LightSP: A Two-Timescale Deep-Broad Learning Framework for Cloud–Edge Spectrum Prediction

Accurate spectrum prediction is essential for proactive radio resource management and dynamic spectrum access (DSA) in 5G/6G Internet of Things (IoT) networks. Although deep learning (DL)-based methods have shown promising prediction performance, their high computational complexity and deployment cost hinder their practical use in resource-constrained edge environments. In contrast, lightweight models such as broad learning (BL) offer improved efficiency, but often suffer from limited generalization capability under nonstationary spectrum conditions. To address these challenges, this article proposes lightweight spectrum prediction (LightSP), a cloud–edge collaborative spectrum prediction framework that balances prediction accuracy, computational efficiency, and local adaptability. Specifically, LightSP employs a high-capacity DL model in the cloud to learn robust global spectral representations from large-scale historical data. At the edge, an adaptive BL-based surrogate model is deployed to enable efficient real-time prediction and dynamically refine the transferred representations using instantaneous local observations. By integrating global knowledge extraction with local adaptive learning, LightSP effectively captures long-term spectrum trends while remaining responsive to rapid local fluctuations. Extensive experiments on real-world spectrum datasets demonstrate that LightSP reduces memory overhead by more than 50% compared with existing methods while maintaining competitive prediction accuracy. These results verify the effectiveness of LightSP as a cost-efficient and scalable solution for intelligent spectrum management in future wireless IoT systems.