Evaluating recurrent neural networks and long short-term memory for air pollution forecasting: mitigating the impact of volatile environmental factors

Abstract

Mitigation is the key to reducing the negative effects caused by air pollution. Forecasting several periods into the future is needed to understand the picture of air pollution as a basis for mitigation. Choosing the right forecasting method is crucial. This research will evaluate two machine learning methods, namely Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) for air pollution forecasting. Air pollution data for the Jakarta area is the object of research. The data is divided into two parts, namely 80% training data and 20% testing data. Both methods were evaluated with Mean Square Error (MSE), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE). The best method is the method that has the smallest MSE, MAE, and RMSE values. We experimented with a combination of hidden layer and epoch values. The results obtained are that air pollution in the Jakarta area is very volatile and is influenced by the COVID-19 pandemic. The correlation between NO2 and CO particles is the highest compared to other particles. The RNN method works well on PM10, O3, and NO2 particles. Meanwhile, the LSTM method works well on SO2 and CO particles. The best hidden layer and epoch values are 50 and 150 and 100 and 200.