The recent success of emission reduction policies in China has significantly lowered sulfur dioxide (SO₂) levels. However, accurately forecasting these concentrations remains challenging due to their inherent non-stationary tendency. This study introduces an innovative hybrid deep learning model, RF-VMD-Seq2Seq, combining the Random Forest (RF) algorithm, Variational Mode Decomposition (VMD), and the Sequence-to-Sequence (Seq2Seq) framework to improve SO₂ concentration forecasting in five coastal cities of northern China. Our results show that the predicted SO₂ concentrations closely align with observed values, effectively capturing fluctuations, outliers, and extreme events—such as sharp declines the Novel Coronavirus Pneumonia (COVID-19) pandemic in 2020—along with the upper 5% of SO₂ levels. The model achieved high coefficients of determination (>0.91) and Pearson’s correlation (>0.96), with low prediction errors (RMSE < 1.35 μg/m³, MAE < 0.94 μg/m³, MAPE < 15%). The low-frequency band decomposing from VMD showed a notable long-term decrease in SO₂ concentrations from 2013 to 2020, with a sharp decline since 2018 during heating seasons, probably due to the ‘Coal-to-Natural Gas’ policy in northern China. The input sequence length of seven steps was recommended for the prediction model, based on high-frequency periodicities extracted through VMD, which significantly improved our model performance. This highlights the critical role of weekly-cycle variations in SO₂ levels, driven by anthropogenic activities, in enhancing the accuracy of one-day-ahead SO₂ predictions across northern China’s coastal regions. The results of the RF model further reveal that CO and NO₂, sharing common anthropogenic sources with SO₂, contribute over 50% to predicting SO₂ concentrations, while meteorological factors—relative humidity (RH) and air temperature—contribute less than 20%. Additionally, the integration of VMD outperformed both the standard Seq2Seq and Ensemble Empirical Mode Decomposition (EEMD)-enhanced Seq2Seq models, showcasing the advantages of VMD in predicting SO₂ decline. This research highlights the potential of the RF-VMD-Seq2Seq model for non-stationary SO₂ prediction and its relevance for environmental protection and public health management.