🟧 INTRODUCTION
In the previous articles of this series, we discussed AI and machine learning applications across various industries, mainly focusing on fuel blending in the refining sector. This final installment presents a case study of developing a Proof of Concept (POC) for an analyzer-less blending system using AI and machine learning.
🟧 ROLE OF AI-BASED MACHINE LEARNING IN FUEL BLENDING
First Principle Blend Models (FPBM) are imperfect, allowing AI-based machine learning to enhance blend quality precision. AI/ML does not replace FPBM but complements it, minimizing residual errors for better precision in blend qualities. This results in increased profitability by reducing quality giveaways.
Additionally, AI/ML enables the creation of an analyzer-less fuel blending system, predicting blend header and final tank qualities without relying on potentially unreliable online analyzers.
The diagram shows all governing mathematical formulations of the First Principle Blend Model (FPBM):
🟧 OBJECTIVE
In a three-tier blend control system—comprising offline optimizer and planning, online control and optimization, and regulatory blend control (RBC)—AI/ML utilizes historical blend data to learn how varying component ratios and properties affect blend qualities. This hybrid approach, combining linear and non-linear models, aims to provide a robust solution for improving prediction accuracy.
🟧 INPUT AND OUTPUT TO AI MODEL
Inputs:
- Optimum or non-optimum component ratios
- Component properties from lab analysis
- Blend product specifications
- Final measured blend qualities (lab or blend header analysis)
- Linear or non-linear blend models
- Prices of products and components, including penalties for quality deviations
Outputs:
- Back-calculated blend header qualities from the final tank analysis
- Optimized component ratios to achieve target blend qualities
- Predicted final tank qualities
- Error bias between AI-predicted and measured values of blend qualities
The AI model requires training with 3-5 years of historical data to capture all process changes due to crude switching and product grade variations. This training is crucial for achieving the desired precision in blend quality predictions.
🟧 CASE STUDY: USA-BASED REFINERY
The case study involves data from 750 gasoline blends from a 300 KB/day refinery. The dataset includes components, recipes, and final tank qualities, covering four gasoline grades from 87 to 92 octane. The AI models developed do not differentiate between gasoline grades.
🟧 CONFIGURATIONS FOR ANALYZER-LESS BLENDING SYSTEM
1. Prediction of Final Product Tank Qualities:
This configuration predicts the final tank qualities based on component qualities, recipes, and lab analyses. AI/ML models can be linear or hybrid, requiring large datasets (5000+ blends) and real-time bias updates for optimal performance.
2. Back-casted Blend Header Qualities:
This approach back-calculates blend header qualities from the final tank analysis. It uses linear models due to their simplicity and minimal error margin. This configuration provides the aggregated blend header qualities necessary for AI/ML training.
3. AI/ML Hybrid Models for Blend Header and Product Tank:
This combines methods to train blend header and product tank qualities using final tank qualities, component qualities, and recipes. It employs mixed linear and non-linear models to create a robust AI/ML hybrid model for an analyzer-less blending system.
🟧 LINEAR VS. HYBRID AI/ML MODELS
Linear models calculate qualities based on component ratios, accounting for about 60% of blend qualities. However, non-linear qualities like RON, MON, RVP, and distillation points require hybrid models. These models use customized parameters specific to refinery processes, providing better accuracy for complex blending scenarios.
🟧 TRAINING THE AI MODEL
The training phase involves predicting back-calculated blend header qualities. While linear models are more straightforward and accessible to implement, hybrid models offer better accuracy for non-linear qualities. Continuous feedback and bias adjustment are essential for refining the model and reducing residual errors.
🟧 ROADMAP TO DEVELOP AI/ML MODELS FOR ANALYZER-LESS FUEL BLENDING SYSTEM
We followed the steps below to develop a Proof of Concept (POC) for an AI/ML-based analyzer-less blending system.
1. Data Collection: Gather 3-5 years of historical blend data, including process changes and crude switches.
2. Data Analysis: Identify and address outliers, inconsistencies, and missing values.
3. Data Consolidation: Integrate all data (qualities, inventories, recipes, lab analyses) into a single interface, often using Excel for optimization.
4. Back-casted Blend Header Qualities: Calculate blend header qualities from final tank analysis to train AI/ML models.
5. Customization of Hybrid Blend Models: Tailor non-linear blend model parameters to specific refinery processes.
6. Training of Linear AI Models: Use component and back-casted blend header qualities for initial model training.
7. Training of Hybrid AI Models: Refine predictions using customized parameters for non-linear blend qualities.
8. Recipe Optimization: Utilize hybrid AI/ML models to optimize blend recipes, incorporating Python-based optimizers.
9. Prediction of Blend Qualities: Compare predicted qualities before and after optimization.
10. Analysis of Quality Giveaways: Validate the AI/ML model’s impact on blend profitability and quality giveaways.
🟧 CONCLUSION
This Proof of Concept demonstrates that a well-trained AI/ML hybrid model can significantly enhance fuel blending systems’ profitability and reliability, providing a comprehensive historical dataset and accurate component qualities. This approach holds promise for refining industry applications, potentially revolutionizing traditional blend control methods.
Disclaimer: OMS eLearning Academy and ChatGPT collaborated as Humans and AI to generate this article for you.
Stay tuned for more groundbreaking publications and enrich your expertise with OMS Academy. 🌐✨
Sign up today to start your learning journey.
➡️ Sign Up Now 🚀
Thank you for being a part of the OMS eLearning Academy community.
If you have any questions or feedback, please contact us.
“At OMS eLearning Academy, excellence in education isn’t just our goal; it’s our promise – a promise to fuel minds and forge futures in Downstream Refining.”