RRecherchily
← Back to projects
ActiveJanuary 2026

Automatic Control of Industrial Processes via Deep Reinforcement Learning with Lyapunov Stability Guarantees

Classical PID controllers struggle to adapt to the nonlinearities and variable disturbances of real industrial processes. This project proposes a control architecture based on deep reinforcement learning (SAC/PPO) with stability constraints integrated via learned Lyapunov functions, applied to industrial furnace temperature control and flow regulation in water distribution networks.

Control systemsReinforcement learningIndustrial controlLyapunov stability

Team

NR
Nadia RAHMANI
Researcher
KT
Karim TOUATI
Researcher

Methodology

  1. Model two simulation environments: industrial resistance furnace (nonlinear thermal dynamics) and water distribution network (EPANET model).
  2. Implement SAC and PPO agents with Lyapunov constraints integrated into the loss function (Lyapunov-constrained policy optimization).
  3. Systematic comparison with auto-tuned PID controllers, linear MPC, and reference fuzzy controllers across 50 disturbance scenarios.
  4. Validation on a physical test bench (pilot furnace at the ENP Oran Fablab) to evaluate sim-to-real transfer.

Evaluation metrics

  • Integral Absolute Error (IAE) — tracking error reduction compared to baseline PID.
  • Safety constraint violation rate (critical temperature overshoot) — target: 0% over 1,000 episodes.
  • Training convergence time and sample efficiency comparison between SAC and PPO.

Budget

$600 for test bench components (sensors, actuators, microcontrollers), $400 in compute credits for RL policy training.

Expected deliverables

  • Gymnasium-compatible simulation environments published as open-source.
  • Documented and reproducible sim-to-real transfer protocol.
  • Paper submitted to ACC 2027 (American Control Conference).
← Back to projects