Air handling unit and method using reinforcement learning model that replicates model predictive control simulation
A method includes generating a mixed air temperature value using a reinforcement learning model running on the edge controller. A temperature setpoint and a weather forecast are inputs to the reinforcement learning model. The method also includes controlling damper positions of an air handling unit to achieve the mixed air temperature value.
1. A method comprising:
training a reinforcement learning model to replicate outputs of a model predictive control algorithm running in a simulation, the model predictive control algorithm determining simulated mixed air temperatures which minimize an objective associated with running a simulated air handling unit in the simulation;
generating a mixed air temperature value using the reinforcement learning model running on an edge controller, wherein an indoor air temperature setpoint and a weather forecast are inputs to the reinforcement learning model; and
controlling damper positions of an air handling unit to achieve the mixed air temperature value.
2. The method of claim 1 , further comprising automatically updating, by the edge controller, the reinforcement learning model based on a reward function comprising a difference between the indoor air temperature setpoint and a measured indoor air temperature.
3. The method of claim 2 , wherein the reward function is further based on an occupancy of a space served by the air handling unit.
4. The method of claim 1 , wherein using the reinforcement learning model comprises performing feature generation on streaming input data based on logic in a functional programming language and using variables resulting from the feature generation as further inputs to the reinforcement learning model.
5. The method of claim 1 , wherein generating the mixed air temperature value using the reinforcement learning model running on the edge controller further comprises processing streams of building and equipment state information at the edge controller.
6. The method of claim 1 , further comprising:
creating the reinforcement learning model on a computing system separate from the edge controller; and
transferring the reinforcement learning model from the computing system to the edge controller after the creating.
7. The method of claim 1 , wherein the inputs to the reinforcement learning model further comprise an occupancy forecast.
8. An air handling unit, comprising:
a plurality of dampers;
a local controller programmed to control the plurality of dampers by:
generating a mixed air temperature value using a reinforcement learning model running on the local controller, wherein a temperature setpoint and a weather forecast are inputs to the reinforcement learning model, wherein the reinforcement learning model is trained to replicate simulated mixed air temperatures output by a model predictive control algorithm running in a simulation, the model predictive control algorithm determining simulated mixed air temperatures which minimize an objective associated with running a simulated air handling unit in the simulation; and
controlling damper positions of the air handling unit to achieve the mixed air temperature value.
9. The air handling unit of claim 8 , wherein the local controller is further programmed to automatically update the reinforcement learning model using reinforcement learning based on a difference between an indoor temperature setpoint for a building served by the air handling unit and a measurement of the indoor temperature setpoint.
10. The air handling unit of claim 9 , wherein the mixed air temperature value is a change in mixed air temperature of the air handling unit to be achieved by controlling the damper positions and the indoor temperature setpoint is a value to be achieved for a building and received by the local controller from an active setpoint management service.
11. The air handling unit of claim 8 , wherein the local controller is further programmed to automatically update the reinforcement learning model using a reward function comprising a difference between one of the inputs to the reinforcement learning and a measurement of a measured condition of a space served by the air handling unit.
12. The air handling unit of claim 8 , the local controller is programed to perform feature generation on streaming input data based on logic in a functional programming language and to use variables resulting from the feature generation as further inputs to the reinforcement learning model.
13. The air handling unit of claim 8 , wherein generating the mixed air temperature value using the reinforcement learning model running on the local controller further comprises processing streams of building and equipment state information at the local controller.
14. The air handling unit of claim 8 , wherein the local controller is further programmed to apply an occupancy forecast as an input to the reinforcement learning model.
15. The air handling unit of claim 8 , wherein the temperature setpoint is an indoor air temperature setpoint in units of degrees.