Meteorological associations of hourly PM₂.₅ in Lhasa, Tibetan Plateau: antecedent weather, trend specification and out-of-sample evaluation

Statistical models that relate PM₂.₅ to meteorology are widely used to separate weather from emission effects, but they are rarely evaluated out of sample in clean, high-altitude cities, and the role of their trend term is seldom examined. This study fitted penalized additive spline models to hourly city-average PM₂.₅ in Lhasa, Tibetan Plateau (about 3,650 m a.s.l.), for 2019–2024 (51,955 valid hours), using ERA5 temperature, relative humidity, wind speed and precipitation as current-hour and 24-hour antecedent predictors with seasonal, diurnal and long-term controls, and benchmarked them against calendar, linear and random forest models. Three results emerge. First, the point-estimated wind and humidity responses are carried mainly by antecedent conditions: in perturbation decompositions, the 24-hour terms carry most of the fitted wind and humidity sensitivity. Second, the trend specification governs both the weather-normalized series and out-of-sample bias: a linear trend forces a near-monotonic normalized decline (10.98 to 9.66 µg/m³), although observed annual means fell from 12.16 to 8.55 µg/m³ by 2022 and rebounded to 10.36 µg/m³ in 2024. Third, a model frozen after 2019–2024 explained 42% of hourly and 60% of daily variance in 2025 (RMSE 5.35 and 2.80 µg/m³; calendar-only benchmark: 32% and 41%) and modestly outperformed the linear benchmark, but errors were concentrated in winter and in a few high-concentration hours; Gutor-evening residuals exceeded the same-year, same-hour winter background in five of the six years of the 2019–2024 record. Weather-normalized trends for Plateau cities should therefore be reported with alternative trend specifications and event-level diagnostics.