Long-Term Prediction for Harmful Gas Concentration Based on RevIN-KANformer Model

Advance prediction of hazardous gases can help monitor trends in air quality and take effective preventive and control measures in advance.In short-term prediction, it is necessary to ensure the accuracy and reliability of the prediction results. And in long-term prediction, it is also crucial to accurately predict the changes of the sequence trend. This paper proposes a RevIN-KANformer model based on the reversible instance normalization method (RevIN), Transformer and KAN for long-term prediction of harmful gas concentrations in the air. The model effectively combines the Transformer and KAN models through an encoder-decoder structure, while the input and output sequences are normalized and denormalized by RevIN method to ensure optimal long-term prediction performance.The experimental results show that RMSE and MAE of the model proposed in this paper are reduced by at least 16% and 9% compared with the baseline model, which provides an effective method for the long-term prediction of harmful gas concentration.