Data-Driven Model Training for Gas Turbines Using Density Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data-driven models for technical systems like gas and wind turbines are inefficient due to uneven data distribution, leading to suboptimal control actions and false alarms, as they often rely on abundant typical operating data and lack sufficient representation of rare or unusual operating conditions, resulting in resource-intensive computations and potential system damage.
Innovation Solution
A method that iteratively learns a data-driven model using a density estimator to select and weight training data sets based on confidence measures and model errors, prioritizing data sets with low confidence and high errors, thereby focusing on novel and less-represented operating conditions, and utilizes neural networks for efficient modeling and control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all training data records are used to train the data-driven model, then the model comprehensively represents the technical system, but the training process becomes computationally expensive and resource-intensive
Solution Approach 1:
The training data is segmented into different density regions using a density estimator. The method divides the state space into high-density regions (common operating conditions) and low-density regions (rare operating conditions), allowing selective processing of different data segments with appropriate weighting to balance model accuracy with computational efficiency
Solution Approach 2:
Different weighting strategies are applied to different regions of the training data based on their density characteristics. High-density regions receive lower weights while low-density regions receive higher weights, creating a non-uniform weighting scheme that optimizes the trade-off between representing common operations and capturing rare but critical operating conditions
2Productivity
If training data is randomly selected to speed up computation, then the training process becomes faster, but relevant information about the technical system is lost
Solution Approach 1:
Before selecting training data for model training, a density estimator is first trained on the complete training dataset to identify high-density and low-density regions. This preliminary analysis enables informed selection of training samples that preserve critical system behavior information while reducing computational load
Solution Approach 2:
The density estimator provides feedback about the distribution characteristics of the training data, which is then used to adjust the sampling or weighting strategy. This feedback mechanism ensures that the selected training subset maintains representativeness of the complete dataset, particularly for rare operating conditions
3Ease of manufacture
If a data-driven model is trained only on typical operating conditions, then the model training is simpler and faster, but the model produces false alarms and incorrect control actions during rare operating modes
Solution Approach 1:
The method changes the parameter distribution of the training data by applying density-based weighting. This transforms the training process from uniform sampling to targeted sampling that emphasizes rare operating conditions, improving the model's ability to handle edge cases without significantly increasing training complexity
Data Source
Figure 1~2
AI summary
The invention relates to a method for the computer-supported generation of a data-driven model (NM) of a technical system, in particular of a gas turbine or wind turbine, based on training data. The method according to the invention is characterized in that the data-driven model is preferably learned in regions of training data having a low data density. According to the invention, it is thus ensured that the data-driven model is generated for information-relevant regions of the training data. The data-driven model generated by the method according to the invention is used in a particularly preferred embodiment for calculating a suitable control and/or regulation model or monitoring model for the technical system. By determining optimization criteria, such as low pollutant emissions or low combustion dynamics of a gas turbine, the service life of the technical system in operation can be extended. The data model generated by the method according to the invention can furthermore be determined quickly and using low computing resources, since not all training data is used for learning the data-driven model.