Multi-Stage Model Optimization for Plant Control Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing control systems in primary industries face challenges in performing online optimization efficiently, particularly when physical modeling is not feasible or sensible, leading to the need for adaptive empirical models like neural networks, but these systems often require offline identification of dynamic process models.
Innovation Solution
The setpoint optimizer in the control system takes into account setpoints, manipulated variables, and actual values from model-based controllers to optimize its internal model, allowing for online optimization by using partial models that may or may not be identical to the controller's models, and can incorporate additional system values for optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If adaptive empirical models like neural networks are used for online optimization, then the system can handle cases where physical modeling is not feasible, but the system requires offline identification of dynamic process models which reduces productivity
Solution Approach 1:
The patent pre-configures the optimizer with a library of candidate models and identification algorithms before online operation. During online optimization, the system directly selects and applies pre-validated models rather than performing full model identification, enabling rapid adaptation without offline delays
Solution Approach 2:
The system dynamically switches between different model types (physical models, empirical models, neural networks) based on real-time system conditions and data availability. This dynamic model selection allows the optimizer to adapt its approach online without requiring offline identification for each scenario
2Measurement precision
If the setpoint optimizer uses a comprehensive model of the entire plant, then optimization accuracy is improved, but the device complexity increases
Solution Approach 1:
The patent divides the plant model into multiple hierarchical levels: a simplified global model for overall optimization and detailed local models for specific plant parts. The setpoint optimizer uses the global model for high-level decisions while local models handle detailed control, achieving comprehensive optimization without managing the full complexity of the entire plant
Solution Approach 2:
The system transitions from using a single comprehensive model to a multi-dimensional modeling approach where models operate at different levels of abstraction. The global model operates at a macro level for strategic optimization, while local models operate at micro levels for tactical control, effectively managing complexity through dimensional separation
3Power
If the setpoint optimizer does not use information from model-based controllers, then the optimizer's computational load is reduced, but the optimization becomes less accurate as it lacks real-time operational data
Solution Approach 1:
The patent extracts only the essential real-time operational data from model-based controllers that are most relevant for optimization, rather than processing all controller information. This selective extraction provides the optimizer with critical data (actual values, setpoints, control actions) while minimizing computational overhead from unnecessary data processing
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a plant (1) of the primary industry, having a number of plant components (2). A target value optimizer (4) of a control system (3) of the plant (1) determines threshold values (X*) for the individual plant components (2) based on predetermined target variables (Z) utilizing an optimizer-internal model (5) of the plant (1). The target value optimizer (4) supplies the respective target values (X*) to a model-based controller (8) of the control system (3) controlling the respective plant component (2), at least for part of the plant components (2). The respective model-based controllers (8) receive current values (X) at least from the respective plant component (2). They determine controlled variables (Y) for the respective plant component (2) based on the target values (X*) fed to them, and based on actual values (X) received by them, utilizing a controller-internal model (9) of the plant component (2) regulated by them, and output the controlled variables (Y) determined by them to the respective plant component (2). A model identifier (11) associated with the respective model-based controller (8) receives the controlled variables (Y) output to the respective plant component (2) by the respective model-based controller (8) and the actual values (X) fed to the respective model-based controller (8), determines parameters (P) based on the controlled variables (Y) and actual values (X) received, and optimizes the controller-internal model (9) of the respective plant component (2) based on the parameters (P). The target value optimizer (4) receives the parameters (P) determined by the model identifiers (11) from the same, and optimizes the optimizer-internal model (5) of the plant (1) utilizing the parameters (P).