GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-05
A simulation that runs without interruption is not necessarily using an uninterrupted record of measured weather. Missing observations may have been replaced before the file reached the design team. If those replacements are invisible, the resulting chart can appear more certain than the evidence supporting it.
Preparing Weather Data for Real-Time Building Energy Simulation, first submitted on 19 November 2020, examines anomaly detection and missing-value imputation. Its neural-network work uses nearby weather stations to estimate temperature and relative humidity at an observation site in Milan. This is historical research about input preparation, not a general validation of every weather variable, climate or project simulation. Research record
Keep a second record beside the values
GAD recommends preserving the original observations and producing a separate quality record for any completed dataset. Mark whether each interval was measured, rejected or estimated. Record the source stations, time alignment and method revision where these contribute to the replacement.
A small gap between ordinary observations and a long outage during unusual conditions should not disappear under the same generic note. Summarise gap lengths and their timing. The design team can then see whether a critical period depends heavily on inferred inputs.
Agree the time convention and units before interpreting anomalies. Apparent inconsistencies may originate in timestamps or conversions. A learned model should not become an elaborate way of hiding an unresolved data-format problem.
Test the outage you actually face
To evaluate an imputation method, withhold known observations in patterns resembling the real gaps and compare the estimates with the hidden values. Keep evaluation periods separate from model fitting. Include longer missing runs if those are part of the intended use; isolated random omissions alone would make a weak test of that condition.
Compare against a simple, clearly specified baseline. Additional modeling complexity is useful only when it improves the relevant task enough to justify its preparation and review effort. Report errors during important periods separately from a whole-record average.
For a hypothetical facade comparison, carry several defensible input scenarios through the same simulation setup. If the preferred option changes, further input validation may be more valuable than refining the facade model. This sensitivity exercise is an editorial recommendation, not a result reported by the paper.
Match the evidence to the claim
A method demonstrated on temperature and humidity should not automatically be described as validated for solar radiation, wind or other inputs. Similarly, completion of a measured series does not turn it into a future-climate scenario or a standard design-year file.
The final study should identify which decisions rely on reconstructed intervals and which conclusions remain stable despite them. GAD has not reproduced the research experiments. The central lesson is straightforward: an AI-filled gap can make a dataset usable while still remaining a gap in direct observation. A trustworthy simulation preserves that distinction through to the design decision.