Research on predicting PM 2.5 concentration in Anhui Province using a VMD-CBiLSTM model based on the shuffled frog-leaping algorithm and attention

Environmental Modeling & Assessment · Springer Nature · 2026-10-29 · DOI: 10.1007/s11869-026-02141-y

Аннотация

Due to the significant non-stationarity, multi-scale fluctuations, and complex temporal dependencies of PM 2.5 concentration time series, traditional deep learning models often exhibit issues such as feature weight averaging and insufficient response during critical periods when making long-term PM 2.5 concentration predictions.In order to solve this problem, this paper proposes a hybrid prediction model (SFLA-VMD-CBiLSTM-AM) that integrates the frog jumping algorithm to optimize variational mode decomposition and attention enhanced deep networks, in order to improve the accuracy and stability of PM 2.5 concentration prediction. Adopting a single step prediction strategy, the dataset is divided into the top 70% training set, the middle 15% validation set, and the bottom 15% test set in chronological order to preserve the temporal dependence of the time series and avoid information leakage. The results show that: 1) the hybrid forecasting model performs well in all three stations, in which the average RMSE decreases by 2.724 and the average R 2 reaches 0.981, which is significantly better than the comparison model; 2) VMD module can effectively improve the correlation between model output and observation and the matching degree of standard deviation, while Attention mechanism has a stable performance enhancement effect in sites with significant pollution fluctuations; 3) Although it is difficult to predict the Huangshan station due to complex terrain, lack of data and low concentration background, the new hybrid prediction model can still achieve high accuracy and robustness, with R 2 reaching 0.959.

Читать у издателя →

Похожие статьи