Commercial building loads account for approximately 25 % of the total end-use load in urban areas, with a substantial portion comprising flexible loads such as air conditioning systems. This endows commercial buildings with significant potential for demand response （DR） regulation. Load disaggregation serves as a means of identifying flexible resources to facilitate the formulation of demand response strategies within the power grid. Nevertheless, existing non-intrusive disaggregation approaches are constrained by variations in the operating frequencies of building systems, whereby single time-scale models exhibit limited capability in effectively capturing the asynchronous characteristics of end-use devices. To overcome these limitations, a non-intrusive load disaggregation method based on multi-timescale deep learning is proposed in this paper. Specifically, a hybrid model integrating convolutional neural networks and long short-term memory networks is constructed to enable the collaborative extraction of spatiotemporal features from dynamic load data. Furthermore, a multi-timescale sampling mechanism is employed to conduct temporal sensitivity analysis and to determine the optimal sampling interval at the edge. Finally, experimental results using real-world commercial building load data demonstrate that the proposed model achieves superior comprehensive performance. When the sampling interval is increased from 30 seconds to 60 seconds, the Matthews correlation coefficient of the proposed model is improved by 7.35 %.