Optimized Ozone Concentration Prediction in Seoul Districts Using ANN and K-means Clustering for Accuracy Enhancement

Abstract Ozone is a dangerous greenhouse gas and air pollutant in urban areas, with significant negative impacts on climate change and human health. Predicting ozone concentrations is a critical factor in environmental issues such as air pollution management, risk assessment, public health, and global warming. Since an early prediction model of ozone is essential for building a warning system, research is needed on indicators that explain whether ozone pollution status will rise or fall. This study proposed a prediction model trained using artificial neural network (ANN)-based classification with training data divided into specific time periods through k-means clustering to predict ozone concentrations. This model lowers the cost of training owing to around 30% of the reduced training data in a specific time period, and is also applicable for a variety of features. Air quality data was collected from 2019 and 2020 in the 25 districts of Seoul, South Korea and used for training and testing on whether the ozone concentration changes after one hour during 07:00 to 18:00. The proposed model yielded 3% higher F1 score and 3-4% higher accuracy in comparison with other models. As a result, the model proposed in this study showed improved performance while reducing the training data in a specific environment.