Multi-Step Prediction of PM2.5 Concentration Using a Sequence-to-Sequence Approach

This paper investigates the use of a sequence-to-sequence approach for multi-step prediction of PM
2.5
concentrations. The aim is to analyze the effects of varying the lengths of input and output sequences, as well as the use of different architectures, specifically LSTM (Long Short Term Memory) and Transformer. The models were trained using three years of PM
2.5
concentration data. The results indicate that the LSTM models, incorporating time series decomposition, achieved the highest performances. The LSTM models outperformed the transformer with multi-head attention architecture, consistent with previous studies that concluded that transformers did not exhibit superior performance for multi-step time series prediction. These findings suggest that LSTM models, with the inclusion of time series decomposition, are valuable for predicting multi-step PM
2.5
concentrations in urban areas grappling with air pollution challenges.