A Hybrid Deep Learning Model for Air Quality Prediction Based on the Time–Frequency Domain Relationship

Deep learning models have been widely used in time-series numerical prediction of atmospheric environmental quality. The fundamental feature this application is to discover the correlation between influencing factors and target parameters through a deep network structure. These relationships original data are affected by several different frequency factors. If adopted without guidance, these correlations may be masked entangled multifrequency data, which will cause problem insufficient extraction difficult model interpretation. Because wavelet transform has ability separate can extracted methods, hybrid combining transformer-like (WTformer) was designed extract time–frequency domain features air 2018–2021 hourly Guilin as benchmark training dataset. Pollutants meteorological variables local dataset decomposed into five bands wavelet. analysis WTformer showed that particulate matter (PM2.5 PM10) had an obvious low-frequency band low high-frequency band. PM2.5 temperature negative positive wind speed results laws could found model, made it possible explain model. experimental show performance established better than multilayer perceptron (MLP), one-dimensional convolutional neural (1D-CNN), gate recurrent unit (GRU), long short-term memory (LSTM) Transformer, all time steps (1, 4, 8, 24 48 h).