Building System Dynamic Modeling for Fast Adaptive Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional system identification methods for model predictive control systems are inadequate as they assume linear models, fail to distinguish between external disturbances and system changes, and require lengthy training periods, leading to suboptimal and unstable control actions.
Innovation Solution
A method that filters training data to remove extraneous disturbances, optimizes model parameters to minimize error cost functions, and recursively updates Kalman gain parameters using extended Kalman filter techniques to adapt to changing systems, allowing for rapid and accurate identification of system parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional system identification methods are used, then the model can be developed, but the training period is long (two days or more) and the model requires additional training to adapt to changing physical systems
Solution Approach 1:
The patent implements a dynamic model adaptation mechanism that automatically adjusts model parameters in response to changing system conditions. The system continuously monitors system behavior and recalibrates the model without requiring lengthy retraining periods, enabling the model to adapt to physical system changes in real-time while maintaining identification speed.
Solution Approach 2:
The system performs self-calibration by automatically detecting changes in system dynamics and adjusting its own parameters. The model identifies its own discrepancies between predicted and actual behavior, then autonomously retunes its parameters without external intervention or lengthy retraining processes, achieving both speed and adaptability.
2Measurement precision
If traditional linear model methods are used, then the model development is simpler, but the model accuracy is reduced due to actuator saturation and non-linear system behavior
Solution Approach 1:
The patent transforms the linear model into a non-linear model by introducing parameter transformations that capture actuator saturation effects. The system maintains mathematical tractability while accurately representing non-linear behavior through parameter redefinitions, achieving high measurement precision without excessive model complexity.
Solution Approach 2:
The patent introduces an intermediate transformation layer between the linear model framework and the non-linear system behavior. This intermediary mechanism uses parameter transformations to bridge the gap, allowing the model to accurately represent non-linear phenomena while retaining the computational simplicity of linear modeling approaches.
3Reliability
If traditional system identification methods are used, then the model can predict future system states, but the control actions become suboptimal and possibly unstable due to model inaccuracies
Solution Approach 1:
The patent implements a feedback mechanism where the system continuously compares predicted outputs with actual system behavior and uses this information to refine model parameters. This closed-loop approach ensures that control actions remain optimal and stable by constantly adapting the model to reflect current system conditions, preventing divergence and instability.
Data Source
AI summary
A controller for a building system receives training data including input data and output data. The output data indicate a state of the building system affected by the input data. The controller pre-processes the training data using a first set of pre-processing options to generate a first set of training data and pre-processes the training data using a second set of pre-processing options to generate a second set of training data. The controller performs a multi-stage optimization process to identify multiple different sets of model parameters of a dynamic model for the building system. The multi-stage optimization process includes a first stage in which the controller uses the first set of training data to identify a first set of model parameters and a second stage in which the controller uses the second set of training data to identify a second set of model parameters. The controller uses the dynamic model to operate the building system.