Based on natural gas consumption, ground conventional meteorological observation data in Beijing in heating season from 2002 to 2018, as well as yearly social statistical information, the inter-annual variation characteristics of natural gas consumption in heating season in Beijing and its impact factors were analyzed by using empirical mode decomposition (EMD) and correlation analysis methods. And on this basis the forecast model of natural gas consumption in heating season was established by using back propagation (BP) neural network method, further the model was tested and evaluated. The results are as follows: (1) The natural gas consumption increased persistently in heating season from 2002 to 2018 in Beijing, and it was decomposed better into social and meteorological consumptions by EMD, which reflected long-term variation trend and short-term fluctuation of natural gas consumption, respectively. (2) The social consumption of natural gas in heating season had significantly positive correlation with GDP, intensive heating supply area and resident population number in Beijing. The meteorological consumption had significantly negative correlation with air temperature and negative accumulative temperature, while it was significantly positive correlated with precipitation and persistent low-temperature days. In heating season, when the air temperature was obviously lower or the continuous low-temperature and strong snowfall processes appeared, the meteorological consumption of natural gas would increase sharply. (3) The forecast model of gas consumption in heating season based on EMD_BP method had better prediction effect in Beijing, the average relative error was 5.6%, especially the model could predict accurately the peak and valley change of meteorological consumption of gas, which could provide scientific reference to a certain extent for energy planning and regulating.