Abstract: Using professional software, the simulation method can accurately calculate the dynamic energy consumption, however, the input parameters are cumbersome and often cannot be changed after the building geometric model is determined. Data mining method is fast and can be applied in various conditions, however, it needs long-time historical training data and is greatly affected by data quality. Based on the above characteristics, an energy consumption prediction model of office buildings was proposed based on the variables extraction from large simulation examples in this paper. EnergyPlus was used to build bulk models of typical office buildings and adjust input parameters to generate a database of millions of data. LightGBM algorithm was used to screen the characteristic factors affecting the load and construct the load forecasting model. Results show that the selected 24 dimensional characteristic variables can ensure that the prediction accuracy of the model is more than 90%. Combined with the energy consumption calculation model of air-conditioning equipment in EnergyPlus, the energy consumption prediction was achieved by python compilation. An office building in Beijing was selected as a research case. The average relative error of daily energy consumption of the measurement period was 8.27%. Then, typical annual weather data was used to calculate annual building energy consumption and a monthly average relative error of 10.37% was obtained. The predicted energy use intensity was 36.25 kW·h/(m²·a), while the real value was 35.20 kW·h/(m²·a), with a relative error of 2.98%.