Conference Year

2016

Keywords

Model Predictive Control, Central Plant, Optimization, Linear Programming

Abstract

A model predictive control (MPC) framework is used to determine how to optimize the distribution of energy resources across a central energy facility including chillers, water heaters, and thermal energy storage; present the results to an operator; and execute the plan. The objective of this MPC framework is to minimize cost in real-time in response to both real-time energy prices and demand charges as well as allow the operator to appropriately interact with the system. Operators must be given the correct intersection points in order to build trust before they are willing to turn the tool over and put it into fully autonomous mode. Once in autonomous mode, operators need to be able to intervene and impute their knowledge of the facilities they are serving into the system without disengaging optimization. For example, an operator may be working on a central energy facility that serves a college campus on Friday night before a home football game. The optimization system is predicting the electrical load, but does not have knowledge of the football game. Rather than try to include every possible factor into the prediction of the loads, a daunting task, the optimization system empowers the operator to make human-in-the-loop decisions in these rare scenarios without exiting autonomous (auto) mode. Without this empowerment, the operator either takes the system out of auto mode or allows the system to make poor decisions. Both scenarios will result in an optimization system that has low “on time” and thus saves little money. A cascaded, model predictive control framework lends itself well to allowing an operator to intervene. The system presented is a four tiered approach to central plant optimization. The first tier is the prediction of the energy loads of the campus; i.e., the inputs to the optimization system. The predictions are made for a week in advance, giving the operator ample time to react to predictions they do not agree with and override the predictions if they feel it necessary. The predictions are inputs to the subplant-level optimization. The subplant-level optimization determines the optimal distribution of energy across major equipment classes (subplants and storage) for the prediction horizon and sends the current distribution to the equipment level optimization. The operators are able to use the subplant-level optimization for “advisory” only and enter their own load distribution into the equipment level optimization. This could be done if they feel that they need to be conservative with the charge of the tank. Finally, the equipment level optimization determines the devices to turn on and their setpoints in each subplant and sends those setpoints to the building automation system. These decisions can be overridden, but should be extremely rare as the system takes device availability, accumulated runtime, etc. as inputs. Building an optimization system that empowers the operator ensures that the campus owner realizes the full potential of his investment. Optimal plant control has shown over 10% savings, for large plants this can translate to savings of more than US $1 million per year.

Share

COinS