Preference Prediction-Based Evolutionary Multiobjective Optimization for Gasoline Blending Scheduling

Background Introduction

With the continuous evolution of the global energy market, gasoline production and blending processes face increasing challenges. As a key product of the oil industry, gasoline’s blending and scheduling processes directly affect product quality and production efficiency. Gasoline blending requires mixing multiple components in varying proportions according to product specifications and performance requirements to produce gasoline of different grades. During this process, multiple performance indicators must be met, such as Octane Number (ON), Reid Vapor Pressure (RVP), lead content, sulfur content, and flash point. This requires stringent quality control while also complying with increasingly stringent environmental regulations.

The gasoline blending and scheduling task is essentially a multi-objective optimization problem. This type of problem involves multiple conflicting objectives, such as improving product quality, maximizing production efficiency, minimizing equipment utilization, and reducing production costs. In addition, it involves complex constraints and non-linear components, making it difficult for traditional mathematical programming methods—such as relaxation, branch-and-bound, and cutting-plane methods—to fully address. Coupled with the inefficiency and unreliability of experience-based manual scheduling, the demand for more intelligent and automated solutions has become urgent.

Against this backdrop, Multiobjective Evolutionary Algorithms (MOEAs) have emerged as potential solutions due to their strong problem-solving capacity with complex optimization challenges. However, the application of MOEAs to gasoline blending scheduling still faces key challenges, including high computational complexity and the effective integration of decision-making preferences. To address these issues, the research team proposed a novel framework called “Preference Prediction-Based Evolutionary Multiobjective Optimization” (PP-EMO).

Research Origin

This paper, titled “Preference Prediction-Based Evolutionary Multiobjective Optimization for Gasoline Blending Scheduling,” was authored by Wenxuan Fang, Wei Du, Guo Yu, Renchu He, Yang Tang, and Yaochu Jin. These researchers are affiliated with multiple universities and research institutions in China, including East China University of Science and Technology, Nanjing Tech University, China University of Petroleum, and Westlake University. The paper was published in the January 2025 issue of IEEE Transactions on Artificial Intelligence (Volume 6, Issue 1).

Research Workflow

The research team conducted comprehensive modeling and optimization for the gasoline blending scheduling problem and proposed an entirely new algorithmic framework. Below is a detailed description of the paper’s workflow and key contributions:

1. Gasoline Blending Scheduling Model

The research team first modeled the gasoline blending scheduling problem using a discrete-time representation. The model assumes that blending tasks begin and end at fixed time intervals, dividing actual operations into multiple time stages. The specific mathematical model encompasses the following components:

• Sets and Decision Variables: The model defines the number of component and product tanks alongside their capacity and flow rate constraints. Binary variables are set to represent the operational status of component tanks and product tanks at specific time intervals.
• Constraints: These include operational constraints (e.g., each component tank can only deliver oil to one product tank at a time), demand constraints (ensuring that every product tank’s demand is met), capacity constraints, and flow constraints.
• Optimization Objectives:
  
  1. Minimize Blending Error: Calculating the squared deviations of gasoline properties ensures compliance with product quality requirements.
     
  2. Minimize Operating Costs: This includes pipeline opening/closing frequency and operational duration. A weighting parameter is introduced in the cost model to better align with practical conditions.

The model serves as a theoretical basis for designing the subsequent optimization algorithm.

2. Proposed PP-EMO Framework

The PP-EMO framework primarily comprises two main components: machine learning-based preference prediction and Preference-Based Multiobjective Evolutionary Algorithms (PBMOEAs).

Preference Prediction Model

The research team employed a Gaussian Process (GP) model as the tool for preference prediction. The GP model’s non-parametric nature, low computational complexity, and ability to handle uncertainty make it an ideal choice. Specifically, the GP model learns the mapping relationship between environmental variables and reference points from historical data, combining numerical uncertainty estimates to provide optimization reference points for the PBMOEA.

The GP model’s performance was assessed through cross-validation experiments, which demonstrated that the GP model outperformed traditional models like linear regression and Support Vector Machine (SVM).

Optimization Algorithm

The PBMOEA is based on the improved R-NSGA-II (Reference-Based Nondominated Sorting Genetic Algorithm II). The newly designed preference guidance mechanism directs the search direction using reference points, while the parameter ε controls the strength of preferences. Additionally, to counteract the impact of incorrect preferences on optimization efficiency, the researchers proposed a dynamic ε adjustment mechanism linked to the GP model’s uncertainty predictions.

3. Simulation Experiments and Validation

The research team validated the PP-EMO framework using data from a real-world medium-scale refinery. Tasks were categorized into three levels: easy, normal, and difficult, depending on differing complexities in equipment utilization rates, product demand, and property discrepancies. The main findings from these experimental results include:

• Optimization Efficiency: PP-EMO significantly outperformed traditional MOEAs like NSGA-II and MOEA/D-DQN (based on deep reinforcement learning), especially for more challenging tasks.
• Solution Robustness: Compared to other algorithms, PP-EMO delivered more stable performance in both convergence efficiency and solution feasibility across tasks of varying difficulty.
• Execution Time: PP-EMO incurred only a marginal increase in computational cost compared to the basic NSGA-II algorithm while drastically reducing computation time compared to complex algorithms like GMOEA and MOEA/D-DQN.

The comparative experiments highlighted that PP-EMO not only enhances optimization performance for gasoline blending scheduling but also substantially shortens computation time, demonstrating its potential for practical industrial application.

Research Conclusion and Significance

Research Conclusions

The PP-EMO framework successfully addresses the multi-objective optimization challenges in gasoline blending scheduling. By predicting operator preferences through a machine learning model and translating them into optimization parameters, this framework achieves efficient scheduling under tight constraints and conflicting objectives.

Experimental results demonstrated that, compared to preference-free traditional methods, PP-EMO reduced operational costs by approximately 25% and blending errors by around 50%. Its performance was particularly notable under complex operational conditions.

Academic and Practical Value

• Academic Value: The PP-EMO framework introduces a low-coupling approach to integrating evolutionary algorithms with machine learning, providing a novel idea for the application of optimization methods in industrial problems.
  
• Practical Value: In the context of declining gasoline demand and increasing competition in the oil market, this framework provides a cost-efficient and effective production optimization strategy for the oil industry.

Research Highlights

• Innovatively applies GP models for preference prediction, combining them with evolutionary algorithms to achieve intelligent and efficient scheduling optimization.
  
• Demonstrates high robustness, achieving superior performance across various levels of task complexity.
  
• Significantly reduces operating costs, providing direct support for decision-making in refinery scheduling.

Potential Improvements and Future Directions

Despite PP-EMO’s promising potential, its performance for larger-scale problems requires further optimization. Additionally, as the current approach only relies on historical data, future improvements could integrate data generated during the optimization process to refine the preference prediction model further. The research team also plans to extend this framework to other real-world scheduling optimization problems of similar nature, thereby broadening its applicability and impact.