A novel bifold-attention-LSTM for analyzing PM2.5 concentration-based multi-station data time series

The prolonged inhalation of poor-quality air is linked to severe respiratory issues, including chronic bronchitis and increased fatalities from heart and lung diseases. Accurate forecasting of air quality time series is crucial for practical preventive actions. This paper introduces the bifold-attention long short-term memory (BA-LSTM) model, which integrates an ensemble attention mechanism with long short-term memory (LSTM) to enhance forecasting accuracy. The attention mechanism extracts vital information from input features in two distinct ways. We divided the data into training and testing sets using hourly PM2.5 air pollutant data from 12 nationally controlled monitoring stations in Beijing, spanning from March 1st, 2013, to February 28th, 2017. The results demonstrate that the BA-LSTM model outperforms four baseline models: multilayer perceptron (MLP), LSTM, attention-LSTM (A-LSTM), and attention Bi-LSTM (A-2LSTM). BA-LSTM achieved the best average RMSE (0.0127) and MAPE (3.8911), with performance improvements ranging from 0.3180 to 4.7186 across various datasets. This indicates the model’s robustness and versatility in multivariate time series analyses. Precise air pollutant forecasts generated by BA-LSTM can assist regulators in mitigating traffic and industrial activities in heavily polluted areas. Future research could extend this approach to other environmental or health-related time series data, such as predicting infectious disease transmission or water quality.