Cloud render farms require efficient scheduling systems to manage variable workloads and improve resource utilization. Despite the fact that probability models such as the exponential distribution can be used to characterize idle behavior in machines, early-stage monitoring provides only partial observations, which makes it difficult to make timely predictions. This paper proposes a Transformer-based framework that predicts workload types from partially observed idle probability sequences, using an exposure ratio to control the visibility of time steps. By learning from incomplete data, the model generalizes from sequences with variable exposure ratios. Experimental results show that the model achieves high accuracy with limited exposure, enabling early and reliable spot resource estimation in high-performance rendering environments.