Data and code for "Input windows, not feature lists, determine leakage safety in renewable supply forecasting" (NASA POWER daily records, South Tangerang, Indonesia, 2006-2024)

Data, code and result artifacts supporting the manuscript "Input windows, not feature lists, determine leakage safety in renewable supply forecasting". The study is a controlled retrospective benchmark on 6,940 daily records derived from NASA POWER reanalysis output for one grid point in South Tangerang, Banten, Indonesia, covering 1 January 2006 to 31 December 2024. The record exists so that every number printed in the manuscript can be traced to the bytes that produced it. It is not a general-purpose forecasting dataset and it is not an operational forecast benchmark; the scope limits are stated at the end of this description. What is in the record The record contains nine files. Four are compressed archives, one is the uncompressed source table, two are documentation readable without downloading anything, and two are checksum manifests. Extracting the four archives and placing the source table at 00_source/S_2006_2024.csv reconstructs a tree of 58 files. zenodo_A_S_2006_2024.csv - group A, the hourly NASA POWER export. 166,560 hourly records. Uploaded uncompressed so that it can be inspected without a download-and-extract step. 20,665,207 bytes. zenodo_B_dataset_turunan.tar.gz - group B, the derived daily dataset. 01_canonical/, 3 files: the canonical table in comma-separated and Parquet form, 6,940 daily records and 41 columns, plus a generation manifest. 2,716,727 bytes. zenodo_C_kode.tar.gz - group C, the code. 02_code/, 35 files: a 19-file supply_forecasting module, 12 architecture and ablation scripts under experiments/positive_control/, the three-arm ablation runner, the linear-probe runner, one test file carrying 13 automated tests, and LICENSE-CODE. The archive also carries README.md and NOTICE, byte-identical to the standalone copies below. 87,543 bytes. zenodo_D_artefak_hasil.tar.gz - group D, the result artifacts. 03_results/, 16 files. Every number reported in the manuscript comes from these. 14,017 bytes. zenodo_E_praregistrasi.tar.gz - group E, the pre-registration note. 04_preregistration/, 1 file recording the seed rule, its scope and a stopping contingency, written before the eight-seed runs were executed. 922 bytes. zenodo_README.md - the full record documentation, including provenance, regeneration instructions, measured reproducibility, exclusions and the licence rationale. 11,759 bytes. zenodo_NOTICE - third-party data attribution for NASA POWER. 823 bytes. zenodo_CHECKSUMS.txt - SHA-256 of all 58 files in the reconstructed tree. 6,338 bytes. zenodo_CHECKSUMS_ARSIP.txt - SHA-256 of the seven uploaded files other than these two manifests. 631 bytes. How to verify a download Verification is two-level, so that a corrupted download is distinguished from a corrupted extraction. Level one, the uploaded files. With all nine files in one directory, run shasum -a 256 -c zenodo_CHECKSUMS_ARSIP.txt. Seven entries should report OK. Level two, the extracted tree. Extract the four archives into one directory, then place the uncompressed source table at 00_source/S_2006_2024.csv, so that the five group directories sit side by side. From the root of that tree run shasum -a 256 -c zenodo_CHECKSUMS.txt. All fifty-eight entries should report OK. To regenerate the derived daily dataset from the source table, invoke the preprocessing entry point as a module from the repository root: python -m supply_forecasting.run_preprocessing. It uses absolute package imports and has no path bootstrap, so a file-path invocation fails. Source data provenance The hourly file was obtained from the NASA POWER archive on 26 December 2025, through service version 2.6.145. POWER versions its data store separately from its services; the data store release in force on the date of access was Data Version 10.2.2, issued 16 December 2025. The next release, 10.2.3, followed on 29 December 2025, after the download. Both identifiers are recorded because the service version alone does not determine the data version a file was generated from. Grid point: latitude 6.358 S, longitude 106.6672 E. The elevation POWER reports for this location, 371.57 m, is documented by the archive as an average over the 0.5 by 0.625 degree grid cell rather than the elevation of the point. Neither the hourly file nor the derived daily file carries embedded location metadata; both begin at the column header row. The coordinate attribution therefore rests on the authors' statement, not on the files. Reproducibility, measured rather than assumed In the authoring environment, regeneration is bit-identical: two consecutive executions reproduce both the comma-separated and the Parquet artifact byte for byte, and an exact-equality comparison across all 41 columns returns no difference. The only field that does not reproduce is the generation timestamp in the manifest, which is written after the data files and cannot enter them. Across environments, bit-exact reproduction is not guaranteed. In a second environment differing in interpreter, numerical library and operating system, the regenerated data agree with the deposited artifact to within four units in the last place, a worst-case relative difference of 9.2e-17. The divergence is confined to columns computed through trigonometric or fractional-power operations, namely the calendar encodings and the hub-height wind chain. The photovoltaic chain and every aggregate statistic in the manifest are bit-identical across both environments. Verify a regenerated copy by SHA-256 where bit equality is expected, and with a tolerance no tighter than 4 ULP otherwise. Note that pandas.testing.assert_frame_equal defaults to an approximate comparison; bit equality requires check_exact=True. What is deliberately not deposited Several scripts present in the project's version history are excluded, and the exclusions are disclosed rather than silent. Four scripts carry a GROUND_TRUTH block of reference values whose originating artifacts are not in the repository; depositing untraceable reference numbers would move a traceability problem into a public record. Five older metrics files are excluded because they are not the artifacts behind any number in the manuscript and two of them disagree with each other by 0.018 on the same quantity. Three further scripts named run_leakage_* are excluded on the same grounds: two direct the reader to compare their output against a value that appears nowhere in the manuscript, all three read from paths outside this record and exit before doing any work, and one evaluates a stub model with its other architectures commented out. None of their outputs appears in 03_results/, and the manuscript names none of the three. All of them remain in the project's version history and can be supplied on request. README.md states each exclusion and its reason in full. One superseded reference is disclosed rather than rewritten. Three fields in 03_results/positive_arm_5seed_results.json, namely day_ahead_arm_mean_r2, day_ahead_arm_mean_rmse and rmse_ratio_day_ahead_over_positive, were carried into that run from constants hardcoded in 02_code/experiments/positive_control/pc_gru_pv_5seed_positive.py, whose originating artifact is not part of this record. They read 0.331024, 26.96952 and 34.648405. The authoritative day-ahead values are the ones in 03_results/day_ahead_arm_5seed_results.json, computed from its own five per-seed runs: 0.318309 and 27.224093 kWh per day. The manuscript uses that pair and reports the ratio as 34.9755. Both files are deposited unaltered, so that what each run actually produced stays visible; this note records which of the two is authoritative. No number printed in the manuscript is affected. Licence This record is released under the Creative Commons CC0 1.0 Universal Public Domain Dedication, with one exception: the contents of 02_code/ are governed by 02_code/LICENSE-CODE, the Apache License, Version 2.0, and are not covered by the CC0 dedication. The reason for the split is stated in README.md and summarised here. The hourly data in 00_source/ are a redistribution of NASA POWER output, for which the POWER project publishes no licence and no formal terms of use; its stated requirements are citation of the service name, version number and date accessed, and notification when the data are passed to other researchers. The authors hold no rights in that material and can grant none, so a licence carrying an attribution condition would purport to impose on downstream users a legal term the upstream source does not impose. CC0 waives only the authors' own rights and therefore cannot over-claim. Code is excluded from the dedication because CC0 has not been approved by the Open Source Initiative and, by its own terms, does not license or otherwise affect any patent rights, whereas Apache-2.0 grants an express patent licence bounded to the contributed work. Citing this record, including the code CC0 removes attribution as a legal condition; it does not displace the scholarly norm. Users are asked to cite the manuscript this record supports and, separately, to acknowledge the NASA POWER project as set out in NOTICE. The code is part of this record and is cited through this record's DOI. 02_code/ is not deposited as a separate software record, so work that reuses the supply_forecasting module, the physics proxy implementations, the three-arm ablation runner or the linear probe should cite this DOI and name the component used, for example "the three-arm ablation runner in 02_code/experiments/ablation_c10/ of [this record]". The code carries its own licence file and its own licence terms; citing the record does not merge the two licences. Where a specific version matters, cite the version DOI rather than the concept DOI. The claim this record supports is the reproducibility of particular printed numbers, and the concept DOI resolves to whatever version is latest. Scope and limits The photovoltaic and wind targets are deterministic physics-derived proxies, not measurements and not independently calibrated estimates of plant output. The photovoltaic target is an exact scalar multiple of daily global horizontal irradiation; the wind target evaluates a normalised cubic power curve at a daily-mean hub-height wind speed, which understates the hourly-summed energy by 56.4 per cent over this record. The meteorological source is a reanalysis product, not an archive of operational forecasts issued at historical prediction times, so the lagged configurations in the study are retrospective information restrictions rather than validated operational day-ahead forecasts. Results cover one tropical location and one chronological test block. The diagnostic workflow proposed in the manuscript has not been validated as a general leakage detector, and no detection or false-positive rate is reported.