Building innovative and complex applications on top of the Internet of Things (IoT) services provided by huge connected devices and software while satisfying quality of service (QoS) parameters has become a challenging topic. Identifying suitable services according to their QoS parameters is one of the main underlying features to enable optimal selection, composition, and self-management of IoT systems. Checking each service to get its accurate QoS is not feasible. QoS prediction has been proposed these last years to try to cope with this issue. Mainly, the existing approaches rely on collaborative filtering methods, which suffer from scalability issues, that can considerably hamper the performance of QoS prediction. To overcome this limit, in this paper, we propose a deep-learning-based QoS prediction approach for IoT services. The approach we propose relies on Long Short-Term Memory (LSTM) to capture the service representation through a service latent vector and on Residual Network (ResNet) for QoS prediction. Unlike existing deeplearning-based approaches that assume a pre-defined static set of services, our approach addresses the QoS prediction problem for dynamic environments where the services are not necessarily fixed in advance.