PM2.5 Concentration Prediction Using Signal Decomposition and Long-short Term Memory Network

The prediction of PM2.5 concentration has garnered notable interest in recent times due to its crucial role in preemptive health safeguarding. In our study, we introduce an innovative method that synergizes ensemble empirical mode decomposition (EEMD) with Long Short-Term Memory (LSTM) networks for predicting PM2.5 levels. This technique involves two primary phases: firstly, the time-series data undergo decomposition into intrinsic features via EEMD; subsequently, these features are analyzed by an LSTM model to uncover patterns within the training data. The experimental results from our study reveal that this integrated method outperforms traditional baseline models, delivering substantial improvements in predictive performance. Specifically, the proposed method achieves improvements with the smallest discrepancies being 8.5% for the Index of Agreement (IOA), 8.3% for Root Mean Square Error (RMSE), 9.8% for Mean Absolute Error (MAE), and 3.3% for Mean Absolute Percentage Error (MAPE), compared to the baseline models. These findings underscore the effectiveness of combining EEMD and LSTM for PM2.5 prediction tasks.