The development of advanced data-driven approaches for building energy management is becoming increasingly essential in the era of big data. Machine learning techniques have gained great popularity in predictive modeling due to their excellence in capturing nonlinear and complicated relationships. However, it is a big challenge for building professionals to fully understand the inference mechanism learnt and put trust into the prediction made, as the models developed are typically of high complexity and low interpretability. To enhance the practical value of advanced machine learning techniques in the building field, this study proposes a comprehensive methodology to explain and evaluate data-driven building energy performance models. The methodology is developed based on the framework of interpretable machine learning. It can help building professionals to understand the inference mechanism learnt, e.g., why a certain prediction is made and what are the supporting and conflicting evidences towards the prediction. A novel metric, i.e., trust, is proposed as an alternative approach other than conventional accuracy metrics to evaluate model performance. The methodology has been validated based on actual building operational data. The results obtained are valuable for the development of intelligent and user-friendly building management systems.