# (199d) Data-Driven Design of Hierarchical MPC Architectures

- Conference: AIChE Annual Meeting
- Year: 2019
- Proceeding: 2019 AIChE Annual Meeting
- Group: Computing and Systems Technology Division
- Session:
- Time:
Monday, November 11, 2019 - 4:24pm-4:42pm

In this work, we present data-driven hierarchical model predictive control (MPC) schemes for systems that exhibit periodicity in state policy over long horizons. Systems driven by exogenous factors with strong periodic components, for instance energy demands and prices in applications involving energy systems, exhibit the periodicity property [7, 8]. The proposed hierarchical schemes are based on a key observation that if the optimal policy of an infinite-horizon problem is periodic (or can be approximated as periodic), a stochastic programming (SP) problem can be used to pose the problem. Under the SP abstraction, the inter-period trajectory of the exogenous factors is interpreted as a realization of a random variable that triggers a periodic trajectory of the states (i.e., the states at the beginning and end of the period are the same). Moreover, the periodic states are interpreted as design variables and operational policies within the periods are interpreted as recourse variables. The SP representation facilitates construction of hierarchical MPC schemes with a long-term (supervisory) MPC controller providing periodic targets to guide a short-term MPC controller operating at finer time resolution [9, 10]. We have shown that under nominal conditions with perfect forecasts, the hierarchical scheme delivers an optimal policy for the infinite horizon problem. For the more relevant case of imperfect forecasts, the hierarchical scheme needs to re-compute periodic targets. We derive retroactive hierarchical MPC schemes under a periodic setting and using statistical approximations, that accumulate real historical data to asymptotically deliver optimal targets [10]. We show that the retroactive design principle offers optimality guarantees and, notably, does not require data forecasts.

Thus the retroactive approach provides key advantages over standard proactive RH schemes which use historical data to compute forecasts, and associated control actions and periodic targets. A fundamental issue with proactive approaches is that no optimality guarantees can be provided unless the forecast is perfect. The targets delivered by the retroactive scheme are used to guide a low-level controller operating at fine time resolutions within the periods. We also derive a specialized retroactive scheme by using incremental cutting-plane (CP) techniques for the case of linear systems [10, 11]. The SP abstraction also provides opportunities to construct retroactive schemes for nonlinear systems, and to obtain the desired stability properties. We demonstrate the concepts using an application of management of stationary battery systems in buildings, where the proposed retroactive hierarchical scheme is used to obtain the optimal charge-discharge policy for the battery system to minimize the peak demand charge of the buildings. We compare the performance of the proposed retroactive hierarchical MPC scheme with a proactive MPC approach for periodic systems.

References:

[1] Rawlings, J.B. and Mayne, D.Q., 2009. *Model predictive control: Theory and design* (pp. 3430-3433). Madison, Wisconsin: Nob Hill Pub..

[2] Braun, J.E., 1990. Reducing energy costs and peak electrical demand through optimal control of building thermal storage. *ASHRAE transactions*, *96*(2), pp.876-888.

[3] Patel, N.R., Risbeck, M.J., Rawlings, J.B., Wenzel, M.J. and Turney, R.D., 2016, July. Distributed economic model predictive control for large-scale building temperature regulation. In *2016 American Control Conference (ACC)* (pp. 895-900). IEEE.

[4] Picasso, B., De Vito, D., Scattolini, R. and Colaneri, P., 2010. An MPC approach to the design of two-layer hierarchical control systems. *Automatica*, *46*(5), pp.823-831.

[5] Scattolini R. and Colaneri P., 2007. Hierarchical model predictive control. In Decision and Control, 2007 46th IEEE Conference on, (pp. 4803-4808). IEEE.

[6] Zavala, V.M., 2016. New architectures for hierarchical predictive control. *IFAC-PapersOnLine*, *49*(7), pp.43-48.

[7] R. Huang, E. Harinath, and L. T. Biegler. Lyapunov stability of economically oriented nmpc for cyclic processes. Journal of Process Control, 21(4):501â€“509, 2011.

[8] Risbeck, M.J., Maravelias, C.T., Rawlings, J.B. and Turney, R.D., 2015, July. Cost optimization of combined building heating/cooling equipment via mixed-integer linear programming. In *2015 American Control Conference (ACC)*(pp. 1689-1694). IEEE.

[9] Kumar, R., Wenzel, M.J., Ellis, M.J., ElBsat, M.N., Drees, K.H. and Zavala, V.M., 2018, June. Handling Long Horizons in MPC: A Stochastic Programming Approach. In *2018 Annual American Control Conference (ACC)* (pp. 715-720). IEEE.

[10] Kumar, R., Wenzel, M.J., Ellis, M.J., ElBsat, M.N., Drees, K.H. and Zavala, V.M., 2019. Hierarchical MPC Schemes for Periodic Systems using Stochastic Programming. *Automatica, Accepted*.

[11] Higle, J.L. and Sen, S., 1991. Stochastic decomposition: An algorithm for two-stage linear programs with recourse. *Mathematics of operations research*, *16*(3), pp.650-669.