This paper presents a two-step approach to forecasting city-wide building energy demand. The initial engineering estimates with input from typical buildings are used as the priori beliefs, and are transformed into the posteriori distributions that describe energy consumption patterns of plots. This two-step approach takes advantage of the engineering estimate in analysis of physical factors that determine building energy consumption and uses demand regression to further correct the priori engineering estimates based on the observed energy consumption of plots. The results of a case study shows that the two-step approach makes the standard deviation of the predicting factors within 0.001 kWh/(m(2).a) compared with 6.186 kWh/(m(2).a) of the engineering estimate or 53.020 kWh/(m(2).a) of the demand regression. It means that the bottom-up two-step approach has a high confidence in forecasting the city-wide building energy demand if the priori engineering estimates are critically accepted and Bayesian analysis is performed according to the observed energy consumption of plots. It is a general methodology and can be applied in most cities to forecast the city-wide building energy demand aiming to enable policymakers to establish the medium or long-term targets related to buildings stock energy consumption and associated CO2 emissions. (C) 2017 Elsevier B.V. All rights reserved.