An ODE based neural network approach for PM2.5 forecasting.

Area of Science:

• Environmental Science
• Computer Science
• Data Science

Background:

• Accurate PM2.5 forecasting is crucial but challenging due to complex influencing factors.
• Traditional deep learning models like LSTM and BiLSTM face limitations in long-term time-series prediction accuracy.
• Recurrent Neural Networks (RNNs) struggle with long-term dependencies and scalability.

Purpose of the Study:

• To develop advanced neural network models for more accurate time-series prediction of PM2.5 concentrations.
• To address the limitations of existing models in handling complex dynamics and long-term dependencies.
• To propose and evaluate novel Ordinary Differential Equation (ODE)-based models for improved PM2.5 forecasting.

Main Methods:

• Proposed two ODE-based models: a transformer-based ODE model and a closed-form ODE model.
• Utilized continuous-time neural networks and differential equations for time-series modeling.
• Conducted empirical evaluations and paired t-tests to compare model performance against LSTM-based models.

Main Results:

• The proposed ODE-based models significantly enhanced prediction accuracy compared to LSTM-based models.
• Improvements in prediction accuracy ranged from 2.91% to 14.15% for 1-hour to 8-hour predictions.
• The proposed model (CCCFC) demonstrated a statistically significant performance advantage over BiLSTM, LSTM, GRU, ODE-LSTM, and PCNN,CNN-LSSTM.

Conclusions:

• ODE-based models, particularly closed-form networks, offer superior scalability and accuracy for PM2.5 time-series forecasting.
• The developed models provide a robust alternative to traditional deep learning approaches for complex environmental data.
• The findings reinforce the effectiveness of the proposed CCCFC model for hourly PM2.5 forecasting.