(324b) Model Identification and Disturbance Forecasting Via Recurrent Neural Networks for Building Applications | AIChE

(324b) Model Identification and Disturbance Forecasting Via Recurrent Neural Networks for Building Applications


Ellis, M. - Presenter, University of California, Davis
Recent developments in machine learning (ML), especially deep learning, have attracted attention in many communities including the control and process-systems engineering (PSE) communities (e.g., [5]). One paradigm for leveraging ML within the context of feedback control is to incorporate ML into existing control methods, which may help address problems that are difficult to solve with well established approaches. To this end, a key implementation challenge of advanced control for some applications is model identification. In fact, within the context of model predictive control (MPC) of buildings, the cost of model development and identification has been widely recognized as the main barrier preventing widespread implementation of MPC (e.g., [3]).

Model identification of building thermal dynamics is particular challenging as significant time-varying and unmeasured heat disturbances are present. Specifically, buildings are subject to significant time-varying and unmeasured heat disturbances due to, for example, solar irradiation, electrical equipment, and occupancy. From a system identification perspective, neglecting the effect of the disturbance in the model will result in biased model parameter estimates (e.g., [4]). Within the literature, a few methods for dealing with unmeasured heat disturbances include pre-filtering the training data (e.g., [2]), performing parameter identification with an augmented model that includes an integrating disturbance model [1], and modeling the disturbance as a second order (linear) disturbance model [4]. As an alternative approach, a recurrent neural network may be trained to model the disturbance, which can provide an estimate of the disturbance to a standard physical-based grey-box model. Moreover, this recurrent neural network may also be used in MPC to provide forecasts of the disturbance over time.

In this work, a framework for performing model identification is proposed. The framework includes a parameter estimation technique to identify the parameters of a grey-box system model and a recurrent neural network to describe the evolution of the disturbances. The approach is demonstrated on a building thermal zone. A comparison of the approach to standard model identification methods is also presented to compare and contrast the proposed framework with standard methods. Specifically, the data requirements to train the models and prediction accuracy are assessed.

[1] Coffman AR, Barooah P. Simultaneous identification of dynamic model and occupant-induced disturbance for commercial buildings. Building and Environment. 2018;128:153-160.
[2] Ellis MJ, Wenzel MJ, and Turney RD. System identification for cascaded model predictive control of building region temperature applications. In: Proceedings of the 4th International High Performance Buildings Conference. 2016; paper number: 3583.
[3] Henze GP. Model predictive control for buildings: a quantum leap? Journal of Building Perfromance Simulation. 2013;6:157-158.
[4] Kim D, Cai J, Ariyur KB, Braun JE. System identification for building thermal systems under the presence of unmeasured disturbances in closed loop operation: Lumped disturbance modeling approach. Building and Environment. 2016;107:169-180.
[5] Lee JH, Shin J, Realff MJ. Machine learning: Overview of the recent progresses and implications for the process systems field. Computers & Chemical Engineering. 2018;114:111-121.