To prepare basic data for the use of machine learning in the building energy field, this study examined the characteristics of each model and compared the prediction performance, calculation efficiency and output result aspects of the machine learning model according to the input parameters. Outdoor temperature was used as a basic input to consider input differences for six machine learning models, MLR, SVM, GPR, ANN, DNN and DT, which are mainly used in the building energy field, and the building energy consumption was predicted and compared depending on whether the indoor temperature was additionally reflected. The predictive performance of most models improved when the outdoor temperature and the indoor temperature were reflected as inputs rather than when the outdoor temperature was reflected as an input in the influence of the input parameters. In the comparison of the predictive performance of the model, DNN(5-Layer) showed the most dominant predictive results with RMSE, MSE, MAE, and R2 (0.190, 0.036, 0.139, 0.88). Next, ANN showed predictive performance of RMSE, MSE, MAE, R2 (0.203, 0.041, 0.142, 0.86), and GPR provided efficient prediction with RMSE, MSE, MAE, R2 (0.211, 0.044, 0.150, 0.85). DNN and ANN improved their prediction performance as the number of hidden layers increased, but the training time increased from 4.8 seconds to 16.5 seconds. In terms of computational efficiency considering training time, MLR showed the best result with 1.4s. As a result, DNN showed 14% better predictive performance than MLR, and MLR were trained 11.8 times faster than DNN. With indoor temperature being further reflected as input parameters, most models better represent actual building energy consumption in aspects of the forecast results. Machine learning model selection should be reviewed not only for predictive performance for errors but also for calculation cost and the discernment provided by predictive results. Since this study was conducted on a single building, research on the selection and development of models with high reproducibility in various models based on big data in terms of utilization should be continued.