Hybrid Process Simulation Models for Operating-Condition Error Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional process simulation systems rely on fixed gain and bias to correct model predictions, which fail to account for non-linear behavior in industrial processes, leading to inaccuracies and inefficiencies.
Innovation Solution
A hybrid process simulation model combining a first principles model with a data-driven model, specifically a neural network or regression model, is used to generate error predictions and correct model predictions as a function of operating conditions, improving prediction accuracy and enabling offline simulation studies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional process simulation systems use fixed gain and bias to correct model predictions, then the correction method is simple and easy to implement, but the prediction accuracy deteriorates due to inability to account for non-linear behavior
Solution Approach 1:
The patent transforms the correction approach from using fixed parameters (gain and bias) to using dynamic parameters that vary with operating conditions. The neural network model learns optimal correction parameters across different operating scenarios, allowing the system to adapt to non-linear behavior while maintaining implementation feasibility through a structured training process.
Solution Approach 2:
The patent introduces dynamics into the correction system by making the correction factors dependent on operating conditions rather than static. The neural network dynamically adjusts correction parameters based on the current operating state, enabling the system to handle non-linearities while preserving the simplicity of the correction framework.
2Measurement precision
If multiple simulations are performed to achieve accurate predictions, then prediction accuracy improves, but computing time and efficiency deteriorate
Solution Approach 1:
The patent performs the computationally intensive work in advance by training the neural network model offline using historical simulation data and ground truth measurements. Once trained, the model can quickly apply learned correction patterns to new predictions without requiring multiple simulations, thus achieving high accuracy with minimal real-time computing overhead.
Solution Approach 2:
The patent creates a corrected version of the simulation model by combining the first principles model with the trained neural network correction model. This hybrid model replicates the accuracy of multiple simulations while executing with the speed of a single simulation, effectively copying the benefits of extensive simulations without the computational burden.
3Measurement precision
If a data-driven model is trained on historical data to correct predictions, then prediction accuracy improves, but model complexity increases
Solution Approach 1:
The patent introduces a neural network model as an intermediary component that bridges the first principles model and the actual process behavior. This intermediary learns the complex non-linear relationships from historical data and translates them into correction factors, adding manageable complexity only where needed to improve accuracy without overcomplicating the entire system.
Data Source
Figure 1
Figure 2
Figure 3A~3B
AI summary
Embodiments of the disclosure provide for generating optimal model prediction for process simulation. Some embodiments receive input dataset associated with the operation of an industrial plant, generates one or more predictions based on the input dataset and using a hybrid process simulation model. The one or more predictions may include FP model-predicted data indicative of predicted value for each of one or more process variables and error prediction indicative of a discrepancy between the FP model-predicted data and ground-truth data for the one or more process variables. Optimal model-predicted data may be generated based on the FP model-predicted data and the error prediction the optimal model-predicted data may be indicative of optimal process simulation output. Performance of one or more prediction-based actions may be initiated based on the optimal model-predicted data.