A Hybrid Method Combining Variational Mode Decomposition and Deep Neural Networks for Predicting PM2.5 Concentration in China

Accurately predicting air pollutants, especially PM2.5 (particles with a diameter of 2.5 micrometers or less) concentration, is crucial for pollution control and public health management. A single deep learning architecture often struggles to achieve satisfactory results for PM2.5 prediction due to the difficulty in accurately capturing hidden spatio-temporal patterns within time series. To address this issue, this study introduces a hybrid parallel method (VDPS) that combines variational mode decomposition (VMD) with single deep neural networks for PM2.5 concentration prediction. The VMD algorithm, which can decompose signals adaptively and possesses a certain level of robustness to noise and interference, decomposes complex time series into intrinsic mode functions (IMFs), which are then used as inputs for training a deep neural network prediction model. The cumulative predictions from all deep network models determine the final PM2.5 predictions for VDPS. The deep neural structures used include recurrent neural networks (RNNs), gated recurrent units (GRUs), long short-term memory (LSTM) networks, and convolutional neural networks (CNNs). To demonstrate the effectiveness of VDPS, we conducted comparative evaluations of different models’ performance on many experimental datasets of PM2.5 concentrations in four cities: Beijing, Shanghai, Guangzhou, and Chengdu. The results confirm that the proposed hybrid structure outperforms single deep neural prediction structures in terms of performance metrics. Specifically, the hybrid structure reduces both the root mean square error (RMSE) and the mean absolute error (MAE) by approximately 20-30% when predicting PM2.5 concentrations. And the combined accuracy (CA) index value has decreased by an average of about 10 percentage points. The results suggest that the proposed hybrid method is effective in capturing the intrinsic spatio-temporal characteristics of time series data compared with many single deep neural structures. This can assist governments in making informed decisions and protecting public health.