A reinforcement learning based Lagrangian relaxation algorithm for multi-energy allocation problem in steel enterprise

Abstract

The integrated iron and steel enterprises are typically characterized by the presence of multiple energy media that are highly coupled, frequent start-stop cycles of energy conversion equipment, and fluctuations in energy supply and demand. In this paper, we address the problem of byproduct gas-steam-electricity scheduling in iron and steel enterprises to achieve optimal energy distribution and conversion and reduce the energy cost. This optimization problem for the multi-period full energy chain is formulated as a mathematical programming model that considers equipment start-stop cycles, with the objective of minimizing energy system operating cost. A Lagrangian relaxation framework is employed to decouple the energy management model into several independent single schedules. To further improve the algorithm performance, a novel reinforcement learning-based Lagrangian relaxation algorithm (RL-LR) is proposed, which can dynamically set step size coefficients during the iteration process. Numerical results are presented demonstrating that the RL-LR algorithm can achieve higher optimization efficiency.