Systems and methods for automated system identification
A controller for performing automated system identification. The controller includes processors and non-transitory computer-readable media storing instructions that, when executed by the processors, cause the processors to perform operations including generating a predictive model to predict system dynamics of a space of a building based on environmental condition inputs and including performing an optimization of a cost function of operating building equipment over a time duration to determine a setpoint for the building equipment. The optimization is performed based on the predictive model. The operations include operating the building equipment based on the setpoint to affect a variable state or condition of the space and include monitoring prediction error metrics over time. The operations include, in response to detecting one of the prediction error metrics exceeds a threshold value, updating the predictive model.
1. A controller for performing automated system identification, the controller comprising:
one or more processors; and
one or more non-transitory computer-readable media storing instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
generating a predictive model to predict one or more system dynamics of a space of a building based on one or more environmental condition inputs;
performing a validation of the predictive model by assessing a variance of internal parameters of the predictive model or comparing the internal parameters of the predictive model to bounds for the internal parameters of the predictive model without requiring inputs to the predictive model or outputs of the predictive model to be used when performing the validation;
performing an optimization of a cost function of operating building equipment over a time duration to determine a setpoint for the building equipment, wherein the optimization is performed based on the predictive model; and
operating the building equipment based on the setpoint to affect a variable state or condition of the space.
2. The controller of claim 1 , further comprising determining a multi-step-ahead prediction error of the predictive model by comparing predicted values of one or more variables generated by the predictive model for multiple time steps subsequent to a given time step to actual values of the one or more variables at the multiple time steps subsequent to the given time step, wherein the predicted values for the multiple time steps subsequent to the given time step are generated by the predictive model without requiring the actual values for any of the multiple time steps subsequent to the given time step.
3. The controller of claim 1 , wherein generating the predictive model comprises:
perturbing the setpoint of the space or a heat duty of the building equipment to excite one or more dynamics of the space;
monitoring one or more effects of perturbing the setpoint or the heat duty; and
generating a set of training data comprising values of the one or more effects.
4. The controller of claim 1 , wherein generating the predictive model comprises:
generating one or more candidate models; and
selecting one of the one or more candidate models based on an estimated accuracy of each of the one or more candidate models.
5. The controller of claim 4 , wherein generating the predictive model comprises:
determining whether the selected candidate model is suitable for use in performing the optimization;
in response to a determination that the selected candidate model is not suitable, generating one or more new candidate models and selecting a new candidate model of the one or more new candidate models; and
in response to a determination that the selected candidate model is suitable, providing the selected candidate model as the predictive model for use in performing the optimization.
6. The controller of claim 1 , wherein the operations further comprise monitoring one or more prediction error metrics by:
calculating one or more statistical characteristics of each of the one or more prediction error metrics;
determining whether any of the one or more statistical characteristics of each of the one or more prediction error metrics exceeds a predetermined threshold value for each of the one or more statistical characteristics; and
triggering a model update process in response to at least one of the one or more statistical characteristics exceeding the predetermined threshold value.
7. The controller of claim 6 , wherein the one or more statistical characteristics comprise at least one of:
a variance;
a moving average; or
a moving standard deviation.
8. A controller for performing automated system identification, the controller comprising:
one or more processors; and
one or more non-transitory computer-readable media storing instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
generating a predictive model to predict one or more system dynamics of a system based on one or more measured inputs;
performing a validation of the predictive model by assessing a variance of internal parameters of the predictive model or comparing the internal parameters of the predictive model to bounds for the internal parameters of the predictive model without requiring outputs of the predictive model to be used when performing the validation; and
in response to the validation indicating that the predictive model is valid, performing an optimization of a function of operating equipment over a time duration to determine a setpoint for the equipment, wherein the optimization is performed based on the predictive model;
operating the equipment based on the setpoint to affect a variable state or condition of the system;
monitoring one or more prediction error metrics over time; and
updating the predictive model in response to one or more of the prediction error metrics exceeding a threshold value.
9. The controller of claim 8 , wherein the equipment comprises manufacturing equipment for producing one or more products, the operations further comprising determining a first prediction error metric of the one or more prediction error metrics based on a difference between expected product quality predicted by the predictive model and actual product quality.
10. The controller of claim 8 , wherein generating the predictive model comprises:
perturbing the setpoint over time to excite the one or more system dynamics;
monitoring one or more effects of perturbing the setpoint; and
generating a set of training data comprising values of the one or more effects.
11. The controller of claim 8 , wherein generating the predictive model comprises:
generating one or more candidate models;
selecting one of the one or more candidate models based on an estimated accuracy of each of the one or more candidate models.
12. The controller of claim 8 , wherein monitoring the one or more prediction error metrics comprises:
calculating one or more statistical characteristics of each of the one or more prediction error metrics;
determining whether any of the one or more statistical characteristics of each of the one or more prediction error metrics exceeds a predetermined threshold value for each of the one or more statistical characteristics; and
triggering a model update process in response to at least one of the one or more statistical characteristics exceeding the predetermined threshold value.
13. The controller of claim 12 , wherein the one or more statistical characteristics comprise at least one of:
a variance;
a moving average; or
a moving standard deviation.
14. A method for triggering an update of a predictive model for use in performing automated system identification, the method comprising:
generating the predictive model to predict one or more system dynamics of a space of a building based on one or more environmental condition inputs;
performing an assessment of an internal property of the predictive model, the internal property comprising a stability, observability, controllability, or condition number of the predictive model without requiring inputs to the predictive model to be used when performing the assessment and without requiring outputs of the predictive model to be used when performing the assessment;
determining whether to use or update the predictive model based on a result of the assessment; and
in response to determining to use the predictive model, performing an optimization of a cost function of operating building equipment over a time duration to determine a setpoint for the building equipment, wherein the optimization is performed based on the predictive model;
operating the building equipment based on the setpoint to affect a variable state or condition of the space;
monitoring one or more prediction error metrics over time; and
updating the predictive model in response to one or more of the prediction error metrics exceeding a threshold value.
15. The method of claim 14 , wherein the one or more prediction error metrics comprise at least one of:
a temperature residual;
a humidity residual;
an air quality residual;
one or more environmental condition residuals; or
a heat load residual.
16. The method of claim 14 , wherein generating the predictive model comprises:
perturbing the setpoint of the space or a heat duty of the building equipment to excite one or more dynamics of the space;
monitoring one or more effects of perturbing the setpoint or the heat duty; and
generating a set of training data comprising values of the one or more effects.
17. The method of claim 14 , wherein generating the predictive model comprises:
generating one or more candidate models;
selecting one of the one or more candidate models based on an estimated accuracy of each of the one or more candidate models.
18. The method of claim 17 , wherein generating the predictive model comprises:
determining whether the selected candidate model is suitable for use in performing the optimization;
in response to a determination that the selected candidate model is not suitable, generating one or more new candidate models and selecting a new candidate model of the one or more new candidate models; and
in response to a determination that the selected candidate model is suitable, providing the selected candidate model as the predictive model for use in performing the optimization.
19. The method of claim 14 , wherein monitoring the one or more prediction error metrics comprises:
calculating one or more statistical characteristics of each of the one or more prediction error metrics;
determining whether any of the one or more statistical characteristics of each of the one or more prediction error metrics exceeds a predetermined threshold value for each of the one or more statistical characteristics; and
triggering a model update process in response to at least one of the one or more statistical characteristics exceeding the predetermined threshold value.
20. The method of claim 19 , wherein performing the assessment comprises assessing the observability or the condition number of the predictive model.