Atmospheric NO2 Concentration Prediction with Statistical and Hybrid Deep Learning Methods

Abstract

Recently, air pollution has become a critical environmental problem in Türkiye as well as in the world. Therefore, governments and scientists are putting a lot of effort into controlling air pollution and reducing its effects on human society. Scientists propose various models and methods for air quality forecasting because accurate estimation of air quality can provide basic decision-making support. This study proposes innovative hybrid models that integrate a Convolutional Neural Network (CNN) with a Long Short-Term Memory (LSTM) neural network and a Gated Recurrent Unit (GRU) to predict one day ahead of NO2 concentration. For this aim, the Time-Series Daily NO2 concentration data obtained between 2015 and 2022 at the Istanbul and Ankara provinces in Türkiye are used. The hybrid CNN-LSTM and CNN-GRU models are compared with various traditional statistical and machine-learning methods such as Autoregressive Moving Average (ARMA), Artificial Neural Network (ANN), CNN, LSTM, GRU, and Adaptive Neuro-Fuzzy Inference System (ANFIS-FCM). The accuracy of the prediction models is assessed using various statistical criteria and visual comparisons. Results show that the proposed hybrid CNN-LSTM and CNN-GRU models in one-day-ahead NO2 concentration predictions yield the best results among all models with R2 accuracy of 0.9547.