Reservoir computing for predicting pm 2.5 dynamics in a metropolis

Abstract

Recently, researchers have used various methods for time-series forecasting based on artificial neural network models. Among these approaches, one of the most effective ones is the Echo State Network (ESN). An ESN is a variant of recurrent neural networks (RNNs) that are used in environmental studies. In this work, we propose models to predict the dynamics of dust particles (PM 2.5) using reservoir computing. The model was based on data on the content of PM 2.5 obtained in Seoul, Republic of Korea, collected between January 2017 and August 2017. Hourly data for this period were averaged over a 6-h interval to reduce variability in the source data. For training, 800 samples of the time series were selected; for the test set, 50 samples (part 1 of the work) and 100 samples (part 2 of the work) were used. Prediction accuracy was assessed using several accuracy indices and a Taylor diagram. The application of the proposed approach demonstrated the effectiveness of reservoir calculations for predicting dust content in megacities. The accuracy and the quality of the models improved from 9 to 67%, depending on the evaluation indicator. It was also found that the accuracy of the model decreased when the predicted time interval exceeded 6% of the training time interval.