Smart rentals are becoming increasingly popular in the Internet of Things (IoT), such as smart hotel reservations and smart vehicle rentals. However, such online reservation services may face the unexpected refusal problems. For instance, a user booked a self-driving car service, but when he arrived at the pickup location, he found that there was no car available. Unexpected refusal will affect users’ experiences over the online reservation platforms, and thus it is necessary to predict unexpected refusals in advance. However, unexpected refusal records are very sparse, and the reasons for unexpected refusals are complex. Therefore, we propose a multi-task, multi-scale time series fusion model (MMTSF) to predict unexpected refusal. In detail, a multi-task learning model is constructed with the main networks as the unexpected refusal orders prediction, and the assistant network as the rejected orders prediction. Besides basic profile attributes, statistical attributes of online booking services are collected as time series. A multi-scale time series fusion model is constructed to better understand relevant features and thus improve the prediction performance. Experimental results indicate the effectiveness of our proposed model.