关键词: 原油短期调度, 代理模型辅助, 多目标优化进化算法, 神经网络

Abstract: The refining production process is characterized by uncertainty, multiple objectives, and numerous constraints, and is an NP-hard problem. In the short-term production scheduling of crude oil processing, each sub-scheduling involves a series of combined operations on equipment such as oil tankers, charging tanks, distillation towers, etc. Consequently, the quantitative relationship between each piece of equipment has a decisive impact on the feasibility of the scheduling solution. For instance, if a production plan necessitates the temporary addition of distillation towers to increase output, or if certain charging tanks require maintenance shutdowns prior to production, this will result in a shortage of charging tanks relative to distillation towers. This will cause the production system to enter a non-equilibrium state, intensifying competition for charging tank resources among related equipment such as pipelines and distillation towers. As a result, the complexity and challenges of production scheduling are significantly heightened.
Currently, there is only a feasibility analysis of crude oil short-term scheduling of non-equilibrium in existing research, without considering the optimization of related costs, and this may result in substantial financial waste during actual production. So, it is necessary to conduct multi-objective optimization research on crude oil short-term scheduling under non-equilibrium conditions between charging tanks and distillation towers. Based on the resource allocation backtracking search method, a multi-objective optimization mathematical model is constructed to find feasible schedules and minimize the five target costs including mixing cost of crude oil in the pipeline and at the bottom of charging tanks, the switching cost of charging tanks, the use cost of charging tanks and the energy consumption cost of crude oil transportation. The feasible scheduling must adhere to all constraints. However, the mathematical model constructed for the problem includes a critical constraint on crude oil residence time. Under conditions of relative equipment resource scarcity, backtracking search is prone to frequent violations of the crude oil residence time constraint, rendering the current scheduling solution infeasible. This issue triggers extensive backtracking, which significantly reduces the efficiency of solving the problem. To improve the quality and efficiency of solving this model, this paper first introduces the Parallel Long-Short Term Memory (PLSTM) module into the Pairwise Comparison Surrogate-Assisted Evolutionary Algorithm (PC-SAEA) and then proposes PLSTM-PC-SAEA. The SAEA leverages computationally efficient surrogate models to approximate optimal values with limited computing resources, showing certain advantages in solving single or multiple objective optimization problems. Within the basic framework of the Surrogate-Assisted Evolutionary Algorithm (SAEA), the surrogate model can be constructed, trained, and tested using solutions that have completed Fitness Evaluation (FE). The model is then utilized to assess the quality of candidate solutions through auxiliary FE. This method selects a portion of better candidate solutions for real FE, conserving computational resources and improving the algorithm’s solving efficiency. The PLSTM model includes a set of parallel LSTM modules with independent parameters, which can make each LSTM network more focus on learning the type features of the corresponding input and clearly distinguish the different permutations and combinations of label 1 and label 0. When the dataset is constructed, based on the principle of pairwise comparison, better and worse solutions are combined into a series of regular tuples and dual tuples according to specific rules. This strategy significantly expands the dataset size and enhances the training effectiveness of the model. Before candidate solution auxiliary evaluation is conducted, the surrogate model undergoes reliability testing, and reliability labels are obtained according to the model management strategy. Then, auxiliary FE is conducted using the corresponding strategy managed by the test result. Additionally, an energy consumption optimization Linear Programming (LP) model is designed to optimize the energy consumption cost, explicitly embedded in the encoding and decoding process of the solution chromosomes.
Finally, the PLSTM-PC-SAEA algorithm is applied to an industrial example of a real crude oil short-term scheduling problem provided in the reference. It is compared with six other comparative algorithms by comparing the metrics like CR, HV, and algorithm operation time. Experimental results demonstrate that the proposed approach achieves favorable overall performance in terms of convergence, solution diversity, and computational efficiency. This provides a valuable reference for solving crude oil short-term scheduling problems under non-equilibrium conditions between charging tanks and distillation towers.

Key words: crude oil short-term scheduling, surrogate-assisted, multi-objective optimization evolutionary algorithm, neural network