GRU for Overcoming Seasonality and Trend in PM2.5 Air Pollution Forecasting

Pollution forecasting is important to research, especially for hazardous particles like PM
2.5
. However, trends and seasonality that can hide in the dataset make the forecasting model’s performance unfavorable. Our research aims to design a forecasting model methodology that involves trend and seasonality analysis using the seasonal decomposition of time series by Loess (STL). We obtained the PM
2.5
concentration dataset from Kaggle. The next step is applying the STL to analyze trends and seasonal components in datasets. Then we compared three forecasting models: gated recurrent unit (GRU), long short-term memory (LSTM), and one-dimensional convolutional LSTM (1D Conv-LSTM). We use the mean absolute error (MAE) for model performance comparisons. The STL test results show a trend component in the dataset that can influence forecasting performances. GRU exhibits a better MAE than LSTM, 1D ConvLSTM, and two state-of-the-art method: Stacked LSTM and BiLSTM. GRU’s MAE score is 13.9. We conclude that GRU is a lightweight solution that can overcome trends and seasonality in PM
2.5
datasets.