Enhancing Air Quality Prediction With A Multi Pollutant Attention-Based Model

Abstract

As the world expands and develops year by year, wars and industries are increasing rapidly, and their remnants negatively affect our environment. So, the air pollution problem has raised and affected various domains such as health, environment, and ecosystem sustainability. Therefore, the need for air quality forecasting has emerged to avoid or reduce this problem. The researchers have worked on this issue, but it still has limitations, especially the need to improve prediction accuracy, model performance and the ability to predict the most common pollutants. In this study, a forecasting model is implemented to predict the concentration of the most common pollutants in the world: Particulate Matter less than 2.5 micrometers in diameter (PM2.5), Particulate Matter less than 10 micrometers in diameter (PM10), Nitrogen Dioxide (NO2), Sulfur Dioxide (SO2), Carbon Monoxide (CO) and Ozone (O3) for three time steps ahead. In the proposed model, a new approach (weight transfer) is implemented to predict all pollutants automatically from just one model. This model is a hybrid deep learning CNN-BiLSTM-Attention model that leverages the strengths of both Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network to capture spatial and temporal dependencies in air quality multivariate data, also applied Bidirectional Long Short-Term Memory (BiLSTM) network and Self-Attention to extract further temporal dependencies. The Huber loss function is used to balance the loss values between Mean Squared Error (MSE) and Mean Absolute Error (MAE) loss functions, aiming for robust model performance. Additionally, a user interface is implemented to display the values and chart of forecasted pollutants. The proposed model is achieved using the dataset of Beijing city in China, which contains historical air quality and meteorological data. It is performed by Python language in the Colab environment, obtained effective results as a regression forecasting type, and the Coefficient of Determination (R²) evaluation metrics for PM2.5, PM10, NO2, SO2, CO, and O3 are 0.951, 0.890, 0.941, 0.918, 0.935, and 0.959, respectively; also, the model is evaluated using Root Mean Squared Error (RMSE) and MAE metrics. In addition, it surpassed some of the state-of-the-art models that used the same dataset. Moreover, the model is applied to the available dataset of Baghdad - Iraq.