Neural Network Control for Gas Turbines
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for controlling complex technical systems, such as reinforcement learning, are not universally applicable and often do not provide sufficient results, limiting their effectiveness in predicting and optimizing the dynamic behavior of technical systems.
Innovation Solution
A method that uses neural networks to model a quality function and learn an action selection rule based on optimality criteria, allowing for the selection of optimal actions in technical systems, with the ability to handle both discrete and continuous actions, and improving data efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If reinforcement learning methods are used to control complex technical systems, then the system can learn optimal control strategies, but the methods are not universally applicable and often do not provide sufficiently good results
Solution Approach 1:
The patent creates a universal control method that combines dynamic programming and reinforcement learning into a single framework (Deep Deterministic Policy Gradient). This approach can be applied to any continuous control problem with a differentiable environment model, making it universally applicable while maintaining high control effectiveness through the integration of value function approximation and policy optimization.
Solution Approach 2:
The patent transforms the control problem by changing parameters from discrete actions to continuous action spaces, and from direct policy learning to gradient-based policy optimization. This parameter transformation enables the method to handle continuous control problems effectively while maintaining universal applicability across different technical systems.
2Reliability
If expert knowledge is used to create automatic control, then the control can be based on human expertise, but it requires extensive domain knowledge and is difficult to generalize
Solution Approach 1:
The patent replaces the mechanical knowledge representation approach (expert rules and heuristics) with a neural network-based functional approximation system. The deep neural network automatically learns the value function and policy from data, substituting explicit knowledge encoding with implicit pattern recognition, thereby reducing the complexity of knowledge representation while maintaining or improving control quality.
Solution Approach 2:
The patent uses neural networks to copy and generalize expert behavior patterns from training data without requiring explicit encoding of expert knowledge. The network learns to replicate optimal control strategies by observing state-action-reward trajectories, enabling generalization to new situations without requiring explicit knowledge transfer.
3Measurement precision
If more data sets are used to learn the action selection rule, then the accuracy of the control policy improves, but the data requirement increases the complexity and cost of the system
Solution Approach 1:
The patent performs preliminary action by pre-training the neural network with a relatively small data set to establish an initial value function approximation. This preliminary learning phase allows the system to achieve reasonable control accuracy with limited data, and subsequent fine-tuning can be performed with even smaller data sets, thereby reducing the overall data volume requirement while maintaining high policy accuracy.
Data Source
Figure 1
Figure 2
AI summary
The invention relates to a method for the computer-aided control and/or regulation of a technical system. The method involves the use of a cooperative learning method and artificial neural networks. In this context, feed-forward networks are linked to one another such that the architecture as a whole meets an optimality criterion. The network approximates the rewards observed to the expected rewards as an appraiser. In this way, exclusively observations which have actually been made are used in optimum fashion to determine a quality function. In the network, the optimum action in respect of the quality function is modelled by a neural network, with this skilled neural network supplying the optimum action selection rule for the given control problem. The invention can be used in any technical systems for regulation and control, a preferred area of application being the regulation and control of turbines, particularly of the gas turbine. In addition, the invention has the advantage that it can be used for control methods with continuous actions.