System identification and model development

Patent metadata

Field	Value
Publication number	US-9235657-B1
Application number	US-201313802233-A
Country	US
Kind code	B1
Filing date	Mar 13, 2013
Priority date	Mar 13, 2013
Publication date	Jan 12, 2016
Grant date	Jan 12, 2016

Abstract

Methods for system identification are presented using model predictive control to frame a gray-box parameterized state space model. System parameters are identified using an optimization procedure to minimize a first error cost function within a range of filtered training data. Disturbances are accounted for using an implicit integrator within the system model, as well as a parameterized Kalman gain. Kalman gain parameters are identified using an optimization procedure to minimize a second error cost function within a range of non-filtered training data. Recursive identification methods are presented to provide model adaptability using an extended Kalman filter to estimate model parameters and a Kalman gain to estimate system states.

First claim

The invention claimed is:

1. A computer-implemented method for identifying model parameters including system parameters and Kalman gain parameters in a dynamic model of a building system, the method comprising:
receiving, at a controller for the building system, training data including input data and output data, wherein the input data includes a setpoint applied as a controlled input to the building system, wherein the output data measures a state or linear combination of states of the building system in response to both the setpoint and an extraneous disturbance, wherein the extraneous disturbance comprises an uncontrolled thermal input to the building system that affects the state or linear combination of states measured by the output data;
performing, by the controller, a two-stage optimization process to identify the system parameters and the Kalman gain parameters, the two-stage optimization process comprising:
a first stage in which the controller filters the training data to remove an effect of the extraneous disturbance from the output data and uses the filtered training data to identify the system parameters, wherein the system parameters describe energy transfer characteristics of the building system, and
a second stage in which the controller uses the non-filtered training data to identify the Kalman gain parameters, wherein the Kalman gain parameters account for the extraneous disturbance; and

using the dynamic model to generate, by the controller, a new setpoint for the building system, wherein the building system uses the new setpoint to affect the state or linear combination of states measured by the output data.

2. The method of claim 1 , wherein the first stage of the two-stage optimization process comprises:
filtering the input data to create filtered input data and filtering the output data to create filtered output data;
generating model-predicted filtered output data based on a set of estimated system parameters and the filtered input data;
defining a first error cost function, wherein the first error cost function expresses a first error cost based on a difference between the filtered output data and the model-predicted filtered output data;
adjusting the estimated system parameters to minimize the first error cost.

3. The method of claim 2 , wherein the training data is filtered using a high-pass filter to remove an effect of a slowly changing disturbance to the building system from the output data.

4. The method of claim 2 , wherein generating model-predicted filtered output data includes applying the filtered input data to the dynamic model of the building system and recording an output of the dynamic model in response to the filtered input data.

5. The method of claim 2 , wherein the first error cost function is defined to mitigate an effect of a filtered output estimation error exceeding a first threshold,
wherein the filtered output estimation error is a difference between a filtered output data point and a model-predicted filtered output data point based on the filtered training data.

6. The method of claim 1 , wherein the second stage of the two-stage optimization process comprises:
generating model-predicted output data based on a set of estimated Kalman gain parameters and the non-filtered input data;
defining a second error cost function, wherein the second error cost function expresses a second error cost based on a difference between the output data and the model-predicted output data;
adjusting the estimated Kalman gain parameters to minimize the second error cost.

7. The method of claim 6 , wherein generating model-predicted output data includes applying the input data to the dynamic model of the building system and recording an output of the dynamic model in response to the input data.

8. The method of claim 6 , wherein the second error cost function is defined to mitigate an effect of a non-filtered output estimation error exceeding a second threshold,
wherein the non-filtered output estimation error is a difference between an output data point and a model-predicted output data point based on the non-filtered training data.

9. The method of claim 1 , wherein the second stage of the two-stage optimization process identifies the Kalman gain parameters while holding the system parameters at constant values.

10. The method of claim 9 , wherein the second stage of the two-stage optimization process holds the system parameters at values identified by the first stage of the two-stage optimization process.

11. The method of claim 1 , wherein the dynamic model is a linear state space model with a parameterized Kalman gain.

12. The method of claim 1 , wherein the dynamic model is a physics-based parameterization of the system.

13. The method of claim 1 , further comprising:
receiving new data including new input data and new output data;
adding the new data to the training data; and
repeating the two-stage optimization process, wherein repeating the two-stage optimization process with the new data included in the training data updates the system parameters and Kalman gain parameters in response to the new data.

14. A computer-implemented method for identifying model parameters in a dynamic model of a building system, the method comprising:
receiving, at a controller for the building system, a set of model parameters for the dynamic model, the model parameters including system parameters and Kalman gain parameters;
estimating, by the controller, updated model parameters for the dynamic model upon receiving a new data measurement from the building system;
determining, by the controller, whether the updated model parameters are stable and robust;
using the dynamic model with the updated model parameters to estimate, by the controller, system states for the building system in response to a determination that the updated model parameters are stable and robust;
reverting the dynamic model to previous model parameters in response to a determination that the updated model parameters are unstable or non-robust; and
using the dynamic model to generate, by the controller, a setpoint for the building system, wherein the building system uses the setpoint to affect a variable state or condition of the building system measured by the data measurement.

15. The method of claim 14 , wherein the updated model parameters are estimated using an extended Kalman filter and the system states are estimated using a Kalman gain.

16. The method of claim 14 , wherein the dynamic model is a linear state space model including an A matrix, a C matrix, and a parameterized Kalman gain, and
wherein and the updated model parameters are determined to be stable if eigenvalues of the expression A-KC are strictly less than one, wherein A and C are the A and C matrices respectively in the linear state space model and K is the parameterized Kalman gain.

17. The method of claim 14 , wherein the updated model parameters are determined to be robust if a difference between a model-predicted output and a measured output is within a threshold, wherein the model-predicted output is based on the updated model parameters and an input applied to the dynamic model.

18. A computer-implemented method for developing a dynamic model of a building system, the method comprising:
receiving, at a controller for the building system, a system of equations which express future states of the building system and outputs of the building system as a function of current states of the building system and controlled inputs to the building system;
accounting, by the co

Assignees

• Johnson Controls Tech Co

Inventors

• Wenzel Michael J
• Turney Robert D

Classifications

• G05B13/048 Primary

  using a predictor · CPC title

• G05B2219/2642

  Domotique, domestic, home control, automation, smart house · CPC title

• G06F30/00

  Computer-aided design [CAD] · CPC title

• G05B13/0265 Primary

  the criterion being a learning criterion · CPC title

• G06F30/13

  Architectural design, e.g. computer-aided architectural design [CAAD] related to design of buildings, bridges, landscapes, production plants or roads · CPC title