An ensemble long short-term memory neural network for hourly PM_2.5 concentration forecasting

Area of Science:

• Environmental Science
• Data Science
• Artificial Intelligence

Background:

• Accurate forecasting of PM2.5 concentrations is crucial for public health protection and early warning systems.
• Existing forecasting models may lack the precision needed for effective environmental monitoring.

Purpose of the Study:

• To propose and evaluate an ensemble long short-term memory neural network (E-LSTM) for enhanced hourly PM2.5 concentration forecasting.
• To compare the performance of the E-LSTM model against single LSTM and feed forward neural networks.

Main Methods:

• Utilized ensemble empirical mode decomposition (EEMD) for multi-modal feature extraction.
• Employed long short-term memory (LSTM) networks for multi-modal feature learning.
• Integrated features using inverse EEMD computation, forecasting PM2.5 modes using historical data and meteorological variables.

Main Results:

• The E-LSTM model demonstrated superior forecasting performance compared to single LSTM and feed forward neural networks.
• Achieved lower mean absolute percentage error (19.604% and 16.929%) and root mean square error (12.077 and 13.983 μg/m³).
• Exhibited higher correlation coefficients (0.994 and 0.991), indicating strong predictive accuracy.

Conclusions:

• The proposed E-LSTM model offers a significant advancement in hourly PM2.5 forecasting accuracy.
• This approach provides a more reliable tool for environmental monitoring and public health advisories.
• Ensemble learning combined with EEMD and LSTM effectively captures complex patterns in PM2.5 data.