Reinforcement Learning for Hanger Rod Systems

  • Wojtala, Maciej (University of Warsaw;IDEAS Research Institute)
  • Tabin, Jakub (IPPT PAN (Polish Academy of Sciences))
  • Prącik, Michał (Cracow University of Technology)
  • Łacny, Łukasz (Cracow University of Technology)

Please login to view abstract download link

Suspended steam boilers in combined heat and power (CHP) plants are supported by networks of vertical hanger rods, whose tension distribution is critical for structural safety and long-term reliability [1]. After maintenance or retrofit operations, re-adjustment of rod forces is often required [2]. In practice, this process is complicated by elastic interaction between rods (bolt crosstalk), whereby adjusting one rod affects the force state of neighbouring rods. Despite this strongly coupled behaviour, regulation procedures in CHP plants remain largely empirical, unlike optimized strategies developed for multi-bolt assemblies [3]. In this study, the tension regulation problem is formulated as a sequential decision-making task and addressed using reinforcement learning (RL) [4]. A physically motivated analytical model of a multi-rod hanger system, representative of real CHP plant configurations, is developed and validated against finite element (FEM) simulations. The analytical model serves as a deterministic environment for training an RL agent. The objective is to minimize the force spread between the most and least loaded rods while achieving the target distribution in a limited number of regulation steps. The RL policy operates in a mixed-action space, selecting both the rod to adjust and the corresponding length increment, and is trained to optimize the final force distribution over a fixed regulation horizon. The approach is compared to empirical regulation strategies, a greedy control scheme, and random sampling baselines. The results show that the RL-based strategy achieves the highest regulation efficiency, reaching near-optimal force uniformity within the first two adjustment steps. Moreover, unlike the second-best method (the sampling scheme), the RL framework shows strong potential for generalization to large multi-rod hanger networks. The work has been supported by the National Centre for Research and Development through Grant No 0317/L-14/2023, “MASy 2.0: system for Monitoring and Adjusting the tenSion force in hanger rods of power boilers”.