Forecasting and analysis of PM2.5 concentration in Hainan Province based on LSTM Neural Networks

Abstract

The change of PM2.5 concentration not only reflects the change of air quality level, but also affects human life and health. However, because the PM2.5 change is caused by wide-ranging factors and the occurrence mechanism is complex, it is difficult to build a stable forecasting model with general physical methods. This study builds three different LSTM models to forecast the PM2.5 concentration of Hainan Province from 2019 to 2021. Model 1 is a fully connected neural network, which is also the base model. Model 2 is a single unit LSTM model. Model 3 is a multi-LSTM layer model. The conclusion indicate that the multi-layer LSTM model has the best forecasting precison for PM2.5 concentration, and the single unit LSTM model quite the opposite. The length of the time series has certain influence on the precision of PM2.5 forecasted by LSTM models, suggesting that a suitable time series length should be selected when building LSTM models.