Neural network forecasting of time series dynamics using the example of dust particle content in atmospheric air

The aim of this study is to create a model for predicting the dynamics of PM 2.5 dust particles in the atmosphere of a megalopolis using reservoir computing based on an artificial neural network. The work was performed on dust particle data collected in Seoul, South Korea. A total of six models based on an artificial echo state neural network (ESN) were built to predict the PM 2.5 concentration dynamics. Application of the proposed approach has shown the effectiveness of ESN-based models for predicting dust content in the atmospheric air of megalopolises. The accuracy and quality of the models were improved from 9% to 67% depending on the evaluation metric compared to the base model. It was found that the accuracy of the model decreases if the predicted time interval exceeds 6% of the training time interval.