Forecasting transboundary PM₁₀ pollution in northern Taiwan using upwind observations and transport timing

Forecasting transboundary particulate matter with an aerodynamic diameter ≤ 10 μm (PM₁₀) requires distinguishing the information carried by upwind observations from the timing information supplied by air-mass trajectories. This study developed an upwind-transition-receptor framework with temporally separated model development and evaluation for 1-4-day PM₁₀ forecasts at three northern Taiwan receptors. Nested models successively introduced local conditions, regional-background measurements, nine continental particulate records, and a transport-timing composite derived from historical trajectories. Nested comparisons attributed most of the added predictive skill to distributed upwind observations, while trajectory timing provided a complementary, receptor-dependent refinement. Across all candidates, the upwind-inclusive CatBoost model achieved the lowest aggregate root mean square error (RMSE; 11.04 μg m^-3), 13.0% below persistence. Multi-height trajectory analysis showed increasing pathway divergence with lookback, identifying lower-tropospheric trajectories as timing evidence whose vertical representativeness decreases over longer lookbacks. Upper-tail discrimination remained informative, with precision reflecting the scarcity of high-PM₁₀ events. The framework clarifies how upwind observations, transport information, and receptor response contribute to forecast performance.