(12e) Efficient Training of Machine-Learning Enhanced Building Models with Noisy Data

In 2020, the combined energy consumption of the residential and commercial building sectors in the United States was around 20 quadrillion Btu [1]. The residential and commercial sectors consumed around 22% and 18%, respectively [1]. With surging energy prices and rising demand [2], efficient building operations are more critical than ever. Model predictive control (MPC) is a potential game-changing control and optimization approach for buildings, and numerous studies have analyzed the potential positive impacts of MPC on building operations [3]. However, MPC has not been widely adopted in buildings, and scalable building model identification has been recognized as the main challenge preventing the widespread adoption of MPC [4]. Model identification is particularly challenging in buildings because of inadequate sensing. Specifically, building spaces have time-varying heat gains, resulting from solar radiation, occupants, lighting, and other electrical equipment, that evolve on similar time scales as that of the relevant building dynamics. The heat gains are difficult to measure. This coupled with a lack of full state measurements results in a coupled state/disturbance estimation and model parameter identification problem.

A few approaches have been proposed for addressing the coupled estimation and identification problem for buildings (e.g., [5], [6], [7]). For example, an output integrating disturbance model was included when identifying the parameters of a physics-based building thermal dynamic model [5]. A simultaneous model parameter and disturbance estimation procedure was proposed [6]. Both modeling approaches result in a building thermal dynamic model and a model or estimator for estimating the unmeasured time-varying heat disturbance. However, the modeling goal for MPC is to obtain a model for predicting the building thermal dynamics over time. Accurately forecasting the time-varying heat disturbance is critical for achieving adequate closed-loop performance under MPC. Thus, as an alternative, a hybrid modeling framework consisting of a simplified physics-based model used to capture the indoor air and building mass thermal dynamics and a machine learning model used to capture the time-varying heat disturbances was proposed [7]. The two underlying models that make up the hybrid model were trained with a simultaneous approach. With noise-free data, the method accurately identified the physics-based model parameters and a machine learning model that could forecast the heat disturbance. A preliminary training strategy was proposed for performing model training.

This work considers the training computational efficiency associated with hybrid model parameter identification using noisy data for buildings. The proposed hybrid model consists of a physics-based model describing the building's thermal dynamics and a neural network model that can forecast the building's unmeasured time-varying heat gains. In the proposed hybrid model, the parameters of the underlying physics-based model and machine learning-based models are trained simultaneously to prevent biased estimates of the physics-based parameters caused by the unmeasured heat gains. The training strategy is formalized, and improvements that address computational efficiency are investigated. To this end, a hybrid training strategy involving gradient descent with BFGS methods is proposed to solve the parameter estimation problem. Parameter regularization and training drop-out are considered to address noisy data. Considering data collected from a building zone, model training results employing the various training strategies proposed in this work are used to evaluate the accuracy of the resulting models and computational efficiency.

References:

[1] D. R. Dunn, J. Zaidi, M. Kopalek, G. Long, R. Berry, G. McGrath, T. Shear, M. Klein, M. Kopalek, A. Gorski, E. Harrison, L. Jamison, K. Nakolan, “Monthly energy review: March 2022”, Technical Report, U.S. Energy Information Administration, 2022.

[2] C. F. Alvarez, G. Molnar, “What is behind soaring energy prices and what happens next?”, Technical Report, International Energy Agency, 2021.

[3] A. Afram, F. Janabi-Sharifi, A. S. Fung, and K. Raahemifar, “Artificial neural network (ANN) based model predictive control (MPC) and optimization of HVAC systems: A state of the art review and case study of a residential HVAC system”, Energy and Buildings 141 (2017), pp. 96–113.

[4] G. P. Henze, “Model predictive control for buildings: a quantum leap?”, Journal of Building Performance Simulation 6 (3) (2013), pp. 157–158.

[5] D. Kim, J. Cai, J. Braun and K. Ariyura, “System identification for building thermal systems under the presence of unmeasured disturbances in closed loop operation: Theoretical analysis and application”, Energy and Buildings 167 (2018), pp. 359–369.

[6] A. R. Coffman and P. Barooah, “Simultaneous identification of dynamic model and occupant-induced disturbance for commercial buildings”, Building and Environment 128 (2018), pp. 153–160.

[7] M. J. Ellis, “Machine Learning Enhanced Grey-Box Modeling for Building Thermal Modeling,” In Proceedings of the American Control Conference, New Orleans LA USA, pp. 3927-3932, 25-28 May, 2021.