Neural Predictive Control with Offline Learning for Nonlinear Plants
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network-based control systems face challenges in properly training and applying to dynamic systems that exhibit nonlinear behavior, as these systems change over time, making it difficult to predict and control processes in industrial settings like petroleum refineries and chemical manufacturing plants.
Innovation Solution
A controller that trains a neural network model offline using historical data to generate a linear predictor for the current state of the plant, allowing for predictive control by adjusting weights to minimize errors and optimize manipulated variables, thereby linearizing nonlinear dynamics and improving control accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network is trained offline using historical plant data to predict nonlinear dynamics, then the control accuracy for dynamic systems is improved, but the ability to adapt to real-time changes in plant behavior deteriorates
Solution Approach 1:
The patent implements a hybrid control system that combines offline-trained neural networks with online adaptive mechanisms. The neural network provides accurate predictions based on historical data, while online adaptation mechanisms continuously adjust the model to accommodate real-time changes in plant behavior, resolving the contradiction between accuracy and adaptability
Solution Approach 2:
The system performs preliminary offline training of the neural network using historical plant data to establish accurate baseline predictions. This preliminary action captures general nonlinear dynamics patterns, while subsequent online operations focus on adapting to specific real-time variations, thereby maintaining both accuracy and adaptability
2Measurement precision
If the neural network is trained to capture complex nonlinear behavior, then the prediction accuracy improves, but the computational complexity and training time increase
Solution Approach 1:
The patent divides the control system into offline training phase and online operation phase. During offline training, the neural network learns complex nonlinear patterns from historical data. During online operation, pre-computed control actions are applied with minimal real-time computation, thereby achieving high prediction accuracy while reducing operational computational complexity
Solution Approach 2:
The system performs extensive training offline to capture comprehensive nonlinear dynamics, then uses this pre-learned knowledge during online operation. This partial action approach (extensive offline, minimal online) achieves high accuracy without requiring continuous heavy computation during plant operation
3Reliability
If the system uses extensive historical data for training, then the model generalization improves, but the training time and data processing requirements increase
Solution Approach 1:
The patent performs comprehensive data processing and model training as preliminary offline actions using extensive historical plant data. This ensures the neural network achieves good generalization capabilities before deployment. The time-consuming operations are completed in advance, allowing fast online operation with minimal additional training time
Data Source
AI summary
A controller for a plant that exhibits nonlinear dynamics includes one or more processors and memory storing instructions that cause the one or more processors to perform operations. The operations include training a neural network model during an offline learning period using historical plant data representing a plurality of different historical states of the plant and using the neural network model during online operation of the plant to generate a linear predictor as a function of a current state of the plant, the linear predictor defining a linearization of the nonlinear dynamics localized at the current state of the plant. The controller controls equipment that operate to affect the current state of the plant by performing a predictive control process that uses the linear predictor to generate values of one or more manipulated variables provided as inputs to the equipment.