(195b) Using Linear Model Decision Tree Surrogates to Integrate Plant Scheduling with Supply Chain Under Disruptions

Supply chains from the chemical industry exhibit high variability and uncertainty due to their multi-step production and tailored product portfolios. This product customization often leads to demand uncertainty, as products may be aligned with individual customers rather than broader market trends. Similarly, the various processing steps that occur in different geographical locations make supply chains from the chemical industry more vulnerable to logistical disruptions.

Decentralized and empirical responses to disruptions often result in suboptimal outcomes. To overcome this challenge, various multiperiod reactive models have been proposed, such as the one introduced by Ovalle et al. (2023) [1]. While these models address large-scale supply chain problems, they frequently do not consider in-site plant scheduling constraints, potentially leading to infeasible operations. Thus, our objective in this study is to integrate supply chain disruption responses with in-site plant scheduling to achieve optimal operations feasible at both supply chain and plant levels.

Supply chain decisions are typically made on a daily or weekly basis, while plant scheduling occurs hourly. Bridging this temporal gap to integrate supply chain planning and plant scheduling under disruptions is a significant challenge [2]. One approach is to construct models operating at finer time discretization, although this can become intractable for larger time horizons and networks. Alternatively, machine learning models, such as deep-neural networks and decision trees, have been successfully employed to surrogate the scheduling feasibility space and integrate it with the supply chain model, as demonstrated in existing literature [3].

To effectively integrate supply chain planning and in-site scheduling in the context of disruptions, we first build from the model proposed in Ovalle et al. (2023)[1] a new model that optimizes the reactive response on arbitrary network topologies. Separately, we develop detailed resource task network (RTN) models for each plant in the system to later sample their feasibility space. Utilizing binary classification linear model decision trees, as suggested by Ammari et al. (2023)[4], we construct surrogate models. Given the potential for a large number of raw materials, intermediate products, and final products within each plant, the number of features required for the surrogates might be large. To address this problem, we integrate the surrogate through aggregation of the key variables (e.g., inventories, demands, production etc.) to reduce the number of features required and the size of the sampling space. Moreover, we introduce an efficient data generation procedure, that proves to be more efficient than standard Monte Carlo schemes [4].

We evaluate the results of our proposed integration pipeline in a case study adapted from Ovalle et al. (2023), which had two plants with 8 and 7 distinct materials respectively. Our results show that the integrated model is able to guarantee in-site feasibility in all scenarios at the expense of a modest reduction in the profit of the operation. Furthermore, the integrated model is computationally tractable as our approach yields one order of magnitude reduction in terms of constraints and variables when compared to monolithic approaches.

References

[1] Ovalle D., Ye Y., Harshbarger K., Bury S., Wassick J. M., Laird C. D. and Grossmann I. E., “Operation optimization of supply chain networks under disruptions in Proceedings of FOCAPO/CPC, 2023

[2] Grossmann I., “Enterprise‐wide Optimization: A New Frontier in Process Systems Engineering.” AIChE Journal, vol. 51, no. 7, July 2005, pp. 1846–57. DOI.org, ttps://doi.org/10.1002/aic.10617

[3] Dias L. S. and Ierapetritou M. G., “Data-Driven Feasibility Analysis for the Integration of Planning and Scheduling Problems.” Optimization and Engineering, vol. 20, no. 4, Dec. 2019, pp. 1029–66, https://doi.org/10.1007/s11081-019-09459-w

[4] Ovalle D., Vyas J., Laird C. D. and Grossmann I. E., “Integration of Plant Scheduling Feasibility with Supply Chain Network Under Disruptions Using Machine Learning Surrogates” to appear in Proceedings of ESCAPE-PSE, 2024