Enhancing reinforcement learning controllers with GAN-generated data and transfer learning
This study addresses the challenge of data scarcity in training reinforcement learning (RL) controllers for power system economic dispatch problems (EDP) by integrating Generative Adversarial Network (GAN)-generated synthetic data and transfer learning (TL). Traditional data collection for power sys...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Taylor & Francis Group
2025-12-01
|
| Series: | SICE Journal of Control, Measurement, and System Integration |
| Subjects: | |
| Online Access: | http://dx.doi.org/10.1080/18824889.2025.2527471 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1839635026040324096 |
|---|---|
| author | Chang Xu Naoki Hayashi Masahiro Inuiguchi |
| author_facet | Chang Xu Naoki Hayashi Masahiro Inuiguchi |
| author_sort | Chang Xu |
| collection | DOAJ |
| description | This study addresses the challenge of data scarcity in training reinforcement learning (RL) controllers for power system economic dispatch problems (EDP) by integrating Generative Adversarial Network (GAN)-generated synthetic data and transfer learning (TL). Traditional data collection for power systems may face limitations like privacy concerns hindering the performance of deep neural network-based controllers. To overcome this, a GAN-based framework is proposed to generate synthetic load demand data, preserving characteristics of real datasets. A TL technique is then employed to fine-tune a Twin Delayed Deep Deterministic Policy Gradient (TD3) agent, pretrained in a synthetic environment, into a target environment with real-world data. Experiments evaluate three GAN-generated datasets, including scenarios with mode collapse, and compare results against regression-based data generation methods. Key findings demonstrate that even low-quality synthetic data, when combined with TL, significantly enhances RL performance. For instance, a mode-collapsed GAN model reduced test operation cost by 54.7% and power unbalance by 89.9% compared to a baseline TD3 agent. This work highlights the potential of synthetic data augmentation and TL in data-scarce power system applications, offering a viable pathway to improve controller performance without additional real-world data collection. |
| format | Article |
| id | doaj-art-0108b0bdebf9473c8e61ce6ac38ae06c |
| institution | Matheson Library |
| issn | 1884-9970 |
| language | English |
| publishDate | 2025-12-01 |
| publisher | Taylor & Francis Group |
| record_format | Article |
| series | SICE Journal of Control, Measurement, and System Integration |
| spelling | doaj-art-0108b0bdebf9473c8e61ce6ac38ae06c2025-07-09T13:40:42ZengTaylor & Francis GroupSICE Journal of Control, Measurement, and System Integration1884-99702025-12-0118110.1080/18824889.2025.25274712527471Enhancing reinforcement learning controllers with GAN-generated data and transfer learningChang Xu0Naoki Hayashi1Masahiro Inuiguchi2Universiti MalayaThe University of OsakaThe University of OsakaThis study addresses the challenge of data scarcity in training reinforcement learning (RL) controllers for power system economic dispatch problems (EDP) by integrating Generative Adversarial Network (GAN)-generated synthetic data and transfer learning (TL). Traditional data collection for power systems may face limitations like privacy concerns hindering the performance of deep neural network-based controllers. To overcome this, a GAN-based framework is proposed to generate synthetic load demand data, preserving characteristics of real datasets. A TL technique is then employed to fine-tune a Twin Delayed Deep Deterministic Policy Gradient (TD3) agent, pretrained in a synthetic environment, into a target environment with real-world data. Experiments evaluate three GAN-generated datasets, including scenarios with mode collapse, and compare results against regression-based data generation methods. Key findings demonstrate that even low-quality synthetic data, when combined with TL, significantly enhances RL performance. For instance, a mode-collapsed GAN model reduced test operation cost by 54.7% and power unbalance by 89.9% compared to a baseline TD3 agent. This work highlights the potential of synthetic data augmentation and TL in data-scarce power system applications, offering a viable pathway to improve controller performance without additional real-world data collection.http://dx.doi.org/10.1080/18824889.2025.2527471economic dispatchgantransfer learningreinforcement learningtwin delayed ddpg |
| spellingShingle | Chang Xu Naoki Hayashi Masahiro Inuiguchi Enhancing reinforcement learning controllers with GAN-generated data and transfer learning SICE Journal of Control, Measurement, and System Integration economic dispatch gan transfer learning reinforcement learning twin delayed ddpg |
| title | Enhancing reinforcement learning controllers with GAN-generated data and transfer learning |
| title_full | Enhancing reinforcement learning controllers with GAN-generated data and transfer learning |
| title_fullStr | Enhancing reinforcement learning controllers with GAN-generated data and transfer learning |
| title_full_unstemmed | Enhancing reinforcement learning controllers with GAN-generated data and transfer learning |
| title_short | Enhancing reinforcement learning controllers with GAN-generated data and transfer learning |
| title_sort | enhancing reinforcement learning controllers with gan generated data and transfer learning |
| topic | economic dispatch gan transfer learning reinforcement learning twin delayed ddpg |
| url | http://dx.doi.org/10.1080/18824889.2025.2527471 |
| work_keys_str_mv | AT changxu enhancingreinforcementlearningcontrollerswithgangenerateddataandtransferlearning AT naokihayashi enhancingreinforcementlearningcontrollerswithgangenerateddataandtransferlearning AT masahiroinuiguchi enhancingreinforcementlearningcontrollerswithgangenerateddataandtransferlearning |