Scientists propose a multi-objective optimization algorithm based on diffusion model to improve the production efficiency of the refining process

At present, in the refinery, the gasoline blending and scheduling problem basically relies on manual scheduling, and the existing research results are not well applied to actual production.

Through the communication with the refinery dispatcher, the team of associate professor Du Wei of East China University of Science and Technology believes that the main reason is that the existing optimization method needs to re-set parameters and strategies when the working conditions of the production process change, which greatly reduces the enthusiasm of the dispatcher to use the optimization algorithm.

图 | 堵威团队（来源：堵威）

Therefore, in a recent study, Du Wei and his team hope to combine diffusion models and multi-objective optimization techniques to solve the problem of industrial process scheduling optimization.

The main purpose of gasoline blending is to optimize the inventory of component oil and refined oil as much as possible from the perspective of the existing resources of the refinery, comprehensively consider the blending capacity of the refining enterprise, the product market demand and the product index, and maximize the economic benefits of the refinery while meeting the upper demand plan.

This problem faces challenges such as large-scale, large number of constraints, mixed integers, and nonlinearity, and previous studies have focused on reducing the complexity of the problem on the modeling side, so that the existing optimizers can solve it.

However, these methods of making additional assumptions about the problem limit the upper limit of the final solution to a large extent, and these assumptions are usually based on specific working conditions, and once applied to different scales, different raw materials, and different pipe structures, the whole method needs to be reversed from the modeling level.

This is undoubtedly not conducive to the universality of the designed method, which also leads to the fact that in the current actual industrial site, dispatchers often rely on manual experience to carry out the work of gasoline blending and scheduling.

Therefore, this study attempts to effectively solve the problem of gasoline blending and scheduling that fits the actual industrial application scenarios. However, the past methods, whether they are mathematical programming optimizers based on traditional operations research or evolutionary algorithms based on stochastic optimization that have emerged in recent years, are difficult to effectively solve large-scale mixed integer optimization problems with a large number of constraints.

Therefore, it is necessary to use a new approach that is very different in principle. In this study, the Diffusion model-based Multiobjective Optimization (DMO) algorithm is used to introduce the diffusion model, a representative model of image generation, into the field of scheduling optimization by taking advantage of the property that "the Gantt chart of the scheduling process is an image".

The research group learns the distribution of historical operational data by using a diffusion model, and iterates to obtain the final solution with the assistance of the target gradient.

The experimental results show that DMO can not only improve the production efficiency of the refining process, but also reduce costs and improve product quality, which also fills the gap of previous research and provides a new idea for industrial scheduling optimization.

DMO starts with the original form of the problem, relying solely on historical operational data, which is very easy to access for the plant, so using DMO does not add additional burden to dispatchers.

At the same time, the DMO runs very quickly, and dispatchers can try it over and over again, reducing the time cost of learning the tool.

In the initial stages of DMO applications, dispatchers can use the optimal solution generated by the DMO as a reference to assist manual scheduling, and finally transition to solving using algorithms.

Therefore, DMO overcomes the shortcomings of previous methods in gasoline blending scheduling, so that it can be widely used in actual production.

For other application scenarios, the research group believes that:

First, DMO also has great potential for applications in other parts of the refinery, such as crude oil scheduling, which is similar to gasoline blending scheduling.

Second, they believe that DMO also has great potential for application in product design problems, which usually have a large number of optimization parameters, complex constraints, multi-objectives, etc., and that DMO can generate a large number of solutions with different characteristics, even if these solutions cannot be truly practical in the end, they can still help and inspire designers to carry out further design.

In fact, in their previous work, the research group has done a lot of research on the problem of gasoline blending and scheduling, and has had many exchanges with the refinery dispatcher.

Having a specific understanding of the characteristics of the problem and the shortcomings of existing methods, they realized that in order for the algorithm to be useful, they needed to follow the following two points:

First, model according to the original data as much as possible to ensure the applicability of the algorithm;

Second, the external environment of the refinery is changing, and the compromise between loss and efficiency is not immutable, so it should be seen as a multi-objective issue.

The former leads to a large number of decision variables and constraints, and the latter requires a large number of solutions to be generated at once to form the Pareto frontier, which increases the complexity of the problem far beyond the capabilities of traditional optimization algorithms and evolutionary algorithms, so they need to find a new approach.

They noted that in recent years, the diffusion model has begun to rise in the field of image generation, and it has rapidly overwhelmed the previously widely used generative adversarial network (GAN) by iteratively generating images from Gaussian noise.

On the one hand, the diffusion model can handle a large number of parameters such as image pixels, and on the other hand, the batch-based method used in deep learning can generate a large number of solutions at the same time, which meets the two requirements mentioned above.

More importantly, the iterative nature of diffusion models is a disadvantage in the field of image generation, which means that it consumes a lot of computing power and takes a long time to compute.

However, in the field of optimization, this process is easily associated with the iteration in the optimization algorithm, the solution in the same batch is like a population in the evolutionary algorithm, and the diffusion model is like the variation operation in the evolutionary algorithm with each iteration of the solution.

This similarity led them to apply the diffusion model to the gasoline blending scheduling optimization problem.

Since the application of diffusion models to optimization problems is a new attempt, there is no ready-made work to refer to.

At the same time, the research on diffusion models in the field of image generation also focuses on reducing the amount of computation and enriching the generation content, and cannot be directly appropriated to the optimization field, so it is difficult to carry out this research "from scratch".

In their research, they refer to the early architecture of the diffusion model, which used a stand-alone discriminator to assist in the generation of the specified content, and an objective function to replace the discriminator.

After the framework is built, there are still many choices, such as whether to use CNNs or Transformer for the network, how to design the weights between goals, how to implement multi-objective optimization, and whether to join the selection mechanism?

After trial and error, they decided on the DMO method they were currently using. In the end, the experimental results proved that their intuition was accurate, and DMO had a significant advantage over the comparison algorithms.

"The collaborators of this work include Academician of the European Academy of Sciences, IEEE Fellow, Professor Jin Yaochu of Westlake University, IEEE Fellow, Professor Gary G. Yen of Oklahoma State University, IEEE Fellow, Professor Tang Yang of East China University of Science and Technology, etc. ”

They are all authoritative scholars in the field of artificial intelligence, and when they first discussed the idea with them, they all found it very interesting, especially by using the fact that "the Gantt chart is an image" to cleverly introduce the generative model technology of diffusion model into the solution of scheduling optimization problems.

This also gave Du Wei and his team great encouragement to realize that their research work is at the forefront of international scholarship.

日前，相关论文以《基于扩散模型的汽油调合调度多目标优化》（Diffusion Model-Based Multiobjective Optimization for Gasoline Blending Scheduling）为题发表在 IEEE Computational Intelligence Magazine（DOI: 10.1109/MCI.2024.3363980）。

Fang Wenxuan, a master's student, is the first author, and Wei Du serves as the corresponding author[1].

In subsequent studies, they plan to combine more actual production data and experimental results to validate and refine their proposed method. In addition, the potential of this method for other industrial processes will be explored for completely different types of optimization problems, such as combinatorial optimization problems.

Resources:

1. Hatps://ArXiv.org/PDF/2402.14600.pdf

Operation/Typesetting: He Chenlong