Dual-Model Control for Technical Systems With Stochastic Inputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
State-of-the-art control systems for complex technical systems, such as wind turbines, face challenges in optimizing control actions due to the dominant influence of stochastic and partially detectable external variables, which impairs the effectiveness of machine learning-based control strategies.
Innovation Solution
The method involves dividing the control system into two models: a first control model for predicting subsequent states and a second control model for predicting the distance between predicted and actual states, allowing for more effective detection of control action influence and reducing training data and time requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a monolithic control model is used to predict subsequent states, then the model structure is simple, but the influence of control actions cannot be effectively detected when external influencing factors dominate
Solution Approach 1:
The control model is divided into two separate models: a first control model that predicts subsequent states without considering control actions, and a second control model that specifically predicts the influence of control actions on those states. This segmentation allows the second model to focus exclusively on detecting control action influences, thereby improving measurement precision while managing complexity through functional decomposition.
2Productivity
If machine learning is used to optimize control strategies, then automatic optimization is achieved, but training is substantially impaired by stochastic external influencing variables
Solution Approach 1:
The first control model extracts and accounts for the dominant external influencing factors by predicting subsequent states based solely on current system states. This extraction removes the confounding effect of stochastic external variables from the training process of the second model, allowing the second model to focus on learning the relationship between control actions and system responses, thereby improving both training efficiency and reliability.
3Measurement precision
If control models are trained with dominant external influencing variables, then prediction of system states is improved, but the influence of control actions becomes undetectable
Solution Approach 1:
The predictive function is segmented into two components: the first control model captures the dominant external influencing variables to achieve accurate system state prediction, while the second control model separately captures the control action influence. This segmentation prevents the loss of control action information by dedicating a specific model component to detect and quantify it, while the first model handles the dominant external factors.
Data Source
AI summary
In order to control a technical system, a system state of the technical system is continually detected. By a trained first control model, a subsequent state of the technical system is predicted on the basis of a sensed system state. Then, a distance value is determined for a distance between the predicted subsequent state and an actually occurring system state. Furthermore, a second control model is trained by the trained first control model to predict the distance value on the basis of a sensed system state and on the basis of a control action for controlling the technical system. A subsequent state predicted by the first control model is then modified on the basis of a distance value predicted by the trained second control model. The modified subsequent state is output in order to control the technical system.