Real-world implementation and evaluation of a Model Predictive Control framework in an office space

Abstract

In this work, Model Predictive Control (MPC) is experimentally implemented on a Heating, Ventilation and Air Conditioning (HVAC) system of a large-scale office space. As controller model, a physics-based Modelica building model is calibrated based on historic data pursuing an iterative, nonlinear optimization approach. For the calibration period with a horizon of seven weeks, the calibrated model exhibits a high model accuracy with a Root Mean Square Error (RMSE) of 0.49 °C between the measured and estimated room temperature. The MPC toolchain includes modules for state estimation and forecasts of disturbances quantities (such as outdoor temperature, solar radiation including calculation of the direct and diffuse fraction, supply temperatures and occupancy). The MPC execution comprises the operation of heating based on radiators and floor heating (via regulation of valve openings) as well as shading of three Venetian blind systems (via regulation of vertical position and slat inclination angle). The experiment is conducted during a heating period with a duration of three weeks from October 21 to November 11, 2022. The heating actuators are controlled considering their typical dynamics and take into account the night setback during unoccupied office periods. User acceptance of the automated shading control is included through additional cost function terms for the shading operation. The field test reveals the predictive control capabilities of the proposed MPC toolchain in a real-life scenario, demonstrating energy-efficient building operation and total average discomfort of 0.53 Kh/d. The MPC formulation provides flexibility regarding adjustability of the control towards energy efficiency, thermal comfort, daylight transmission and non-oscillating shading control. Finally, the disturbance forecast accuracies for outdoor temperature and the solar radiation quantities are evaluated and the MPC control performance is compared against a conventional control approach.