PM2.5 Concentration Prediction Using a CNN-Optimized LSTM-GRU Integrated Model

With the intensification of air pollution issues, accurate prediction of pollutant concentration variations such as PM2.5 has become critical for air quality management. However, single models exhibit limited predictive capabilities when handling complex time-series data. To address this, this paper proposes a CNN-optimized LSTM-GRU integrated model to enhance PM2.5 prediction accuracy. The model synergistically combines the advantages of Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). First, CNN extracts local features from time-series data to identify short-term dependencies. Subsequently, bidirectional LSTM processes the CNN-extracted features to capture long-term trends. Finally, GRU further refines the LSTM outputs, simplifying computational processes while strengthening temporal dependency modeling. Through this hierarchical integration, the model comprehensively captures the characteristics of pollutant concentration variations. Experiments were conducted on PM2.5 prediction in Dalian City, with comparisons to existing research models. Results demonstrate that the CNN-optimized LSTM-GRU integrated model achieves a significantly lower Root Mean Square Error (RMSE) than other models, indicating superior predictive performance. The proposed model exhibits high accuracy and stability in PM2.5 prediction tasks, validating its advantages in processing time-series data. Compared to traditional single models, this integrated approach demonstrates enhanced temporal feature-capturing capabilities, providing novel insights and tools for future air quality prediction and related applications.