HVAC control system with model driven deep learning

1. A method, comprising:

operating equipment to affect a variable state or condition of a space;

determining a set of learned weights for a neural network by modeling an estimated cost of operating the equipment over a plurality of simulated scenarios, each simulated scenario comprising simulated measurements relating to the space, the neural network configured to generate simulated control dispatches for the equipment based on the simulated measurements;

configuring the neural network for online control by applying the set of learned weights;

applying actual measurements relating to the space to the neural network to generate a control dispatch for the equipment; and

controlling the equipment in accordance with the control dispatch.

2. The method of claim 1 , wherein the set of learned weights are determined as a set of weights that minimize the estimated cost of operating the equipment over the plurality of simulated scenarios.

3. The method of claim 1 , wherein determining the set of learned weights comprises:

identifying a state-space thermal model for the space;

defining a cost function using the state-space thermal model;

generating, by the neural network for each scenario, a simulated control dispatch based the simulated measurements for the scenario and a set of weights; and

calculating, using the cost function, the estimated cost of operating the equipment over a simulated time period for the scenario given the simulated control dispatch and the simulated measurements.

4. The method of claim 3 , wherein determining the set of learned weights further comprises:

modifying the set of weights to drive the estimated cost toward a minimum of the cost function; and

determining the set of learned weights as the set of weights that results in a minimum cost over the plurality of simulated scenarios.

5. The method of claim 1 , wherein the control dispatch comprises one or more of a temperature setpoint, temperature schedule, humidity setpoint, airflow setpoint, power level, on/off setting, damper position, fan speed, compressor frequency, or resource consumption allocation.

6. The method of claim 1 , wherein the equipment comprises one or more of an airside system or a waterside system.

7. The method of claim 1 , wherein the equipment comprises one or more of a variable refrigerant flow system, a room air conditioner, or a packaged air conditioner.

8. A system, comprising:

HVAC equipment operable to affect a variable state or condition of a space;

one or more sensors configured to collect measurements relating to the space;

an offline training system configured to determine a set of learned weights for a neural network by modeling an estimated cost of operating the HVAC equipment over a plurality of simulated scenarios, each simulated scenario comprising simulated measurements relating to the space, the neural network configured to generate simulated control dispatches for the HVAC equipment based on the simulated measurements; and

an online control circuit configured to:

apply the measurements from the one or more sensors to the neural network to generate a control dispatch for the HVAC equipment, the neural network configured in accordance with the set of learned weights; and

control the HVAC equipment in accordance with the control dispatch.

9. The system of claim 8 , wherein the offline training system is configured to determine the set of learned weights as a set of weights that minimize the estimated cost of operating the equipment over the plurality of simulated scenarios.

10. The system of claim 8 , wherein the offline training system is configured to:

identify a state-space thermal model for the space;

define a cost function using the state-space thermal model;

generate, with the neural network and for each scenario, a simulated control dispatch based the simulated measurements for the scenario and a set of weights; and

calculate, using the cost function, the estimated cost of operating the HVAC equipment over a simulated time period for the scenario given the simulated control dispatch and the simulated measurements.

11. The system of claim 10 , wherein the offline training system is configured to:

modify the set of weights to drive the estimated cost toward a minimum of the cost function; and

determine the set of learned weights as the set of weights that results in a minimum cost over the plurality of simulated scenarios.

12. The system of claim 8 , wherein the HVAC equipment comprises one or more of an airside system or a waterside system.

13. The system of claim 8 , wherein the online control circuit is included locally with the HVAC equipment and the offline training system comprises one or more cloud-computing resources.

14. A system, comprising:

a cooling device operable to affect a temperature of a space;

one or more sensors configured to collect measurements relating to the space;

an offline training system configured to determine a set of learned weights for a neural network by modeling an estimated cost of operating the cooling device over a plurality of simulated scenarios, each simulated scenario comprising simulated measurements relating to the space, the neural network configured to generate simulated control dispatches for the cooling device based on the simulated measurements; and

an online control circuit configured to:

apply the measurements from the one or more sensors to the neural network to generate a control dispatch for the cooling device, the neural network configured in accordance with the set of learned weights; and

control the cooling device in accordance with the control dispatch.

15. The system of claim 14 , wherein the offline training system is configured to determine the set of learned weights as a set of weights that minimize the estimated cost of operating the cooling device over the plurality of simulated scenarios.

16. The system of claim 14 , wherein the offline training system is configured to:

identify a state-space thermal model for the space;

define a cost function using the state-space thermal model;

generate, with the neural network and for each scenario, a simulated control dispatch based the simulated measurements for the scenario and a set of weights; and

calculate, using the cost function, the estimated cost of operating the cooling device over a simulated time period for the scenario given the simulated control dispatch and of simulated measurements.

17. The system of claim 16 , wherein the offline training system is configured to:

modify the set of weights to drive the estimated cost toward a minimum of the cost function; and

determine the set of learned weights as the set of weights that results in a minimum cost over the plurality of simulated scenarios.

18. The system of claim 14 , wherein the cooling device comprises one or more of a room air conditioner, a packaged air conditioner, or a variable refrigerant flow device.

19. The system of claim 14 , the online control circuit is included locally with the cooling device and the offline training system comprises one or more cloud-computing resources.

20. The system of claim 14 , wherein the control dispatch comprises a temperature setpoint.