This paper addresses the problem of highly volatile and bursty network loads, which pose a challenge to maintaining Quality of Service (QoS) in crowdsourcing cloud-edge platforms (CCPs) due to the surge in streaming services. Existing predictive scheduling architectures suffer from the problem of minimizing mean absolute error, leading to SLA violations during peak hours, or conservative overloading strategies that waste resources. To address this, this paper proposes a scheduling-aware hybrid representation framework (HRS) that integrates numerical and image-based representations to better capture extreme load dynamics. Furthermore, we introduce a scheduling-aware loss (SAL) that captures the asymmetric impact of prediction errors, resulting in predictions that better support scheduling decisions. Extensive experiments on four real-world datasets demonstrate that HRS outperforms ten baseline models, achieving state-of-the-art performance by reducing SLA violations by 63.1% and gross profit loss by 32.3%.
