Code and source data for "UC-PPO: Universal Cell-wise Proximal Policy Optimization for Passive Cell Balancing in Battery Packs"
Description
This dataset contains Python code, trained reinforcement-learning models, evaluation outputs, and source data supporting the manuscript "UC-PPO for Passive Cell Balancing in Battery Packs with Variable Series Configurations." The archive covers comparisons with no-balancing, rule-based BMS, and pack-specific deep-reinforcement-learning baselines; direct deployment of the same 8S-trained UC-PPO policy to different series configurations; computational-cost evaluation; action-related reward ablation; input-feature masking and mutual-information analysis; integrated-gradients and permutation-sensitivity analysis; voltage-rank retraining ablation; and standardized large-pack baseline comparisons. A figure-to-file map, checksums, environment information, and reproduction notes are included.
Files
Steps to reproduce
1. Download README.md and UC-PPO_Mendeley_Upload.zip, and extract the archive with a UTF-8-compatible tool. Keep the original directory structure and Chinese file names unchanged. 2. Use 64-bit Python 3.10. From the extracted package root, install the required dependencies using: pip install -r environment/requirements.txt 3. Change the working directory to: project/universal_model 4. The retained UC-PPO model is PPO_Universal_no_gate_8S_80Ah_4M.zip. It can be loaded directly with Stable-Baselines3 2.7.1. The corresponding environment and cell-wise vector wrapper are implemented in fast_rc_env.py and vec_env_wrapper.py. 5. The correspondence between manuscript figures/tables, numerical source data and plotting scripts is documented in FIGURE_SOURCE_DATA_MAP.md. Replotting the included CSV files is recommended before rerunning the computationally expensive simulations. 6. Cross-series evaluation and plotting are implemented in plot-universal-model/test-different-series-battery.py. Reward and input-feature ablation scripts are provided in the 消融实验 directory. Integrated-gradient and permutation-sensitivity analyses are provided in 特征贡献分析/梯度归因分析. Computational-cost evaluation is provided in 计算资源分析. 7. Pack-specific DRL models, reward-ablation models, the voltage-rank-ablation model and the corresponding evaluation data are included. Full retraining is stochastic and may not produce bitwise-identical model weights. 8. File integrity can be verified using the SHA-256 checksums in MANIFEST_SHA256.csv.
Institutions
- Lanzhou University of TechnologyGansu, Lanzhou