A statistical learning framework for spatial-temporal feature selection and application to air quality index forecasting

Accurate air quality index (AQI) forecasting makes a difference to public health, local economic development, and ecological environment. As typical geographical datum, the spatial autocorrelation (SAC) of AQI is often ignored, which may violate assumptions some models, such as machine learning requires variables be independent identically distributed. Considering strong SAC AQI, this study proposes novel statistical framework integrating variables, feature selection, support vector regression (SVR) for prediction in correlation analysis time series are used extract spatial-temporal features. In addition, historical target site adjusted by using trigonometric eliminate non-stationarity. To further improve accuracy, selection method combining reinforcement with heuristic algorithm adopted. demonstrate effectiveness our proposed framework, we select data 34 cities from Yangtze River Delta, one most polluted areas eastern China, focus on three largest cities, Nanjing, Hangzhou, Shanghai. We compared several baselines, experiment illustrates that accuracy significantly better than baselines at all selected key sites can provide accurate predictions quality.