BALANCE-ST: distilling deep reinforcement learning policies into PLC logic for multi-VFD load balancing

Industrial multi-VFD load balancing on programmable logic controllers (PLCs) requires control policies that are accurate, transparent, and fast enough for deterministic scan cycles. Deep reinforcement learning can learn effective balancing strategies, but neural policies are difficult to inspect and deploy as IEC 61131-3 Structured Text (ST). We propose BALANCE-ST, a simulation-to-PLC pipeline that trains a deep Q -network (DQN) teacher, distils it into a depth-2 decision tree selected by held-out validation parsimony, renders the tree as PLC-oriented ST, and wraps the resulting policy with a deterministic guard enforcing command-bound, ramp-rate, dwell-time, and overload checks. In an in-distribution post-training evaluation across eight demand families and ten seeds, BALANCE-ST achieves mean load-imbalance coefficient of variation 0.0697 versus 0.0696 for the ten-teacher DQN aggregate, while reducing imbalance by \(34.4\%\) , \(52.6\%\) , and \(15.0\%\) relative to equal-share, threshold-staging, and PI baselines, respectively. On the interquartile mean, the reduction versus threshold staging is \(43.2\%\) . A guard ablation and hard-stress evaluation show that the runtime guard reduces realised violations at the cost of higher imbalance under actuator dropout. Compared with beam-search model predictive control, BALANCE-ST yields higher imbalance but much lower online decision cost and a compact policy with four leaves and three effective actions. These results support BALANCE-ST as a practical trade-off between control performance, inspectability, and deployment-oriented simplicity for PLC-based multi-VFD control.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-15
DOI
https://doi.org/10.1038/s41598-026-80105-z
Primary Topic
Industrial Automation and Control Systems
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

BALANCE-ST: distilling deep reinforcement learning policies into PLC logic for multi-VFD load balancing

Lei Li, Guopeng Zhang
Scientific Reports
Industrial Automation and Control Systems
article

BALANCE-ST: distilling deep reinforcement learning policies into PLC logic for multi-VFD load balancing

Lei Li, Guopeng Zhang
article en

Abstract

Industrial multi-VFD load balancing on programmable logic controllers (PLCs) requires control policies that are accurate, transparent, and fast enough for deterministic scan cycles. Deep reinforcement learning can learn effective balancing strategies, but neural policies are difficult to inspect and deploy as IEC 61131-3 Structured Text (ST). We propose BALANCE-ST, a simulation-to-PLC pipeline that trains a deep Q -network (DQN) teacher, distils it into a depth-2 decision tree selected by held-out validation parsimony, renders the tree as PLC-oriented ST, and wraps the resulting policy with a deterministic guard enforcing command-bound, ramp-rate, dwell-time, and overload checks. In an in-distribution post-training evaluation across eight demand families and ten seeds, BALANCE-ST achieves mean load-imbalance coefficient of variation 0.0697 versus 0.0696 for the ten-teacher DQN aggregate, while reducing imbalance by \(34.4\%\) , \(52.6\%\) , and \(15.0\%\) relative to equal-share, threshold-staging, and PI baselines, respectively. On the interquartile mean, the reduction versus threshold staging is \(43.2\%\) . A guard ablation and hard-stress evaluation show that the runtime guard reduces realised violations at the cost of higher imbalance under actuator dropout. Compared with beam-search model predictive control, BALANCE-ST yields higher imbalance but much lower online decision cost and a compact policy with four leaves and three effective actions. These results support BALANCE-ST as a practical trade-off between control performance, inspectability, and deployment-oriented simplicity for PLC-based multi-VFD control.

Scientific Reports
Henan Polytechnic University (CN), Henan College of Industry and Information Technology (CN)
Industrial Automation and Control Systems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon