Similarity-Controlled Fourier Surrogates and Robustness Evaluation Results for Building Electricity Load Forecasting Models

This dataset is the complete analysis archive for the reported stress-test and downstream results accompanying the study “Robustness Evaluation of Building Electricity Load Forecasting Models Using Similarity-Controlled Fourier Surrogates”. Core contents: • 60 non-residential buildings and 525,600 hourly observations from the 2017 test year. • 4,200 exact archived Fourier-surrogate series: seven nominal stress levels (S1–S7), ten realizations per level, and 60 buildings, stored as compressed NPZ arrays. • 42,600 model-evaluation records: 600 unchanged-reference records and 42,000 surrogate-evaluation records covering five forecasting models, two target offsets, seven stress levels, and ten realizations. • 600 building–model–offset robustness summaries. • Model-selection and ranking-stability outputs, including strict winner reversals, 1% and 5% practical-margin reversals, equivalence sets, Top-2 stability, common-information comparisons, and learned-model-only sensitivity analyses. • Complete Friedman omnibus tests and Wilcoxon–Holm paired comparisons with rank-biserial effect sizes. • Exact archived surrogate arrays, corrected metadata, audit files, and SHA-256 manifests for the archived arrays and building-level NPZ files. • Independent zero-clipping sensitivity diagnostics covering amplitude-spectrum changes, autocorrelation changes at lags 1, 24, and 168 h, and daily- and weekly-band spectral-energy changes. • Observed-similarity AUC integration points, coordinate-sorting rules, level-index AUC results, and tie-handling documentation. • Feature-timing, target-alignment, training-matrix completeness, missing-value, finite-row, and common-information-availability audits. • Executable analysis software, validation utilities, unit tests, figures, tables, and data dictionaries. The source building-load data originate from the Building Data Genome Project 2: Miller, C.; Kathirgamanathan, A.; Picchetti, B.; Arjunan, P.; Park, J.Y.; Nagy, Z.; Raftery, P.; Hobson, B.W.; Shi, Z.; Meggers, F. The Building Data Genome Project 2, energy meter data from the ASHRAE Great Energy Predictor III competition. Scientific Data 2020, 7, 368. https://doi.org/10.1038/s41597-020-00712-x. Scope and reproducibility note: This record is a complete analysis archive for the reported stress-test and downstream results. It includes the exact 2017 source series used for testing, all 4,200 surrogate arrays, model-evaluation outputs, statistical analyses, metadata, software, and audit records. It does not include unretained intermediate processed arrays from the 2016 training period. The exact surrogate arrays used in the study are permanently archived and protected by cryptographic hashes. Owing to a legacy inconsistency in the original seed-to-realization mapping, the historical surrogate arrays cannot be regenerated bit-for-bit solely from the original recorded seeds. The archived arrays nevertheless permit complete reproduction of the reported evaluations, statistical analyses, tables, and figures. Canonical stress definition: S7 is the weakest perturbation level and uses a shortest preserved period cutoff of 6 h. S1 is the strongest perturbation level and uses a cutoff of 720 h. The canonical analysis order from the unchanged reference to increasing perturbation strength is: REF → S7 → S6 → S5 → S4 → S3 → S2 → S1. The seven settings and retained-phase counts are: S7: cutoff 6 h, m = 1460S6: cutoff 12 h, m = 730S5: cutoff 24 h, m = 365S4: cutoff 72 h, m = 121S3: cutoff 168 h, m = 52S2: cutoff 336 h, m = 26S1: cutoff 720 h, m = 12 Forecast-alignment note: The target offsets of 1 and 24 h correspond to effective prediction leads of 2 and 25 h from the latest load feature. For the offset-24 common-information comparison, Seasonal Naive-168, Ridge, HistGradientBoosting, and LightGBM share compatible information availability. Seasonal Naive-24 uses an observation one hour later than the learned-model cutoff and is therefore retained only as an unequal-information descriptive result at that offset. License: Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0).