Bilevel optimisation with embedded neural networks: Application to scheduling and control integration

TL;DR

This paper introduces a novel approach using ReLU neural networks to transform bilevel scheduling and control problems into a single-level MILP, improving computational efficiency and solution accuracy.

Contribution

It proposes a new neural network-based reformulation of bilevel problems into MILP, enabling efficient and accurate integrated scheduling and control optimization.

Findings

Outperforms traditional hierarchical and monolithic approaches in computational time.

Maintains high solution accuracy with neural network-based reformulation.

Embedding neural networks into MILP has minimal impact on solver performance.

Abstract

Scheduling problems requires to explicitly account for control considerations in their optimisation. The literature proposes two traditional ways to solve this integrated problem: hierarchical and monolithic. The monolithic approach ignores the control level's objective and incorporates it as a constraint into the upper level at the cost of suboptimality. The hierarchical approach requires solving a mathematically complex bilevel problem with the scheduling acting as the leader and control as the follower. The linking variables between both levels belong to a small subset of scheduling and control decision variables. For this subset of variables, data-driven surrogate models have been used to learn follower responses to different leader decisions. In this work, we propose to use ReLU neural networks for the control level. Consequently, the bilevel problem is collapsed into a single-level MILP that is still able to account for the control level's objective. This single-level MILP reformulation is compared with the monolithic approach and benchmarked against embedding a nonlinear expression of the neural networks into the optimisation. Moreover, a neural network is used to predict control level feasibility. The case studies involve batch reactor and sequential batch process scheduling problems. The proposed methodology finds optimal solutions while largely outperforming both approaches in terms of computational time. Additionally, due to well-developed MILP solvers, adding ReLU neural networks in a MILP form marginally impacts the computational time. The solution's error due to prediction accuracy is correlated with the neural network training error. Overall, we expose how - by using an existing big-M reformulation and being careful about integrating machine learning and optimisation pipelines - we can more efficiently solve the bilevel scheduling-control problem with high accuracy.