Adaptive model predictive control of building HVAC using moving horizon estimation
A differentiable physics model of a building is used that defines thermodynamic relationships between zones of the building and a heating, ventilation, and air-conditioning (HVAC) system. A physics-constrained, data driven model learns behaviors of controlled components of the HVAC system. For each of a series of times during online operation of the HVAC system, past state values are recorded representing a performance of the HVAC system in the building and past inputs to the HVAC system to maintain the states. The past state values and the past inputs are input into the differentiable physics model and the data driven model to: jointly update first parameters of the differentiable physics model and second parameters of the data driven model, e.g., using moving horizon estimation; and determine a current input to the controlled components, e.g., using model predictive control.
1 . A method comprising:
determining a differentiable physics model of a building, the differentiable physics model defining thermodynamic relationships between zones of the building and a heating, ventilation, and air-conditioning (HVAC) system;
forming a physics-constrained, data driven model that learns behaviors of controlled components of the HVAC system;
for each of a series of times during online operation of the HVAC system, recording past state values representing a performance of the HVAC system in the building and recording past inputs to the HVAC system to maintain the states;
inputting the past state values and the past inputs into the differentiable physics model and the data driven model to: jointly update first parameters of the differentiable physics model and second parameters of the data driven model; and determine a current input to the controlled components; and
applying the current input to the controlled components.
2 . The method of claim 1 , wherein the inputs to the controlled components are determined via a supervisory model predictive controller (MPC), wherein MPC controller determines setpoints for the controlled components to minimize an overall power consumption of the HVAC system while adhering to occupant comfort constraints.
3 . The method of claim 2 , wherein the differentiable physics model and the data driven model are continuously estimated before each MPC iteration.
4 . The method of claim 1 , wherein the first and second parameters are updated using moving horizon estimation.
5 . The method of claim 1 , wherein the controlled components comprise any combination of heating and cooling units.
6 . The method of claim 5 , wherein the cooling units comprise a central cooling system and the heating units comprise zone-specific variable air volume reheat boxes.
7 . The method of claim 1 , wherein the controlled components each comprises fine-tuned proportional integral (PI) controllers, the physics-constrained, data driven model modeling behaviors of the PI controllers.
8 . The method of claim 1 , wherein the data driven model is physics-constrained to limit temperatures of various locations in the building to predefined ranges.
9 . The method of claim 1 , wherein the data driven model includes a feedforward neural network.
10 . The method of claim 9 , wherein the feedforward neural network comprises a plurality of feedforward neural networks, each of the plurality of feedforward neural networks associated with one of the controlled components.
11 . The method of claim 1 , wherein jointly updating the first and second parameters comprises computing first and second order derivatives of the differentiable physics model and the data driven model with automatic differentiation, the first and second order derivatives being passed to a non-linear solver that updates the first and second parameters using a non-linear optimization.
12 . A system comprising:
a plurality of sensors that collect state information from zones of a heating, ventilation, and air-conditioning (HVAC) of a building;
a plurality of controlled components of the HVAC system coupled to receive inputs that affect the state information; and
a controller comprising a data interface that receives the state information and sends the inputs to controlled components of the HVAC system, the controller comprising a data storage unit that stores:
a differentiable physics model of the building that defines thermodynamic relationships between the zones of the building and the HVAC system, the differentiable physics model having first parameters that are updatable during operation of the HVAC system; and
a physics-constrained, data driven model that learns behaviors of the controlled components, the data driven model having second parameters that are updatable during operation of the HVAC system;
the controller further comprising a processor coupled to the data interface and the data storage unit, the processor configured to perform:
for each of a series of times during online operation of the HVAC system, recording past state values representing a performance of the HVAC system in the building and recording past inputs to the HVAC system to maintain the states;
inputting the past state values and the past inputs into the differentiable physics model and the physics-constrained, data driven model to jointly update the first parameters and the second parameters and to determine a current input to the controlled components; and
applying the current input to the controlled components.
13 . The system of claim 12 , wherein the inputs to the controlled components are determined via a supervisory model predictive controller (MPC), wherein MPC controller determines setpoints for the controlled components to minimize an overall power consumption of the HVAC system, and adhering to occupant comfort constraints.
14 . The system of claim 12 , wherein the first and second parameters are updated using a moving horizon estimate.
15 . The system of claim 12 , wherein the controlled components comprise a combination of a central cooling system and zone-specific variable air volume reheat boxes.
16 . The system of claim 12 , wherein the controlled components each comprises fine-tuned proportional integral (PI) controllers, the physics-constrained, data driven model modeling behaviors of the PI controllers.
17 . The system of claim 12 , wherein the data driven model includes a plurality of feedforward neural networks, each of the plurality of feedforward neural networks associated with one of the controlled components.
18 . The system of claim 12 , wherein jointly updating the first and second parameters comprises computing first and second order derivatives of the differentiable physics model and the data driven model with automatic differentiation, the first and second order derivatives being passed to a non-linear solver that updates the first and second parameters using a non-linear optimization.