(596g) Recurrent Neural Network-Based Economic Model Predictive Control Framework Applied to a Building HVAC System

Economic model predictive control (EMPC) is a unique control strategy that combines economic optimization and control in a unified framework (e.g., [2]). Over the last ten years, EMPC has proved to be well suited to handle features commonly arising in the context of energy systems including time-varying electric rates, electric peak demand charges, and time-varying demands. As such, EMPC has been adopted in industry and commercialized (e.g., [6]). For buildings and HVAC systems, a well documented source of substantial deployment cost of EMPC systems is model construction/identification (e.g., [3], [4]). Given the renewed interest in machine learning within the control community (e.g., [1], [6]), it is important to consider if recurrent neural network (RNN)-based modeling may reduce model development cost in EMPC deployments compared to conventional control modeling approaches. While recent work on artificial neural network- or RNN-based MPC strategies proposed in the literature (e.g., [1], [6]) has shown potential, limited work has been done on the end-to-end design of RNN-based EMPC that addresses important practical considerations including state estimation, disturbance/exogenous input forecasting, and automatic EMPC problem generation.

To this end, an EMPC framework is developed, which incorporates a sequence-to-sequence RNN model as the predictive model. Specifically, the model training methodology and the overall EMPC framework design are presented. The framework utilizes the trained model for state estimation and for the predictive model used within the EMPC problem. In this sense, an interpretation of the model as a black-box nonlinear state space model is given. The automatic EMPC problem construction and solution strategy, which leverages a multiple shooting approach, is also provided. Finally, the overall framework is applied to a building HVAC system simulated using EnergyPlus to demonstrate the approach. Issues arise from using an RNN-based model within an EMPC framework are also discussed.

[1] Afram A, F Janabi-sharifi, AS Fung, 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. 2017; 141: 96-113.
[2] M. Ellis, H. Durand, and P. D. Christofides. A tutorial review of economic model predictive control methods. Journal of Process Control. 2014; 24: 1156-1178.
[3] Henze GP. Model predictive control for buildings: a quantum leap? Journal of Building Performance Simulation. 2013; 6: 157-158.
[4] Killian, M, M Kozek. Ten questions concerning model predictive control for energy efficient buildings. Building and Environment. 2016; 105: 403-412.
[5] Wenzel MJ, MN ElBsat, MJ Ellis, MJ Asmus, AJ Przybyiski, R Baumgartner, JH Burroughs, G Willmott, KH Drees, and RD Turney. Large scale optimization problems for central energy facilities with distributed energy storage. In: Proceedings of the 5th International High Performance Buildings Conference, Paper 3560, West Lafayette, IN, 2018.
[6] Wu, Z, A Tran, D Rincon, PD Christofides. Machine learning-based predictive control of nonlinear processes. Part I: Theory. AIChE Journal. 2019; 65: e16729.