Automating the data-driven predictive control design process for building thermal management

Decarbonisation of the energy sector has brought about a need for intelligent building energy management strategies that can respond optimally to changing external conditions and user requirements. Data-driven modelling methods that can reduce the implementation effort of such strategies have recently gained significant attention, but the inherent variability of building design can mean that a non-trivial parameter selection and tuning process remains. These parameters include continuous and granular variables. This paper proposes a method for automating the control design process for building energy management, enabling predictive control implementation without bespoke system identification or extensive controller tuning effort. This is achieved by combining Data-enabled Predictive Control (DeePC) methods with a parameter selection and tuning procedure using the Mesh Adaptive Direct Search (MADS) algorithm. This selection and tuning procedure is carried out prior to the implementation of the DeePC using data measurements from the building under standard control. The impact of parameter selection on the control performance is shown through a set of case studies using both real data and a simulation environment in which predictive control is used to maintain thermal comfort in a building. Using the autotuned DeePC approach, a reduction of 74% in thermal comfort violation is achieved compared to a classical PI approach, while a reduction of 28% is achieved relative to a DeePC control without autotuning.