Plant Control Optimization Using Pre-Trained Simulation Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing control optimization methods for complex dynamic systems, such as gas turbines and power plants, are inflexible and time-consuming, particularly when adapting to changes in control criteria, due to the complexity of interacting parameters and the need for extensive neural network training.
Innovation Solution
An interactive assistance system that uses a simulation module and optimization module to rapidly optimize control criteria by simulating action sequences, determining rewards, and generating reward-optimizing action sequences, allowing for flexible modification of reward functions and control criteria, leveraging stochastic optimization methods and pre-trained neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional neural network training methods are used to optimize control criteria, then the system can learn optimal control actions, but the training process is time-consuming and slow to adapt to changed control criteria
Solution Approach 1:
The patent pre-trains a neural network model on general system dynamics and control relationships before actual optimization is needed. This preliminary training establishes a baseline understanding of system behavior that can be quickly refined through reinforcement learning with reward functions, avoiding the need for complete retraining when control criteria change
Solution Approach 2:
The patent changes the approach from training the entire neural network from scratch to adjusting specific parameters (reward function weights) that guide optimization. By modifying reward function parameters rather than retraining the whole network, the system rapidly adapts to new control criteria while maintaining previously learned system dynamics knowledge
2Adaptability or versatility
If control criteria are frequently changed to adapt to different application situations, then the system becomes more versatile, but the optimization process becomes increasingly time-consuming
Solution Approach 1:
The system performs preliminary training on general system dynamics once, creating a reusable neural network model. When control criteria change, only the reward function parameters need adjustment rather than complete retraining, enabling rapid adaptation to new situations
Solution Approach 2:
The neural network is trained to understand general system dynamics and control relationships that are universally applicable across different control criteria. This universal understanding allows the same base model to serve multiple different optimization goals by simply changing reward function parameters
3Measurement precision
If a large number of control parameters are considered to accurately represent complex dynamic systems, then the system model becomes more accurate, but the complexity of finding optimal control actions increases significantly
Solution Approach 1:
The patent introduces a neural network as an intermediary that processes the large number of control parameters and system states. The neural network learns to map complex parameter relationships to optimal control actions, simplifying the optimization problem while maintaining accuracy in representing system dynamics
4Reliability
If extensive neural network training is performed to achieve optimal control, then control accuracy improves, but the cost and time required for optimization increase
Solution Approach 1:
The system performs extensive training once in advance to build a robust neural network model. After this preliminary training, subsequent optimizations require minimal additional training time because the network already understands system dynamics, needing only parameter adjustments via reward functions
Data Source
Figure 1
AI summary
The invention relates to an interactive assistance system and method for computer-aided control optimisation for a technical system, e.g. a gas or wind turbine, in particular for optimising the action sequence (A) or the control variables of the plant (e.g. gas supply, compression), wherein an input terminal is provided for reading at least one status parameter providing a first system status (So) of the technical system, and at least one setting parameter (W) for adapting a reward function (RF). A simulation module (SIM) having a pre-trained neuronal network, simulating the plant, serves to simulate an action sequence (A) on the technical system, starting from the first system status (So) and to the prediction of the resulting statuses (S) of the technical system. In addition, an optimisation module (OPT) is provided for adapting the reward function (RF) based on the setting parameter (W), for generating a plurality of action sequences (A) for the first system status (So), for transmitting the action sequences (A) to the simulation module (SIM), and for receiving the resulting statuses (S). Furthermore, the optimisation module (OPT) is provided for determining the rewards of the simulation results using the adapted reward function (RF), and for ascertaining a reward-optimising action sequence (A'). An output terminal (OUT) is provided for issuing a system status (S') resulting from the reward-optimising action sequence (A').