AirTFT: A Novel Transformer-Based Approach for Air Quality Prediction

Weather planning is vital in public health and environmental conservation thus a necessity for probable accurate air quality forecasting. Benchmarks like ARIMA and SVR derived through statistical methods and machine learning approaches do not work well with non-linear patterns and long term dependencies. Long Short-Term Memory and Gated Recurrent Unit learning algorithms produce higher accuracy than the traditional versions but are not well suited for multiple steps of prediction. This paper presents AirTFT (Air Temporal Fusion Transformer), a new model derived from the Temporal Fusion Transformer (TFT) to overcome these challenges. The used attention mechanism in AirTFT is the dynamic attention which learns to attend to the necessary time steps only for both short and long term dependencies. We benchmark AirTFT on an actual dataset involving hourly air quality recordings, and compare that against ARIMA, SVR, RF, LSTM, GRU, and CNN-LSTM based on MAE, RMSE, and MAPE measures. AirTFT performs better than all models with RMSE and MAPE reduction rates of 35% and 40% compared to LSTM respectively. Furthermore, the results reveal that AirTFT has 50% less MAPE compared to MAPE of ARIMA model and, therefore, AirTFT is superior to all in multi-horizon air quality forecasting. More precisely, AirTFT provides higher accuracy and achieves the goal of predicting temporal related factors, and therefore, can be considered as an efficient instrument to control and improve air quality in real time.