Research and Application of the Mode Decomposition-Recombination Technique Based on Sample-Fuzzy Entropy and K-Means for Air Pollution Forecasting

Abstract

Practical forecasting of air pollution components is important for monitoring and providing early warning. The accurate prediction of pollutant concentrations remains a challenging issue owing to the inherent complexity and volatility of pollutant series. In this study, a novel hybrid forecasting method for hourly pollutant concentration prediction that comprises a mode decomposition-recombination technique and a deep learning approach was designed. First, a Hampel filter was used to remove outliers from the original data. Subsequently, complete ensemble empirical mode decomposition adaptive noise (CEEMDAN) is employed to divide the original pollution data into a finite set of intrinsic mode function (IMF) components. Further, a feature extraction method based on sample-fuzzy entropy and K-means is proposed to reconstruct the main features of IMFs. In conclusion, a deterministic forecasting model based on long short-term memory (LSTM) was established for pollutant prediction. The empirical results of six-hourly pollutant concentrations from Baoding illustrate that the proposed decomposition-recombination technique can effectively handle nonlinear and highly volatile pollution data. The developed hybrid model is significantly better than other comparative models, which is promising for early air quality warning systems.