Comprehensive Analysis of Deep Learning Approaches for PM2.5 Forecasting

Abstract

AbstractAir pollution is causing massive damage to human health. PM2.5 in particular, has been shown to have a significant effect on human health. So, forecasting of PM2.5 is essential. Approaches like the ARIMA model used for time series forecasting. The invention of Deep Learning, especially the Recurrent Neural Networks, revolutionized the methods of forecasting the time series to achieve predictions that are more precise. Variants of RNN like LSTM, GRU which had long term dependencies unlike basic RNN gives more accurate predictions. Temporal Convolutional Network, which is a synthesis of 1D Fully Convolutional Network and Causal Convolutions, came into the picture during 2018 and also provided successful results in sequence learning and Forecasting time series. We compared deep learning approaches LSTM, GRU, CovLSTM and Temporal convolutional networks using three types of losses. After comprehensive analysis, our results proved that TCN also gives comparable results for time series forecasting as LSTM and, GRU.KeywordsTime series forecastingRecurrent neural networks (RNN)Gated recurrent unit (GRU)Temporal convolutional networks (TCN)Dilated convolutions