<p>Data-driven methods for air quality forecasting are beneficial to early-warning of air pollution, especially in small and medium-sized cities. In this study, the Prophet model is used for extracting time-related features, such as trend, annual cycles, and weekly cycles, from the PM_2.5 series, and the gradient boosting regression trees (GBM) and long short-term memory networks (LSTM) are used to build forecasting models utilizing predictors including time-related features. All predictors are grouped into time-related features, antecedent air pollutant concentration, and meteorological parameters, and permutation feature importance (PFI) is used for measuring the importance of these feature groups. A series of forecasting experiments on a middle city in China are carried out, and the results are analyzed for a better understanding of the mechanisms of the forecasting skill. The result shows that time-related features extracted by Prophet can enhance the forecasting skill significantly, and this advantage is more apparent when the forecasting horizon is large. PM_2.5 time series has short memory and a rapid response to changes in meteorological conditions, and only meteorological conditions in the previous 2 days have useful signals for forecasting. The rapid response and short memory of PM_2.5 make LSTM show no advantage over GBM in this case study. We show that feature engineering is important for forecasting PM_2.5, and the time-related features and antecedent pollutant concentrations can be seen as proxy variables for some variables that are difficult to obtain, such as pollutant emission and missing meteorological conditions.</p>