Residual reinforcement learning cuts modelled stack wear in a fuel-cell mini-excavator: Co-simulation benchmark against tuned energy managers data and code

This record contains the data and code of a simulation study on energy management and storage sizing for a 5-t fuel-cell hybrid mini-excavator. The companion article has been submitted to Automation in Construction.

**Study in brief.** The machine has a proton-exchange-membrane fuel cell, a 9.9 kWh LFP battery and a 20.5 F supercapacitor on a 650 V DC bus; its boom, arm and bucket are driven by electric planetary-roller-screw cylinders. The plant is a Simcenter Amesim 2504 model exported as FMI 2.0 co-simulation units (FMUs). A residual reinforcement-learning energy manager (MMEG-TD3), which corrects six parameters of a charge-sustaining expert controller, is compared with six tuned rule- and optimisation-based managers (including adaptive ECMS and MPC) and three direct deep-RL learners. The comparison uses three training seeds, a protocol fixed before training and a single evaluation on frozen test records. The cost combines equivalent hydrogen with modelled fuel-cell degradation and battery wear. A 3×3 battery–supercapacitor sizing map is included.

**Files**

- `1_code_models_data.zip` (317 MB): analysis code that regenerates every table and figure of the article (`code_v15/`); simulation, training and evaluation code; FMUs of all plants; load records with construction manifests and operating-phase labels; the phase classifier and all learner checkpoints with their selection files; training logs; the protocol and freeze lists (SHA-256); and the scores of every tuning, validation, test and sizing-map run.

- `2_trajectories_test_part1.zip` to `2_trajectories_test_part6.zip` (2.2 GB in total): 10-ms trajectories of the single test evaluation (24 records) for every controller and seed, split into six parts for upload; unpack all parts into the same folder.

- `README.md`: folder layout and step-by-step reproduction.

- `SHA256SUMS.txt`: checksums of every archive and of every file inside them.

- `LICENSE_code_MIT.txt`: licence of the code.

**Reproduction.** Unpack all archives into one folder. The statistics, tables and figures of the article are regenerated from `1_code_models_data.zip` alone in a few minutes (Python 3.8+, NumPy, pandas, SciPy, Matplotlib); the generated tables are identical to those of the article. The test scores can be recomputed from the trajectories in the six `2_trajectories_test_part*.zip` archives, and any trajectory can be re-simulated with the included FMUs and checkpoints (Python, PyTorch, FMPy; Windows 64-bit). Trajectories of tuning, validation and sizing-map runs are not included; their scores are, and they can be regenerated with the same code.

**Load records.** Simulated with the Simcenter Amesim demonstration model Hydraulic_Excavator_LS (a full-size hydraulic excavator) and scaled to the drives of the mini-excavator; they are not measured duty cycles.

**Licences.** Data: CC BY 4.0. Code (.py, .ps1): MIT.

**Funding.** Basic Science Program through the National Research Foundation of Korea (NRF), funded by the Ministry of Science and ICT (NRF-RS-2026-25619102); ANCHOR program through the Ulsan ANCHOR Center, funded by the Ministry of Education (MOE) and the Ulsan Metropolitan City, Republic of Korea (2026-ANCHOR-07-001).

**Contact.** Kyoung Kwan Ahn, University of Ulsan (kkahn@ulsan.ac.kr).