UC Davis

The energy consumed by the residential and commercial building sectors in the United States has been increasing at around 1.3\% per year over the past decade, making efficient building operations more crucial than ever. Model predictive control (MPC), which is a model-based control method, has been proposed as a solution for the control and optimization of building operations due to its ability to optimize control actions based on constraints such as cost and energy. However, widespread adoption of MPC in buildings is limited by the challenges in developing and training a control-oriented building model. Building modeling is a challenging task due to the presence of unmeasured time-varying heat disturbances due to people, lighting, and electricity, and the lack of full state measurements, resulting in a coupled state, disturbance, and model parameter estimation problem.

Despite being unmeasured, these time-varying heat disturbances are correlated to certain time-features like the time of the day and the day of the week for several building types and occupancy patterns. Hence, hybrid models, which combine physics-based models, to capture the underlying dynamics of the system, and data-driven models, used to forecast the disturbances, have been proposed as a potential method for control-oriented modeling of buildings. In our previous work, a low-order thermal resistance-capacitance network was formulated to capture the dynamics of the building space and a feedforward neural network (FNN) was used to forecast the time-varying unmeasured disturbances.

This thesis presents a generalized hybrid modeling framework to identify models for systems that are subject to unmeasured time-varying disturbances. The proposed hybrid modeling framework combines a parameterized low-order physics-based model and a feedforward neural network (FNN) and utilizes a novel three-step training methodology to simultaneously estimate both the physics-based and FNN model parameters. The aim of the three-step training methodology is to provide better model predictions compared to the predictions made by the same model trained with alternative strategies. A model validation approach is also provided as part of the training methodology. The effectiveness of the proposed modeling and training approach is demonstrated by applying it to model the thermal dynamics of a building space. The time features, which provide the desired model predictions, are first determined. The superiority of the three-step training methodology is demonstrated by comparing the predictions generated by the models trained with alternative strategies to those generated by the model trained using the three-step training methodology. These results demonstrate that the hybrid modeling framework is suitable for modeling systems with unmeasured time-varying disturbances, and that the three-step training methodology results in models with minimal prediction errors, with fewer number of iterations as compared to its alternatives. The impact of unavailability of full state measurements is studied. Finally, the ability for the hybrid modeling framework to reproduce the results is evaluated.