Ozone forecasts stop learning raw series: two-stage decomposition plus two-scale feature selection cuts Beijing daily ozone MAE to 0.535

Ozone forecasting no longer feeds the raw concentration series straight into the predictor; the series is first split into more regular subsequences. Ozone series are volatile and noisy, and direct prediction carries visible error; earlier work mostly used a single decomposition stage, and the first modal component IMF1 stayed highly complex. On daily ozone data from Beijing and Shanghai, adding ICEEMDAN decomposition lowered MAE, RMSE and SMAPE of the comparison models by 15.5186, 21.1828 and 15.09% on average.

After decomposition, the input is filtered again so that only lagged terms useful for prediction are kept. Decomposition multiplies the number of series, inflating input dimension, which invites overfitting and lengthens training. In the Beijing setting, adding PACF and random forest two-scale selection cut MAE, RMSE and SMAPE by a further 1.5283, 2.0335 and 2.4284%; the Shanghai setting dropped by 2.222, 2.990 and 2.3189%.

The pipeline chains decomposition, reconstruction, selection and prediction, with a temporal convolutional network producing the final output. Single-stage decomposition with one predictor struggles to suppress noise and dimension at the same time. On 2495 daily ozone observations each for Beijing and Shanghai, the combined model reached MAE, RMSE and SMAPE of 0.53509618, 0.733460954 and 0.843300731% in Beijing, beating ELM, LSSVM and LSTM controls across four comparison experiments.

The model targets daily-scale ozone concentration forecasting, validated on daily ozone data from the Tiantan station in Beijing and the Zhangjiang station in Pudong, Shanghai, from January 1, 2015 to November 21, 2021, 2495 observations each, split 80% training and 20% testing. The intended readers are researchers and engineers working on air quality forecasting and time series modeling. The source proposes future work that adds meteorological factors such as temperature and wind speed, and explores applying the model in other fields.

The source states that the study did not consider other meteorological factors such as temperature and wind speed, and that the model performs well in empirical analysis but can still be improved. All metrics come from retrospective experiments in Beijing and Shanghai, with no reported performance for other cities or real-time forecasting settings. The conclusion section cites 0.6038, 0.7992 and 0.70% as the model's overall best metrics; these three values correspond to the MAE, RMSE and SMAPE of the Shanghai ICEEMDAN-VMD-FE-PACF-RF-TCN row, while the Beijing row for the same model gives 0.53509618, 0.733460954 and 0.843300731%, and the conclusion does not label which city each set belongs to.