A Particulate Matter 2.5 Concentration Forecasting Method Based on Multi-Input LSTM

With the expansion of urban area and population, air pollution in many areas is becoming more and more serious. PM
_2.5
has the characteristics of complexity and randomness, because its concentration is affected by many factors. Researchers usually input a large number of data into random forest, Neuro-Fuzzy model or neural network model for prediction. However, there are many factors affecting PM
_2.5
, some factors can promote the production of PM
_2.5
, while other factors may hinder its production. In the process of training, these factors will not only bring extra noise to the prediction, but also increase the complexity of training. In this study, a method for predicting PM
_2.5
concentration based on Multi-Input LSTM model was proposed. First, we design a Multi-Input LSTM model, and then the features are divided into two groups through the Spearman's rank correlation coefficient, which can be used as auxiliary input to LSTM, and add attention mechanism to reduce the noise in the data, and extract valuable features. The experimental results show that this method achieves 97.3% accuracy on the dataset better than the baseline method.