a beneficial effect on ANNs and that ANNs show superior performance compared to LSTMs. On the contrary, LSTMs do not improve their prediction ability by using engineered features. We validate the derived ANN through real-life experiments, in which one multi-person office and three single offices are controlled by an MPC framework. A novel method to impose temperature constraints based on the predicted probability of occupancy is presented and used to define occupancy-dependent temperature constraints on the MPC. The experiments demonstrate reduced energy usage in comparison to both, a conventional RBC and a reference MPC experiment using a scheduled occupancy profile. Comfort constraint violations are reduced by 15% by the novel occupancy-informed MPC framework in comparison to the reference MPC experiment