Site-Wide Set-Point Optimization Using MILP Surrogate Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems struggle to perform real-time adjustments for optimal set points in production processes due to the lack of plant simulators and reliance on historical data, especially in big data operations, limiting the ability to optimize industrial operations effectively.
Innovation Solution
A computer-implemented method using mixed regression models with mixed control variables, including decision tree regression models reformulated into MILP, to solve nonlinear optimization problems, enabling real-time set-point adjustments through decomposition and federated learning, ensuring data privacy and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If surrogate models are used for optimizing production processes during design phase, then optimization capability is provided, but real-time adjustment ability is lost
Solution Approach 1:
The system performs preliminary actions by training multiple regression models (polynomial, neural network, support vector regression) during the design phase to capture complex nonlinear relationships. These pre-trained models are then reused in real-time operations through an ensemble approach, eliminating the need for real-time model training while maintaining both optimization capability and real-time responsiveness
Solution Approach 2:
The patent creates a universal optimization system that handles multiple types of control variables (continuous, discrete, categorical) and multiple model types within a single framework. The ensemble model serves multiple functions: it provides optimization during design phase, enables real-time adjustments, and works across different production scenarios without requiring separate systems for each function
2Extent of automation
If plant simulator based on surrogate model is not available, then real-time optimization execution is impossible, but historical data only approach lacks automatic execution capability
Solution Approach 1:
The system enables self-service by using the trained regression models to automatically execute real-time optimizations without requiring external plant simulators. The models take current operating conditions as input and directly generate optimized setpoints, providing automatic execution capability while maintaining reliability through the robustness of the pre-trained ensemble models
Solution Approach 2:
The patent introduces the ensemble of regression models as an intermediary between historical data and real-time optimization execution. These models serve as mediators that translate historical patterns into actionable real-time recommendations, enabling automatic execution without requiring a full plant simulator while maintaining optimization reliability
3Measurement precision
If deep neural network is used for predicting controllable variable status, then prediction accuracy is improved, but computational overhead and training complexity increase
Solution Approach 1:
Instead of using a single complex deep neural network, the patent applies partial action by combining multiple simpler regression models (polynomial, neural network, support vector regression) in an ensemble. Each model captures different aspects of the nonlinear relationships, and their combination achieves accuracy comparable to deep neural networks while significantly reducing training complexity and computational overhead through parallel, independent training of simpler models
Data Source
AI summary
A computer-implemented method for site-wide prediction optimization includes training a plurality of a mixed type of regression models with a mixed type of control variables for identifying control set-points of a site-wide operation. A decision tree regression model is trained to predict a status of the plurality of initial set-points for non-linear regression functions. The decision tree regression model is reformulated into a mixed-integer linear program (MILP) and solved by an MILP solver to find a global solution. An MILP surrogate is determined for a nonlinear optimization problem to provide a best solution for one or more of the non-linear regression functions using the best solution as a starting point for solving non-linear regression functions, and a set-point of the mixed control variables is recommended to control a throughput of the site-wide operation by executing a decomposition operation or a federated learning algorithm.