AI-Native Predictive Scheduling for Computing Power Networks: A Multi-Scale Learning Framework for Ultra-Low Latency 6G Applications

Computing Power Networks (CPNs) represent a paradigm shift from traditional cloud-edge architectures to unified computing fabrics essential for sixth-generation (6G) wireless networks supporting ultra-low latency AI applications. However, existing CPN frameworks rely on reactive resource management that cannot meet stringent latency requirements of time-critical workloads. This paper introduces ProactiveCPN, the first AI-native predictive scheduling framework that leverages multi-scale machine learning to anticipate system states and proactively manage AI task allocation in CPNs. Our approach integrates Long Short-Term Memory (LSTM) networks for short-term prediction, Graph Neural Networks for mobility-aware placement, and Transformer networks for long-term forecasting utilizing supervised learning models with empirically validated convergence. Through comprehensive evaluation using ITU-R compliant 6G scenarios, ProactiveCPN significantly outperforms state-of-the-art reactive methods in average latency, deadline satisfaction, energy efficiency, and system throughput while maintaining robust performance across diverse mobility patterns and system loads. The framework provides a foundation for autonomous network resource management, enabling next-generation applications requiring ultra-low latency and extreme reliability.