A Study on the Supplementary Method between RNN and CNN Model to Improve the Accuracy of Fine Dust Concentration Inference

Abstract

This paper proposes a method to compensate with CNN models strong in image classification to improve the accuracy of fine dust concentration inference using time series data and RNN models. The longer the prediction time, the more the predictions converge on the average concentration due to the loss function characteristics of the RNN model. To solve this problem, we devised a method to compensate for the predicted value of RNN by inferring the fine dust grade in the prediction time to the CNN model. In this paper, we show the possibility that the CNN model can distinguish metaphysical images with complex spatiotemporal relationships rather than human-identifiable images such as dogs and cats, and we think they can be used to infer the source of fine dust in the future.