Naphtha Cracking Scheduling With Multi-Agent Reinforcement Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing manufacturing system scheduling methods, particularly in naphtha cracking centers, face challenges due to complex production constraints that are difficult to define mathematically, limiting the application of optimization techniques like mixed-integer linear programming, and relying heavily on expert experience which is variable and hard to verify.

Innovation Solution

A method using reinforcement learning with multi-agents to autonomously determine optimal scheduling for naphtha cracking centers, where each agent handles specific processes, learning through a shared reward mechanism to maximize profit while adhering to constraints.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If mixed-integer linear programming is used for scheduling optimization, then mathematical optimization can be achieved, but it cannot be applied in reality due to complex and difficult-to-define production constraints

Engineering Contradiction:
Improvescheduling optimization accuracyVSAvoidapplicability to real production constraints
Core Design Contradiction:
Manufacturing precisionVSAdaptability or versatility

Solution Approach 1:

The patent replaces traditional mathematical optimization methods (mixed-integer linear programming) with reinforcement learning-based artificial intelligence. This substitution allows the system to handle complex, difficult-to-define production constraints that cannot be easily modeled mathematically, while still achieving scheduling optimization. The RL agents learn optimal scheduling policies through interaction with the production environment, naturally adapting to complex constraints without requiring explicit mathematical formulation.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Ease of operation

If expert experience is used for scheduling decisions, then practical knowledge can be applied, but the results are variable and hard to verify

Engineering Contradiction:
Improvepractical applicabilityVSAvoidconsistency and verifiability
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The patent implements a self-learning system where reinforcement learning agents automatically learn optimal scheduling strategies through interaction with the production environment. The system rewards themselves based on scheduling performance (profit maximization, constraint satisfaction), enabling autonomous improvement without relying on variable expert judgment. This self-service mechanism ensures consistent, verifiable, and reproducible scheduling decisions.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent incorporates feedback mechanisms where the RL agents receive reward signals based on scheduling outcomes (profit, constraint satisfaction). This feedback loop allows the system to learn from past decisions and continuously improve scheduling performance. The verifiable reward structure provides objective criteria for evaluating and comparing different scheduling policies, ensuring reliability and consistency.

Inventive Principle:
Principle #23Feedback

3Productivity

If reinforcement learning with multi-agents is used, then scheduling optimization and profitability increase can be achieved, but system complexity increases

Engineering Contradiction:
Improveprofitability and scheduling efficiencyVSAvoidsystem architecture complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent divides the scheduling problem into multiple independent but coordinated agents, each responsible for specific scheduling decisions (e.g., different production units, time periods, or resource types). This segmentation allows parallel learning and computation, improving efficiency while managing complexity through modular architecture. Each agent learns its specialized policy independently, reducing the overall computational burden compared to a single monolithic agent.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4679341A1Scheduling optimization method and apparatus using reinforcement learning
Publication Date: 2026.01.14 LG MANAGEMENT DEV INST CO LTD
  • EP4679341A1 patent drawingFigure 1
  • EP4679341A1 patent drawingFigure 2
  • EP4679341A1 patent drawingFigure 3A~3B

AI summary

A method for scheduling a naphtha cracking center by at least one processor includes: obtaining input information; determining incoming tank information using a first agent based on the input information; determining mixing tank combination information using a second agent; and determining cracking furnace operation information using a third agent.