Building energy modelling (BEM) is crucial for achieving energy conservation in buildings, but occupant energy-related behaviour is often oversimplified in traditional engineering simulation methods and thus causes a significant deviation between energy prediction and actual consumption. Moreover, the conventional fixed schedule-setting method is not applicable to the recently developed data-driven BEM which requires a more flexible and data-related multi-timescales schedule-setting method to boost its performance. In this paper, a data-based schedule setting method is developed by applying K-medoid clustering with Principal Component Analysis (PCA) dimensional reduction and Dynamic Time Warping (DTW) distance measurement to a comprehensive building energy historical dataset, partitioning the data into three different time scales to explore energy usage profile patterns. The Year–Month data were partitioned into two clusters; the Week–Day data were partitioned into three clusters; the Day–Hour data were partitioned into two clusters, and the schedule-setting matrix was developed based on the clustering result. We have compared the performance of the proposed data-driven schedule-setting matrix with default settings and calendar data using a single-layer neural network (NN) model. The findings show that for the data-driven predictive BEM, the clustering results-based data-driven schedule setting performs significantly better than the conventional fixed schedule setting (with a 25.7% improvement) and is more advantageous than the calendar data (with a 9.2% improvement). In conclusion, this study demonstrates that a data-related multi-timescales schedule matrix setting method based on cluster results of building energy profiles can be more suitable for data-driven BEM establishment and can improve the data-driven BEMs performance.