Feature Selection Techniques for Building Robust Air Quality Prediction Model

The availability of clean air is of utmost importance for sustaining life, promoting good health, and preserving the environment. Unpolluted air not only ensures the well-being of ecosystems but also supports agricultural productivity and enhances overall quality of life. Predicting air pollution levels is crucial as it enables proactive measures to mitigate its adverse effects, protecting public health and reducing environmental harm. Effective prediction models rely on the critical process of feature selection, which is instrumental in pinpointing the most relevant and informative features. The study is centered on the examination and comparison of various attribute selection methods, encompassing both filter-based and wrapper-based methods. A thorough examination and comparative evaluation of diverse deep learning algorithms were carried out in the process of constructing an air quality index (AQI) prediction model. Neural network architectures, including long short-term memory (LSTM), bidirectional long short-term memory (BILSTM), and gated recurring unit (GRU), were employed for this purpose. The results indicate that the GRU-based model, when coupled with a filter-based feature selection technique, outperforms other deep learning models in the realm of air quality prediction. Notably, it achieves an impressive mean absolute error (MAE) of 0.4305, a root mean square error (RMSE) of 0.5527, and a mean squared error (MSE) of 0.3054. Additionally, the model attains an R-squared (R2) value of 0.7017, which underscores its potential for delivering highly accurate and effective predictions for air pollution levels.