Hybrid Deep Learning Model Improves PM2.5 Forecasts in Anhui Province

Researchers have developed a hybrid deep learning model that predicts PM2.5 concentrations at monitoring stations in Anhui Province, China, with markedly higher accuracy than conventional approaches. The model, called SFLA-VMD-CBiLSTM-AM, combines a nature-inspired optimization algorithm, a signal decomposition technique, a bidirectional recurrent network, and an attention mechanism to handle the notoriously erratic behavior of fine particulate pollution data.

PM2.5 time series are characterized by strong non-stationarity, multi-scale fluctuations, and complex temporal dependencies. According to the study’s authors, traditional deep learning models often average feature weights unevenly across these patterns and respond insufficiently during critical pollution episodes, which limits their usefulness for long-term forecasting. The new framework addresses this by using the shuffled frog-leaping algorithm to optimize variational mode decomposition, which breaks the raw signal into components the network can learn more effectively, before feeding them into a convolutional bidirectional LSTM enhanced with attention.

The team adopted a single-step prediction strategy and split the dataset chronologically into a top 70% training set, a middle 15% validation set, and a bottom 15% test set.

The reported results show the hybrid model performing well at all three stations examined. Across the comparison, the average RMSE decreased by 2.724 relative to benchmark models, and the average coefficient of determination (R2) reached 0.981, which the authors describe as significantly better than the comparison models. The variational mode decomposition module improved the correlation between model output and observations as well as the matching degree of standard deviation, while the attention mechanism provided a stable performance boost at sites with significant pollution fluctuations.

One station proved especially challenging. Huangshan, with its complex terrain, sparse data, and low background PM2.5 concentrations, was difficult to predict, yet the hybrid model still achieved an R2 of 0.959 there, which the authors present as evidence of the model’s robustness under unfavorable conditions. The findings, published in Air Quality, Atmosphere & Health, suggest that combining optimized decomposition with attention-enhanced recurrent networks may offer a practical route to more reliable fine particulate matter forecasting in regions with varied pollution dynamics.

Subject of Research: Machine learning prediction of PM2.5 air pollution concentrations in Anhui Province, China

Article Title: Research on predicting PM2.5 concentration in Anhui Province using a VMD-CBiLSTM model based on the shuffled frog-leaping algorithm and attention

Article References: Dong, J., Su, T., Cui, X., Wang, L., Liu, D., Li, X., Kong, D., Liu, S., & Chen, J. (2026). Research on predicting PM2.5 concentration in Anhui Province using a VMD-CBiLSTM model based on the shuffled frog-leaping algorithm and attention. Air Quality, Atmosphere & Health, 19(11), Article 250. https://doi.org/10.1007/s11869-026-02141-y

Image Credits: AI Generated

DOI: 10.1007/s11869-026-02141-y

Keywords: PM2.5, air quality forecasting, deep learning, variational mode decomposition, attention mechanism, LSTM, Anhui Province, shuffled frog-leaping algorithm, time series prediction, China, Research, predicting