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

This dataset is the manuscript-aligned v1.3.0 analysis archive accompanying the study “Robustness Evaluation of Building Electricity Load Forecasting Models Using Similarity-Controlled Fourier Surrogates.” Core contents Exact 2017 source load series for 60 non-residential buildings, comprising 525,600 hourly observations. 4,200 archived Fourier-surrogate series generated from seven nominal stress settings (S1-S7), ten realizations per setting, and 60 buildings. Complete model-evaluation records comprising 600 unchanged-reference records and 42,000 surrogate-evaluation records (42,600 model-task rows in total) across five forecasting models and two target offsets. 600 building-model-offset robustness summaries and detailed model-selection outputs, including strict winner reversals, practical-margin reversals, equivalence sets, ranking stability, and information-fair comparison sets. Recovered historical artifacts comprising 60 processed training arrays and 600 saved model objects: 360 learned-model fitted objects and 240 seasonal-baseline model-specification/prediction objects. Complete reference-test point predictions, execution logs, metadata, SHA-256 manifests, the recovered historical software-environment lock, and surviving historical source modules. Friedman omnibus tests, Wilcoxon-Holm paired comparisons, rank-biserial effect sizes, observed-similarity AUC integration outputs, site-cluster sensitivity analyses, zero-clipping diagnostics, normalization sensitivities, surrogate-count sensitivity, and calendar-feature-timing sensitivity. Executable analysis software, validation utilities, unit tests, figures, tables, audit outputs, and a manuscript claim-to-file map. Scope and reproducibility note This archive contains the exact historical surrogate arrays and downstream outputs used for the reported stress-test analyses. The archived arrays are the canonical inputs for reproducing the reported model evaluations, statistical analyses, tables, and figures. The historical top-level run_all.py orchestrator was not recovered byte-for-byte. Revision-stage orchestration and sensitivity-analysis code are therefore explicitly labelled as reconstructed where applicable. Recovered historical artifacts and reconstructed revision-stage components are distinguished in the archive documentation and Supplementary Materials. Because the executed Fourier operator applies zero clipping after inverse transformation, reported post-correction results correspond to the joint Fourier-phase-randomization-plus-zero-clipping operator rather than to a pure phase-randomization effect. Canonical stress definition S7 is the weakest nominal perturbation setting and uses a shortest preserved-period threshold of 6 h. S1 is the strongest nominal setting and uses a threshold of 720 h. The canonical analysis order from the unchanged reference to increasing nominal perturbation is: REF → S7 → S6 → S5 → S4 → S3 → S2 → S1 The seven settings are: S7: 6 h, m = 1460S6: 12 h, m = 730S5: 24 h, m = 365S4: 72 h, m = 121S3: 168 h, m = 52S2: 336 h, m = 26S1: 720 h, m = 12 These settings are generator controls rather than equally spaced physical stress doses. The surrogate scenarios are controlled sensitivity tests and should not be interpreted as probabilities or causal simulations of future deployment drift. Forecast-alignment note The two target offsets are 1 h and 24 h from the feature anchor. Because the most recent load feature is y(t−1), these correspond to effective forecast leads of 2 h and 25 h from the latest observed load feature. At the 25 h effective lead, the primary model-selection comparison uses the four models with common information availability; the three learned models are also reported as a sensitivity set. Source data 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. DOI: 10.1038/s41597-020-00712-x.