• • At workshop sites with stable conditions, the LSTM model achieved high goodness-of-fit (R²) and low errors for VOCs, NH3, H2S, and CH3SH, with 5-second resolution data enabling minute-level predictions.
• • At boundary sites, peak episodes caused significantly amplified errors for OU and VOCs due to plume arrival delays and non-stationary diffusion-dilution, leading to unstable skill scores relative to persistence baseline.
• • Stratified error analysis consistently showed non-peak periods outperformed peak periods, identifying event-driven fluctuations as the primary error source, with implications for early-warning system design.
• • The study recommends incorporating exogenous variables (wind speed/direction, ventilation, gate access, operational rhythms) and peak-sensitive loss functions to improve pulse-process characterization and prediction accuracy.