A Multi-Pollutant Time-Series Prediction Model Based on Long Short-Term Memory Networks
To meet the minute-level early-warning requirements for odor and multi-pollutant emissions at waste treatment facilities, this study proposed a multivariate short-term time-series prediction framework applicable to multi-tier scenarios covering source and boundary points (i.e., workshops and plant boundaries). Based on continuous online monitoring data with a 5-second resolution, a long short-term memory (LSTM) model using a sliding-window and recursive multi-step prediction strategy was constructed to jointly model odor concentration (OU) and pollutants including VOCs, NH3, H2S, and CH3SH (mg/m³). An evaluation protocol aligned with environmental supervision practice was established, incorporating mean absolute error (MAE), root mean square error (RMSE), goodness-of-fit (R²), skill scores (SS) relative to a persistence baseline, and threshold-based error stratification to characterize uncertainty during peak emission periods. The results showed that at workshop monitoring sites with relatively stable operating conditions, VOCs, NH3, H2S, and CH3SH exhibited a high goodness of fit and low prediction errors. In contrast, at boundary sites affected by plume arrival delays and diffusion-dilution non-stationarity, OU and VOCs displayed significantly amplified errors during peak episodes, and the skill score advantage over the baseline became unstable at certain sites. Stratified analysis consistently revealed that non-peak periods outperformed peak periods, indicating that event-driven fluctuations were the main sources of error. Accordingly, this study suggested incorporating exogenous variables such as wind speed and direction, ventilation and gate access control, and operational rhythms, along with peak-sensitive loss functions, into the model to enhance its capacity to characterize and provide early warnings for transient emission pulses. Overall, this study established a reusable methodological baseline and evaluation paradigm for minute-scale multi-pollutant prediction, providing quantitative support for the operational management and source-to-boundary coordinated control of waste treatment facilities.