In the era of data-driven intelligent Internet, efficient utilization of computing power is paramount. Yet, current cloud-edge collaboration architectures face challenges with computing power isolation, adversely affecting efficiency and user experience. Computing Power Networks (CPNs) leverage networks with cloud-native applications to connect and manage resources, offering a blueprint for a collaborative computing ecosystem. In the CPN, scheduling stands as a pivotal function. However, conventional scheduling often neglects the interplay between computing and networks. To rectify this, we present a collaborative task scheduling system in CPNs that simultaneously contemplates the selections of computing nodes and network links. Aiming to maintain a balanced load for both computing and network resources, we formulate the scheduling challenge as a Constrained Markov Decision Process (CMDP). This approach focuses on optimizing both execution delay and success rate of computing tasks with load balancing constraints in CPNs. To facilitate the resolution of the CMDP, we introduce a Lyapunov-optimized Deep Reinforcement Learning (DRL) algorithm, which reconfigures the long-term constraint into immediate optimization. We provide numerical results to demonstrate the effectiveness of our suggested policy and algorithm.