Online System Identification with Automatic Training Data Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing building control systems face challenges in generating accurate predictive models due to complex, nonlinear system dynamics, and current methods lack automation in selecting suitable training data for system identification, often resulting in unreliable models.
Innovation Solution
A method and system that assess various characteristics of data segments during a training period to automatically select and exclude data, ensuring correlation between setpoints and measurements, sufficient load conditions, and proper frequency of setpoint changes, thereby training a system model for accurate predictive control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If automatic training data selection is implemented through multiple assessments, then model accuracy is improved, but system complexity increases
Solution Approach 1:
The training period is divided into multiple segments, and each segment is independently assessed against multiple criteria (correlation assessment, load assessment, setpoint change assessment). This segmentation allows the complex selection process to be broken down into manageable, modular assessments that can be systematically applied to improve model accuracy without overwhelming system complexity.
Solution Approach 2:
The system performs preliminary assessments of training data segments before actual model training occurs. By pre-evaluating data quality through multiple assessments (correlation, load, setpoint changes) and selecting only suitable segments, the system ensures high-quality training data is prepared in advance, improving model accuracy while keeping the training process itself simpler.
2Reliability
If multiple assessments are performed on training data segments, then data quality is improved, but processing time increases
Solution Approach 1:
The training period is divided into multiple segments that can be independently assessed. This segmentation allows parallel processing of different segments through the multiple assessments (correlation, load, setpoint change), improving data quality while reducing overall processing time compared to sequential evaluation of entire training datasets.
Solution Approach 2:
The system performs assessments on partial segments of training data rather than evaluating entire datasets. By selectively assessing and selecting only the most suitable segments that meet the multiple criteria, the system achieves high data quality while avoiding the time cost of processing unnecessary or low-quality data portions.
3Reliability
If strict assessment criteria are applied to training data, then model reliability is improved, but data quantity available for training decreases
Solution Approach 1:
By dividing the training period into multiple segments and applying strict assessment criteria to each segment independently, the system can identify and select multiple high-quality segments. This segmentation approach ensures that only reliable data from segments meeting all criteria (correlation, load, setpoint changes) is used, improving model reliability while potentially finding sufficient quantity across multiple selected segments.
Solution Approach 2:
The system changes the parameters of training data segments by selecting only those that meet strict assessment criteria. This parameter-based selection (filtering segments based on correlation coefficients, load levels, and setpoint change frequencies) ensures high model reliability by using only data with appropriate characteristics, while the system can accumulate sufficient data quantity by selecting multiple segments that individually meet the criteria.
Data Source
AI summary
A method includes performing a plurality of assessments of different characteristics of data for a segment of the training period, including the data for the segment in a set of training data or excluding the data from the set of training data based on results of the plurality of assessments for the segment, repeating the plurality of assessments for additional data of a plurality of additional segments of the training period, including the additional data in the set of training data or excluding the additional data from the set of training data based on results of the plurality of assessments for the plurality of additional segments, training a system model for a system using the set of training data, controlling the system using the system model.