Derived data and code for fixed-hourly cross-condition PEMFC voltage forecasting under dynamic ripple-current ageing
Description
This dataset contains the derived data, code, model weights, protocol records, figure and table data, and audit files supporting the manuscript “Unsupervised domain-adaptive Transformer for remaining useful life prediction of proton exchange membrane fuel cells under dynamic ripple current ageing,” prepared for Measurement Science and Technology. The released protocol evaluates fixed-hourly PEMFC voltage forecasting under dynamic ripple-current ageing, with H24 as the primary forecast horizon and H6/H48 as auxiliary horizons. The package contains 546 protocol-condition/model score jobs and 99,660 scoreable point records. These repeated method, seed, and information-budget records are not independent physical experiments. FC1 and FC2 are the source and retrospective target records, while D06 is an external target record. The raw FC1/FC2 and D06 repositories are cited but are not redistributed in this package: https://doi.org/10.17028/rd.lboro.3518141.v1 https://doi.org/10.5281/zenodo.20715542 The release includes CSV and JSON result files, figure and table source data, frozen protocol and execution records, source-model weights, preprocessing, forecasting and scoring code, and reproducibility audits. Missing or ineligible hours are retained in the coverage fields rather than silently imputed. A full clean rerun requires lawful access to the cited source repositories and local path configuration.
Files
Steps to reproduce
1. Obtain the FC1/FC2 and D06 source records from the cited repositories under their applicable terms. The raw source files are not included in this dataset. 2. Download and unpack this dataset. Read 03_PROTOCOL/protocol_v2.json, 03_PROTOCOL/experiment_records/dependencies.json, 09_RELEASE_METADATA/DATA_DICTIONARY.md, and 09_RELEASE_METADATA/MANUSCRIPT_DATA_MAP.csv before running the workflow. 3. Use the recorded scientific environment: Python 3.13.12, NumPy 2.4.6, pandas 3.0.3, PyTorch 2.12.0, scikit-learn 1.8.0, and SciPy 1.17.1. Use CPU execution and the deterministic settings specified in the package records. 4. Place the legally obtained source files in local paths and update only the local path settings. Keep the frozen protocol, source-target information boundaries, and execution lock unchanged. 5. Run the included pipeline qualification and causal preprocessing/eligibility modules identified by extv2_pipeline_*. Then run the source-fitting, target-prediction, and scoring modules identified by extv2_experiment_* in the order specified by the protocol and execution records. Future target responses must not be used for source fitting or target issuance. 6. For direct verification, inspect 05_DERIVED_DATA/v2_score_metrics.csv, v2_score_points.csv, and the corresponding v2_score_json files. Compare the reported counts with 546 score jobs and 99,660 scoreable point records. Inspect the figure and table source data in 06_FIGURE_DATA and the supplementary outputs in 07_V1_SUPPLEMENTARY_RESULTS. 7. Verify file integrity using 09_RELEASE_METADATA/SHA256SUMS.txt. Check coverage and eligibility fields before aggregation. The repeated methods, seeds, and information-budget conditions are not independent physical replicates. 8. Recreate the figures and tables using the included scripts and data mappings. Record the software versions, random seeds, input hashes, and output hashes. Exact clean reproduction depends on lawful access to the cited source repositories and the recorded computational environment.