This study utilizes machine learning techniques to conduct an in-depth analysis of time-series historical data on energy consumption in buildings. A generalized model identification method was developed using an optimization algorithm based on black-box models. The final identification model was determined after optimizing three machine learning methods, including polynomial regression, artificial neural networks, and extreme gradient boosting. A near-zero energy office building in Beijing is the primary focus of this study. Using historical building data and simulation data of the heating system in TRNSYS, load prediction and equipment energy consumption models were established using the developed model identification method. During deployment, the predicted R² value and total energy consumption deviation were 0.87 and 5.18%, respectively. The results demonstrate that the prediction models established through this method possess high accuracy, providing a reliable basis for subsequent system energy consumption optimization.