Assessment of daytime ozone using a baseline-deviation multivariate linear regression framework: a long-term analysis...

Ground-level ozone (O₃) shows strong diurnal and seasonal variability, but interpretable station-scale assessments are often constrained by the limited availability of measurements beyond routine monitoring. This study developed multivariate linear regression (MLR) models to assess daytime O₃ at the Zhongli Air Quality Monitoring Station, a suburban site in northern Taiwan. Using a 10-year day-hour baseline to define deviation variables, the models estimated hourly daytime O₃ (07:00-18:00) by modeling deviations from the baseline and reconstructing absolute O₃ as the sum of the baseline and its estimated deviation. A predictor set was selected using correlation filtering, variance inflation factor checks, and stepwise selection guided by the Bayesian information criterion (BIC). An overlapping training window across month boundaries was applied while retaining atypical periods and events, allowing the models to capture a broad range of 10-year variability in O₃ and its predictors. The models achieved an adjusted R² of 0.765 for the 2014-2023 training period. External testing using an independent external test period from 2024 to 2025 yielded an adjusted R² of 0.674, with a root mean square error (RMSE) of 7.00 ppb and a mean absolute error (MAE) of 5.39 ppb. This framework provides a practical and interpretable station-scale approach for assessing daytime O₃ variability and complements more complex machine learning models.